PDA Vol. 77, Issue 3 - VinaGMP

vinagmp.
com | Group Facebook: EU GMP in Viet Nam
PDA Journal of
Pharmaceutical
Science and
Technology
2023
May/June
Volume 77
Number 3
vinagmp.com | Group Facebook: EU GMP in Viet Nam
PDA Journal of Pharmaceutical Science and Technology

PDA JPST is the primary source of peer-reviewed scientific and technical papers on topics related to pharmaceutical/biopharmaceutical
manufacturing, sterile product production, aseptic processing, pharmaceutical microbiology, quality, packaging science, and other
topics relevant to PDA members. PDA JPST is an internationally recognized source that receives over a quarter of a million visitors
annually. Starting in 2008, JPST is published online only.
Acting Editor-In-Chief PDA, Inc., Leadership

Shanker Gupta, PhD Officers
4350 East West Highway Chair: Susan Schniepp, Regulatory Compliance Associates Inc.
Suite 600 Chair-Elect: Anil Sawant, PhD, Merck & Co./Merck Sharp & Dohme
Bethesda, MD 20814 USA Treasurer: Melissa Seymour, Biogen, Inc.
Telephone: +1 301-728-6045 Secretary: Emma Ramnarine, Genentech/Roche
Email: journal@pda.org Imm. Past Chair: Jette Christensen, Novo Nordisk
Manuscripts must be submitted online at https://submitjournal.pda.org.

Directors
Bettine Boltres, PhD, West Ivy Louis, VIENNI Training &
Associate Editor Pharmaceutical Services Consulting LLP
Michael (Mike) Sadowski Cristiana Campa, PhD, GSK Amy McDaniel, PhD, Bristol
Javier Camposano, PhD, Celltrion Myers Squibb
Journal Editorial Board Mirko Gabriele, PhD, Thermo Cylia Chen Ooi, Amgen
Fisher Scientific Mathias Romacker (ret.)
James Akers, PhD, Akers Kennedy & Associates Marc Gloglovsky, ValSource Inc. Osamu Shirokizawa,
Jose Caraballo, MBA/ME, Kite Pharma, a Gilead Company Andrew Hopkins, Abbvie Life Scientia Ltd.
Dayue Chen, PhD, Genentech, a Member of the Roche Group Stephan O. Krause, PhD,
H. Gregg Claycamp, PhD, Consultant (U.S. FDA, Retired) Bristol Myers Squibb
Tony Cundell, PhD, Microbiological Consulting, LLC
Fenton Fong, MSc, xCellarate Technology Consulting Inc.
Ghada Haddad, PhD, Merck & Co. Staff
Ankur Kulshrestha, PhD, Bristol Myers Squibb
Kelvin Lee, PhD, NIIMBL Glenn E. Wright, President & CEO
Richard V. Levy, PhD, PDA (ret.) Walter Morris, Sr. Director of Publishing and Press Relations
Duncan Low, PhD, Claymore Biopharm LLC Tehya Enger, Editorial Assistant
Hanns-Christian Mahler, PhD, ten23 health AG
Vision
Murali Muralidhara, PhD, RVAC Medicines
Neha Pardeshi, PhD, Amgen To be an essential global medium for the dissemination of key
Jai Pathak, PhD, Octapharma scientific discovery that creates the foundation for pharmaceutical
Michael (Mike) Sadowski, Baxter Healthcare Corporation and biopharmaceutical quality and regulatory best practices.
Wendy R. Saffell-Clemmer, MS, Baxter International Inc.
Tim Sandle, PhD, Bio Products Laboratory Limited Mission
Alicja Sobantka, PhD, Octapharma To communicate and publish scientific ideas and technological
István Szikora, MBA/MS, Novartis applications promoting the advancement of quality and regulatory
Edward C. Tidswell, PhD, Merck best practices throughout the global pharmaceutical and
Henrietta Vinnerås, MSc, Fresenius Kabi AB
biopharmaceutical industry in support of PDA’s overall mission.
Kelly Waldron, PhD, ValSource Inc.
Wesley Workman, PhD, Workman Biotech Consultants, LLC Access
Glenn Wright, Parenteral Drug Association
Earl Zablackis, PhD, Sanofi Pasteur All active PDA members receive a free license to access the two most
Steve Zdravkovic, Baxter International Inc. current volume years of PDA JPST as part of their member benefits.
Qingyu Zeng, PhD, West Pharmaceutical Services, Inc. PDA JPST is also available by subscription to industry, university, and
Daniele Zuccato, MS, SCHOTT AG public libraries, as well as to government agencies. Nonmembers can
purchase access to articles.
Copyright © PDA, Inc. 1994 – 2023

ISSN 1079-7440
May–June 2023 Volume 77, No. 3
PDA Journal of
Pharmaceutical Science and Technology
CONTENTS
EDITORIAL
Artificial Intelligence and Pharmaceutical Production 145
Shanker Gupta
RESEARCH
CPV of the Future: AI-Powered Continued Process Verification for Bioreactor Processes 146
Andrej Ondracka, Arnau Gasset, Xavier Garcı́a-Ortega, David Hubmayr,
Joeri van Wijngaarden, José Luis Montesinos-Seguı́, Francisco Valero, and Toni Manzano
Flexible Loading Pattern Approach in Overkill Steam Sterilization Based on the Physical 166
Properties of Steam and Thermodynamics of Sterilization
Arnan Ben-David
Tolerance Interval Approach for the Determination of Overfill of Liquid Parenteral Drug Products 181
Bernhard Schmelzer and Marc Sutter
Determination of ICH-Q3D Elemental Impurity Leachables in Glass Vials by Inductively 197
Coupled Plasma Mass Spectrometry
Lydia Breckenridge, Yusuf Oni, Christina Evans, Jason Franck, Sharla Wood,
Meng Xu, Erinc Sahin, and Brian Zacour
TECHNOLOGY/APPLICATION
Rapid Sterility Test Systems in the Pharmaceutical Industry: Applying a Structured 211
Approach to Their Evaluation, Validation and Global Implementation
Sven Deutschmann, Mousumi Paul, Marja Claassen-Willemse, Jonas van den Berg,
Pieta Ijzerman-Boon, Viviane Grunert da Fonseca, Ellen Brunbech, Lynn Johnson,
Chris Knutsen, Lucile Plourde, Joanny Salvas, Philip Villari, and Lisa Wysocki
CASE STUDIES
Multisite Qualification of an Automated Incubator and Colony Counter for 236
Environmental and Bioburden Applications in Pharmaceutical Microbiology
Hans Joachim Anders, Daniel Männle, William Carpenter, Wolfgang Eder,
Ivana Heckel, Tobias Gøtzen, Corinne Oechslin, Cedric Joossen, Maria Eugenia Giribets
Parra, Jason Rose, Vaishali Shah, and David L Jones
COMMENTARY
Something for Nothing 248
James Agalloco
Published by PDA, Inc.

CODEN:JPHTEU 77(3) (2023)
EDITORIAL
Artificial Intelligence and Pharmaceutical Production

The production of pharmaceuticals, food ingredients and liquid parenteral products, and determination of leach-
industrial chemicals is big business in the United States. ables in glass vials. The article by Ben-David speaks to
There have been advances in the use of Big Data for loading patterns in an autoclave running overkill steam
continuous monitoring of biopharmaceutical processing sterilization cycles. The tolerance interval during filling
leading to process intensification, which we addressed in
of liquid parenteral drug products is addressed by
this journal in the past issues. Having said that, pharma-
Schmelzer and Sutter. And Breckenridge, et al., present
ceutical production is a complex and highly regulated
industry. There have been considerable advances in on leachables from glass vials and their identification
genetic engineering, proteomics, metabolomics, and using inductively coupled plasma mass spectrometry.
machine learning that have not translated into pharma-
ceutical production and biopharmaceutical processing in Lastly, a word about the 2022 Frederick D. Simon Pa-
a meaningful way. per of the Year award. The Journal Editorial Board and
selection committee chose the paper by Lenger et al.,
PDA and the PDA JPST are always looking for the 100% Control of Controlled Ice Nucleation Vials by
most recent advances to bring to our readers. And the Camera-Supported Optical Inspection in Freeze-Drying
current issue is no exception. We lead off with a study as the awardee. The “best” in scientific literature is a very
that uses AI-powered continuous process verification difficult decision, however. There were very close sec-
of bioreactor processing. This study outlines how contin- onds—papers by Franzese, et al., Coleman, et. al., Stan-
uous process verification can be achieved using machine ley, et. al., and Boltres. These manuscripts were
learning and artificial intelligence. (See also, Rathore, AS
published in the PDA JPST Volumes 75(6), 76(1) and 76
& Fernandez-Lahore M, J. Chem. Technol. Biotechnol
(2). My congratulations to all the authors for their well-
2022:97:2287.)
done studies and their submissions to the Journal. I am
grateful for their support.
Some of the other research articles shed light on such
topics of parenteral production as sterilization, filling of
I look forward to your engagement and contributions.
The success of this Journal relies on you!
doi: 10.5731/pdajpst.2023.001423
Shanker Gupta
Vol. 77, No. 3, May--June 2023 145

RESEARCH
CPV of the Future: AI-Powered Continued Process

Verification for Bioreactor Processes
ANDREJ ONDRACKA1,#, ARNAU GASSET2,#, XAVIER GARCÍA-ORTEGA3, DAVID HUBMAYR4,
JOERI VAN WIJNGAARDEN1, JOSÉ LUIS MONTESINOS-SEGUÍ2, FRANCISCO VALERO2, and
TONI MANZANO1,*
1
Aizon, Córcega 301, 08008 Barcelona, Spain; 2Department of Chemical, Biological and Environmental Engineering,
School of Engineering, Universitat Autònoma de Barcelona, 08193 Bellaterra (Barcelona), Spain; 3QuBilab, Departament
de Biocie`ncies, Facultat de Cie`ncies i Tecnologia, Universitat de Vic – Universitat Central de Catalunya, Vic, Spain; and
4
Process Development & Breakthrough Technologies R&D, CSL Behring AG, Wankdorfstrasse 10, 3014 Bern, Switzerland
© PDA, Inc. 2023
ABSTRACT: According to the standard guidelines by the FDA, process validation in biopharma manufacturing encom-
passes a life cycle consisting of three stages: process design (PD), process qualification (PQ), and continued process
verification (CPV). The validity and efficiency of the analytics methods employed during the CPV require extensive
knowledge of the process. However, for new processes and new drugs, such knowledge is often not available from Pro-
cess performance qualification and Validation (PPQV). In this work, the suitability of methods based on machine learn-
ing/artificial intelligence (ML/AI) for the CPV applied in bioprocess monitoring and cell physiological control of the
yeast Pichia pastoris (Komagataella phaffii) was studied with limited historical data. In particular, the production of
recombinant Candida rugosa lipase 1 (Crl1) under hypoxic conditions in fed-batch cultures was considered as a case study.
Supervised and unsupervised machine learning models using data from fed-batch bioprocesses with different gene dosage
clones under normoxic and hypoxic conditions were evaluated. Firstly, a multivariate anomaly detection (isolation forest)
model was applied to the batch phase of the bioprocess. Secondly, a supervised random forest model for prediction of
required operator’s control actions during the semiautomated fed-batch phase under hypoxic conditions was assessed to
maintain the respiratory quotient (RQ) within the desired range for maximizing the specific production rate (qP). The per-
formance of these models was tested on historical data using independent evaluation of the process by the process control
engineer (subject matter expert—SME), and on real-time data in the case of manual action prediction, where the model was
implemented to guide the control of the bioprocess. The work presented here constitutes a proof-of-concept that multivariate
analytics methods, based on machine learning, can be a valuable tool for real-time monitoring and control of biopharma
manufacturing bioprocesses to improve its efficiency and to assure product quality.
KEYWORDS: Bioprocess engineering, Bioreactor, Pichia pastoris, Artificial intelligence (AI), Machine learning,
Anomaly detection, Random Forest.
Introduction attributes and process parameters, process capability, man-

ufacturing and process control technologies, and quality
The PhRMA Quality Technical Committee defined systems infrastructure. This definition is crucial to
manufacturing science in the pharmaceutical field in understand the scope of the research described in this
2003 as the body of knowledge available for a specific work. In biopharmaceuticals, the expected requirements
product and process, including critical-to-quality product need to be met to ensure product quality. However, ro-
bust processes are synonymous with operational control,
and robust processes represent an additional warranty to
* Corresponding Author: Aizon, Córcega 301, 08008
produce the expected quality. Advanced Process Control
Barcelona, Spain; Telephone:+34 93627454605; E-mail:
(APC) is a term used to describe procedures and technol-
toni.manzano@aizon.ai
# Andrej Ondracka and Arnau Gasset contributed equally ogies applied in industrial frameworks to ensure that
doi: 10.5731/pdajpst.2021.012665 processes are under control when there is a multivariable
approach by more than one measurement, more than one
146 PDA Journal of Pharmaceutical Science and Technology

Figure 1
Process validation schema based on the FDA guidance (FDA, 2011). The two arrows indicating the possibility
to bring feedback to previous stages are an interpretation proposed by the authors based on the goal of this
research applied to the FDA guidance.
final control element, the control addresses the interaction critical material attributes (CMAs), critical process pa-
between process variables and incorporates functions such rameters (CPPs), key process parameters (KPPs), and
as constraint control that are intended to optimize process critical quality attributes (CQAs), are not sufficient to
operations (1). Unfortunately, in pharma biomanufactur- describe a real and complete picture of the drug manufac-
ing, the APC systems are still managing processes by uni- turing bioprocess as they were designed at Stage 1. New
variate controls (2). Multivariate analysis is a set of results or additional factors not considered during Stage 1
techniques used to understand and interpret the complex- or Stage 2 that would be relevant could appear during the
ity embedded in a system governed by multiple dimen- production. Stage 3 is typically a long manufacturing
sions, and it is useful in the biotech field (3). Although phase in which extensive data is accumulated, trended,
multivariate techniques are broadly accepted in biophar- and analyzed to ensure that the bioprocess is always
maceutical environments, artificial intelligence (AI) is still under control. The drug manufacturing and the imple-
not adopted in Good Manufacturing Practices (GMP) mented industrial processes are verified at Stage 2, but
environments, even though it has a demonstrated value in the inherent variability of the huge number of elements
the industry (4). involved in the industrialization of the drug design is
always providing uncertainty to the process. Depending
When the Food and Drug Administration (FDA) pub- on the robustness of the initial quality by design process,
lished the first version of the Process Validation guide- a backward movement to the previous stage or even to
lines (FDA, 2011) (5), the administration established the Stage 1 often has to be considered (Figure 1).
path for implementing the expected quality, safety, and
efficacy by design in the product, as well as in the pro- In general, chemical processes are usually considered
cess. These guidelines establish the continuation of the much easier to be on-line monitored than bioprocesses.
initial proposal established by the International Confer- Therefore, the implementation of CPV in pharmaceuti-
ence on Harmonisation (ICH) guidances for industry, Q8 cal industries is easier for products obtained by chemi-
(R2) Pharmaceutical Development, Q9 Quality Risk cal transformations rather than biological processes or
Management, and Q10 Pharmaceutical Quality System. biotransformations. It is due to the lower complexity of
process validation encompasses a life cycle approach that the process and the need not to include so many key
includes three stages: process design (Stage 1), process process variables. Moreover, CPV has been considered
qualification (Stage 2), and continued process verification as an essential mechanism to implement APC in bio-
(CPV, Stage 3) (Figure 1). manufacturing processes (2) and a relevant procedure
to ensure consistency of biologics manufacturing (3).
Established conditions (EC) are defined as legally binding
information (or approved matters) considered necessary The methylotrophic yeast Pichia pastoris (Komagataella
to assure product quality (6). The ECs, which include phaffii) is a widely used microbial cell factory for the
Vol. 77, No. 3, May--June 2023 147

metabolites and recombinant proteins production (RPP), makes the likelihood of finding a single model that cap-
including both biopharmaceutical and industrial enzymes tures the information necessary for a proper predictive
(7–11). Most of them are secreted to the culture broth model rather low (22). Hence, the use of data-driven
facilitating the subsequent downstream processes due to models in the sense of learning from experience with
the low secretion of native proteins and other metabolites. measurement data represents a suitable alternative. In
Although the inducible alcohol oxidase 1 promoter addition, the run-to-run variability on bioprocess model
(PAOX1) has been the classical promoter used with the parameters and predictions that present the currently
expression regulated by the presence of methanol, devel- available macroscopic models (24, 25) is not low enough
opment of alternative new methanol-free expression sys- for a successful use in process control. Thus, as an alter-
tems is considered of great interest during the last years native, the relationships between state variables such as
to avoid the drawbacks and costs associated with the stor- biomass, substrate, and product concentrations and the
age and handling of methanol as well as the operational measured variables can be modeled to a sufficient degree
problems caused by the methanol metabolism (12). The of accuracy with modern data-driven AI methods devel-
most used methanol-free alternative is the widely used oped by the machine learning (ML) community.
constitutive glyceraldehyde-3-phosphate dehydrogenase
promoter (PGAP) (13–15). Consequently, great efforts AI algorithms have been promoted as valid analytical
have been undertaken to improve bioprocess efficiency methods, and good practices have been proposed for bio-
with PGAP, boosting the production yields and rates to pharmaceuticals manufacturing. Two AI algorithms
become a cost-effective alternative to PAOX1 for the indus- (Neural Networks and Support Vector Machines) were
trial production of recombinant proteins. introduced by the European Pharmacopoeia (26) as valid
chemometric methods applied to analytical data in
It has been reported in previous works that cellular pharma contexts. The FDA proposed different tools and
stress in P. pastoris and other yeasts can trigger an techniques to reduce the variability and to understand the
increase in production parameters (12, 16). Therefore, effects of the unavoidable and unpredictable events on
implementation of hypoxic conditions as a novel and commercial manufacturing processes. Unknown factors
nonconventional cultivation strategy can lead to signif- can lead to byproducts formation due to perturbations on
icant increases in terms of specific production rate (qP) operating conditions or nonconsidered process variables,
and productivity (17–19). This environmental stress which may affect bioprocess efficiency and product qual-
leads to a shift from a respiratory to respiro-fermenta- ity. A more recent publication elaborated by the FDA
tive metabolism, which can be monitored by either described a methodology for AI application in medical
ethanol production and/or an increase in the respiratory devices (FDA, 2021). Thus, the CPV highly recommends
quotient (RQ) as indirect reporting parameters. There- bioprocess automation, use of process analytical technol-
fore, these parameters must be properly controlled not ogies (PAT), risk assessment, and a deep knowledge of
only to maximize protein production but also to keep a the biomanufacturing process and drug product attributes.
constant protein quality to standardize the process and They are considered and evaluated individually, although
to guarantee the reproducibility between lots. most of them are part of a multivariate and complex real-
ity. Statistics and multivariable analysis can be comple-
Process modeling can be applied to perform an opti- mented with AI to bring more knowledge at this stage.
mized, feasible, precise, and robust process control.
Mechanistic and data-driven modeling possess separate Many AI algorithms have been developed and can be
unique advantages as well as disadvantages and limita- applied to tasks such as describing the relationships
tions due to their respective model structures. No gen- between variables in the bioreactor. The support vector
eral statement can be made about whether mechanistic, machine regression (SVR) techniques are one example,
data-driven, or hybrid approaches (20–22) are more neural networks and relevance vector machines are
suitable, because the selection is strongly dependent on other alternatives (20, 27). Some recent examples of
the available process knowledge and measurement sys- applications to P. pastoris bioprocesses include a neu-
tems (off-line/on-line) as well as the number of data ral network-based feedback control for methanol feed-
sets and data points (23). However, the fact that P. pas- ing (28) and final yield prediction (29) in case of
toris bioprocesses have distinct process phases such as recombinant antibody production, and a support vector
batch, transition, and fed-batch phases, with more than machine (SVM) regression-based yield prediction in an
one substrate and time-varying key process rates, insulinase production system (30).

However, another class of AI/ML algorithms that are Materials and Methods
particularly suitable for such tasks are algorithms based
on decision trees, such as random forest regression, Risk Assessment
which has also been implemented due to its versatility,
no need for complex parameterization, no need for Risk assessment was performed before the start of the
variable normalization or scaling, relative insensitivity study. The results of the risk assessment are in the
to outliers, and ability to extract the important predic- Table A-I at the end of the document.
tors by feature importance, making the resulting mod-
els more interpretable (27). Strain and Cultivation Methods
In addition to supervised learning methods described Clone Construction and Gene Dosage Determination:
previously, which aim to model the relationship Two clones of P. pastoris harboring one (single-copy
between independent variables and a target variable clone, SCC) and five (multicopy clone, MCC) copies
(typically a critical quality attribute, CQA), another of the C. rugosa lipase 1 (CRL1) expression cassette
potentially useful application of AI methods is outlier regulated by PGAP were tested in chemostat and fed-
detection. One of the main root causes of issues in drug batch mode. The clone construction and the gene dos-
manufacturing are mechanical human interactions age determination have been previously described in
within the production chain (31). The routine actions other works (39, 40).
executed by process operators are often a source of
failures. In contrast to univariate outlier detection, AI
Chemostat and Fed-Batch Cultivation: Chemostat
algorithms for outlier detection (or anomaly detection)
cultivations with these clones (SCC and MCC) had
can detect outliers in a dataset in a multivariate man-
been performed before this study to evaluate the effect
ner, considering not only data points that are anoma-
of oxygen limitation as described in previous published
lous in a single variable but also detecting potential
works (19).
anomalous combinations of variable values. Several
ML algorithms have been adapted and developed for
For the work presented hereafter, the producer clones
multivariate anomaly detection, such as one-class sup-
were grown in carbon-limited fed-batch cultures with a
port vector machine (SVM), local outlier factor, and
preprogrammed exponential glucose feeding profile at
isolation forest (32–34).
a constant specific growth rate (l) of 0.10 h1 in both
normoxic and hypoxic conditions. This strategy allows
AI can replicate simple manual tasks of human cogni-
a pseudostationary state to be reached as has been
tion in an automated way, and as such, it is suitable
described elsewhere (14, 41). Prior to the fed-batch
for controlling multiple processes with a high number
phase, a glycerol batch phase was conducted (14). The
of variables to make fast decisions. AI is already
extensively used in operations such as image recogni- composition of both batch and fed-batch media have
tion, multivariate prediction, or fast classifications been previously reported (19). Temperature and pH
(35, 36). The overall goal of the present work is to de- were controlled at 25˚C and 6.0, respectively. The inlet
velop and implement AI for CPV to demonstrate that gas flow rate was 2 L min1. The fermentations were
AI can be used as not only a multivariable tool for done in a 5 L Biostat B Fermenter (Sartorius Stedim,
detecting bioprocess anomalies but also for estimating Goettingen, Germany), equipped with MFCS/win 3.0
the most suitable operating conditions for bioprocess Process Control software for the monitoring and con-
intensification, including bioprocess efficiency and trol of the fermentation.
product quality in real time. The previously described
P. pastoris cell factory was selected as a case study Analytical Methods
producing Candida rugosa lipase 1 (Crl1) regulated
by the constitutive PGAP, one of the most promising Biomass Concentration and Composition: Biomass
industrial lipase enzymes for biocatalysis (37, 38). dry cell weight (DCW) analysis was performed as
Specifically, the use of AI modeling to detect anoma- described previously (42). Biomass elemental composi-
lies in high automated bioprocesses and to develop a tion was also determined as it was done in other studies
novel physiological control strategy based on AI has (43). In both cases, the relative standard deviation
been explored. (RSD) of the measures was below 5%.
Vol. 77, No. 3, May--June 2023 149

Lipase Production: Product quantification was per- Ideal Agitation Prediction Models: The random forest
formed through a lipolytic assay, using p-nitrophenyl model for stirring rate was trained using the ranger
butyrate (pNPB) as a substrate for Crl1, as previously function from the ranger package in R software. Before
described (44). Product titer is presented as activity training the models, data were filtered so that only the
units·mL1, in which one activity unit is defined as the data points that were considered in the ideal regime
amount of enzyme needed to hydrolyze 1 lmol ester were kept. The number of trees constructed for the ran-
bond per min. RSD of the analysis was below 1%. dom forest models was 500, variance was used as the
splitting rule, and no constraints were put on maximal
tree depth. Impurity was used to calculate the variable
Carbon Source and By-Products Quantification: The
importance.
concentrations of substrates and potential fermentation
by-products were analyzed through high-performance
Data Governance and Deployment of AI Models: The
liquid chromatography (HPLC) for each sample. The
raw data were stored and managed, and the models
software for the quantification and the column specifi-
were deployed in the AWS cloud using the Aizon plat-
cations are described elsewhere (45). The RSD of the
form (Aizon, Barcelona, Spain) to ensure data integrity
method was below 1%. In addition, ethanol concentra-
and compliance.
tion was measured on-line with a Methanol Sensor Sys-
tem (Raven Biotech Inc., Vancouver, Canada) used in Results
previous studies (46, 47). This sensor system is sensi-
tive to other volatile compounds, and it was adapted to Bioprocess Background, Experimental Setup, and Data
ethanol measurement, being calibrated in each fed- Gathering
batch integrating HPLC data with the electric signal of
the probe to obtain an on-line ethanol concentration Previously, the application of hypoxic conditions in the
profile. production of an antibody fragment (Fab) increased up
to threefold the specific production rate (qP) and spe-
Inlet- and off-Gas Analyses: To analyze CO2 and O2 cific productivity (QP) (17–19). In the present work,
mole fractions of both the inlet- and the off-gas, a this culture strategy has been applied to the recombi-
BlueInOne FERM (BlueSens, Herten, Germany) ana- nant production of Crl1.
lyzer linked to the software BlueVis (BlueSens, Herten,
Germany) was used. As described in previous works To gather information about the system, a systematic
study of the effect of hypoxic conditions has been
(19), the recorded data was used to calculate the
made in chemostat cultures at a dilution rate of
respirometric parameters: carbon dioxide emission rate
0.10 h1 with both producer clones (SCC and MCC).
(CER), oxygen uptake rate (OUR), and respiratory quo-
Chemostat experiments have shown that at around 12%
tient (RQ). The RSD was below 5%. The gas analyzer
O2 molar fraction in the inlet-gas can be found the tran-
was recalibrated for every fed-batch to assure an accu-
sition between normoxic and hypoxic conditions.
rate and reliable data set.
Accordingly, and regardless of the gene dosage of the
clones, RQ increased linearly from a constant value
AI Algorithms close to 1.2 in normoxic conditions up to 1.6 once the
culture achieved the most restrictive oxygen condition
Anomaly Detection Model: The anomaly detection tested. Similar behavior was observed for ethanol-spe-
models were trained using the isolation.forest function cific production rate (qEthOH) from 0 under normoxic
from the isotree package in R software, with default pa- conditions to the highest value of 0032 gEtOH gX1 h1
rameters, except for where the column weights were under the most severe hypoxic condition.
specified. The data were split in hourly intervals from
the beginning of the fermentation, and a separate model On the other hand, it must be noticed that the pattern
was trained for every interval. Finally, the trained mod- observed for qP with respect to oxygen limitation dif-
els were applied to each data point to compute the fered from that observed with RQ or qEtOH. Although
anomaly scores. For visual clarity, the anomaly scores the value of these last two parameters increased line-
(output of the model) were binned and averaged in 1- arly when oxygen limitation was applied, qP values
hour intervals for visual clarity. showed a high increase when applying hypoxia but

remained constant regardless of the hypoxic level. It

Summary of Crl1 and Ethanol Specific Production Rates, and RQ for the Stationary State at Different % O2 in the Inlet Gas Obtained in Chemostat Cultures
0,001 6 0,001 0,012 6 0,002 0,032 6 0,005

1,56 6 0,07
can be stated that, in terms of qP, hypoxia acted as an
935 6 85
on/off switch, generating a fivefold higher qP for SCC
8
and a threefold higher qP for MCC comparing hypoxic
with normoxic conditions. As stated previously, the
condition corresponding to 12% of oxygen in the inlet-
Multicopy Clone (MCC)
1,29 6 0,05
921 6 31 gas acted as a transition between these two states,
showing higher variability in qP values between repli-
10
cates. In Table I, a summary of the main process pa-

rameters reached in chemostat culture is presented.
1,49 6 0,03 1,12 6 0,01 1,13 6 0,02
483 6 181
From these results, the next step was to design a culture

strategy in fed-batch mode with P. pastoris under
12
hypoxic conditions to reach the higher specific produc-

tion rate observed under these operational conditions in
chemostat cultures.
374 6 36
n.d.
21
The experimental setup of the bioprocess is presented

in Figure 2. A feeding glucose solution was added into
the fermenter following a preprogrammed exponential
0,012 6 0,002 0,026 6 0,001
profile through an automatic microburette to maintain a

489 6 42
constant specific growth rate of 0.10 h1. The out-gas

8
was conducted through a gas analyzer to determine the

O2 and CO2 molar fractions. Then, CER, OUR, and RQ
(calculated as a ratio of CER and OUR) were calcu-
Single Copy Clone (SCC)
1,12 6 0,001 1,12 6 0,01 1,30 6 0,01
lated on-line from these measurements. In addition, a

482 6 56
sensor of volatile compounds was coupled to the fer-

10
menter to monitor ethanol concentration. In Figure 2B,

the collection of on-line and off-line data of the biopro-
cess is presented.
194 6 100
n.d.
12
Physiological Control Strategy
From the previous bioprocess results obtained, RQ and

qEtOH were selected as the most promising candidates
106 6 6
n.d.
to be controlled with the goal to maximize specific pro-

21
duction rate (qP). Although both parameters can be

monitored on-line, RQ was selected as the controlled
variable, because it was considered more feasible and
scalable as the control of ethanol production would be
% O2 in the input gas in chemostat
more complex and require some estimation techniques

or software sensors. Considering an industrial scenario,
the control was performed by modifying the stirring
rate instead of the gas flow rate and composition,
mainly with the aim of reducing costs, because lower-
(gEtOH · gX-1· h-1)
ing agitation power would decrease energy demands

(AU · gX-1· h-1)
without the use of extra gases to modify gas composi-

tion in the inlet gas stream.
qp CRL1
qp EtOH
culture
Table I
Thus, to maintain the RQ in the corresponding set point

RQ
value or desired range, the agitation was modified
Vol. 77, No. 3, May--June 2023 151

Figure 2
(A) Experimental setup to perform RQ control in fed-batch bioprocess. (B) Classification of off-line (white
boxes) and on-line (blue boxes) data obtained from the bioprocess. Variables are separated depending on
whether they are direct/indirect measurements (primary variables) or calculated parameters (secondary varia-
bles). Arrows indicate which measurements are used for the calculation of each parameter.
following heuristic rules: to increase RQ, agitation was obtainment of higher specific production rates. However,
reduced to decrease O2 transfer rate to the culture because the work aims to implement a feasible and
media and therefore decrease OUR. To reduce RQ, the robust control strategy, an estimation algorithm for the
opposite action was carried out. Additionally, ethanol proper stirring rate is required.
concentration was also monitored on-line using a Rav-
en’s probe, to double-check hypoxic level conditions. Exploratory Analysis
In accordance with chemostat results previously described A set of 11 fed-batch experiments was carried out at a
(48), preliminary experiments through manual action on constant specific growth rate of 0.10 h1 with the single
stirring rate based on heuristic rules demonstrated that and multicopy clones under hypoxic and normoxic
the control of RQ in fed-batch fermentations allows the conditions. In Table II, the set of available experiments

Table II
Set of Experiments Performed in This Work, Classified According to the Clone Tested: Single-Copy Clone or
Multicopy Clone (First Letter of the Fermentation Code, S/M); and Oxygen Supply Condition: Normoxic and
Hypoxic (Second Letter of the Fermentation Code, N/H). Arrows Indicate the Good (:), Average (!) or Bad (;)
Performance of Each Phase (Batch, Transition, Early Fed-Batch, and Later Fed-Batch)
Fermentation Fermentation
Code Phase Performance Code Phase Performance
Early Later Early Later
Transi- Fed- Fed- Transi- Fed- Fed-
Batch tion Batch Batch Batch tion Batch Batch
NORMOXIC SN-1 : : : fi MN-1 fi : : :
SN-2 : : : : MN-2 : fi : :
SN-3 : : : :
HYPOXIC SH-1 : : ; ; MH-1 : : fi :
SH-2 : : fi : MH-2 : : fi :
SH-3 fi : : : MH-3-V : : : :
SINGLE-COPY CLONE (SCC) MULTICOPY CLONE (SCC)
to feed AI algorithms is presented. MH-3-V was the control actions on the manipulated variable (stirring
experiment to validate the general algorithm under rate) required due to the slow dynamics of the culture.
hypoxic conditions with the MCC.
Phase 4 Later fed-batch. The manipulated variable
However, because the number of experiments to feed AI requires continued actions (increasing its value) to
algorithms was quite limited, each fed-batch experiment maintain the RQ set point. Normally, this phase starts
was divided into four different phases as shown in Figure when the biomass concentration is around 50 g L1.
3. Furthermore, the process was split into phases because
they represent different subprocesses, with significantly Finally, the different phases were classified by the pro-
different dynamics or time evolution characteristics due to cess control engineer as “good”, “average”, and “bad”,
the carbon source used, concentrations of substrates and according to their bioprocess expertise (Table II). The
biomass, and consumption and production rates. phase was classified as “good” if RQ was controlled
within the desired range and the values of all parame-
The main characteristics of the different phases for ters (especially biomass and ethanol concentrations and
each process are detailed as follows: product titer) were those expected. On the other hand,
the phase was classified as “average” if RQ was con-
Phase 1 Batch. This phase aims to generate a signifi- trolled within the desired range, but the value of any
cant amount of biomass growing on glycerol. In the other parameter was significantly different from the
conditions used, the reached biomass concentration is expected (but without having a significant impact on
about 25 g L1. No off-line data is collected in this biomass and ethanol concentrations and product titer).
phase, and no manual actions are made; it can be con- Finally, the phase was classified as “bad” if RQ was
sidered as a routine phase before starting the fed-batch not controlled within the desired range and/or a signifi-
stage. The duration is around 20-24 h. cant impact on biomass and ethanol concentrations and
product titer was observed. For the anomaly detections
Phase 2 Transition. This phase runs between the end of in phase 1, the experiments marked both as “good” and
the batch, once glycerol is completely depleted, and “average” were used, with a goal to detect the “aver-
the beginning of the fed-batch, when the culture is fed age” experiments as anomalies (with suboptimal per-
with a glucose solution. formance) in this phase. To calibrate the AI algorithms
for the operator’s control action to maintain RQ in its
Phase 3 Early fed-batch. This phase lasts 10 h from the set point during phase 4, only the experiments under
start of the fed-batch, and it is characterized by fewer hypoxic conditions marked as “good” and “average” in
Vol. 77, No. 3, May--June 2023 153

Figure 3
The different phases of the bioprocess SH-3 as an example. The vertical lines represent the transitions between
the 4 phases: Batch, Transition, Early fed-batch, Late fed-batch.
phases 3 and 4 were used, not including experiment manually inspected and validated by the process con-
MH-3-V, which was only used for the model validation trol engineer. MH-3-V was not included in this set.
step. Eight of them were labeled as “good” and unlikely to
contain anomalous data points. In contrast, two were
Anomaly Detection: As an alternative to univariate suboptimal, labeled as “average”, and it was hypothe-
anomaly detection, whereby each process variable is sized that these two fed-batches might contain data val-
monitored individually, a multivariate statistics method ues that would be detected as anomalies.
to efficiently summarize the entire bioprocess by trans-
forming all the relevant variables into a single metric Because the behavior of the variables is dynamic, com-
was selected. Ideally, this metric would indicate bining all of the data points into a single model would
whether, and with what certainty, the bioprocess is not be an appropriate strategy, as the isolation forest
anomalous. With this aim, the isolation forest algo- model works by majority rule. Thus, the areas where
rithm (IFA) to implement the outlier detection was variables change faster would be flagged as outliers.
applied (32). The IFA was selected because it has been Because the outlier detection using isolation forest
previously qualified for use for anomaly detection (49). requires stationarity, it was necessary to split the data
into subintervals for which the bioprocess can roughly
Even though there are a few manual operations such as be assumed to be constant. Taking into account that the
inoculation, sampling, eventual foaming correction, batch phase lasts approximately 1 day and there are
and other minor operator’s actions, the batch phase of only slight differences in the duration between differ-
the bioprocess is considered fully automated for this ent batches, it was assumed that all the experiments
analysis. Thus, without considering manual actions, proceeded at the same process rates and consequently,
possible anomalies arise due to system perturbations, the experiments were aligned by absolute time. Thus,
improper functioning of the equipment, and/or unex- data were split into 1-h intervals and isolated forest
pected performance of the microorganism not detected. models for each interval were trained. The training was
The dataset contains 10 fed-batches that were initially made using multiple variables that included operating

Figure 4
Distribution of anomaly scores for each different batch culture colored by the expert according to the classifica-
tion of the run. The batches were sorted by median anomaly score. The vertical line represents the global me-
dian anomaly score.
parameters (primary variables) such as pH, tempera- consistent with the previous knowledge that, except for
ture, pO2, and O2 - CO2 molar fraction in the off-gas the period during which the pH dropped in the time
stream as well as the respirometry parameters (second- window around 15 h, the experiment was labeled as
ary variables): CER, OUR, and RQ. Because no signifi- good.
cant differences were observed when the weights for
individual variables were varied, the models were Overall, the anomaly score based on the isolation forest
trained using equal weight for all variables. Comparing model accurately reflects known anomalous behavior.
the distribution of anomaly scores of individual experi-
ments revealed that some cultures consistently scored AI Modeling Strategy to Predict Stirring Rate for RQ
higher than others. Thus, these cultivations were fur- Control: In contrast to the batch phase, where the bio-
ther investigated. process is considered fully automated, the fed-batch
stages (phases 3 and 4) rely on manual actions to main-
In Figure 4, the distribution of anomalies scored are tain the cell physiological state, especially when the
presented for all experiments, and in Figure 5, the culture is kept in hypoxic conditions. To maximize the
anomaly score over time is plotted. The highest anom- specific production rate in hypoxic bioprocesses, RQ
aly score was obtained in SH-3; pO2 values were over was selected as a controlled variable that was kept con-
100% in the first hours, due to an undetected overpres- stant, and the stirring rate was manually controlled.
sure in the fermenter. In the experiment MH-2, which Under hypoxic conditions, the suitable RQ values were
had the second highest anomaly scores, no untypical selected to be in the wide range of 1.3–1.7, with 1.5 as
behavior was detected by the process control engineer. the RQ set point value.
The third-highest scoring experiment (MN-1) was also
flagged as suboptimal by the process control engineer Due to the manual actions required during the fed-
operator: a malfunction of the base addition pump to batch phase, it is expected to observe higher variability
keep the pH at the set point of 6 was detected. It is among the individual runs. Therefore, the application
Vol. 77, No. 3, May--June 2023 155

Figure 5
Anomaly score of the 10 analyzed batch cultures. The suboptimal experiments are plotted in light blue. The
dashed lines represent the period for which the pH dropped below 5.5 during MN-1. Anomaly scores are aver-
aged by hourly intervals for a cleaner visualization.
of anomaly detection in an unsupervised manner would culture: RQ, CER, OUR, ethanol concentration, and
not be a suitable strategy. Instead, a supervised model- substrate feeding rate.
ing strategy was implemented. The approach was to
give a dataset of successfully completed manually con- A first random regressor model (Model_1) was trained
trolled fed-batches to formulate the problem of finding with three historical datasets from hypoxic fed-batches
the ideal stirring rate at any moment as a supervised including SCC and MCC (SH-3, MH-1, and MH-2).
learning problem. Because the exact relationships The validation of the model was made with data from
between the concerned variables are unknown and very the SH-2 experiment. The predicted stirring rate is pre-
difficult to model without additional information, a sented in Figure 6A showing a similar trend compared
deterministic model that does not account for empiric with the manual actions of the operator, and the mean
observations would likely yield inaccurate predictions. absolute error of the prediction is 57 rpm. However, the
predicted stirring rate is systematically higher than that
This model was built based on the following assump- chosen by the process control engineer (mean error =
tions: 1. Ideal stirring rate at a given time can be pre- +38 rpm), and the model has a clear bias as the error
dicted only from the current state of the bioreactor, distribution is not symmetric around 0 (Figure 6B).
without considering the time evolution of bioprocess This bias of the model predictions likely represents a
variables; 2. no explicit previously known relationship systematic error due to the run-to-run variability.
between variables is needed; 3. in the successful man-
ually operated fed-batches, the actual stirring rate is To evaluate this possibility, a second random forest
considered ideal as long as the RQ is kept successfully regressor model (Model_2), which was trained on all
between 1.3 and 1.7, and the data points where the four historical hypoxic fed-batches (SH-2, SH-3, MH-
value of RQ was out of the range were excluded. 1, and MH-2), was built and evaluated. The same crite-
ria for the inclusion of the data points were used con-
Random forest regressor models were trained to predict sidering them to be in the ideal regime. However, this
the stirring rate based on the five critical variables that time the dataset was evaluated on 25% of the data
are concerned with the physiological state of the cell points that were randomly selected from all four fed-

Figure 6
(A) Comparison of the stirring rate implemented by the process control engineer with the predicted values
based on Model_1 for the SH-2 experiment dataset. (B) Distribution of the error from Model_1 evaluated on
the unseen data (the entire SH-2 dataset).
batches, whereas the remaining 75% of the data points were training of both models were analyzed. The importance
used for training the model. With this evaluation approach, of a feature (variable) is defined as the sum over the
the error of the model on the test set was much lower than number of splits (across all trees in the constructed ran-
for Model_1 (mean absolute error = 5.6 rpm), and the distri- dom forest model) that include the feature, normalized
bution showed no bias (mean error = 0.2 rpm) as can be by the number of samples it splits. As expected, the
observed in Figure 7. highest importances were found for CER, OUR, and
substrate feeding rate, because these three variables
To understand how the model makes predictions and correlate very well with stirring rate due to the relation
what features contribute to its performance, the relative among oxygen transfer rate (OTR), governed by the
importances of each independent variable used in the stirring rate, and the other variables concerned with
Vol. 77, No. 3, May--June 2023 157

Figure 7
Distribution of the error from Model_2 evaluated on the unseen data (randomly sampled data points from all
four historical fed-batch datasets).
cell activity. In contrast, the RQ, which fluctuates broth, and the changes on stirring rate are more frequent
around a rather constant value during the bioprocess, according to the faster dynamics of the bioprocess.
contributes almost nothing to the prediction of the Using the stirring rate predicted by the model, the RQ
model, which is expected for small fluctuations without was successfully kept within the desired operation range
a trend. Additionally, it is important to bear in mind (Figure 9). Bioprocess efficiency in terms of specific
that the RQ is directly related to CER and OUR, production rate of Crl1 was similar to that obtained with
because it is calculated as the quotient between them. heuristic controlled fed-batches (Table III). This result
The ethanol concentration showed a moderate relative demonstrates that the developed AI ML-based modeling
importance. Because it is an indicator parameter of the strategy can be satisfactorily applied to the prediction of
cell metabolism, it is expected that a higher increase is operating conditions such as stirring rate in a bioreactor
produced when a higher limited oxygen transfer rate is to maximize the bioprocess efficiency through the appli-
applied and consequently higher hypoxic conditions cation of physiological control.
are implemented. The ethanol concentration is the vari-
able that fluctuates the most between fed-batches, In Figure 10, the evolution of RQ for all hypoxic fed-
highly contributing to the systematic bias on the model batch experiments and the stirring rates predicted with
prediction (Figure 8). the AI-ML are compared with the ones carried out fol-
lowing heuristic rules. The four manually controlled
Real-Time Prediction and Implementation of the fed-batch fermentations varied significantly around the
Stirring Rate: Finally, to drive the prediction stirring RQ set point of 1.5, although in all the fed-batches the
rate in real time, Model_1 was evaluated in real time to RQ was mostly kept within the interval 1.3–1.7. How-
control the bioprocess. The prediction of the required ever, in MH-3-V, the validation experiment, the fluctua-
stirring rate using the model every 30 min during a 6 h tions of the RQ were lower, within the interval between
time window of late fed-batch (phase 4) was imple- 1.5 and 1.7, normally working at RQ not lower than 1.5.
mented in the validation process MH-3-V. The last 6 h In fact, the trend seems to be slightly above the set point.
of the fed-batch were controlled by the model, because The predicted stirring rates are lower than those applied
in this latter phase the control is most demanding, in most heuristic-guided experiments. Thus, the AI-ML
because the changes on RQ are more important due to strategy can be considered more efficient while keeping
the high biomass concentration present in the culture the RQ within the optimal interval.

Figure 8
Relative importance of the critical variables selected in the Random forest regressor models, Model_1 and
Model_2.
Figure 9
AI-Model prediction of stirring rate to maintain RQ in a mean set point of 1.5 in the validation fed-batch pro-
cess MH-3-V (gray) compared with a heuristic control MH-2 (blue).
Conclusions Revolution (50). The digitization, big data, and comput-

ing power considered as a trigger of the last industrial
The industry in general has identified a great value in the revolution brings process insights, reliable control, and
adoption of the technology provided by the Fourth operation improvement. The use of Internet of Things
Vol. 77, No. 3, May--June 2023 159

TABLE III
Comparison between Heuristic Controlled Fed-Batches and AI ML Model Validation Fed-Batch
Performance COMPARISON
Fermentation code MH-1 MH-2 MH-3-V
qp Crl1 997 6 29 1110 6 8 1061 ± 17
(AU · gX-1· h-1)
RQ 1,37 6 0,13 1,30 6 0,16 1,60 ± 0,13
(IoT), cloud computing, big data technologies, and AI has processes. Thus, these results demonstrate that even
accelerated the mechanisms to orchestrate the huge with scarce data, the IFA can be used as an analytical
amount of data generated in the supply chain. The trans- tool to reliably detect anomalous batches in near real
formation of data into information and finally into knowl- time.
edge is the pinnacle of the digitization journey.
Additionally, an AI/ML modeling strategy based on
The level of control and process understanding needed random forest regressor models has been shown to be a
to deploy a CPV strategy around the biotech process useful tool in the bioprocessing area, especially in mul-
requires a high maturity of digitization, and the techni- tivariate systems requiring feedback in near real time.
ques described in this article support this assumption. In this work, this approach has been successfully
The final goal of a total CPV implementation is to keep applied to the prediction of the stirring rate to maintain
the full process under control with the ability of near the RQ set point, maximizing the specific production
real-time interaction with the system. A first approach rate of a heterologous protein with the P. pastoris cell
to this vision has been deployed in this work, consoli- factory.
dating these ideas and establishing the foundation for a
broad scope in which a full control would drive the bio- This study is related to the first phase of a broad initia-
process using adaptive mechanisms provided by AI. It tive led by the PDA where the final purpose is to
is important to highlight that the AI strategy proposed empirically demonstrate the effectiveness of the appli-
in this research is not replacing the classical univariate cation of digital twins in upstream biomanufacturing
or multivariate statistical process control tools used in operations for drug manufacturing. Digital twin can be
CPV. The implementation of AI techniques is useful to defined as a cyber copy of a physical system that is
complement the existing and validated methods already replicated by a virtual system that interacts with the
in place. Actually, the mechanisms described in the rest of the elements in a similar way as the real object
study explore better process control in combination does. AI is the mechanism that orchestrates the digital
with existing mathematical models. twin, preventing issues and proposing optimal condi-
tions. Under a biopharma perspective, the regulatory
The complexity inherent to biological systems makes it approach is required, and this article is part of the pre-
difficult to manage and control upstream processes liminary approach that will be developed in a second
from a pure analytical perspective and even more so phase, including the compliant vision.
when manual actions are required to keep the process
under control. Although the critical process parameters Acknowledgments
and the expected quality attributes are well-known by
the subject matter experts, the cell metabolism could This work was funded by the Spanish Ministry of
react unexpectedly in light of unknown factors or as a Science and Innovation (Project PID2019-104666GB-
consequence of the complex combination of variables 100) and the Product Quality Research Institute (PQRI).
that are governing the bioreaction. The authors are members of the group Continued Process
Verification of the Future of the Parenteral Drug
The experimental process described in this article Association (PDA). AGF acknowledges the award of
has shown that AI works successfully as a modeling a scholarship (FI-DGR 2019 from Generalitat de
mechanism to detect anomalies in highly automated Catalunya).

Figure 10
Results comparison achieved with the AI-ML model-strategy in fed-batch MH-3-V and the four heuristic-
guided fed-batches. The top graph represents stirring rate (actuation variable), the bottom graph represents
RQ (controlled variable). The dashed lines represent the selected range of RQ values. The AI-ML model-based
prediction of the stirring rate was implemented from 12 h after the beginning of the fed-batch (dotted line).
LOESS smoothing, span = 0.1, was used for the RQ plot.
Vol. 77, No. 3, May--June 2023 161

Conflict of Interest Declaration 10. Vogl, T.; Glieder, A. Regulation of Pichia pastoris
Promoters and Its Consequences for Protein Produc-
The authors have no conflicts of interest to declare. tion. New Biotechnol. 2013, 30 (4), 385–404.
References 11. Cos, O.; Ramón, R.; Montesinos, J. L.; Valero, F.

Operational Strategies, Monitoring and Control of
1. Smith, C. L. Advanced Process Control: Beyond Heterologous Protein Production in the Methylotro-
Single Loop Control; John Wiley & Sons, 2011. phic Yeast Pichia pastoris under Different Promoters:
A Review. Microb. Cell Fact. 2006, 5 (1), 17.
2. Huang, J.; O’Connor, T.; Ahmed, K.; Chatter-
jee, S.; Garvin, C.; Ghosh, K.; Ierapetritou, M.; 12. Ortega, X.; Cámara, E.; Ferrer, P.; Albiol, J.; Mon-
Jeffers, M.; Pla, D. L.; Lee, S. L.; Lovett, D.; tesinos-Seguí, J. L.; Valero, F. Rational Development
Lyngberg, O.; Mack, J.; McManus, E.; Romero- of Bioprocess Engineering Strategies for Recombinant
Torres, S.; Undey, C.; Venkatasubramanian, V.; Protein Production in Pichia pastoris (Komagataella
Warman, M. AIChE PD2MAdvanced Process phaffi) Using the Methanol-Free GAP Promoter. Where
Control Workshop-Moving APCForward in the Do We Stand? New Biotechnol. 2019, 53, 24–34.
Pharmaceutical Industry. J.Adv. Manuf. Process.
2021, 3 (1), e10071. 13. Zhang, A.-L.; Luo, J.-X.; Zhang, T.-Y.; Pan,
Y.-W.; Tan, Y.-H.; Fu, C.-Y.; Tu, F.-Z. Recent
3. Mora, A.; Zhang, S. S.; Carson, G.; Nabiswa, B.; Advances on the GAP Promoter Derived Expres-
Hossler, P.; Yoon, S. Sustaining an Efficient and sion System of Pichia pastoris. Mol. Biol. Rep.
Effective CHO Cell Line Development Platform by 2009, 36 (6), 1611–1619.
Incorporation of 24-Deep Well Plate Screening and
Multivariate Analysis. Biotechnol. Prog. 2018, 34 (1), 14. Garcia-Ortega, X.; Ferrer, P.; Montesinos, J. L.;
175–186. https://doi.org/10.1002/btpr.2584. Valero, F. Fed-Batch Operational Strategies for
Recombinant Fab Production with Pichia pastoris
4. Masood, A.; Hashmi, A. AI Use Cases in the Using the Constitutive GAP Promoter. Biochem.
Industry. In: Cognitive Computing Recipes; Apr- Eng. J. 2013, 79, 172–181.
ess: Berkeley, CA, 2019., pp 383–396.
€ G€unes, H.; Massahi, A.;
15. C alık, P.; Ata, O.;
5. FDA. Process validation: general principles and Boy, E.; Keskin, A.; Ozt€€ urk, S.; Zerze, G. H.;
practices. Guidance for Industry, 2011. €
Ozdamar, T. H. Recombinant Protein Production
in Pichia pastoris under Glyceraldehyde-3-Phos-
6. International Conference for Harmonisation, Qual- phate Dehydrogenase Promoter: from Carbon Source
ity Guideline Q12: Technical and Regulatory Con- Metabolism to Bioreactor Operation Parameters. Bio-
siderations for Pharmaceutical Product Lifecycle chem. Eng. J. 2015, 95, 20–36.
Management. ICH: Geneva, 2017.
16. Huang, M.; Bao, J.; Nielsen, J. Biopharmaceuti-
7. Peña, D. A.; Gasser, B.; Zanghellini, J.; Steiger, cal Protein Production by Saccharomyces cerevi-
M. G.; Mattanovich, D. Metabolic Engineering of siae: Current State and Future Prospects. Pharm.
Pichia pastoris. Metab. Eng. 2018, 50, 2–15. Bioprocess. 2014, 2, 167–182.
8. Juturu, V.; Wu, J. C. Heterologous Protein Expres- 17. Baumann, K.; Maurer, M.; Dragosits, M.; Cos,
sion in Pichia pastoris: Latest Research Progress and O.; Ferrer, P.; Mattanovich, D. Hypoxic Fed-Batch
Applications. ChemBioChem 2018, 19 (1), 7–21. Cultivation of Pichia pastoris Increases Specific and
Volumetric Productivity of Recombinant Proteins.
9. Looser, V.; Bruhlmann, B.; Bumbak, F.; Stenger, Biotechnol. Bioeng. 2008, 100 (1), 177–183.
C.; Costa, M.; Camattari, A.; Fotiadis, D.; Kovar,
K. Cultivation Strategies to Enhance Productivity 18. Baumann, K.; Carnicer, M.; Dragosits, M.; Graf,
of Pichia pastoris: A Review. Biotechnol. Adv. A. B.; Stadlmann, J.; Jouhten, P.; Maaheimo, H.;
2015, 33 (6 Part 2), 1177–1193. Gasser, B.; Albiol, J.; Mattanovich, D.; Ferrer, P.

A Multi-Level Study of Recombinant Pichia pasto- 28. Beiroti, A.; Hosseini, S. N.; Aghasadeghi, M. R.;
ris in Different Oxygen Conditions. BMC Syst. Biol. Norouzian, D. Comparative Study of l -Stat Metha-
2010, 4, 141. nol Feeding Control in Fed-Batch Fermentation of
Pichia pastoris Producing HBsAg: An Open-Loop
19. Garcia-Ortega, X.; Valero, F.; Montesinos-Seguí, J. L. Control versus Recurrent Artificial Neural Network-
Physiological State as Transferable Operating Criterion Based Feedback Control. J. Chem. Technol. Biotech-
to Improve Recombinant Protein Production in Pichia nol. 2019, 94 (12), 3924–3931.
pastoris through Oxygen Limitation. J. Chem. Technol.
Biotechnol. 2017, 92 (10), 2573–2582. 29. Hosseini, S. N.; Javidanbardan, A.; Khatami, M.
Accurate and Cost-Effective Prediction of HBsAg
20. Simutis, R.; L€ubbert, A. Hybrid Approach to Titer in Industrial Scale Fermentation Process of
State Estimation for Bioprocess Control. Bioen- Recombinant Pichia pastoris by Using Neural
gineering 2017, 4 (1), 21. Network Based Soft Sensor. Biotechnol. Appl.
Biochem. 2019, 66 (4), 681–689.
21. Bayer, B.; Striedner, G.; Duerkop, M. Hybrid
Modeling and Intensified DoE: An Approach to 30. Wang, X.; Guo, T.; Hao, W.; Guo, Q. Predicting
Accelerate Upstream Process Characterization. Bio- Model Based on LS-SVM for Inulinase Concentra-
technol. J. 2020, 15 (9), e2000121. tion during Pichia pastoris’ Fermentation Process.
2019 Chinese Control Conference (CCC); Guangz-
22. Brunner, V.; Siegl, M.; Geier, D.; Becker, T. Bio- hou, China; July 27–30, 2019; pp 1531–1536.
mass Soft Sensor for a Pichia pastoris Fed-Batch
Process Based on Phase Detection and Hybrid 31. Cintron, R. Human Factors Analysis and Classifi-
Modeling. Biotechnol. Bioeng. 2020, 117 (9), cation System Interrater Reliability for Biophar-
2749–2759. maceutical Manufacturing Investigations. Ph.D.
dissertation, Walden University, 2015.
23. Solle, D.; Hitzmann, B.; Herwig, C.; Pereira
Remelhe, M.; Ulonska, S.; Wuerth, L.; Prata, 32. Liu, F. T.; Ting, K. M.; Zhou, Z.-H. Isolation-
A.; Steckenreiter, T. Between the Poles of Data- Based Anomaly Detection. ACM Trans. Knowl.
Driven and Mechanistic Modeling for Process Discov. Data 2012, 6 (1), 1–39.
Operation. Chem. Ing. Tech. 2017, 89 (5), 542–
561. 33. Breunig, M. M.; Kriegel, H.-P.; Ng, R. T.; Sander,
J. LOF: Identifying Density-Based Local Outliers.
24. Hong, M. S.; Velez-Suberbie, M. L.; Maloney, SIGMOD Record 2000, 29 (2), 93–104.
A. J.; Biedermann, A.; Love, K. R.; Love, J. C.;
Mukhopadhyay, T. K.; Braatz, R. D. Macroscopic 34. Schölkopf, B.; Platt, J. C.; Shawe-Taylor, J.;
Modeling of Bioreactors for Recombinant Protein Smola, A. J.; Williamson, R. C. Estimating the
Producing Pichia pastoris in Defined Medium. Bio- Support of a High-Dimensional Distribution. Neu-
technol. Bioeng. 2021, 118 (3), 1199–1212. ral Comput. 2001, 13 (7), 1443–1471.
25. Barrigon, J. M.; Valero, F.; Montesinos, J. L. A 35. Lee, J.; Davari, H.; Singh, J.; Pandhare, V. Indus-
Macrokinetic Model-Based Comparative Meta- trial Artificial Intelligence for Industry 4.0-Based
Analysis of Recombinant Protein Production by Manufacturing Systems. Manuf. Lett. 2018, 18, 20–
Pichia pastoris under AOX1 Promoter. Biotech- 23.
nol. Bioeng. 2015, 112 (6), 1132–1145.
36. Dwivedi, Y. K.; Hughes, L.; Ismagilova, E.; Aarts,
26. Council of Europe. European Pharmacopoeia (Ph. G.; Coombs, C.; Crick, T.; Duan, Y.; Dwivedi, R.;
Eur.), 9th edition, Council of Europe: Strasbourg, Edwards, J.; Eirug, A.; Galanos, V.; Ilavarasan, P. V.;
France, 2017. Janssen, M.; Jones, P.; Kar, A. K.; Kizgin, H.; Kro-
nemann, B.; Lal, B.; Lucini, B.; Medaglia, R.;
27. Breiman, L. Random Forests. Mach. Learn. 2001, Le Meunier-FitzHugh, K.; Le Meunier-FitzHugh,
45 (1), 5–32. L. C.; Misra, S.; Mogaji, E.; Sharma, S. K.; Singh,
Vol. 77, No. 3, May--June 2023 163

J. B.; Raghavan, V.; Raman, R.; Rana, N. P.; Samo- pastoris Growing at Different Oxygen Levels. Microb.
thrakis, S.; Spencer, J.; Tamilmani, K.; Tubadji, A.; Cell Fact. 2009, 8 (1), 65.
Walton, P.; Williams, M. D. Artificial Intelligence
(AI): Multidisciplinary Perspectives on Emerging 44. Garrigós-Martínez, J.; Nieto-Taype, M. A.; Gasset-
Challenges, Opportunities, and Agenda for Research, Franch, A.; Montesinos-Seguí, J. L.; Garcia-Ortega,
Practice and Policy. Int. J. Inf. Manage. 2021, 57, X.; Valero, F. Specific Growth Rate Governs AOX1
101994. Gene Expression, Affecting the Production Kinetics
of Pichia pastoris (Komagataella phaffii) PAOX1-
37. Sánchez, A.; Ferrer, P.; Serrano, A.; Pernas, M. A.; Driven Recombinant Producer Strains with Differ-
Valero, F.; Rúa, M. L.; Casas, C.; Solà, C. Charac- ent Target Gene Dosage. Microb. Cell Fact. 2019,
terization of the Lipase and Esterase Multiple Forms 18 (1), 187.
in an Enzyme Preparation from a Candida rugosa
Pilot-Plant Scale Fed-Batch Fermentation. Enzyme 45. Jordà, J.; de Jesus, S. S.; Peltier, S.; Ferrer, P.;
Microb. Technol. 1999, 25 (3–5), 214–223. Albiol, J. Metabolic Flux Analysis of Recombinant
Pichia pastoris Growing on Different Glycerol/
38. Ugo, K. A..; Vivian Amara, A.; Kenechuwku, U.; Methanol Mixtures by Iterative Fitting of NMR-
Cn, I. Microbial Lipases: A Prospect for Biotechno- Derived 13C-Labelling Data from Proteinogenic
logical Industrial Catalysis for Green Products: A Amino Acids. New Biotechnol. 2014, 31 (1), 120–
Review. Ferment. Technol. 2017, 6 (2), 1000144. 132.
39. Nieto-Taype, M. A.; Garrigós-Martínez, J.; Sán- 46. Ponte, X.; Montesinos-Seguí, J. L.; Valero, F.
chez-Farrando, M.; Valero, F.; Garcia-Ortega, X.; Bioprocess Efficiency in Rhizopus oryzae Lipase
Montesinos-Seguí, J. L. Rationale-Based Selection Production by Pichia pastoris under the Control
of Optimal Operating Strategies and Gene Dosage of PAOX1is Oxygen Tension Dependent. Process
Impact on Recombinant Protein Production in Biochem. 2016, 51 (12), 1954–1963.
Komagataella phaffii (Pichia pastoris). Microb. Bio-
technol. 2020, 13 (2), 315–327. 47. Ponte, X.; Barrigón, J. M.; Maurer, M.; Matta-
novich, D.; Valero, F.; Montesinos-Seguí, J. L.
40. Cámara, E.; Albiol, J.; Ferrer, P. Droplet Digital Towards Optimal Substrate Feeding for Heterol-
PCR-Aided Screening and Characterization of ogous Protein Production in Pichia pastoris
Pichia pastoris Multiple Gene Copy Strains. Bio- (Komagataella Spp) Fed-Batch Processes under
technol. Bioeng. 2016, 113 (7), 1542–1551. PAOX1 Control: A Modeling Aided Approach. J.
Chem. Technol. Biotechnol. 2018, 93 (11), 3208–
41. Garcia-Ortega, X.; Adelantado, N.; Ferrer, P.; Mon- 3218.
tesinos, J. L.; Valero, F. A Step Forward to Improve
Recombinant Protein Production in Pichia pastoris: 48. Gasset, A.; Garcia-Ortega, X.; Garrigós-Martínez,
From Specific Growth Rate Effect on Protein Secre- J.; Valero, F.; Montesinos-Seguí, J. L. Innovative
tion to Carbon-Starving Conditions as Advanced Bioprocess Strategies Combining Physiological Con-
Strategy. Process Biochem. 2016, 51 (6), 681–691. trol and Strain Engineering of Pichia pastoris to
Improve Recombinant Protein Production. Front.
42. Cos, O.; Serrano, A.; Montesinos, J. L.; Ferrer, P.; Bioeng. Biotechnol. 2022, 10, 818434.
Cregg, J. M.; Valero, F. Combined Effect of the
Methanol Utilization (Mut) Phenotype and Gene 49. Manzano, T.; Fernàndez, C.; Ruiz, T.; Richard,
Dosage on Recombinant Protein Production in H. Artificial Intelligence Algorithm Qualification:
Pichia pastoris Fed-Batch Cultures. J. Biotechnol. A Quality by Design Approach to Apply Artificial
2005, 116 (4), 321–335. Intelligence in Pharma. PDA J. Pharm. Sci. Tech-
nol. 2021, 75 (1), 100–118.
43. Carnicer, M.; Baumann, K.; Töplitz, I.; Sánchez-
Ferrando, F.; Mattanovich, D.; Ferrer, P.; Albiol, 50. Hahn, G. J. Industry 4.0: A Supply Chain Innova-
J. Macromolecular and Elemental Composition Anal- tion Perspective. Int. J. Prod. Res. 2020, 58 (5),
ysis and Extracellular Metabolite Balances of Pichia 1425–1441.

APPENDIX
TABLE A-I
The Table of Issues and Mitigation Plans Identified during the Risk Assessment
Operations That May Have Impact

Issue Severity Mitigation plan
Equipment is not properly calibrated Critical Large scale: Introduce non-intrusive sensors (e.g.,
vibration) that could lead to the anomalous function.
Small scale: Working with experimented opera-
tors that should not be a problem
No good data quality is available to allow Critical When large time series are used as input data for AI,
application of AI models afterwards it is necessary to get frequent CQA measurements
close to the real process in order to have better
models. Otherwise, if only one CQA measure is
sampled at the end of the process, all the time series
must be collapsed, losing the expected data quality.
Instrumentation not properly calibrated High Use common sense to establish the proper actions to
keep the instrumentation calibrated.
There are manual operations (cleaning, setup, High Add recommendations supplied by AI recommender
sampling) affecting critical operations during the systems
process Automate all those manual actions considered as
medium or high risk.
Assign a risk to each manual action
Take measures of manual actions (operator, dura-
tion, process, task, etc.)
Emphasize trainings for complex and critical tasks
Vol. 77, No. 3, May--June 2023 165

RESEARCH
Flexible Loading Pattern Approach in Overkill Steam

Sterilization Based on the Physical Properties of Steam and
Thermodynamics of Sterilization
ARNAN BEN-DAVID*
10AR – Sterile Engineering Solutions, 7 Yona Volach St., Rosh Haain, Israel © PDA, Inc. 2023
ABSTRACT: Because overkill steam sterilization processes in autoclaves are considered critical, they are highly scruti-
nized, and the use of autoclaves in fixed loading patterns is a common approach to the interpretation of regulatory
requirements. Many such regulations are attributed to tradition and a buildup of restrictions that aim to improve the lev-
els of assurance of the process and minimize risk. However, these measures complicate the operation and qualification
of autoclaves, becoming cumbersome, time-consuming, and costly. In actuality, overkill sterilization is one of several
processes in the pharmaceutical industry that provides the highest levels of sterility assurance. This method provides a
minimum reduction of highly durable spore populations of 12 logs, achieving a probability of a nonsterile unit (PNSU)
of 106. Because these spores are far sturdier than the common microorganisms that can be found in pharmaceutical
facilities, overkill sterilization effects significantly lower PNSU values for the latter. The physical properties of steam
and the thermodynamics of steam sterilization constitute a predictable and repeatable process that can be monitored and
verified. The high assurance level of overkill sterilization, combined with the properties of steam, actualizes a high
safety margin that encompasses nearly every load type and load configuration when the cycle is performed under cer-
tain basic rules. The aim of this article is to present data that advocate and favor an approach that allows greater free-
dom and variability in arranging items in an autoclave when running overkill cycles, without the need to qualify each
configuration.
KEYWORDS: Overkill, Saturated steam sterilization, PNSU, Lethality, Loading pattern, Steam properties, Thermodynamics.
On the Safe Side of Safe Side An overkill cycle is defined in PDA TR-1 (1) as
follows:
Overkill Cycles
“a cycle designed with the overkill design
Steam sterilization is one of the most common and im- approach can be defined as a sterilization cycle
portant practices in parenteral manufacturing, bound that is demonstrated to deliver an FPHY and
by many rules and regulations. In a routine inspection, FBIO of at least 12 min to the items being
the validation, maintenance, and operation of an auto- sterilized.”
clave are usually examined thoroughly. Without down-
grading the criticality of steam sterilization, many require- In other words, it is a cycle with a lethality value (F0)
ments are derived from the definition of “critical”, even that achieves a probability of a nonsterile unit (PNSU)
when there is no scientific basis for the requirements. In of 106 for a spore population of 106 with a D121˚C1 of
other cases, the requirements are extended far beyond 1.0 min—that is, an overkill cycle should demonstrate
what is really needed to be “on the safe side”. 12-log reductions in spores with a D121˚C of 1.0 min. A
cycle with a lethality of 12 min (F0 = 12 min) would
1
* For a steam sterilization processes, the D-value represents the dura-
Corresponding Author: 10AR – Sterile Engineering bility of a microorganism in an environment of steam at a given temper-
Solutions, 7 Yona Volach St., Rosh Haain, Israel; Tele- ature. It refers to the time (in minutes) required at the given temperature
phone: (+972) 54-888-5265: E-mail: arnan@10ar.co.il to achieve a log reduction, that is, to kill 90% of the relevant microor-
ganism population. D121˚C is the D-value when steam temperature is
doi: 10.5731/pdajpst.2021.012708 121˚C.

attain this goal. EU regulations have set forth minimal calculated for a BI and a common microorganism. The
requirements - 15 min at 121˚C for items that are suita- resulting PNSU that is calculated for the common
ble for such conditions (2)—that effect a minimal microorganism is 101598!
lethality value of F0 = 15 min and decrease the PNSU
to <109. Clearly, based on sterility assurance levels, there is no
difference between PNSUs of 106 and 1074, and there
The use of biological indicators (BIs) with D121˚C = 1.0 is no practical benefit in obtaining a PNSU of below
min already establishes a safety margin, because in 106; such a practice appears increasingly academic.
actual situations, few spore-producing microorganisms However, The mathematical calculations reflect the
that are found in pharmaceutical manufacturing facili- combined effect of all safety margins applied. The
ties have a D121˚C that exceeds 0.5 min (1) (such micro- resulting values demonstrate how the desire for safety
organisms will be referred to as facility-durable has elevated and exaggerated the process grossly, adding
microorganisms). Although it might appear that the no real value but imposing massive costs to the industry
PNSU is reduced by half for standard microorganisms, and environment.
in point of fact, a change from a D121˚C of 0.5 to 1 min
decreases the PNSU to 1 trillionth that of the more No other process in the pharmaceutical industry (with
durable microorganism—that is, a cycle that delivers the exception of depyrogenation, in which the same
an F0 of 12 min, as discussed previously, prompting a overly conservative philosophy applies) aims to attain
12-log reduction for a microorganism with D121˚C = such excessive safety levels.
1 min will effect a 24-log reduction for facility-
durable microorganisms with a D121˚C of 0.5 min In terminal sterilization, in which a heat-sensitive prod-
(Calculation - 12/0.5 = 24) . Thus, at 12 min, a uct is sterilized, the cycle is designed to effect signifi-
safety factor is already built into the overkill cycle cantly lower lethality—for example, targeting the F0 to
to allow for the rare (or even unrealistic) presence 8 min. Such a cycle will provide only 8-log reductions
of a resistant spore-forming microbe in the autoclave for a BI with D121˚C of 1 min and a PNSU of 102
load at a concentration of 106, which is the concen- when starting with a population of 106. For facility-
tration of BIs that is used for validation. Com- durable microorganisms with a maximum D121˚C of 0.5
pounded with the additional 3 min of lethality in the min, it will result in 16-log reductions, yielding a
EU requirements, as referenced earlier, we essen- PNSU of <1010 for such microorganisms when start-
tially apply a safety factor to another safety factor ing with the same population of 106. This example sup-
that already buffers a third safety factor. ports the belief that regulatory bodies are willing to
accept lower assurance levels even for the sterilization
The premise of the overkill cycle was extended again, of final products, not considering it a risk.
and instead of being calculated based on a D121˚C of 1
min, the typical Geobacillus stearothermophilus BI The rationale of the overkill cycle was that it would
that is in use has a D121˚C in the range of 1.6–2.5 min allow more flexibility in sterilization, due to the high
(3). If D121˚C is 2.5 min in the calculations, a 12-log safety margins that it provides to compensate for
reduction will require a cycle with F0 = 30 min. Such a potential variation in this process. The premise of over-
cycle will yield a PNSU of 10−54 for facility-durable kill cycles is that the bioburden population and its re-
microorganisms (Calculation: for those microorgan- sistance can be ignored, because they are irrelevant to
isms with a D121˚C of 0.5 min, a cycle of 30 min will re- the outcome. This genuine benefit is eliminated,
alize 30/0.5 = 60-log reduction; thus, when starting at a because in actuality, overkill sterilization cycles are re-
population of 106, 60-log reductions will achieve a stricted by additional regulatory requirements and
PNSU of 1054). “expectations” to inflate the safety margin. The
requirement for fixed loading patterns is one of the
In certain facilities, an F0 value of 30 min is the mini- most burdensome of these principles.
mum allowed value, and to be “on the safe side”, the
cycles are designed at an F0 of 40 min, yielding a Requirements of Fixed Loading Patterns
PNSU of 1074 for facility-durable microorganisms.
The USP (4) provides such an example, in which the In a fixed loading pattern, the items that are to be
PNSU of a sterilization cycle with F0 = 8 min is sterilized are arranged according to a scheme or
Vol. 77, No. 3, May--June 2023 167

depiction that shows their exact number, size, posi- The requirement for fixed loading patterns is triggered
tion, and orientation in the chamber. This arrangement by the belief that there is an interaction between items,
is validated, and if all the acceptance criteria are met, that one item can influence or interact with an adjacent
the load is approved for production. If a new item is item, and that the item per se might not be sterilized in
added to a validated load, the load is considered to be a different area in the autoclave—that is, that items can
new, and the validation must be repeated. The same “steal” steam from other items, creating a cold spot or
process applies to items that have been validated in restricting steam flow to other items. For example,
other loads and, due to changing production require- heavy items could be believed to be “steam-consum-
ments, must now be sterilized with other items in a ing” items that reduce the supply of steam to other
different load. items. This suspicion then leads to the requirement that
many loading patterns be validated and maintained
The roots of this method date back to the 1970s, when through revalidation activities.
many practices that were related to large-volume par-
enteral (LVP) sterilization were universally adopted by The standard validation procedure that is used by cer-
the small-volume parenteral (SVP) industry, regardless tain firms is to run three distribution cycles and three
of their use in a different setting (5). At that time, when penetration cycles for each new load, rendering the
gravity displacement autoclaves were used in many workload and downtime of the facility to be significant
facilities, this principle was justified. Since then, auto- and costly. Reducing nonvalue-added validation activ-
clave design has evolved considerably, with the advent ities can create significant savings for facilities, with
of vacuum and steam pulses to remove air from porous no adverse impact on patient safety.
loads and greater control over temperature and pressure
in the chamber. In contrast, validation practices have Several studies have examined the requirements for
remained almost unchanged. fixed loads in reducing the workload. Agalloco, Akers,
and Madsen have described such requirements to be a
Such regulations as the EU Annex 1 on Sterile Manu- myth (8). Pavell and Hughes (9) discussed this topic,
facturing (6) require validated loading patterns to be based on maximal load, using mass as the criterion.
applied: These works have provided a means to incorporate
more flexible loading but have nevertheless failed to
“Validated loading patterns should be estab- address certain complexities and limitations.
lished for all sterilization processes and should
be subject to periodic revalidation. Maximum Notably, Pavell and Hughes (9) state that:
and minimum loads should also be considered
as part of the overall load validation strategy. . . “Lacking definitive data on how much variation
Routine operating parameters should be estab- is allowable without affecting sterilization effi-
lished and adhered to for all sterilization proc- cacy, this is a difficult if not impossible ques-
esses, eg, physical parameters and loading tion to adequately answer and thus presents a
patterns.” significant obstacle to effective training of equip-
ment preparation operators. Although reasonable
Further, the WHO (7) states that “validated loading and rational scientific thought supports the abil-
patterns should be established for all sterilization ity to relocate but not reorient objects, the fixed-
processes.” load approach initially utilized did not provide
sufficient data to identify process boundary
In certain firms, only exact, validated loads are allowed conditions.”
to be run—if for example, just 1 of the 20 items that
are validated in the full load is needed, the entire load This article will attempt to challenge this difficult ques-
must be prepared (i.e., washed, bagged, and arranged) tion with data and convincing arguments, suggesting
and sterilized. To this end, some firms use maximal an approach that allows greater freedom and variability
and minimal load cycles (Max-Min approach), which in arranging items in an autoclave when running over-
provides more flexibility. But, when a new item is kill cycles, without the need to qualify each configura-
required or validated loads must be combined, the new tion or any adverse impact on patient safety or the
load needs to be validated again before use. required sterility assurance levels.

Physical Properties of Pure Saturated Steam and the The only interaction that might occur between items
Thermodynamics of Sterilization is the condensate, which can drip from heavy nonpo-
rous items onto lower items; this situation can be
Steam sterilization relies on the physical properties of resolved easily by placing the former on the lower
pure saturated steam and the thermodynamics of the shelf of the autoclave.
sterilization process. Any claim that relates to the abil-
ity of steam to sterilize should be evaluated in light of 5. Heat can be transferred between objects only when
its properties, as described following. there is a difference in temperature (12). Once the
items in the autoclave attain the temperature of the
1. Steam releases a large amount of energy relative to steam, no more energy is transferred to the item, and
its weight when it condenses. A small amount of further condensation stops—an equilibrium is reached.
steam can heat a large load, as well as the metal Additional small quantities of steam enter the autoclave
structure of the autoclave. As it condenses, 1 kg of only to compensate for the heat loss to the doors and
saturated steam releases 2200 kJ, which, when con- chamber walls (even when the jacket is heated, because
densed, can heat over 40 kg of stainless steel (SS) the temperature of the jacket is lower than that of the
from 20˚C to 121˚C (10); thus, heating “heavy” chamber).
metal items does not require large amounts of
steam. 6. After an equilibrium is reached (after the heating
phase and at the outset of the sterilization phase),
2. When saturated steam is condensed, its volume pressure across the chamber will be equal, according
decreases significantly (1 L of saturated steam turns to Pascal’s law.
into approximately 1.2 mL of condensate) (11).
Whenever condensation occurs (when steam makes 7. Steam and air (primarily nitrogen and oxygen) are
contact with cold areas and condenses), the reduction small molecules (>1 nm); thus, air removal and
in volume will bring more steam to cold sections due steam penetration can also be achieved through nar-
to the decrease in pressure until the item reaches the row channels or gaps. Large quantities of steam
same temperature as the steam and no further heating require wider passageways, but even with small
is needed. gaps, steam will eventually penetrate all parts of an
object.
3. There is a 1:1 correlation between saturated steam
pressure and its temperature in an enclosed space With regard to load, other physical properties are
with constant volume. After the pressure stabilizes, involved in steam sterilization. Table I lists the load prop-
temperatures in the chamber become steady and will erties of items and their effects on sterilization. Combin-
be maintained throughout the chamber as a result of ing these properties simplifies the decision over which
the stabilization of pressure. items are easy or more challenging to sterilize. Table II
lists examples of common items. Note that the table
4. Steam that is under pressure in an enclosed space merely compares items and does not definitively state
does not flow like air; thus, there are no steam flow whether an item is challenging for an overkill cycle.
patterns in the autoclave. Items do not obstruct steam
or prevent it from entering other objects. Providing a In summary:
load does not block the passage of steam com-
pletely—steam will reach all items. As indicated in 1. Steam can heat a heavy load at low quantities due to
Point 2, whenever condensation arises, the steam that high latent heat of condensation (Point 1). Fresh
is nearest to the condensation point will move toward steam will fill the low-pressure zones that are created
it. These movements occur on a molecular level and due to condensation (Point 2), supplying more
are thus rapid; consequently, steam molecules are not energy. Heavy loads might elongate the heating
influenced by the other items. Combining this inher- phase, but this extra time is not included in the sterili-
ent property of steam with the property in Point 3 zation phase time. Eventually, after the load reaches
ensures that wherever items are positioned in the the sterilization temperature, there is no difference
chamber, each element will be subjected to the same between heavy and light loads, because there is no
sterilization conditions. further transfer of energy to the load (Point 5). Mass
Vol. 77, No. 3, May--June 2023 169

Table I
Load Item Properties and Their Effect on Sterilization
Higher
Property Meaning Values Effect on sterilization
Mass Mass of item Higher Greater energy to heat the item and
mass higher steam demand
Specific The amount of energy required to elevate Higher heat Greater energy to heat the item and
heat the temperature of an item by 1˚C per unit capacity higher steam demand
capacity weight.
Specific How energy is transferred in an item, for Higher heat Faster heating of the item; heat
thermal example, from the outside surface to the transfer penetrates the item much faster.
conductivity inside of the item or from the steam to the
item
Material Internal pores, channels, or lumens an Higher Restricts air removal and steam
porosity item has porosity penetration. The length of channels also
- more has an effect, and long lumens are more
smaller difficult to sterilize.
pores, Porous materials may also retain
longer condensate within.
lumens
Air volume How much air is entrapped within a Higher Air removal and steam inlet will take
inside item hollow item with only vents to allow air volume longer time.
removal and steam inlet Condensate may accumulate inside.
Table II
Decision Whether an Item is Easy or Hard to Sterilize, Based on an Overall Assessment
Specific Specific
Heat Thermal Material
Item Mass Capacity Conductivity Porosity Overall Assessment
Stopper bowl H L H L Relatively easy to sterilize but will drip condensate
Filter L M L H Relatively hard-to-sterilize item
Long silicone L M L H Relatively hard-to-sterilize item
tube
Arthroscopic L L H H Relatively medium–hard-to-sterilize item
metallic
instruments
Isolator glove L M L L No issues if left open. Glove rubber is not permeable
to steam, so in case the glove is pressed or bent,
steam penetration may be restricted.
Closed SS M–H L H L Vessel should be equipped with enough venting
vessel with ports to allow more rapid air removal and steam
few vents entrance. Air removal process should include more
pulses to allow efficient air removal.
If the steps above are done properly, the item is
relatively easy to sterilize.
L, Low; M, Moderate; H, High.

should not be the chief parameter for selecting chal- Steam purity is influenced by how steam is produced and
lenging items. The results of the tests in Pavell and transferred and by efficient air removal from the chamber
Hughes (9) also showed that the more challenging and load. Assuming that steam is tested on a regular basis
items were the filter and the cleaning hose with a for low levels of NCGs2, proper air removal should be
valve on one end, neither of which was among the ensured before starting the sterilization phase. In most
heaviest items. cases, air removal is more important than additional sterili-
zation time and increased temperature. The PNSU calcula-
2. One of the most common myths that concern steam tions in the first section are valid only with efficient air
sterilization is that the high lethality of steam is removal so that the full lethality of steam is applied.
attributed to the high energy transfer to the sterilized
object during condensation (13): “Direct steam con- Air removal and steam penetration are complemen-
tact with the surface of the object to be sterilized is tary, because the same channel that allows air to be
required for the steam to transfer its stored energy to removed from internal voids allows steam to enter
the object.” Direct contact is definitely required, these voids.
because the moisture in the steam must make contact
with the item, but the last section of this statement The level at which air is not considered to be harmful
becomes irrelevant when the setpoint temperature depends on the load. Small air bubbles that are dis-
has been attained by all load items. As indicated in solved in steam are not a problem, but their accumula-
Point 5, energy transfer occurs only during the heat- tion in closed packs or pockets can become an issue
ing phase; throughout the sterilization phase, there is (16).
no energy transfer, and microorganisms are eradi-
cated by maintaining the temperature and water mol- Effective air removal and steam penetration are
ecules contact for the required time. achieved by the repeated evacuation and entrance of
steam into the voids. Collectively, this process entails
3. From Points 2, 4, 6, and 7, I conclude that there is dilution of the air in steam and subsequent extrusion of
no impact of the location of items in the autoclave the mixture. The efficacy of the process is dictated by
with regard to achieving the same extent of sterili- two factors:
zation. The only parameter that requires attention
relates to the prevention of condensate dripping. 1. Values of pressure for the vacuum and steam pulses.
This conclusion was also the outcome of Pavell and Larger differences between these values affect
Hughes (9) greater efficacy. The values for the remaining air in
the chamber (not inside of the load) can be calculated
Air Removal and Steam Penetration theoretically using the pressure values and can pro-
vide a good indication of the effectiveness of the pro-
Steam sterilization is based on the use of pure steam. cess. (In many modern autoclaves, the vacuum pump
When air or other noncondensable gases (NCGs) are continues to operate throughout the complete cycle
present, they will affect the sterilization on several to remove condensate, which improves results com-
levels: pared with the calculated value. To compensate for
the pumping, more steam is forced into the chamber,
1. Steam will follow the physical properties described mixed with the remaining air in it, and eventually
above only when it is pure. When it is impure, the sucked out by the pump.) For example, a process
outcome is not predicted as easily. with an initial vacuum of 10 kPa and 4 pulses in the
range of 90 to 20 kPa will leave 0.03% air. The same
2. Effective steam sterilization is predicated on the process with pulses in the range of 120 to 20 kPa will
steam being in contact with the item. NCGs create leave only 0.01% air in the chamber.
isolation and can limit the destructive effectiveness
of steam.
2
The requirements for the NCGs content in the steam supply are strict
(14) and stipulate that the level of NCGs in steam be below 0.003% (M/
3. Steam lethality is attributed to the presence of water M) (15). This value is calculated from the reference value of 3.5%,
molecules. NCGs reduce steam lethality because which is a volume-to-mass ratio (volume of NCGs vs. mass of
condensate).
they do not include water molecules.
Vol. 77, No. 3, May--June 2023 171

Figure 1
Ten pulses air removal process.
2. Pulse time. Slow pulses are more effective in remov- Item Preparation, Wrapping, and Orientation
ing air from challenging items, such as long lumens,
porous loads, and packed items, because physical Preparation, wrapping, and orientation are significant
restrictions slow the evacuation and steam penetra- factors in the success of the sterilization cycle, because
tion. This property is the chief reason that a minimal they impact air removal, steam penetration, and con-
load is required in an overkill cycle. In minimal loads, densate removal (8). An item that is validated after
the evacuation time and steam pulse are shorter, being prepared, wrapped, and oriented in the chamber
impeding the air removal from and steam penetration according to specific instructions should always be
of challenging items. In certain autoclaves, pulses du- sterilized in the manner that it was validated. For
ration can be set and be equal, regardless of load size, example, a tube coiling of a specific length should be
but in many autoclaves, this option does not exist. In inserted into a double bag and placed flat on the auto-
such cases, this issue can be overcome through redun- clave shelf. A validated item includes its preparation,
dancy by adding more pulses or increasing the differ- wrapping, and orientation.
ence between the vacuum and steam pressures,
rendering even shorter pulses sufficiently effective.
Data Collection and Interpretation
An effective cycle for even the most difficult loads is a
10-pulse cycle using 5 subatmospheric pulses and 5 Data were collected from actual validation runs that
superatmospheric pulses, such as that in Figure 1. were performed at a specific facility from 2017 to
2019, as presented in Table III. The data are used to
This process will reduce air to a level of <0.004% (cal- assess differences between maximal and minimal load
culated value), help heat the load before entering the results.
sterilization phase, and provides sufficient redundancy
in the case of shorter pulses. The maximal load is prepared by collecting the more
challenging items (items with a minimal F0 value)
In an overkill cycle, air removal is the main parameter from all validated loadings.
when determining whether an item is simple or difficult
to sterilize. When air removal is conducted properly, The minimal load usually includes one item that is
steam will penetrate all voids, eliminating “hard-to- found to be more challenging during the run of the
sterilize or challenging items.” maximal load.

Figure 2 shows examples of maximal and minimal Note that the validation runs were performed using
loads. the “minimal cycle” concept, wherein the cycle times
were shortened or the sterilization temperature set
Data were collected from: point was reduced by 1˚C from the routine production
parameters.
6 autoclaves, varying in size and age
Figure 3 shows four types of air removal processes that
4 types of loads—garments, gloves, tools, and rubber were included in the data collection.
stoppers
The decision on using a specific air removal process is
4 types of air removal processes related to the autoclave age and load. Older autoclaves
use the simple but less effective process of 3 pulses. In
Cycles from 2017, 2018, and 2019 certain cases, the more effective “3 pulses + steam
injection” process was adopted.
Each line in the data collection table relates to a valida-
tion activity that was performed for a specific load Subsequently, the effective “10-pulse” process was
type. The maximal and minimal F0 values that were found to be required for more complex loads to comply
obtained for the maximal and minimal loads were with the validation requirements of equilibration time
collected. and was thus chosen for most new loading patterns.
According to the F0 values that were obtained, a corre- The “3 pulses reduced vacuum rate” process is applied
lated temperature (CT) was calculated. The CT is a sin- only for the sterilization of rubber stoppers.
gle temperature value that provides an equivalent F0
value for the entire sterilization phase. There are several insights from the data collection:
The CT is calculated per the following formula: 1. Minimal load F0 values were within the values that
CT ¼ ðlog10 F0 log10 Ts Þ 10 þ 121:1 were obtained in the maximal load (the highest and
lowest F0 values were found in the maximal load
where Ts is the sterilization phase time in minutes.
cycles) (Figure 4). These data indicate that faster air
removal and faster heating, which characterize the
Example:
minimal cycle, do not necessarily provide the lowest
lethality values.
An item attains an F0 value of 43.5 min in a steriliza-
tion cycle of 40 min.
2. In certain cases, the differences in F0 values

CT ¼ ðlog10 43:5 log10 40Þ 10 þ 121:1 ¼ 121:46 C between the hottest and coldest points in a specific
cycle might appear to be significant, but when CT
values are considered, the differences are small
Check: (Figure 5).
F0 ¼ 40 10ð Þ
121:46121:1
10 ¼ 43:5 min: 3. Because the thermocouples that are used for the PQ
study are calibrated to 60.5˚C, some differences can
The difference between the two CT values of the be related to calibration tolerances; when this factor
maximal and minimal F0 values was calculated (CT is considered, the results indicate excellent tempera-
DT). ture uniformity inside all of the loads.
Equilibration time (Equil. time)3 and air removal phase 4. The correlation was low (0.26) between the equili-
time were also collected. bration time and correlated temperature differences.
It would be expected that a short equilibration time
3
Equilibration time– the period that elapses between the attainment of indicates faster heating of the items, thus reducing
the sterilization temperature in the sterilizer chamber and that at all temperature differences, but the low correlation does
points within the load (17).
not support this relationship. Pavell and Hughes (9)
Vol. 77, No. 3, May--June 2023 173

Table III
174
Data Collection from Actual Validation Runs
Sterilization Cycle Details Maximal Load Minimal Load

Air Air
Steril. Steril. Max Min CT Equil. Removal Max Min CT Equil. Removal
Load Temp Time Air Removal F0 CT F0 CT DT time Phase F0 CT F0 CT DT time Phase
Autoclave Year Type [˚C] [min] Phase Type [min] [˚C] [min] [˚C] [˚C] [sec] Time [min] [˚C] [min] [˚C] [˚C] [sec.] Time
1 2017 Garments 124 20 3 pulses 46.9 124.8 43.6 124.5 0.3 19 0:20:05 47.6 124.9 46 124.7 0.1 14 0:10:51
820L 2018 Group 46.4 124.8 44.2 124.5 0.2 22 0:17:36 46.3 124.7 42.8 124.4 0.3 0 0:15:27
2019 45.3 124.6 43.9 124.5 0.1 2 0:20:44 44.9 124.6 43.9 124.5 0.1 0 0:12:50
2017 Gloves 124 20 10 pulses 50.5 125.1 43.4 124.5 0.7 6 0:20:26 44.9 124.6 44.6 124.6 0 2 0:20:27
2018 Group 44.1 124.5 40.6 124.2 0.4 0 0:24:58 42.9 124.4 38.6 124 0.5 0 0:20:21
2019 46.8 124.8 42 124.3 0.5 2 0:25:28 42.9 124.4 42.1 124.3 0.1 1 0:20:48
2017 Tools 122 30 3 Pulses + 46.6 123 40.8 122.4 0.6 34 0:20:36 43.7 122.7 39.7 122.3 0.4 2 0:23:09
2018 Group Steam 44.7 122.8 39.5 122.3 0.5 2 0:22:43 42.1 122.6 40.7 122.4 0.2 2 0:22:43
2019 injection 47 123 40.8 122.4 0.6 4 0:22:16 42.2 122.6 41.4 122.5 0.1 5 0:21:17
2 2017 Garments 121 40 10 pulses 41.6 121.3 39.3 121 0.3 1 0:36:25 40.1 121.1 39.4 121 0.1 0 0:22:37
820L 2018 Group 44 121.5 41.5 121.3 0.3 7 0:36:44 43.3 121.4 42.8 121.4 0.1 4 0:25:21
2019 43.4 121.5 38.5 120.9 0.5 12 0:34:09 41.4 121.2 40.3 121.1 0.1 4 0:20:57
2017 Tools 121 40 10 pulses 41.1 121.2 38.3 120.9 0.3 2 0:30:03 39.7 121.1 39 121 0.1 5 0:18:20
2018 Group 45.9 121.7 39.8 121.1 0.6 14 0:30:40 44.2 121.5 41.4 121.2 0.3 7 0:21:47
2019 43.8 121.5 39 121 0.5 10 0:29:00 43.7 121.5 40.9 121.2 0.3 7 0:15:11
3 2017 Garments 121 40 3 pulses 47.3 121.8 44.3 121.5 0.3 6 0:20:36 46.1 121.7 46 121.7 0 1 0:14:33
1,500 L 2018 Group 47.7 121.9 45.3 121.6 0.2 10 0:20:47 48.3 121.9 47.2 121.8 0.1 1 0:14:41
2017 Gloves 121 40 3 pulses 47.7 121.9 44.9 121.6 0.3 6 0:16:54 46 121.7 45.7 121.7 0 1 0:13:43
2018 Group 47.6 121.9 45.6 121.7 0.2 5 0:14:53 47.2 121.8 46.2 121.7 0.1 NA 0:14:29
2019 48.7 122 47.1 121.8 0.1 4 0:19:09 49.3 122 47.8 121.9 0.1 1 0:19:41
2017 Tools 121 40 3 pulses 49.2 122 45 121.6 0.4 39 0:18:35 47.8 121.9 45.4 121.6 0.2 3 0:13:44
2018 Group 48.7 122 45.1 121.6 0.3 1 0:17:08 47.1 121.8 45.6 121.7 0.1 6 0:12:34
2019 48.8 122 45.5 121.7 0.3 13 0:20:30 48.8 122 46.9 121.8 0.2 2 0:14:26
4 2017 Garments 122 27 10 pulses 38 122.6 36.3 122.4 0.2 4 0:34:06 37.3 122.5 37.2 122.5 0 0 0:21:45
820 L 2018 Group 39.6 122.8 37.8 122.6 0.2 2 0:31:55 39.2 122.7 38.2 122.6 0.1 1 0:23:31
2019 41 122.9 37.5 122.5 0.4 14 0:37:33 40.6 122.9 39.1 122.7 0.2 6 0:22:56
2017 Tools 122 27 10 pulses 39.7 122.8 36.2 122.4 0.4 10 0:29:00 38.6 122.6 34.7 122.2 0.5 11 0:23:11
2019 Group 41.8 123 38.4 122.6 0.4 24 0:23:24 41.7 123 39.8 122.8 0.2 3 0:19:46

Table III
(continued)
Sterilization Cycle Details Maximal Load Minimal Load

Air Air
Steril. Steril. Max Min CT Equil. Removal Max Min CT Equil. Removal
Load Temp Time Air Removal F0 CT F0 CT DT time Phase F0 CT F0 CT DT time Phase
Autoclave Year Type [˚C] [min] Phase Type [min] [˚C] [min] [˚C] [˚C] [sec] Time [min] [˚C] [min] [˚C] [˚C] [sec.] Time
5 2017 Garments 121 40 3 pulses 50.3 122.1 48 121.9 0.2 0 0:27:36 49.9 122.1 48.3 121.9 0.1 1 0:19:21
1,750 L 2018 Group 50.2 122.1 48.3 121.9 0.2 3 0:23:09 48.9 122 47.6 121.9 0.1 2 0:17:33
Vol. 77, No. 3, May--June 2023

2019 50.6 122.1 47 121.8 0.3 1 0:23:03 51 122.2 49.3 122 0.2 1 0:16:57
2017 Gloves 121 40 3 Pulses + 49.4 122 46.9 121.8 0.2 16 0:33:52 50.7 122.1 48.9 122 0.2 1 0:33:11
2019 Group Steam 50.7 122.1 48.7 122 0.2 2 0:33:17 50.9 122.1 48.7 122 0.2 3 0:32:07
injection
2017 Tools 121 40 3 Pulses + 55.5 122.5 46.2 121.7 0.8 9 0:34:36 49.1 122 46.7 121.8 0.2 2 0:29:46
2018 Group Steam 50.6 122.1 47.6 121.9 0.3 4 0:34:39 50.5 122.1 47.1 121.8 0.3 1 0:29:32
2019 injection 58.8 122.8 48.8 122 0.8 8 0:35:11 50.8 122.1 47.9 121.9 0.3 1 0:29:25
2018 Stoppers 121 40 3 pulses 42.8 121.4 40.1 121.1 0.3 3 0:26:49 42.6 121.4 40.8 121.2 0.2 2 0:12:26
2019 Group 1 44.7 121.6 41.7 121.3 0.3 4 0:15:08 44.8 121.6 43.9 121.5 0.1 2 0:13:45
2017 Stoppers 121 40 3 pulses 41.6 121.3 39.6 121.1 0.2 3 1:07:11 40.8 121.2 39.7 121.1 0.1 2 0:31:21
2018 Group 2 reduced 45.1 121.6 41 121.2 0.4 2 0:55:12 42.5 121.4 40.9 121.2 0.2 1 0:35:32
2019 vacuum rate 44.8 121.6 42 121.3 0.3 20 0:55:23 44.8 121.6 43.6 121.5 0.1 3 0:35:40
6 2017 Garments 121 40 3 pulses 40.7 121.2 39.4 121 0.1 1 0:16:19 41.2 121.2 40.3 121.1 0.1 0 0:12:57
435 L 2018 Group 121 40 3 pulses 43.4 121.5 40.2 121.1 0.3 0 0:16:01 41.7 121.3 40.5 121.2 0.1 0 0:15:09
2019 121 40 3 pulses 41.6 121.3 41 121.2 0.1 1 0:14:48 46.5 121.8 43 121.4 0.3 2 0:12:09
2017 Tools 121 40 3 Pulses + 43.5 121.5 41.9 121.3 0.2 2 0:24:52 43.4 121.5 42.2 121.3 0.1 2 0:24:39
Group Steam
injection
2018 121 40 3 Pulses + 45.8 121.7 41.8 121.3 0.4 0 0:25:21 44.6 121.6 42.2 121.3 0.2 0 0:25:48
Steam
injection
2019 121 40 3 Pulses + 44.5 121.6 42.6 121.4 0.2 1 0:24:59 44.5 121.6 44.4 121.6 0 1 0:25:20
Steam
injection
175
Figure 2
Examples of maximal and minimal loads.
also observed this finding. It would be interesting to and between items in the load are small. The large dif-
understand the reason, scientifically. Practically ferences in the mass or density of arrangement of the
speaking, the lack of correlation has no effect on the items between the maximal and minimal loads did not
overall conclusions that are related to the overkill have any effect or direction that was related to the
cycle. lethality that was gained by the items.
The results of the data collection and analysis show When evaluating the results for an overkill cycle, the
clearly that the differences in lethality between loads differences are negligible. No item can be considered

“hard to sterilize” or “challenging”, because all items

achieved lethality values far beyond the initial require-
ments of the overkill cycle (F0 ≥ 15 min, which also
covers the EU requirements) and the sterility assurance
level that was required.
Conclusions
Overkill cycles, as currently used in industry, are ex-

cessive, attaching overly conservative safeguards to
safeguards. The approach of allowing flexible loads
for validated items, if implemented with the princi-
ples following, will result in safe and effective steri-
lization cycles at far lower cost to industry and the
environment.
The overkill cycle provides safe and robust sterilization

for any load configuration if certain simple and basic
Figure 3 principles are always followed:
Four types of air removal processes used in the 1. The cycle should be an extended overkill cycle. To
data collection. allow load flexibility, more robust overkill cycles,
Figure 4
Distribution chart of F0 values obtained during the validation runs.
Vol. 77, No. 3, May--June 2023 177

Figure 5
Distribution chart of the correlated temperature differences in maximal and minimal loads.
with a calculated F0 value >25 minutes,4 are recom- 5. The air removal stage should be designed to provide
mended to ensure the destruction of Geobacillus redundancy, such that even with fast pulses, it will be
stearothermophilus BIs. efficient. (If the autoclave is capable of controlling
the timing of the pulses to be equal for maximal and
2. A validated item is one that has been validated by minimal loads, the minimal load cycle validation is
three penetration cycles (using thermocouples and no longer relevant, and it could be skipped or be per-
BIs that are inserted into adjacent locations inside formed only once to ease regulatory questions [19]).
of the item) in any loading pattern. The term “item”
includes the specific wrapping that is used when 6. Revised or new loads are created under a change con-
validated. trol procedure or a detailed SOP that includes full
documentation and pictures of the revised or new
3. The item must be placed in the same orientation in load.
which it was validated and in a manner that will not
block air or steam passage to other items. The overall conclusion can be summarized as follows:
4. Due to greater condensation from heavy metal items,
If the preceding principles are followed, an item that has
they should be placed on the lower shelves to avoid
been validated in a specific overkill cycle will be effec-
dripping onto another item.
tively sterilized in any other load configuration using the
same sterilization cycle parameters (including the air
4
The 25 minutes was selected from the kill time calculation of Geoba- removal phase parameters), wrapping, and orientation.
cillus stearothermophilus BI. The kill time for a BI is calculated as:
Kill time = (log10 N0 + 4) x D121˚C (18), where N0 is the initial popula-
tion. When taking 106 for the N0 and D121˚C = 2.5, the kill time = No additional validation is required to ensure that the
(log10(106)+4) x 2.5 = (6+4) x 2.5 = 25 minutes.
combination of such items will be sterilized properly.

Figure 6
Flexibility options for new loading patterns.
Consequently, several options are created when apply- Conflict of Interest Declaration
ing the preceding principles, such as the following:
The author declares that he has no competing interests.
1. A new item can be tested alone, running 3 penetra-
tion cycles, after which it can be multiply or com- References
bined with any other validated item.
1. Parenteral Drug Association Inc. PDA Technical
2. A combination of any validated item in any number Report No. 1: Validation of Moist Heat Steriliza-
is allowed tion Processes: Cycle Design, Development, Quali-
fication and Ongoing Control; Bethesda, MD,
The two options are demonstrated in Figure 6. 2007.
2. European Medicines Agency, Decision Tree for

Allowing facilities to use the preceding approach will
Sterilisation Choices for Aqueous Products. In
provide flexibility in their operation and can signifi-
Guideline on Sterilisation of the Medicinal
cantly reduce the validation workload without any Product, Active Substance, Excipient and Pri-
adverse impact on patient safety. mary Container. EMA: London, 2015; Chapter
5, Figure 1, EMA/CHMP/CVMP/QWP/850374/
2015.
Acknowledgments
3. U.S. Food and Drug Administration, Guidance for
The author would like to express his sincere gratitude to Industry and FDA Staff. Biological Indicator (BI)
James Agalloco for his guidance and support in writing Premarket Notification [510(k)] Submissions. Cen-
this article, providing insightful comments and resources. ter for Devices and Radiological health. U.S.
Special thanks to Karen Ginsbury and Phil DeSantis for Department of Health and Human Services: Rock-
their technical review and important input. ville, MD, 2007.
Vol. 77, No. 3, May--June 2023 179

4. U.S. Pharmacopeial Convention, General Chapter Michigan State University: East Lansing, MI,
<1229.2> Moist Heat Sterilization of Aqueous 2012.
Liquids. In USP 40—NF 35, USP: Rockville, MD,
2017. 12. Stewart, K. “Laws of thermodynamics”. Encyclope-
dia Britannica, 6 Oct. 2022, https://www.britannica.
5. Agalloco, J. P. A Tale of Two Sterilizers. PDA J. com/science/laws-of-thermodynamics (accessed April
Pharm. Sci. Technol. 2020, 74 (1), 162–169. 26, 2023).
6. European Commission, Annex 1, Manufacturing 13. Dion, M.; Parker, W. L. Steam Sterilization Prin-
of Sterile Medicinal Products, Eudralex Volume ciples. Pharm. Eng. 2013, 33 (6), 60–69.
4. European Commission: Brussels, 2022.
14. Standard EN 285:2021 Sterilization. Steam Steri-
7. World Health Organization. WHO Technical lizers. Large Sterilizers. (2021). https://www.en-
Report Series No. 961: Annex 6—WHO Good standard.eu/bs-en-285-2015-sterilization-steam-
Manufacturing Practices for Sterile Pharmaceuti- sterilizers-large-sterilizers/, 13.3.2
cal Products. WHO, 2011.
15. Kaiser U., Effectes of Non-Condensable Gasses
8. Agalloco, J. P.; Akers, J. E.; Madsen, R. E. (NCGs) on Steam Sterilization Processes, Central
Moist Heat Sterilization—Myths and Realities. Service, Vol. 13, 2005, P 395.
PDA J. Pharm. Sci. Technol. 1998, 52 (6), 346–
350. 16. Kaiser, U. Effects of Non-Condensable Gases
(NCGs) on Steam Sterilisation Processes. Zentral-
9. Pavell, A.; Hughes, K. A. A Risk-Based sterilisation—Central Service 2005, 13 (1), 45–50.
Approach to Variable Load Configuration Valida-
tion in Steam Sterilization: Application of PDA 17. Standard EN 285:2021 Sterilization. Steam Steri-
Technical Report 1 Load Equivalence Topic. lizers. Large Sterilizers. (2021). https://www.en-
PDA J. Pharm. Sci. Technol. 2010, 64 (2), 124– standard.eu/bs-en-285-2015-sterilization-steam-
136. sterilizers-large-sterilizers/ ,8.3.1.2
10. AISI 316 Stainless Steel Properties, SS 316 Grade 18. Gillis, J. R. What does the Survival/Kill Calcula-
Density, Composition, Yield Strength, Thermal tion on the Certificate of Analysis Really Indicate?
Conductivity (theworldmaterial.com), https://www. Spore News 2005, 2 (3).
theworldmaterial.com/316-stainless-steel/ (accessed
April 26, 2023). 19. Agalloco, J. A. Maximum and Minimum Loads in
Steam Sterilization. PharmaManufacturing, February
11. Somerton, C. W. Phase Change Substance Property 19, 2018. https://www.pharmamanufacturing.com/
Tables in SI units for ME 201 section 001 Spring articles/2018/maximum-and-minimum-loads-in-steam-
2012. Department of Mechanical Engineering, sterilization/ (accessed February 19, 2018).

RESEARCH
Tolerance Interval Approach for the Determination of Overfill

of Liquid Parenteral Drug Products
BERNHARD SCHMELZER1* and MARC SUTTER2
1
Sandoz GmbH, Biochemiestraße 10, 6336 Langkampfen, Austria; and 2Novartis Pharma AG, Klybeckstrasse 141, 4057
Basel, Switzerland © PDA, Inc. 2023
ABSTRACT: Liquid parenteral products contain an overfill to ensure withdrawal of the declared volume. The overfill
must be sufficiently high to compensate for the expected loss during product preparation and administration, but it
should also be minimized to prevent accidental overdosing and unforeseen dose splitting of single-dose products. Find-
ing the right balance between too much and too little overfill with an acceptable risk of product failure is challenging
and requires consideration of the relevant sources of variability of the extractable volume. This article provides a novel
approach for the calculation of the required overfill based on tolerance interval methodology. In a first step, a tolerance
interval multiplier from the literature is proposed, and a simulation study is conducted to assess the appropriateness of
its use for overfill determination. In a second step, this multiplier is adapted to cover operator-to-operator variability in
the loss data and compared with other multipliers via a second simulation study. Use of a tolerance interval multiplier
enables adaptation of the overfill such that the risk of not reaching the minimum extractable volume fulfills predefined
acceptance criteria. By this, the scientific justification of the selected overfill is strengthened and control over a critical
quality attribute is improved.
KEYWORDS: Overfill, Hold-up volume, Void volume, Liquid parenteral product, Tolerance interval, Manufacturing.
1. Introduction dose splitting of single-dose liquid parenteral products.

For these reasons, an overfill that appears to be unusually
Selection of an appropriate overfill volume is an impor- high compared with those of other products or compared
tant decision in the development of a liquid parenteral with pharmacopeia requirements (e.g., [1]) will likely be
product. The overfill is added during manufacturing of challenged by health authorities.
the product to ensure extraction of the declared volume
when the product is used. As trivial as its function is, Performing a series of extraction experiments with itera-
defining the overfill requires careful balancing between tive changes of the overfill is very helpful for finding an
too little and too much. If a too low overfill volume is appropriate overfill. Established knowledge about the
selected, there is a risk that the declared volume cannot expected hold-up volume or loss can effectively reduce
be extracted, which again could lead to complaints and the number of iterations required in such extraction
to underdosing of patients in the worst case. On the experiments (2). However, consideration of the loss and
other hand, the manufacturing site will try to optimize its variability alone is not necessarily sufficient to define
the manufacturing process of a commercial product for an overfill that ensures consistent extraction of the
highest yield and in the course of this, a high overfill declared volume. For a robust definition of the overfill,
might be questioned. Furthermore, health authorities aim all possible sources of variability of the extractable vol-
at preventing possible medication errors and adverse ume, from product manufacturing until administration to
events related to overdosing and also want to prevent the patient, should be taken into account (3).
In this article, an approach for calculating the target fill

* Corresponding author: Sandoz GmbH, Biochemiestraße volume for liquid parenteral products is presented,
10, 6336 Langkampfen, Austria; Telephone: +43 664 which takes four factors into account: the minimum
88709209; E-mail: bernhard.schmelzer@novartis.com acceptable extractable volume, the loss (nonextractable
doi: 10.5731/pdajpst.2022.012743 volume), the variability of the loss, and the variability
of product fill weight. The minimum acceptable
Vol. 77, No. 3, May--June 2023 181

extractable volume is selectable. The factors loss, vari- equation is proposed for overfill calculation. In addition,
ability of the loss, and variability of product fill weight a tolerance interval approach for the determination of
together define the required overfill. They can be deter- the multiplier is introduced to enable a product-specific
mined from experimental data during development and calculation and justification of the overfill.
during product transfer to the designated commercial
manufacturing site. The proposed calculation does not The article is organized as follows. Section 2 presents
include the factor product fill weight in addition to the an approach to compute the target fill volume (or
fill weight variability as described by Dasnoy et al. (3), equivalently, the overfill) by using a tolerance interval
because it is assumed that the filling machine is con- multiplier. Section 3 focusses on the question how to
trolled and steered to reach the target fill weight very compute the tolerance interval multiplier (Section 3.1).
closely during batch filling. As a result of this, the dif- It includes a simulation study to confirm the appropri-
ference between the achieved average fill weight of ateness of the suggested approach (Section 3.2) and a
batches and the target fill weight should be negligibly sample computation to demonstrate its application
small for the overfill calculation approach presented (Section 3.3). Section 4 is dedicated to the question
here. To calculate a robust target fill volume with suffi- how the computation of the multiplier changes when
cient overfill, a tolerance interval multiplier is included taking into account the (potential) operator-to-operator
in the proposed calculation. The value of the multiplier variability in the loss data. Seven approaches are pre-
depends on the sample size of the experimental loss sented in Section 4.1. Their performance is assessed in
data, the sample size of the experimental fill weight Section 4.2 by a simulation study, and their application
data, as well as the selected acceptable failure rate and is demonstrated in Section 4.3 by revisiting the exam-
confidence level.
ple from Section 3.3. Conclusions are in Section 5.
Details on the statistics used to derive the multipliers
The proposed calculation of target fill volume is intended
are provided in Appendix A.
to be applied product specifically, that is, for a particular
liquid parenteral product in its container and packaging.
2. Definition of Overfill to Meet Minimum
Within this frame, it includes the right factors to take into Extractable Volume
account all sources of variability that could affect the ex-
tractable volume from manufacturing to administration to For the correct dosing of patients, a minimum extracta-
the patient. The approach presented in this article does not ble volume (EVmin) corresponding to the nominal vol-
attempt to replace experimental work for overfill determi- ume of a liquid parenteral product must be ensured.
nation by a prediction model as described by Dasnoy et al. Overfill must be added to account for the variability of
(3). Nevertheless, it can help to reduce the number of ex- filling during manufacturing of the product and for
perimental iterations that are needed until the final overfill unavoidable loss during preparation and conduction of
and target fill volume are set. It can also help development the administration according to the recommended pro-
teams to assess the expected effect of changes on the ex- cedure. The target fill volume is the sum of the mini-
tractable volume, for example, when a product is trans- mum extractable volume and the required overfill. The
ferred to a new filling line with different fill weight fill volume and the loss can be seen as random varia-
variability. Finally, it enables development teams to define bles, each with a certain average and a certain standard
what failure rate for the extractable volume is still accept- deviation with any difference between the average fill
able and to adapt the overfill accordingly. This again pro- volume and the target fill volume being negligible. The
vides improved justification of the selected overfill. extractable
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi volume standard deviation is then equal to
s2F þ s2L with sF and sL denoting the standard devia-
The approach described in this article is complemen- tion of filling and loss, respectively. The idea for the
tary to recent publications about overfill, which have computation of the target fill volume is to add to EVmin
proposed prediction models for estimating hold-up vol- the average loss and a multiple of the (theoretical) ex-
ume (2) and overfill (3). Dasnoy et al. (3) proposed a tractable volume standard deviation. The multiplier is
basic equation for the calculation of the overfill, which determined such that the risk of too low extractable vol-
takes into account the filling variability, the analytical ume is deemed acceptable. Taking into account that the
variability, and a fixed multiplier to ensure withdrawal delivery of the nominal volume to the patient is of major
of the nominal volume. In this article, the same basic importance, the acceptable risk level should be low. For

example, if a failure rate of not more than 0.01% with k: tolerance interval multiplier. It should be com-
95% confidence is selected, this means that with 95% con- puted as the multiplier of a one-sided tolerance inter-
fidence 99.99% of all extractable volume values are not val for the difference of two normally distributed
less than EVmin. This corresponds to a one-sided 99.99%/ random variables (fill volume and loss) with a certain
95% lower tolerance limit (LTL) for extractable volume population content (e.g., 99.99%) and a certain con-
matching EVmin. The advantage of using a tolerance inter- fidence level (e.g., 95%). The multiplier depends on
val multiplier over using a fixed multiplier (e.g., 3 as sug- the sample size of the fill weight data, the sample
gested in [3]) is that it can be precisely adapted to the risk size of the loss data, and the ratio between the stand-
level and provides a confidence statement by accounting ard deviations of the fill weights and loss. The
for the uncertainty related to the estimates of filling vari- smaller the sample sizes, the bigger the multiplier.
ability, average loss, and loss variability.
3. Determination of Tolerance Interval Multiplier
Using the average loss and the extractable volume
standard deviation to determine the overfill leads to the 3.1. Tolerance Interval Multiplier by Guo & Krishnamoorthy
following equation for the target fill volume:
For a certain proportion of the population (e.g., batch or
Target filling volume
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi product), extractable volume should be above the mini-
¼ EVmin þ Losstotal þ k s2F þ s2L (1) mum extractable volume (label claim). To ensure this, a
where: tolerance interval on extractable volume is an appropriate
statistical approach. The extractable volume is the differ-
EVmin: Minimum acceptable extractable volume corre- ence between the fill volume and the loss during extraction
sponding to the nominal fill volume (or label claim) (hold-up volume, dead volume), which are both subject to
random variation. These two quantities can be seen as ran-
Losstotal: Average total loss, including void volume dom variables for which the assumption of independence
in, for example, primary packaging, needles, and is reasonable, because the loss should not depend on the
syringes used for preparation for administration and fill volume. Assuming that the fill volume and the loss fol-
administration. The average total loss should be low normal distributions with means (expectations) lF and
experimentally determined by repeated extraction lL and variances r2F and r2L , respectively, the extractable
experiments using the materials and the extraction volume follows a normal distribution with mean (expecta-
procedure intended for administration of the product tion) lEV ¼ lF lL and variance r2EV ¼ r2F þ r2L . The
to the patient. If products require the use of closed lower p-quantile of this distribution is the value above
system transfer devices (CSTDs), extraction experi- which a proportion p of the extractable volume population
ments should be performed with these devices in lies, that is,
addition to extraction experiments with standard qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
needles and syringes. CSTDs may substantially Qp ¼ lEV zp rEV ¼ lF lL zp r2F þ r2L
increase the loss compared with that with standard (2)
needles and syringes (4) and because of this, the cal-
culated overfill may even exceed pharmacopeia lim- where zp is the p-quantile of the standard normal distri-
its for vials (1). In this situation, the decision for an bution. As the (theoretical) means and variances are
overfill is difficult to make and cannot be based only not known, samples need to be taken to estimate these
on the calculation. parameters. Fill volume or fill weight data of one or
several batches can be used to estimate the average and
sL: Loss variability (independent of fill weight) standard deviation of filling. For loss (dead volume),
expressed by the standard deviation computed from extraction experiments can be performed.
the repeated extraction experiments. It comprises
within-operator variability and between-operator (oper- Assume that samples of size nF and nL are available for
ator-to-operator) variability. the fill volume and the loss (dead volume; e.g., from
extraction experiments), respectively. Let the average and
sF: Filling variability expressed by the standard devi- the (sample) standard deviation of the fill volume data be
ation of fill volumes during manufacturing of the denoted by x F and sF , respectively, and let the average
product. and the (sample) standard deviation of the loss data be
Vol. 77, No. 3, May--June 2023 183

denoted by x L and sL , respectively. Then a lower toler- to Appendix A1 for more information and background
ance limit (LTL) on the extractable volume is of the form on the preceding formulas.
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
LTL ¼ x F x L k s2F þ s2L (3) 3.2. Simulation Study to Check the Appropriateness of
the Multiplier
The multiplier k depends on the population content p and
the confidence level c. The lower tolerance limit repre- A simulation study was performed to evaluate the
sents a lower confidence limit on the quantile Qp (5). appropriateness of kGK defined in eq 4. The sample size
nL of the loss data was varied between 30 and 120,
In Reference 5, one can find several approaches to because this covers a typical number of extractions in
obtain approximations to the tolerance interval multi- experiments. The sample size nF of the filling data was
plier k for the difference of two independent random varied between 30 and 300,000. This covers a wide
variables following a normal distribution. In the fol- range of possible batch sizes, as well as the possibil-
lowing, the focus is on the approach published in Guo ities that fill weight in-process control (IPC) is per-
& Krishnamoorthy (6) for which the index “GK” is formed in defined time intervals or continuously (100%
used subsequently. It is based on a multiplier derived fill weight IPC). The variance ratio RF;L ¼ r2F =r2L was
in Reference 7 for which the index “Hall” is used sub- varied between 0.01 and 100, which corresponds to one
sequently. More precisely, standard deviation being up to ten times as big as the
other one. The population content p was varied from
kGK ¼ maxfkHall;F ; kHall;L g (4)
0.9 to 0.9999. In total, 5175 combinations of values for
the four variables nF , nL , RF;L , and p were generated. In
with 2675 of the combinations, the population content was
pffiffiffiffiffi 1 fixed to 0.9, 0.95, 0.99, 0.999, or 0.9999 (in equal
kHall; ¼ tf^ ;c ðzp ^m Þ pffiffiffiffiffi (5)
^m share). In 175 out of these 2675 cases, nF was fixed to
30, 3000, or 300,000; nL was fixed to 30, 60, or 120;
ðR^L;F þ 1Þ2 and the variance ratio RF;L was fixed to 0.01, 1, or 100.
f^F ¼ ;
1=ðnF 1Þ þ R^ =ðnL 1Þ
2
L;F In the other 2500 of the 2675 cases (where p was
fixed), the latter three variables were varied using Latin
1 þ R^L;F
^
mF ¼ ; Hypercube Sampling. For the remaining 2500 out of
1=nF þ R^L;F =nL the 5175 cases, all four variables were varied using
s2 ðnF 3Þ Latin Hypercube Sampling. Because the multiplier kGK
R^L;F ¼ 2L
sF ðnF 1Þ from eq 4 does not depend on the means (of fill volume
and loss) and does not directly depend on the variances
but only on their ratio, the simulations were performed
ðR^F;L þ 1Þ2
f^L ¼ ; with lF ¼ lL ¼ 0 and r2L ¼ 1. For each combination of
R^F;L =ðnF 1Þ þ 1=ðnL 1Þ
2
the four variables, the quantile Qp (see eq 2) was com-
puted and 10,000 random samples for fill volume and
1 þ R^F;L
^
mL ¼ ; loss were generated using normal distributions. For
R^F;L =nF þ 1=nL
each of the generated data sets, the lower tolerance
s2 ðnL 3Þ limit was computed according to eq 3 using kGK from
R^F;L ¼ F2
sL ðnL 1Þ eq 4 and a confidence level of c = 0.95. The proportion
of simulated data sets for which the tolerance limit
does not exceed the theoretical quantile Qp (i.e., the
where zp is the p-quantile of the standard normal distri- tolerance interval contains Qp ) can be called achieved
pffiffiffiffiffi
bution and tf^ ;c ðzp ^m Þ is the c-quantile of the noncen- confidence, and ideally, it is equal to or at least close to
tral t-distribution with f^ degrees of freedom and the nominal confidence level of 0.95.
pffiffiffiffiffi
noncentrality parameter zp ^m with ^m representing the
effective sample size. The placeholder * stands for ei- The scatterplots in Figures 1 and 2 show the achieved
ther F or L, depending on which of the variance ratio confidence in dependence of the variables varied. Fig-
estimates (R^L;F or R^F;L ) is used. The reader is referred ure 1 shows that most of the simulated values for

Figure 1
Results of the simulation study in Section 3.2—scatterplot of achieved confidence versus population content;
each blue dot represents one configuration of the four variables (filling sample size, loss sample size, filling-to-
loss variance ratio, population content); red dashed line represents nominal confidence of 0.95.
achieved confidence scatter closely around the nominal From the simulation results for achieved confidence
value of 0.95. Note that deviations in a magnitude of and the results for the multiplier kGK , there is no need
0.01 can be due to the number of simulation runs. to restrict the approach in its practical application with
Deviations of more than 0.01 only show up on the regard to nF and nL or to avoid use of a high population
lower side, that is, there is not a single case with content up to 99.99%.
achieved confidence above 0.96, but there are cases
with achieved confidence below 0.94, and the devia-
tions become larger when the population content 3.3. Example Calculations
increases. This means that the multiplier kGK from eq 4
tends to be slightly too liberal for some of the cases The computations presented in Section 3.1 are applied
considered in the simulation. Figure 2 shows the to two Novartis biopharmaceutical liquid in vial prod-
achieved confidence in dependence of the variance ra- ucts. The first one has a label claim of 10.0 mL, the sec-
tio and the sample size ratio for different values of pop- ond one has a smaller label claim of 4.0 mL.
ulation content p. It can be seen that the lowest values
of achieved confidence show up when the sample size For the product with the 10.0 mL label claim, the filling
of fill volume is much larger than the sample size of variability (standard deviation) was estimated as
losses and the fill volume variance is close to or not sF ¼ 0:0254 mL based on a data set comprising
more than ten times the loss variance. Nevertheless, the nF ¼ 694 values. An extraction experiment was con-
multiplier from eq 4 is deemed to work sufficiently ducted according to the procedure in the product leaflet
well for all relevant cases because all simulated values comprising nL ¼ 40 extractions resulting in an average
for achieved confidence range between 0.92 and 0.96. total loss of x L ¼ 0:2203 mL and a standard deviation
of sL ¼ 0:0397 mL. Thepextractable volume standard
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Figure 3 shows the values of the multiplier kGK for the deviation is thus sEV ¼ 0:0254 þ 0:03972 ¼ 0:0471
2
same parameter combinations as used in the simula- mL. The variance ratio estimates result in R^L;F ¼ 2:4359
tions for achieved confidence (note that values of RF;L and R^F;L ¼ 0:3884. The degrees of freedom and the effec-
and RL;F ¼ 1=RF;L have been used instead of R^F;L and tive sample sizes are computed as f^F ¼ 76:8647,
R^L;F , respectively, to compute kGK from eq 4). One can f^L ¼ 74:5457, ^m F ¼ 55:1169, and ^m L ¼ 54:32. Using a
see that the formula in eq 4 results in reasonable multi- population content of p = 0.9999 and a confidence level of
pliers. The smallest multipliers are obtained when the c = 0.95 results in a multiplier of kGK ¼ maxfkHall;F ;
filling-to-loss variance ratio is high and the sample size kHall;L g ¼ maxf4:342; 4:353g ¼ 4:353 (with values of the
of filling is high at the same time. quantile of the noncentral t-distribution of 32.238 and
Vol. 77, No. 3, May--June 2023 185

Figure 2
Results of the simulation study in Section 3.2—gray scale plots of achieved confidence (ac) versus sample size
ratio and variance ratio for four different values of population content; each dot represents one configuration
of the four variables (filling sample size, loss sample size, filling-to-loss variance ratio, population content).
32.0804, respectively). Applying eq 1 results in a target fill <1151> (1) for liquid in vials, which, depending on vis-
volume of 10:0 þ 0:2203 þ 4:353 0:0471 ¼ 10:428 mL. cosity, permit overfills of up to 0.50 mL and 0.70 ml for
10.0 mL label claim and up to 0.30 mL and 0.50 mL for
For the product with the 4.0 mL label claim, the filling 5.0 mL label claim (which is the closest to the 4.0 mL
variability (standard deviation) was estimated as sF ¼ label claim in the second example). If the calculated over-
0:0164 mL based on a data set comprising nF ¼ 740 val- fill for a liquid in vial product should exceed these require-
ues. An extraction experiment was conducted according ments, the presented approach enables identification of the
to the procedure in the product leaflet comprising nL ¼ 90 source of variability that eventually leads to a high over-
extractions resulting in an average total loss of x L ¼
fill. One can also readily assess the effect of lower varia-
0:1913 mL and a standard deviation of sL ¼ 0:0203 mL.
bilities on the calculated overfill. This information can be
The extractable volume standard deviation is thus sEV ¼
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi used to improve the filling process or the extraction proce-
0:01642 þ 0:02032 ¼ 0:0261 mL. The variance ratio
dure to remain within a defined variability, which leads to
estimates result in R^L;F ¼ 1:528 and R^F;L ¼ 0:638. The
degrees of freedom and the effective sample sizes are com- a calculated overfill within USP <1151> (1) criteria.
puted as f^F ¼ 231:6617, f^L ¼ 227:6332, ^m F ¼ 137:9225,
and ^ m L ¼ 136:8047. Using a population content of 4. Taking into Account Operator-to-Operator Variability
p = 0.9999 and a confidence level of c = 0.95 results in a
multiplier of kGK ¼ maxfkHall;F ; kHall;L g ¼ maxf4:064; Usually there are several operators involved in extrac-
4:067g ¼ 4:067 (with values of the quantile of the non- tion experiments. The induced operator-to-operator
central t-distribution of 47.7288 and 47.5703, respec- variability can take a substantial part of the overall var-
tively). Applying eq 1 results in a target fill volume of iability of the loss (dead volume). The basic approach
4:0 þ 0:1913 þ 4:067 0:0261 ¼ 4:298 mL. presented in the previous section does not take into account
the operator-to-operator variability explicitly. However,
In these examples, the calculated overfills of 0.428 mL the approach presented in (7) and (6) can be extended in
and 0.298 mL correspond to the requirements of USP such a way that operator-to-operator variability is covered

Figure 3
gray scale plots of the multiplier from eq 4 (k) versus sample size ratio and variance ratio for four different
values of population content; each dot represents one configuration of the four variables (filling sample size,
loss sample size, filling-to-loss variance ratio, population content) in the simulation study in Section 4.2
by an own variability term, which means that the loss vari-

1
X
m X
l
ability is composed of operator-to-operator variability r2op s2e ¼ ðyij y i Þ2 (6)

nL m
and within-operator (or residual) variability r2e , that is, i¼1 j¼1
r2L ¼ r2op þ r2e . Consequently, the variability of the extractable volume

can be estimated as
As these variances are (usually) unknown, extraction
1
experiments need to be performed to estimate them. s2EV ¼ s2F þ s2L ¼ s2F þ s2op þ 1 s2e
l
Assume that there are m operators involved in the
extraction experiments, each of them performing l Note that this quantity is a variance; taking the square
extractions. Then the total number of extractions is nL= root yields the respective standard deviation. Further-
m·l. Let yij denote the loss observed with the jth extrac- more, note that s2e is an unbiased estimate for the
tion performed by operator i, let y i denote the average within-operator variance, but s2op is not an unbiased
loss observed with operator i and let y denote the aver- estimate for the operator-to-operator variance. In order
age loss observed across all operators. Then the total to obtain an unbiased estimate for r2op , s2e =l needs to be
loss variability is computed by subtracted from s2op , which can result in a negative esti-
mate of r2op . In this case, one should not use the
1
s2L ¼ s2op þ 1 s2e approach described in this section but rather use the
l
tolerance interval multiplier from eq 4 in Section 3.1.
with One can also perform an F-test to test whether r2op is
1
X
m significantly different from zero to decide whether to
s2op ¼ ðy i y Þ2 account for operator-to-operator variability with an
m1 own variability term or not. For additional explanations
i¼1
and background information on operator-to-operator
and variance, the reader is referred to Appendix A2.
Vol. 77, No. 3, May--June 2023 187

4.1. Options for Computing the Multiplier One can see from these formulas that it is not meaning-
ful to use this approach when the number of operators
Motivated by (6), the multiplier to compute the toler- in the extraction experiment is only 3, because the esti-
ance interval on extractable volume can be defined as mates R^F;op and Rê;op vanish. One way out could be to
kGK ¼ maxfkHall;F ; kHall;op ; kHall;e g use (instead of kGK ) the multiplier kHall;0 where
unbiased estimates for the variances are used (instead
with kHall; being defined as in eq 5 and of unbiased estimates for variance or mean square
2 ratios; see also [8]), that is,
1 þ Rôp;F þ 1 1l Rê;F
f^F ¼ 2
Rôp;F
2 ^2 ;
1 2 R e;F s2F þ s2op þ 1 1l s2e
1
nF 1 þ m1 þ 1 l nL m f^0 ¼ s4 ;
s4op
1 2 s4e
nF 1 þ m1 þ 1 l
F
1 þ Rôp;F þ 1 1l Rê;F nL m
^
mF ¼ ;
Rôp;F
1
nF þ m
s2F þ s2op þ 1 1l s2e
s2op ðnF 3Þ ^m 0 ¼ s2op
s2F
Rôp;F ¼ ; nF þ m
s2F ðnF 1Þ
However, the performance of this multiplier is low (see
s2 ðnF 3Þ
Rê;F ¼ 2e ; Section 4.2). Other alternatives are approximations
sF ðnF 1Þ
2 based on References 9 and 10, which have been

R^F;op þ 1 þ 1 1l Rê;op derived as tolerance interval approximations for ran-
fôp ¼ 2 ^2 ; dom or mixed effect models. In Reference10, the non-
R^F;op 1 2 R e;op
nF 1 þ m1 þ 1 l
1
nL m central t-distribution in eq 5 is replaced by an
approximation suggested by Reference 11 resulting in
R^F;op þ 1 þ 1 1l Rê;op
^
m op ¼ ; sffiffiffiffiffiffiffiffiffiffi rffiffiffiffiffi
R^F;op
nF þ m
1
UCL 1
kHH ¼ zp þ zc
sEV
2 ^
m
s2 ðm 3Þ 0
R^F;op ¼ 2F ;
sop ðm 1Þ This formula involves an estimate and an upper confi-
s2e ðm 3Þ dence limit (UCL) of the total variance, that is the var-
Rê;op ¼ ; iance of the extractable volume. For the computation
s2op ðm 1Þ
2 of the UCL, two approximations are considered: the
R^F;e þ Rôp;e þ 1 1l Satterthwaite approximation (12, 13) and the modified
fê ¼ 2 ; large sample (MLS) procedure (14, 15). Using the Sat-
R^F;e Rôp;e
2
1 2 1
nF 1 þ m1 þ 1 l nL m terthwaite approximation leads to

R^F;e þ Rôp;e þ 1 1l
^
me ¼ ; f0 ^
R^F;e Rôp;e UCLSatt ¼ s2EV
nF þ m vf^ ;1c
2
0
s ðnL m 2Þ
2
R^F;e ¼ F 2 ; where v2f^ ;1c is the (1-c)-quantile of the chi-square dis-
se ðnL mÞ
tribution with f^0 degrees of freedom. The resulting tol-
0
s2op ðnL m 2Þ erance interval multiplier is denoted kHH;Satt . Using the

Rôp;e ¼ ;
s2e ðnL mÞ MLS procedure results in the following approximation
for the UCL:
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
!2 !2
u 2 2
u nF 1 m1 1 nL m
UCLMLS ¼ s2EV t
þ sF 2
4 1 þ sop 2 4 1 þ 1 s4e 2 1
vnF 1;1c vm1;1c l vnL m;1c

The resulting tolerance interval multiplier is denoted We are mutually independent random variables follow-
kHH;MLS . ing chi-square distributions with nF 1, m 1, and
ml m degrees of freedom, respectively. Replacing
Motivated by Reference 16, Hoffman & Kringle (9) lF , lL , r2F , and r2L in eq 2 by GPQlF , GPQlL , GPQr2F ,
provide an approximation for the multiplier of a two- and GPQr2L , respectively, results in a GPQ for Qp . The
sided tolerance interval for a mixed effect model. By generalized confidence limit for Qp is obtained via sim-
replacing the (1+p)/2-quantile zð1þpÞ=2 by zp , it can be ulation. Let N denote the number of simulations runs
used as an approximation in the one-sided case: (say N = 10,000). Then for each i ¼ 1; 2; . . . ; N one
ðiÞ ðiÞ ðiÞ
rffiffiffiffiffiffiffiffiffiffiffiffiffi sffiffiffiffiffiffiffiffiffiffi generates (independent) random samples ZF , ZL , WF ,
ðiÞ ðiÞ
1 UCL Wop , and We for ZF , ZL , WF , Wop , and We , respec-
kHK ¼ zp 1 þ
^m 0 s2EV tively, and computes the GPQ for Qp by the formula
0 sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1
ðnF 1Þs2F A
GPQQp ¼@x F ZF
ðiÞ ðiÞ
ðiÞ
The multiplier obtained when using the Satterthwaite n F WF
approximation to compute the UCL is denoted by 0 vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1
u
kHK;Satt , the multiplier using the MLS procedure is B uð
ðiÞ t m1 sop C
Þ2
denoted by kHK;MLS . @x L ZL ðÞ
A
mWopi
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
As an additional alternative approach, one can consider ðnF 1Þs2F ðm1Þs2op 1 ðmlmÞs2e
a tolerance interval based on general pivotal quantities zp þ þ 1
W
ðiÞ
Wop
ðiÞ l We
ðiÞ
(GPQs).1 In Reference 5, GPQs are extensively used to F
derive simulation algorithms to obtain tolerance inter-

vals for mixed and random effects models. A lower tol-
erance limit for extractable volume can be computed This yields N random samples for the GPQ of Qp . Taking
by a generalized lower confidence interval (i.e., a confi- the 1 c quantile of these N samples provides (an esti-
dence interval based on GPQs) for the theoretical quan- mate of) the lower confidence limit on Qp and thus the
tile Qp from eq 2. Motivated by the discussions in ðp; 1 cÞ lower tolerance limit LTLGPQ for extractable
Sections 1.4 and 4.3.3 in Reference 5, one can define volume. The respective tolerance interval multiplier kGPQ
the following GPQs for lF , lL , r2F , and r2L : is obtained by dividing the difference between the mean
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi value difference and the lower tolerance limit by the esti-
ðnF 1Þs2F
GPQlF ¼ x F ZF mate of the extractable volume standard deviation (note
nF WF that fill and loss averages cancel out and are thus irrele-
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi vant for computing the multiplier):
ðm 1Þs2op
GPQlL ¼ x L ZL
m Wop x F x L LTLGPQ
kGPQ ¼
sEV
ðnF 1Þs2F
GPQr2F ¼
WF 4.2. Simulation Study to Assess Appropriateness of
Multipliers

ðm 1Þs2op 1 ðml mÞs2e
GPQr2L ¼ þ 1 (7) A simulation study was conducted to assess the per-
Wop l We
formance of the seven multipliers. Six variables were
where ZF and ZL are independent random variables fol- varied in this study: The variables nF , RF;L ¼ r2F =r2L
lowing the standard normal distribution,WF ; Wop , and and p were varied as in the simulation study in Section
3.2. In addition, the number of operators was varied
1. This approach is basically also applicable to the situation in Section
3.1. It is only mentioned in this section because the multiplier by Guo from 3 to 9, the number of extractions per operator was
& Krishnamoorthy (GK) yields satisfactory results in the simulation varied from 6 to 40, and the intraclass correlation of
study in Section 3.2 and might be easier to implement. The simulation
study in Section 3.2 was also conducted for the GPQ multiplier. Values operator-to-operator variability
of achieved confidence were very similar to those obtained with the GK
multiplier although slightly higher, resulting in more conservative tol- r2op r2op Rop;e
erance intervals. qop ¼ ¼ ¼
r2L r2op þ r2e Rop;e þ 1
Vol. 77, No. 3, May--June 2023 189

was varied from 0.01 to 0.99. In total, 15,000 combina- considerable number of results below 0.9. The
tions of values for p, nF , RF;L , m, l, and qop were gener- achieved confidence of kGK is slightly biased to the
ated. This list of combinations is a mixture of lower side, but the spread is relatively small, that is, all
combinations for which the variables were fixed to cer- results are above 0.9. The multiplier kGPQ has a bias of
tain values (3, 4, 5, 6, and 9 for m; 6, 10, 25, and 40 for similar magnitude, but to the upper side, and it has the
l; 0.01, 0.5, and 0.99 for qop ) on the one hand, and var- smallest spread. Overall, kGPQ yields the best results
ied randomly by Latin Hypercube sampling on the and the results of kGK appear acceptable as well. It
other hand. It should be noted that the case l = 6 is only should be noted that it is not possible to identify one of
considered in combination with m = 5 to correspond to the multipliers as optimal for all situations, because the
a minimum of nL ¼ 30; this refers to 4225 out of the
achieved confidence changes with the values of the var-
15,000 combinations. In the other 10,775 combina-
iables. For example, the achieved confidence of kGK
tions, l was varied from 10 to 40 and the proportion for
and kHH;Satt is lower than 0.95 for moderate and high
which p was fixed to 0.999 or 0.9999 and m was fixed
values of qop whereas kHK;MLS is quite stable with
to 3 or 4 was considerably higher than for the other val-
ues of p and m. For each combination of the six varia- respect to qop but implies values lower than 0.95 with
bles, the quantile Qp (see eq 2 where r2L ¼ r2op þ r2e ) increasing RF;L .
was computed and 10,000 random samples for fill vol-
ume and loss were generated using normal distribu- Figure 5 shows the values of the multipliers for popula-
tions. For each of the generated data sets, the lower tion content p = 0.9999 and confidence c = 0.95 for the
tolerance limit was computed using the above seven same parameter combinations as used in the simula-
multipliers (with number N of simulation runs for kGPQ tions for achieved confidence. The left picture shows
set to 10,000) and a confidence level of 0.95, and the the multipliers when the number of operators is 3, and
achieved confidence was determined. the right picture shows the multipliers when the num-
ber of operators is 4 to 9. One can see that all seven
The left picture in Figure 4 shows boxplots and indi- approaches can produce very large multipliers (and
vidual values (blue dots) of achieved confidence of the still, in many cases, they are too small to achieve the
seven multipliers for all simulation cases for which nominal confidence of 0.95 as shown by Figure 4).
m = 3. One can see that the multiplier kGK yields unsat- These large values belong to the cases where the fill-
isfactory results. Looking at the other six multipliers, ing-to-loss variance ratio RF;L takes values of 4 and
kGPQ seems to provide the best performance. The val- more. A more detailed look into the simulation results
ues for achieved confidence show the smallest spread also reveals that the multiplier kGK increases with the
even though there is a small bias to a too high achieved intraclass correlation qop and slightly decreases with
confidence, that is, the GPQ approach tends to result in
the number of operators m (within the investigated
slightly too conservative tolerance intervals. The multi-
range of 3 to 9). The median of the multipliers kGK in
plier kHH;MLS shows a slightly higher bias to the upper
the cases with more than three operators is approxi-
side and a slightly wider spread. The median achieved
mately 8.2, which seems to be unreasonably high, and
confidence of kHK;MLS is very close to the nominal value
the medians of the other multipliers are similar. For the
of 0.95, but results show a very wide spread with quite
some values below 0.9. Achieved confidence results number of operators investigated in this simulation
for kHH;Satt and kHK;Satt show an unacceptable bias and a study (and likely usually available for extraction
wide spread. Therefore, it cannot be recommended to experiments), care should be taken when using the tol-
use these two multipliers in general, although in some erance interval multipliers from Section 4.1. They
cases they might yield more acceptable results than might yield reasonable multipliers in individual cases.
other multipliers, for example kHH;Satt yields acceptable Nevertheless, for obtaining a suitable and functional
results if filling variability is high (compared to loss overfill one can restrict to using the multiplier pre-
variability) and operator-to-operator variability is low. sented in Section 3.1 by default, and using the loss var-
iability s2L computed by eq 6 also when operator-to-
The right picture in Figure 4 shows boxplots of operator variability is significant.
achieved confidence of the seven multipliers for all
simulation cases for which the number m of operators 4.3. Example Calculation
is greater than 3. One can see that the median achieved
confidence of kHK;MLS matches the nominal confidence This section describes how the results for the two
of 0.95, but the spread of the results is very wide with a examples in Section 3.3 change when accounting for

Figure 4
Results of the simulation study in Section 4.2— boxplots of achieved confidence per multiplier when the num-
ber of operators is three (A) and when the number of operators varies from four to nine (B); each blue dot rep-
resents one configuration of the six variables (population content, filling sample size, filling-to-loss variance
ratio, number of operators, number of extractions per operator, operator-to-operator intraclass correlation
coefficient); red dashed line represents the nominal confidence level of 0.95.
operator-to-operator variability in the losses by using In the second example, the 90 extractions were actually
the approaches presented in Section 4.1. performed by three operators (each of them performing
30 extractions). Computing the standard deviation of
In the first example, the 40 extractions were actually per- the total loss according to eq 6 results in sL ¼ 0:0211
formed by four operators (each of them performing ten mL with s2op ¼ 0:00011692 and s2e ¼ 0:00033917. As
extractions). Computing the standard deviation of the total the F-statistic for testing r2op ¼ 0 (see Appendix A2)
loss according to eq 6 results in sL ¼ 0:0419 mL with yields a value of 10.34, which corresponds to a p-value
s2op ¼ 0:00087532 and s2e ¼ 0:00097469. As the F-statistic of 0.00009, the operator-to-operator variance r2op is
for testing r2op ¼ 0 (see Appendix A2) yields a value of significant, and it is reasonable to use an own variabili-
8.98, which corresponds to a p-value of 0.00014, the oper- ty term. The standard deviation of the extractable vol-
ume (sEV ) is 0.0267 mL. The approximate degrees of
ator-to-operator variance r2op is significant, and it is rea-
freedom needed to compute the tolerance interval multi-
sonable to use an own variability term. Note that there is a
pliers are computed as f^0 = 62.3645, f^F = 62.488, fôp =
slight increase in sL compared with the result in Section
2, and fê = 63.2643. The effective sample sizes are com-
3.3. The standard deviation of the extractable volume
puted as ^m 0 = 18.1444, ^m F = 18.1625, ^m op = 3, and ^m e =
(sEV ) is 0.049 mL. The approximate degrees of freedom 18.3406. The upper confidence limit on the extractable
needed to compute the tolerance interval multipliers are volume variance is 0.000984 when using the Sat-
computed as f^0 = 20.7266, f^F = 20.7587, fôp = 7.4159, and terthwaite approximation, whereas applying the MLS
fê = 21.4284. The effective sample sizes are computed as procedure results in 0.002878. Table I shows the multi-
^
m 0 = 10.9106, ^m F = 10.9189, ^m op = 6.31, and ^m e = 11.1454. pliers and the resulting overfills when using a population
The upper confidence limit on the extractable volume var- content of p = 0.9999 and a confidence level of c = 0.95
iance is 0.00437 when using the Satterthwaite approxima- (with the multipliers needed for kGK being
tion whereas applying the MLS procedure results in kHall;F ¼ 4:494, kHall;op ¼ 16:598, kHall;e ¼ 4:488Þ.
0.009007. Table I shows the multipliers and the resulting
overfills when using a population content of p = 0.9999 and From both examples, one can see that the multipliers
a confidence level of c = 0.95 (with the multipliers needed differ substantially. For the multipliers kHK;MLS , kHK;Satt ,
for kGK being kHall;F ¼ 5:146, kHall;op ¼ 6:684, kHall;e ¼ kHH;MLS , and kHH;Satt , this is induced by the fact that the
5:117). approximations for the upper confidence limit on
Vol. 77, No. 3, May--June 2023 191

Figure 5
Boxplots of each multiplier when the number of operators is three (A) and when the number of operators
varies from four to nine (B); each blue dot represents one configuration of the six variables (population content,
filling sample size, filling-to-loss variance ratio, number of operators, number of extractions per operator, oper-
ator-to-operator intraclass correlation coefficient) in the simulation study in Section 4.2.
extractable volume standard deviation yield very dif- 5. Conclusions

ferent results. In the first example, the multiplier kGK is
somewhere in between, as it is highly influenced by the In this article, a tolerance interval approach has been
variance (mean square) ratios involving the smallest presented to set the overfill of liquid parenteral prod-
degrees of freedom (those with s2op in the denominator). ucts. It assures with high confidence that the extracta-
For the second example, the multiplier kGK yields an ble volume of the filled units is not less than the label
unusable value, because there are only three operators. claim, and it provides a possibility to define the accept-
The multiplier kHall;0 in both examples is close to kHall;F able remaining risk of failure resulting in a better pro-
and kHall;e ; the multiplier kGPQ is close to kHK;MLS and tection against an extractable volume below the label
kHH;MLS in both examples. For the first example, all claim compared with using a fixed multiplier (e.g., 3 as
multipliers introduced in Section 4.1 are substantially suggested in [3]). It has been demonstrated by a simu-
larger than the value obtained when not taking into lation study that application of the approach from Guo
account operator-to-operator variability (see Section
& Krishnamoorthy (6) yields highly appropriate results
3.3), whereas for the second example, the multipliers
(when not accounting for operator-to-operator variabil-
kHall;0 , kHH;Satt , and kHK;Satt are only slightly higher than
ity by an own variability term), because the achieved
the multiplier in Section 3.3. For both examples, using
confidence is very close to the nominal value of 0.95.
the multipliers kHall;0 , kHH;Satt , and kHK;Satt results in
Furthermore, the approach from Reference 6 has been
overfill values close to the recommended excess vol-
umes in USP <1151> (1) for liquid in vials. Using the extended to cover operator-to-operator variability of
loss variability computed by eq 6 and applying the the loss by an own variability term. The so obtained
multiplier from eq 4 yields a multiplier of 4.373 and an multiplier has been compared with alternative approxi-
overfill of 0.434 mL for the first example, and a multi- mations—among them a multiplier based on general
plier of 4.076 and an overfill of 0.300 mL for the sec- pivotal quantities—via a simulation study, which
ond example. For easier comparison, these results and revealed that it is unusable in the case of only three
the results obtained in Section 3.3 are also displayed in operators. In case of more than three operators, it yields
Table I. Furthermore, results when using the fixed mul- acceptable performance in terms of achieved confi-
tiplier 3 have been added to demonstrate that it yields dence. However, it has been demonstrated that the mul-
significantly lower overfills. tiplier based on general pivotal quantities yields best

Table I
Multipliers and Resulting Overfills for the Two Examples When Using a Fixed Multiplier, the Multiplier from
Section 3.1 (GK_Simple) and the Multipliers from Section 4.1 (Accounting for Operator-to-Operator Variability—
GK_op, GPQ, Hall0, HH_Satt, HH_MLS, HK_Satt, HK_MLS); for the Fixed and the Simple Multiplier Both Cases
of Determining the Loss Standard Deviation Are Considered: Simple Computation as Suggested in Section 3.1 and
Computation according to eq. 6 (Accounting for Operator-to-Operator Variability by an Own Variance
Component)
Example 1 (10.0 mL Label Claim) Example 2 (4.0 mL Label Claim)

Multiplier type Loss SD Multiplier k Overfill [mL] Multiplier k Overfill [mL]
GK_simple simple 4.353 0.428 4.067 0.298
GK_simple op 4.373 0.434 4.076 0.300
GK_op op 6.684 0.548 16.598 0.634
GPQ op 7.323 0.579 7.597 0.394
Hall0 op 5.147 0.473 4.494 0.311
HH_Satt op 5.515 0.491 4.755 0.318
HH_MLS op 7.701 0.598 7.859 0.401
HK_Satt op 5.242 0.477 4.488 0.311
HK_MLS op 7.526 0.589 7.676 0.396
Fixed multiplier simple 3 0.362 3 0.270
Fixed multiplier op 3 0.367 3 0.271
confidence levels for all numbers of operators consid- during product development and, if considered neces-
ered in this article. sary, to increase the calculated overfill based on this
knowledge.
From a practical perspective, it will usually be adequate to
derive a suitable and functional overfill using the approach Acknowledgements
from Guo & Krishnamoorthy (6) without taking into
account the operator-to-operator variability in the compu- The authors would like to thank the following colleagues
tation of the multiplier. Using an own variability term for for their support in the establishment of the concept of
operator-to-operator variability might in general only work overfill determination: Sigrun Löw, Thomas Schröder,
well when ten or more operators are involved in the extrac- Aleksandra Stojkovic, Robert Hormes, Fabien Canepa,
tion experiments. Even when experimental work can be Alexander Z€urn, Robert M€uhlbacher, and Caroline Hilbert.
conducted with many operators, it is recommended to use
this approach complementary to the simpler approach not
taking into account operator-to-operator variability and to Conflict of Interest Declaration
make the final decision on the overfill based on a careful
interpretation of the outcomes of both approaches. Bernhard Schmelzer and Marc Sutter declare no com-
peting interests. All authors are employees of their re-
The presented tolerance interval approach is applicable spective group of companies.
to single-dose liquid parenteral products in different
presentations, such as vials or prefilled syringes. It can References
be adapted for multidose vial presentations, which was
not shown in this article. Losses that occur occasion- 1. U.S. Pharmacopeial Convention, General Chapter
ally, such as drop loss when removing needle shield <1151> Pharmaceutical Dosage Forms. In USP
caps of devices, are not necessarily known and do not 42—NF 37, USP: Rockville, MD, 2019.
necessarily occur when conducting the extraction
experiments. Therefore, they cannot be covered by the 2. Mehta, S. B.; Subramanian, S.; Brown, R.; D’Mello,
proposed tolerance interval approach. It is advisable to R.; Brisbane, C.; Roy, S. Use of a Predictive Regres-
collect data about such losses over a longer time period sion Model for Estimating Hold-Up Volume for
Vol. 77, No. 3, May--June 2023 193

Biologic Drug Product Presentations. PDA J. Pharm. 13. Satterthwaite, F. E. An Approximate Distribution
Sci. Technol. 2020, 74 (3), 290–300. of Estimates of Variance Components. Biom. Bull.
1946, 2 (6), 110–114.
3. Dasnoy, S.; Simonin, L.; Radulovic, S.; White,
A.; Decoster, J.-F.; Denis, L. A Predictive Model- 14. Graybill, F. A.; Wang, C. M. Confidence Intervals
ing Approach to Support the Overfill Volume Def- on Nonnegative Linear Combinations of Varian-
inition of Liquid-in-Vial Drug Products. PDA J. ces. J. Am. Stat. Assoc. 1980, 75 (372), 869–873.
Pharm. Sci. Technol. 2022, 76 (5), 384–394.
15. Burdick, R. K.; Graybill, F. A. Confidence Inter-
4. Besheer, A.; Burton, L.; Galas, R. J. Jr; Gokhale, vals on Variance Components; Marcel-Dekker,
K.; Goldbach, P.; Hu, Q.; Mathews, L.; Muthura- Inc.: New York, 1992.
nia, K.; Narasimhan, C.; Singh, S. N.; Stokes,
E. S. E.; Weiser, S.; Zamiri, C.; Zhou, S. An 16. Howe, W. G. Two-Sided Tolerance Limits for
Industry Perspective on Compatibility Assessment Normal Populations— Some Improvements. J.
of Closed System Drug-Transfer Devices for Bio- Am. Stat. Assoc. 1969, 64, 610–620.
logics. J. Pharm. Sci. 2021, 110 (2), 610–614.
Appendices
5. Krishnamoorthy, K.; Mathew, T. Statistical Toler-
ance Regions: Theory, Applications, and Compu-
tation; John Wiley & Sons, 2009. Appendix A: Details on the Tolerance Interval
Approach
6. Guo, H.; Krishnamoorthy, K. New Approximate
Inferential Methods for the Reliability Parameter Appendix A1: Hall’s Approach for the Difference of Two
in a Stress-Strength Model: The Normal Case. Independent Random Variables Following a Normal
Commun. Stat. Theory Methods 2004, 33 (7), Distribution
1715–1731. Let X and Y denote independent random variables both
following a normal distribution with means (expectations) l1
and l2 , respectively, and variances r21 and r22 , respectively.
7. Hall, I. J. Approximate One-Sided Tolerance Lim- Then the difference D = X – Y follows a normal distribution
its for the Difference or Sum of Two Independent with mean (expectation) l1 l2 and variance r2D ¼ r21 þ r22 .
Normal Variates. J. Qual. Technol 1984, 16 (1), Let X1 ; ; Xn1 and Y1 ; ; Yn2 denote random samples
15–19. (i.e., samples that are independently and identically distributed)
for X and Y, let X and Y denote the sample means, and let S12
and S22 denote the sample variances. The mean difference is
8. Reiser, B.; Guttman, I. Statistical Inference for Pr given by D ¼ X Y and the variance of D is estimated by
(Y<X): the Normal Case. Technometrics 1986, SD2
¼ S12 þ S22 . Hall (7) suggests a lower tolerance limit for D
28 (3), 253–257. of the form
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
9. Hoffman, D.; Kringle, R. Two-Sided Tolerance D k SD ¼ X Y k S21 þ S22
Intervals for Balanced and Unbalanced Random with the multiplier k being defined as in the case of a simple
Effects Models. J. Biopharm. Stat. 2005, 15 (2), normal random variable (e.g., see [5]), but the sample size
283–293. being replaced with the effective (or equivalent) sample size m,
that is,
10. Hoffman, D. One-Sided Tolerance Limits for Bal- pffiffiffi 1
kHall ¼ tf ;c ðzp mÞ pffiffiffi (A1)
anced and Unbalanced Random Effects Models. m
Technometrics 2010, 52 (3), 303–312. where zp is pffiffithe
ffi p-quantile of the standard normal distribution
and tf ;c ðzp mÞ is the c-quantile of the noncentral t-distribution
pffiffiffi
11. Halperin, M. Approximations to the Non-Central with f degrees of freedom and noncentrality parameter zp m.
The effective sample size is the ratio between the variance of D
t, with Applications. Technometrics 1963, 5 (3),
and the variance of D. The latter is given by
295–305.
r21 r22
r2D ¼ þ
12. Satterthwaite, F. E. Synthesis of Variance. Psy- n1 n2
chometrika 1941, 6, 309–316. Consequently, the effective (or equivalent) sample size is

r2D r21 þ r22 into account operator-to-operator variability. Let Yij denote the
m¼ ¼ r2 r2 loss observed with the jth extraction performed by operator
r2D 1
þ 2
n1 n2 number i. Then Yij can be modeled by the one-way random
effect model
The degrees of freedom f involved in eq A1 is approximated
by using the Satterthwaite approximation (12, 13): Yij ¼ lL þ Oi þ eij

2 2 where lL represents the (theoretical) average loss across all
r4D r21 þ r2
f ¼ r21 r2
¼ r21 r2
operators, Oi represents the random effect introduced by
n1 1 þ n2 1
2
n1 1 þ n2 1
2
operator i, and eij represents the random error term
associated with the jth extraction performed by operator i.
Hall (7) expresses the effective sample size and the degrees Under the assumption that the random terms Oi and eij are
of freedom by using the variance ratio R1;2 ¼ r21 =r22 . As the mutually independent and follow a normal distribution with
variances are (usually) unknown, Hall (7) suggests to replace mean 0 and variance r2op and r2e , respectively, the losses
the unknown ratio R1;2 by the unbiased estimate follow a normal distribution with mean lL and variance
r2L ¼ r2op þ r2e . Note that r2op corresponds to operator-to-
S2 ðn2 3Þ
R^1;2 ¼ 12 operator variability and r2e corresponds to within-operator
S2 ðn2 1Þ variability.
resulting in the multiplier Assume that extractions are performed by m operators, and
that each operator performs the same number of extractions,
pffiffiffiffiffi say l. Then the total number of extractions is nL ¼ m l. Let Y i
1
kHall;2 ¼ tf^2 ;c ðzp ^m 2 Þ pffiffiffiffiffi denote the average loss associated with operator i and let Y
^m 2 denote the average loss across all operators, that is,
where 1
X
l
Yi ¼ Yij
ðR^1;2 þ 1Þ2 R^1;2 þ 1 l
f^2 ¼
j¼1
; m^ 2 ¼
R^1;2 =ðn1 1Þ þ 1=ðn2 1Þ
2 ^
R 1;2 =n1 þ 1=n2
and
Guo & Krishnamoorthy (6) extended Hall’s approach by 1

X
m

considering also the converse variance ratio R2;1 ¼ r22 =r21 for Y ¼ Yi
m
which an unbiased estimate is given by i¼1
S ðn1 3Þ
2
R^2;1 ¼ 22
S1 ðn1 1Þ
In order to estimate operator-to-operator and within-operator
resulting in the multiplier
variability, the following mean squares are considered (see for
example [15])
pffiffiffiffiffi
kHall;1 ¼ tf^1 ;c ðzp
1
^m 1 Þ pffiffiffiffiffi 1
X
m
^m 1 S2op ¼

ðY i Y Þ2
m1
i¼1
where
1
X
m X
l
ð1 þ R^2;1 Þ2 1 þ R^2;1 S2e ¼ ðYij Y i Þ2

f^1 ¼ ; ^m 1 ¼ nL m
1=ðn1 1Þ þ R^2;1 =ðn2 1Þ
2
1=n1 þ R^2;1 =n2 i¼1 j¼1
Guo & Krishnamoorthy (6) observed that using the greater

of the two multipliers for computing the tolerance limit yields The expected values of these mean squares are
satisfactory results in terms of coverage. (Note that the
r2e
indexing is different as in Reference 6 as the notation above is hop ¼ EðS2op Þ ¼ r2op þ
beneficial for extending the approach to the case when l
operator-to-operator variability is present, see Appendix A2.)
he ¼ EðS2e Þ ¼ r2e
Appendix A2: Derivation of an Extension Accounting for

Operator-to-Operator Variability in the Multiplier
In this section, the theoretical background is provided for how Thus, unbiased estimates for the variances r2op , r2e , and r2L
to adapt the approach presented in References 4 and 5 to take are given by
Vol. 77, No. 3, May--June 2023 195

S2e 2 approximation for deriving the approximate degrees of

^ 2op ¼ S2op
r ^ E ¼ S2e ; r
;r ^ 2L ¼ r
^ 2op þ r
^ 2e freedom f. More precisely, f r^ 2D =r2D approximately follows a
l
chi-square distribution with f degrees of freedom where
1 2
¼ Sop þ 1 Se
2
l r4D
f ¼ h2e
and the variances associated with the means Y i and Y are h2F h2op
nF 1 þ m1 þ 1 l nL m
1
r2Y ¼ hop and r2 ¼ hop =m.
i Y 2
Under the assumptions made previously, the distributions of hF þ hop þ 1 1l he
the mean squares Se2 and Sop 2
are related to chi-square ¼ 2 2
hF h2op 1 2 he
distributions. More precisely, ðnL mÞSe2 =he follows a chi-
nF 1 þ m1 þ 1 l nL m
square distribution with nL m degrees of freedom and
ðm 1ÞSop 2
=hop follows a chi-square distribution with m – 1 The formula for f involves the (usually) unknown variances
degrees of freedom. As Sop 2
and Se2 are independent, the statistic or, equivalently, the expectations of the mean squares.
l Sop =Se follows an F-distribution with m – 1 and nL m
2 2
Following References 4 and 5, ratios of expected mean squares
degrees of freedom. This can be used to perform a hypothesis are replaced by unbiased estimates. As there are three mean
test with null hypothesis r2op ¼ 0.
squares, there are three possibilities for considering ratios:
Note that if not all operators perform the same number of
extractions, the variable l in the formula for hop and the F-
statistic l Sop
2
=Se2 needs to be replaced by the harmonic mean 1. use ratios RF;e ¼ hF =he , Rop;e ¼ hop =he and replace
of the number
h Xmof extractions performed by each operator, that them with
is l ¼ m= i¼1 1=li where li is the number of extractions
performed by operator number i, and the distribution associated
with Sop 2
can only be approximated by using a chi-square S ðnL m 2Þ ^
2 S2op ðnL m 2Þ
R^F;e ¼ F 2 ; R op;e ¼ ;
distribution with m – 1 degrees of freedom, see Reference 15 Se ðnL mÞ S2e ðnL mÞ
for example.
If X is a random variable representing the fill volume and
following a normal distribution with mean (expectation) lF respectively. The resulting multiplier is denoted by
and variance r2F , the difference D = X – Y representing
extractable volume follows a normal distribution with mean
kHall;e
(expectation) lD ¼ lF lL and variance r2D ¼ r2F þ r2L ¼
r2F þ r2op þ r2e . An unbiased estimate for r2D is 2. use ratios RF;op ¼ hF =hop , Re;op ¼ he =hop and replace
them with
1
^ 2D ¼ r
r ^ 2F þ r
^ 2op þ r
^ 2e ¼ S2F þ S2op þ 1 S2e
l S2 ðm 3Þ ^ S2 ðm 3Þ
Similar to the situation of Appendix A1, a lower tolerance R^F;op ¼ 2F ; R op;e ¼ 2e ;
Sop ðm 1Þ Sop ðm 1Þ
limit for D is of the form
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
1 respectively. The resulting multiplier is denoted by
^ D ¼ X Y k S2F þ S2op þ 1 S2e
Dkr
l kHall;op
pffiffiffi pffiffiffi
with k ¼ tf ;c ðzp mÞ= m and the effective sample size m ¼
3. use ratios RF;op ¼ hF =hop , Re;op ¼ he =hop and replace
rD =rD . As hF ¼ r2F and
2 2
them with
2 hF hop r2F r2op r2e
r2D ¼ rX þ r2Y ¼ þ ¼ þ þ S2op ðnF 3Þ
nF m nF m ml S2e ðnF 3Þ
Rôp;F ¼ ; Rê;F ¼ ;
the formula for the effective sample size reads as S2F ðnF 1Þ S2F ðnF 1Þ

r2D hF þ hop þ 1 1l he r2F þ r2op þ r2e respectively. The resulting multiplier is denoted by
m¼ ¼ h
¼ r2 r2 kHall;F
r2D hF
þ op F r2
þ op þ e
nF m nF m ml
Note that compared with the situation in Appendix A1, the Following Reference 6, the multiplier to compute the tolerance
role of the variances r21 and r22 is taken by hF , hop , and he as interval on extractable volume is then the greatest of the three
these are the expected values of the mean squares SF2 , Sop2
, and multipliers, that is kGK ¼ maxfkHall;e ; kHall;op ; kHall;F g. This
Se . This is particularly relevant when using the Satterthwaite
2
results in the formulas presented in Section 4.1.

RESEARCH
Determination of ICH-Q3D Elemental Impurity Leachables in

Glass Vials by Inductively Coupled Plasma Mass
Spectrometry
LYDIA BRECKENRIDGE*, YUSUF ONI, CHRISTINA EVANS, JASON FRANCK, SHARLA WOOD, MENG XU,
ERINC SAHIN, and BRIAN ZACOUR
Bristol Myers Squibb, 1 Squibb Drive, New Brunswick, NJ 08903 © PDA, Inc. 2023
ABSTRACT: Container closure systems that are used for packaging pharmaceutical products are required to satisfy
numerous safety requirements. Maximum permitted limits on the concentrations of numerous toxic elemental impur-
ities that potentially leach from the packaging are one such requirement. The implementation of ICH-Q3D Guideline
for Elemental Impurities, in conjunction with the 2018 publication of USP <232> Elemental Impurities—Limits and
USP <233> Elemental Impurities—Procedures, requires a critical risk assessment of all container closure systems to
evaluate their contribution of certain elemental impurities to the enclosed drug product. ICH-Q3D has established limits
for each specific elemental impurity that considers relevant toxicological data and administration route (oral, parenteral,
or inhalation) and presents them as permitted daily exposures based on the maximum daily dosage of the final drug
product. A study was undertaken to assess the degree of elemental impurity leaching from one type of pharmaceutical
glass vial under specific, fixed environmental controls. Multiple buffer systems representing a broad spectrum of possi-
ble parenteral drug product formulations were used in the study. Resulting buffer solutions that had been in contact
with a single type of glass vial under specific conditions were subsequently analyzed using an inductively coupled
plasma mass spectrometry (ICP-MS) method developed and validated specifically for the purpose of quantifying ele-
mental impurity leachables in a variety of parenteral formulations. Results indicated that the degree of elemental impu-
rity leachables imparted by the specific type of glass vial evaluated during this study posed no risk to patient safety,
regardless of the drug product buffer formulation. Following this evaluation, the ICP-MS method developed for the
determination of elemental impurities leachables has been successfully applied to the assessment of elemental impur-
ities in a number of different biological parenteral drug product formulations currently under development. These data
can be leveraged for inclusion in elemental impurities component ICH-Q3D risk assessments to satisfy the container
closure system contribution.
KEYWORDS: Elemental impurities, ICH-Q3D, Leachables, ICP-MS, Glass vials, Metals.
Introduction related (e.g., global quality policies, image, market trends

and requirements, and manufacturing economics). The
Container closure systems (CCSs) used for packaging of 1999 Food and Drug Administration guidance document
drug products must satisfy many requirements. Generally, on CCSs for packaging human drugs and biologics sum-
these requirements can be functional and performance marized most of these requirements as they relate to metals
related (e.g., protection, stability, sterilization, dosing and in four sections: performance, protection, compatibility,
patient compliance, security/tamper evidence and manu- and safety (1).
facturing), regulatory (e.g., legislation, guidance, stand-
ards, and compendia), or commercial and corporate Although toxic metals have been a concern in the phar-
maceutical industry for decades, it was not until the
implementation of ICH-Q3D Elemental Impurities, in
* Corresponding Author: 1 Squibb Drive, New Bruns-
conjunction with the 2018 publication of USP <232>
wick, NJ 08903; Telephone: (732) 227-7540; E-mail:
Elemental Impurities—Limits and USP <233> Ele-
Lydia.Breckenridge@bms.com
doi: 10.5731/pdajpst.2021.012655 mental Impurities—Procedures, that the contribution of
these elements by CCSs was required to undergo a
Vol. 77, No. 3, May--June 2023 197

2B and 3 elements include Pd, Ru, and Li, among

Table I
others, and present reduced probability of occurrence
ICH-Q3D Permitted Daily Exposure (PDE) Limits for
in drug product owing to their lower natural abun-
Individual Elemental Impurities in Parenteral Drug
dance, although they may be intentionally added during
Products with Dosages of 10 g/day (4) and 100 g/day
manufacturing (e.g., as catalysts). The maximum limits
Parenteral Parenteral attributed to each of these elemental impurities are
Concentrationa Concentrationb based on toxicological data and administration route
Element Class (mg/g) (mg/g) (oral, parenteral, or inhalation) and is presented as per-
Cd 1 0.2 0.02 mitted daily exposures (PDEs) based on the maximum
Pb 1 0.5 0.05 daily dosage of the drug product (Table I).
As 1 1.5 0.15
Hg 1 0.3 0.03 There are four options suggested in ICH-Q3D by which to
perform an elemental impurities risk assessment; options
Co 2A 0.5 0.05
1, 2A, and 2B involve assessing all the different chemical
V 2A 1 0.1
components used to make the final formulation that can
Ni 2A 2 0.2
potentially contribute elemental impurities to a final drug
Tl 2B 0.8 0.08
product. The options differ mostly by the dosage used to
Au 2B 10 1
calculate the individual component’s control threshold
Pd 2B 1 0.1
limits. Option 3 provides for testing in the final drug prod-
Ir 2B 1 0.1 uct, which would, in theory, include elemental impurities
Os 2B 1 0.1 contributions from all components. When following any
Rh 2B 1 0.1 of the ‘component’ approaches (options 1, 2A, and 2B),
Ru 2B 1 0.1 the following contributors to elemental impurities must be
Se 2B 8 0.8 considered during a risk assessment: the drug substance or
Ag 2B 1 0.1 active pharmaceutical product (API), each excipient of the
Pt 2B 1 0.1 formulation, processing and synthesis water quality, man-
Li 3 25 2.5 ufacturing equipment, and CCS. It is expected that the
Sb 3 9 0.9 CCS should present minimal risk for contributing elemen-
Ba 3 70 7 tal impurities. Indeed, for orally administered drug prod-
Mo 3 150 15 ucts (e.g., tablets), this risk is deemed negligible and not
Cu 3 30 3 included in the scope of ICH-Q3D (4–5). However, paren-
Sn 3 60 6 terally administered products do present a greater risk of
Cr 3 110 11 leaching elemental impurities from the primary packaging
a components in which they are stored (6). This is likely
Based-on ICH-Q3D Option 1 default 10 g/day dosages.
b related to the higher extraction kinetics generally present
Based on a higher daily dosage of 100 g/day to
encompass a broader range of potential drug products. for a liquid medium. Indeed, the leaching of substances
from container closure systems into a solution is well
described in the literature (7–10). In addition, the elemen-
critical risk assessment (2–5). ICH-Q3D establishes a tal impurities threshold limits for parenteral products are
list of 24 elements, collectively referred to as Elemen- lower for many elements than those for products intended
tal Impurities, which should be assessed for their pro- for oral administration, thereby further increasing the
pensity to leach from CCSs, including vials and potential risk of leachable concentrations exceeding ac-
stoppers, that are used for parenteral drug formulations. ceptable exposure thresholds.
This list categorizes the elements into different classes
based on their toxicity and likelihood to be introduced. Although CCSs can refer to a variety of modalities,
Class 1 elements include As, Pb, Cd, and Hg, which including vials, ampoules, syringes, cartridges, bottles,
are elements that present significant human toxicity and stoppers, this article will focus solely on glass vials
and have little-to-no expectation to be innately present that are used to transport and store parenteral pharma-
in a drug product. Class 2 A elements include Ni, V, and ceutical products. Further, only one type of glass vial
Co, which have a probability of presence in a drug product from a single manufacturer was assessed; further stud-
owing to the manufacturing equipment composition. Class ies to compare different vial types and vendors are

ongoing. There are numerous manufacturers of glass the knowledge base, covering four of the six buffers
vials, and each will employ a different proprietary for- most commonly selected for the formulation of the ma-
mulation for their specific glass composition. These dif- jority of parenteral drug products (15). Some of these
ferences in formulation may involve the types of glass buffers serve a secondary purpose as possible ‘worst
former, intermediates, modifiers, colorants, and/or fining case scenarios’ due to the aggressive nature of their
agents (11). Each of these glass formulation variations interactions with components of the glass formulation
can impact the elemental impurity leaching profile of the (e.g., the metal chelation potential of citrate).
glass vial, which is further compounded by variations in
the parenteral drug product’s chemical and biological In addition to various buffers, this study investigated the
formulations. The exact chemistry of glass leaching, or leachable impact of other commonly used parenteral exci-
corrosion, is a complex phenomenon (12). There are pients such as DTPA, sucrose, polysorbate 80 (PS80), me-
multiple mechanisms by which elements may migrate thionine, arginine, and sodium chloride. Beyond their use
from the glass network into the interior solution, includ- in formulations, these materials also provide test cases for
ing congruent and incongruent dissolution, reactive cor- strong metal chelators and high ionic strength environ-
rosion, and alkali leaching. The most significant variable ments, both of which are chemically relevant scenarios in
that impacts the leaching mechanism is the pH of the generating leachables knowledge.
contact solution, which may in turn be variable over
time as the concentration of ionic leachates increases. The glass vial storage conditions of the various buffer
Additionally, leachable concentrations may be further formulations are also expected to have an impact on
exacerbated by other drug formulation components such the extent of elemental impurities leachables. For prac-
as metal chelators (e.g., pentetic acid [DTPA]). Owing tical purposes, aligning a leaching study with the drug
to the broad spectrum of possible parenteral pharmaceu- product’s long-term stability studies (LTSSs) would
tical formulations, it cannot be assumed that two drug present a convenient opportunity to monitor trends
products of different formulations will elicit the same associated with both real-world and exaggerated stor-
elemental impurities leachable profile from the same age conditions over the life cycle of the drug. For the
vial. Other factors such as storage conditions (solution purpose of this investigation, and as a means of reduc-
residence time and temperature) and vial type (molded ing the number of compounding variables to only the
vs. tube, volume) may also contribute to the leachable formulations, the storage conditions for results pre-
variability in glass vial–drug formulation systems (13, sented here represent only those solutions exposed to
14). glass vials for 6 months at 25˚C.
Histidine is a frequently used buffer component in many In order to assess the extent of elemental impurity leach-
commercial monoclonal antibody (mAb) formulations ables attributed to any specific variable, a highly sensi-
owing to its physiologically relevant buffering range as tive analytical technique must be employed. Inductively
well as its freeze/thaw and lyophilization-friendly phase coupled plasma mass spectrometry (ICP-MS) has been
behavior, as opposed to buffers such as phosphate and demonstrated to be an ideal instrument for accurate,
acetate. Approximately one-third of ready-to-use liquid quantitative leachable work (16–20). ICP-MS is a solu-
and lyophilized commercialized mAb drug products are tion-based analytical technique that provides both quali-
reported to use histidine as an ingredient (15). As a tative and quantitative multielement analysis with the
result, generation of data and knowledge on the elemen- possibility of sub-ppb detection limits.
tal impurities leachable potential of histidine buffers
would be of paramount interest to inform formulation- This study is separated into two parts, the first being a
induced packaging risks over the shelf life of many par- broad design of experiment (DoE) for the purpose of
enteral drug products. studying the extent of leaching from a single, specific type
of glass vial as a function of exposure to 14 different
However, despite it being one of the dominant buffer buffer/excipient formulations, all stored under the same
selections for parenteral drug products, generating conditions for 6 months. All the solutions employed in
leachable knowledge from only histidine formulations this part of the study were placebos and did not contain
would be a vastly incomplete coverage of available any actual drug product. All ICH-Q3D elements (except
buffer options. The addition of citrate, phosphate, and Os) were assessed as potential leachables. Several non-
tris to the experimental design space greatly expands ICH-Q3D elemental impurity elements were also included
Vol. 77, No. 3, May--June 2023 199

in this part of the study: Ti, Zn, Zr, and Al were included simultaneously to test the extremes of the CCS (in this
because, at the time of the study design, these elements case, buffer formulation composition) on the observed
were under consideration for inclusion in extractable studies leachable results. The samples were tested similar to a
of USP <661.1> Plastic Materials of Construction and full design approach. This design approach assumes that
were considered relevant to glass vial packaging leachable the stability of any intermediate level is within the sta-
studies. Additionally, although Al and Zn are not elements bility results observed for the extreme samples tested. If
with prescribed ICH-Q3D permitted daily exposures differences in leachables are observed across the range
(PDEs), they are identified as elements that may have studied, the individual factor causing the difference can-
patient-specific or regional toxicological limits (4). Other not be identified statistically. However, if the entire
elements that are major components of glass construction range of leachables observed are within acceptable lim-
(e.g., Si or B) but that do not pose toxicological concerns its, then it can be concluded that the entire range of
were not considered in this study; however, they were factors studied has no practical effect on elemental
assessed in a parallel study of glass delamination propensity. leachables.
The second part of this study involved leveraging the ana- In order to design the histidine formulations that would
lytical parameters employed during the buffer leachable most effectively capture the most aggressive scenarios,
DoE previously described and developing a novel, gen- a factorial design was created using JMP software. A
eral, elemental impurities ICP-MS method that can be factorial design is a type of experiment that allows
applied for the assessment of only ICH-Q3D elemental researchers to study the effects of several different fac-
impurity leachables in authentic drug product formula- tors on a response. When conducting a DoE, varying
tions (as opposed to only placebo buffer formulation). the levels of all factors within the design (instead of
This general method was refined and validated on a single across several designs) enables the researcher to study
representative buffer formulation with the expectation the interactions between the factors.
that it can be applied to any biologic formulation follow-
ing minimal matrix compatibility confirmation, a process This study utilized a 23 full factorial design with three
referred to as bridging. It has been demonstrated to pro- identical center points (center points are the midway
vide accurate and precise quantitative data at low limits points between the low and high settings for a factor).
for all ICH-Q3D elemental impurities (except Os, which Center point replicates are included to determine the
is monitored qualitatively for safety reasons). The general experimental repeatability at the center of the experi-
method subsequently has been successfully applied to mental space, which can be compared with the change
many different parenteral drug product formulations, in in responses at different experimental conditions to test
the commercial and development space, to assess for for factor significance. It is assumed that the experi-
leachables at various defined storage time points and con- mental repeatability is similar across the experimental
ditions. Each different drug product formulation has been range. The center point runs also enable the analyst to
‘bridged’ to the method by performing a spike-and-recov- test for nonlinearity across the design, although if non-
ery study at three concentration levels that encompass the linearity is found, the factor causing the nonlinearity
PDE of each element. Successful recovery of these spikes cannot be identified statistically.
ensures that any difference in matrix effects relative to
the representative buffer used to validate the method are Histidine Formulations for Buffer Leachable DoE
negligible and confirms that the method is appropriate for
use. Data collected for each drug product following this Buffers are major excipients in developing a robust
method can then be applied to an ICH-Q3D Elemental drug formulation. Specifically, buffer systems help to
Impurities risk assessment to satisfy the contributions of control the pH, an important factor during drug product
the glass vial CCS component. formulation and processing. In addition to the buffer,
there are eight factors (excipients) that are generally
Experimental Method considered in the formulation: osmolality modifier,
chelator, surfactant, pH, antioxidant, viscosity modi-
Experimental Design for Buffer Leachable DoE fier, and an ionic strength modifier.
A DoE was conducted using a bracketing variable These factors were varied in the design as three inde-
approach, in which multiple factors are changed pendent variables: pH, NaCl, and a combined factor

that includes the remaining six factors. The pH and that was not used to fill vials was stored in Nalgene
NaCl factors are hypothesized to most likely have an plastic containers and submitted to the analytical test-
effect on elemental impurities leachables and were var- ing group for use to matrix-match calibration stand-
ied independently in the study to estimate their inde- ards. Because these formulations were not exposed
pendent effects. The combined set of factors are to any glassware, they could be considered control
hypothesized to have a low likelihood of affecting ele- samples.
mental impurities and thus treated as a single factor.
For example, when the combined factor is at its low Analytical Experimental Details
level ( 1), all of the factors in this group will be at the
low level depicted in Table II. If an effect on elemental Buffer Leachable DoE: For analysis, solutions of inter-
impurities is found with the combined factor, further est were removed from the vials using standard polypro-
experiments will need to be executed to identify which pylene syringes with a stainless-steel needle, typically
individual factor caused the effect. If no effect is iden- used in the pharmaceutical industry. This sampling pro-
tified, it can be concluded that none of the individual cedure was preferred to the procedure of decrimping the
factors have an effect on elemental impurities over the Al cap and removal of the stopper to pour the solution
ranges studied. As described previously, a 23 full facto- over the vial lip into a polypropylene tube. This latter
rial design with 3 center points DoE used for this study preparation was found to impart subvisible, Al particles
leads to 11 histidine buffer formulations with variable
generated by the decrimping process into the solution,
excipient levels. This design has many merits. It can fit
which in turn completely dissolved in the more aggres-
a full model with all main effects (all 2-factor interac-
sive buffer formulations. Solutions were received and
tions) as well as provide the opportunity to test for non-
stored before preparation in polypropylene tubes (SCP
linearity. It is also a high-power design with lower
Sciences) to ensure that no further leaching occurred.
financial and resource burden compared with an eight-
factor design.
All sample solutions were prepared using a 10 dilu-
tion in deionized (DI) water (Millipore, 18.2 MX-cm)
Solution Sample Preparation
to minimize matrix effects and to protect the ICP-MS
cones from the salt buildup that would occur if the
Triplicate vials (Valor, 30 R, Corning) were filled with
samples were to be analyzed at high concentrations.
the various buffer formulations in an ISO 5 laboratory
Ge, Y, and Ho were used as internal standards at a con-
environment. Prior to fill, vials were washed with
centration of 10 ng/mL with the further addition of
water-for-injection (WFI) and depyrogenated under
standard conditions. Fourteen bulk buffer solutions of 10 ng/mL of Au as a stabilizing agent for Hg. Calibra-
varying formulations were prepared using a variety of tion standards were prepared by matrix-matching the
excipient components (histidine [Avantor JT Baker], calibration standards to the buffer solutions through the
sucrose [Pfanstiehl], citrate [Spectrum], phosphate addition of 10% of the corresponding control buffer
[Thermo Fisher], tris [Spectrum], methionine [Sigma formulation (e.g., if citrate buffer solutions were being
Aldrich], arginine hydrochloric acid [Avantor JT analyzed, calibration standards would be prepared with
Baker], sodium chloride [Avantor JT Baker], DTPA the same concentration of citrate as in the samples, fol-
[EMD Millipore], and PS80 [NOF]). The range of com- lowing dilution). Such incorporation of the control (not
ponents in each formulation, if not the exact concentra- exposed to glass) formulation into calibration standards
tions, are provided in Table II. At the onset of the not only served to matrix-match the standards to the
study, triplicate glass vials were filled with 33 mL of samples and thereby increase method accuracy but also
each formulation. The vial type selected for this study to negate the impact any elemental impurities native to
was Valor 30 R, manufactured by Corning. Vials were the formulation would have on the sample solution
stoppered with West/Daikyo 20 mm stoppers and results. With this approach, subtraction of the control
capped with West 20 mm Al caps using a Genesis cap- sample from the sample solutions was not necessary,
ping station. No container was filled with more than and all elements detected in the sample solutions could
one individual formulation. Filled vials were subse- be attributed to glass leaching. Multi-element stock
quently stored upright in monitored and controlled ICP standards (Inorganic Ventures) were used to pre-
(63˚C) stability chambers at 25˚C for 6 months before pare calibration standards in concentrations of 100 and
withdrawal for testing. The remaining bulk formulation 200 ng/mL of each element.
Vol. 77, No. 3, May--June 2023 201

202
Table II
The Formulations Explored in the Glass Vial Leachables DoE
Ionic Strength Osmolality Chelator, Surfactant, Antioxidant, Viscosity,

Formulation Combined Modifier, Modifier, DTPA PS80 (%w/ Methionine Arginine
Number Pattern Factor pH NaCl (mM) Buffer Sucrose (mM) (mM) w) (mM) HCl (mM)
Type Level
(mM)
1 000 0 6 300 Histidine
2 1 5 200 Histidine
3 +++ 1 7 400 Histidine
4 000 0 6 300 Histidine
5 + + 1 7 200 Histidine
6 000 0 6 300 Histidine 250–500 50–100 0.05–0.2 20–50 80–150
20–50
7 + 1 5 400 Histidine
8 ++ 1 5 400 Histidine
11 ++ 1 7 400 Histidine
12 NA NA 3 400 Citrate
13 NA NA 8 400 Phosphate
14 NA NA 9 400 Tris
DoE, design of experiment; DTPA, pentetic acid; NA, not available; PS80, polysorbate 80.

Table III
ICP-MS Operating Conditions for Leachable Elemental Impurities for Both the Buffer Study and the General
Method
Parameter Setting
Element isotopesa 7
Li, 27Al, 47,48Ti, 51V, 52,53Cr, 58,60Ni, 59Co, 63,65Cu, 64,66,68Zn, 75As, 76,77,78Se, 90,92,94Zr,
95,96,98
Mo, 101,102,104Ru, 103Rh, 105,106,108Pd, 107,109Ag, 111Cd, 116,118,120Sn, 121,123Sb,
136,137,138
Ba, 191,193Ir, 194,195,196Pt, 197Au, 200,201,202Hg, 203,205Tl, 206,207,208Pb
74
Internal standard isotopes Ge, 89Y, 165Ho
Interference reduction Kinetic energy discrimination (KED)
configuration
Sampling cone Nickel
Spray chamber Quartz cyclonic
Measurement replicates 3
Dwell time 0.05 seconds
Resolution Normal
Other settings Determined by daily tuning report
a
Ti, Zr, Zn, and Al were included in the buffer leachable study; however, they are omitted from the subsequently developed
General Method as these elements are not considered ICH-Q3D Class 1–3 elements.
ICP-MS, inductively coupled plasma mass spectrometry.
Samples were analyzed in groups according to the Elemental Impurities Leachables General ICP-MS
buffer formulation using a Thermo iCAP RQ ICP-MS Method Development: Following the assessment of ele-
(Table III). All elements listed in ICH-Q3D (except mental impurities leachables from glass vials in the vari-
Os) were included in the analysis, with four additional ous formulations, the analytical work was leveraged for
elements of interest (Al, Zn, Ti, and Zr) commonly the development and validation of a general elemental
used by various packaging manufacturers. The most impurities leachable method. The Histidine 4 buffer for-
sensitive isotopes for each element were selected. Lim- mulation was selected as a permanent, representative
its of detection (LODs) and limits of quantification buffer for calibration standard matrix-matching (10%
(LOQs) for each element isotope were calculated as 3r Histidine 4:90% DI Water) in the method. This buffer
and 10r of the blank solution, respectively, for each would be used regardless of the formulation of the par-
matrix (Table IV). However, to normalize the mini- enteral drug product for which the general method was
mum quantifiable limit (MQL) to a common reporting applied provided a spike-and-recovery bridging study on
limit for all isotopes, elements, and matrices, a practi- that specific matrix was successful. Parenteral drug
cal MQL of 1 ng/mL (on instrument, 10 ng/mL in sam- product samples exposed to glass vials that were ana-
ple following dilution correction) was adopted. lyzed following the general method were weighed to
yield 10 mg/mL sample solutions. Providing data rela-
For the leachable assessment, data was exported from tive to weight (as opposed to volume, as for the buffer
the ICP-MS and processed off-line using Excel. All leachables DoE study) permitted the results to be
isotopes were confirmed to have satisfied quality con- reported in lg/g in alignment with ICH-Q3D. The
trol criteria of standard checks (620%) and those suf- method LOQ for each element was calculated as 30% of
fering from significant interferences were omitted from the target limits, verified by spike recovery. Elemental
calculation (e.g., 114Cd was removed from considera- impurities limits were, in turn, determined from ICH-
tion due to a Sn interference). In cases in which multi- Q3D parenteral PDE limits with one modification: a
ple isotopes of an element were used, the isotopes were blanket daily dose of 100 g/day was utilized to ensure
averaged before calculation of the dilution-corrected that the method would be applicable to a wider variety
sample concentration. Triplicate vials of the same vari- of potential drug product formulations. This results in
able were assessed for extreme variability and the limits that were a factor of ten lower than those detailed
results averaged. Results were presented in ng/mL. in ICH-Q3D for a 10 g/day dose.
Vol. 77, No. 3, May--June 2023 203

Table IV
The On-Instrument Limits of Detection, Limits of Quantification and Practical Limits of Quantification (Those That
Are Used to Report Data) Are Presented for a Representative Histidine Formulation. For Elements with Multiple
Isotopes, Only One Is Presented
Histidine 4
On-Instrument Dilution Corrected
Practical LOQ Practical LOQ
LOD (ng/mL) LOQ (ng/mL) (ng/mL) (ng/mL)
7
Li 0.053 0.057 1.00 10.0
27
Al 0.334 0.748 1.00 10.0
48
Ti 0.096 0.179 1.00 10.0
51
V 0.007 0.013 1.00 10.0
53
Cr 0.171 0.031 1.00 10.0
59
Co 0.002 0.004 1.00 10.0
60
Ni 0.042 0.013 1.00 10.0
65
Cu 0.011 0.024 1.00 10.0
68
Zn 0.052 0.017 1.00 10.0
75
As 0.002 0.013 1.00 10.0
78
Se 0.111 0.305 1.00 10.0
90
Zr 0.004 0.045 1.00 10.0
98
Mo 0.001 0.047 1.00 10.0
103
Rh 0.000 0.000 1.00 10.0
104
Ru 0.002 0.003 1.00 10.0
108
Pd 0.003 0.005 1.00 10.0
109
Ag 0.021 0.001 1.00 10.0
111
Cd 0.001 0.002 1.00 10.0
120
Sn 0.002 0.004 1.00 10.0
123
Sb 0.003 0.023 1.00 10.0
138
Ba 0.007 0.003 1.00 10.0
193
Ir 0.000 0.001 1.00 10.0
196
Pt 0.001 0.004 1.00 10.0
197
Au 0.018 0.055 1.00 10.0
202
Hg 0.015 0.003 1.00 10.0
205
Tl 0.000 0.000 1.00 10.0
208
Pb 0.001 0.003 1.00 10.0
LOD, limit of detection; LOQ, limit of quantification.
Results and Discussion glass used in the formulation of the test vials was made
from an alumino-silicate matrix and the proprietary
Assessment of Buffer Leaching Profiles manufacturing process used SnO2 as a fining agent
(21), the leaching of these two elements was not unex-
Results for the elements that exhibited leaching pected. Furthermore, of these two elements, Sn is the
>10 ng/mL for at least one or more buffer systems are only element controlled with specific limits by ICH-
presented in Table V. Of all the elements investigated, Q3D. Al is subjected to possible regional regulatory
only Al, Cu, Zn, and Sn were found to leach from the limits. Despite this, neither element approached the
glass vials, and of these, only Al and Sn were found to 100 g/day parenteral drug product limit utilized in this
leach from all 14 buffer formulations. Because the study as an alert concentration. The results indicate

Table V Table VI
Elements Detected in Each of the Formulations Al Concentration Variability Determined in
Following Dilution Correction Are Presented Relative Triplicate Vials for Phosphate and Tris Formulations
to the LOQ of 10 ng/mL (Corrected for 10 dilution). Demonstrating the Extent of Leachable
Al and Sn Were the Only Leachable Elements Detected Reproducibility
in All Formulations
Al Concentration (ng/mL)
Analyte Concentration (ng/mL)a Phosphate Tris
Al Cu Zn Sn
Vial 1 750 520
Histidine 1 150 ND ND 38 Vial 2 800 650
Histidine 2 73 ND ND 32 Vial 3 800 590
Histidine 3 260 ND ND 40
Histidine 4 140 22 18 36
Some Cu and Zn leachates (>10 ng/mL) were observed
Histidine 5 220 13 ND 38 in one of the histidine formulations. However, this was
Histidine 6 160 ND ND 39 limited to only one of the triplicate vials (which also
Histidine 7 73 ND ND 39 exhibited trace quantities of Hg and Pb, <10 ng/mL).
Histidine 8 100 ND ND 34 Thus, this vial was designated an outlier as it is not rep-
Histidine 9 240 ND ND 39 resentative of typical histidine leaching under these
Histidine 10 77 ND ND 32 conditions. The detection of these elements is likely
Histidine 11 300 ND ND 43 due to environmental contamination or spurious, anom-
Citrate 83 ND ND 43 alous leaching effects. Because this work was not per-
Phosphate 780 ND ND 79 formed in a designated clean room and an analytical
Tris 590 ND ND 48 technique with sub-ppb sensitivity was employed for
a
Average of three vials. detection, infrequent ultratrace contamination is an
LOQ, limit of quantification; ND, not detected— unfortunate reality. In order to ensure that such trace
concentrations below the LOQ. contaminations do not skew the data with artificial con-
centrations, it is vital that multiple vials (triplicate, at
minimum) be interrogated. If such anomalous or unex-
that none of these formulations stored in the type of pected results arise at similar levels in all replicates,
glass vial under consideration would present a concern then further investigation would be warranted.
for patient safety or product quality resulting from
leachable elemental impurities. The higher Al concentrations observed in the phosphate
and tris buffer systems were not attributable to Al con-
The results indicated that under the conditions in this tamination from the cap. Table VI demonstrates the
study, the phosphate leached the highest concentrations repeatability of the Al concentrations for triplicate indi-
of Al and Sn of the 14 buffer formulations studied. Tris vidual vial measurements, indicating little variability.
also demonstrated greater Al leaching relative to citrate When Al contributions from particles liberated from the
or the 11 histidine formulations, all of which exhibited cap were observed, they tended to be highly variable
very similar leaching patterns. Generally, glass attack is from vial-to-vial and exhibited obvious anomalously
dependent on glass formulation and pH. Ion exchange or high concentrations in excess of 1000 ng/mL. As such,
hydrolysis occurs in acidic and neutral formulations, the Al concentrations demonstrated in Table VI are con-
whereas dissolution occurs in alkaline media. Thus, sidered to be solely from the glass vial. Contributions
more aggressive reactions occur at alkaline conditions. from the stopper were deemed negligible, because the
The phosphate and tris buffer solutions were formulated vials were stored in the upright position and contact
at pH of 8 and 9, respectively, aligning with the between the solution and stopper was minimized if not
observed data. The differences noticed in metal leach- completely excluded. In order to avoid this potential
ates between both solutions may be related to the nega- conflict of possible Al contributors in similar leachable
tive pH shifts at higher temperatures in tris buffer and studies, it is recommended to either forego the use of Al
the chelation/coordination abilities of phosphate vs. tris caps or remove the solution for analysis using a syringe
with metals in glass. ensuring no contact with the Al rim of the vial. This
Vol. 77, No. 3, May--June 2023 205

potentially be contributed to the liquid solution from

Table VII
the glass vial used for storage. It is expected that any
Among the Different Histidine Formulations Used in
elemental impurities present in components used to
the Evaluation Study, the Formulation Designated as
prepare the formulations, and therefore the final formu-
‘Histidine 4’ Was Assessed to Be the Most
lations themselves, have been evaluated separately and
Representative of Commonly Used Parenteral
Formulations and As Such Was Selected as the deemed to comply with ICH-Q3D elemental compo-
Formulation to Semi-Matrix Match Calibration nent limits before use. Had any of the formulations
Standards for a General Elemental Impurities used in this study contained trace levels, then the incor-
Leachable Method Applicable to Other Histidine or poration of the same formulation, at the same concen-
Non-Histidine Parenteral Drug Products tration, into the calibration standards (exact matrix-
matching) would negate those contributions. The use
Formulation: Representative Buffer Histidine 4 of exact or close-to-exact matrix-matched standards for
Excipienta Function ICP-MS quantification ensures that all leachables
Histidine Buffer determined during the study were attributable to only
Sucrose Stabilizer & tonicity modifier the glass vials, and that results for different formula-
Methionine Antioxidant tions can be compared directly without normalization
Arginine HCl Viscosity modifier for the formulation.
NaCl Viscosity modifier
a However, if formal leachable studies are to be per-
Excipients shown are those with no less than 0.5%
formed on parenteral drug products (with active ingre-
concentration by weight.
dients, not just buffer systems), as recommended by
Note: Trace amounts of surfactants and chelators also
contribute to the formulation. ICH-Q3D, replication of a matrix-matched study during
development may not be possible or practical. Formu-
lated drug product may not be available in large enough
prevents potential for microscopic Al particles that may quantities for incorporation into the calibration standards
subsequently be dissolved into the solution. necessary to provide matrix-matched quantitative analy-
sis. Additionally, full validation studies would be required
Although the kinetic properties of leaching are expected on each separate drug product that utilized matrix-
to be similar for the same formulations of glass subjected matched standards to provide results of acceptable accu-
to the same conditions, some variabilities can be expected racy. Based on the results of this buffer leachable DoE
in the leached elements among triplicate samples of the and the determination that the 14 different buffer formu-
same formulation. Although this is certainly the case for
lations exhibited similar behavior, a singular, general
Al, the Sn leaching between vials of the same formulation
elemental impurities leachable method, applicable to an
was remarkably consistent. Furthermore, for the histidine
array of different parenteral drug products, was devel-
and citrate formulations, the Sn concentrations were
oped using a single, representative buffer formulation as
reproducible between formulations as well as within.
the matrix with which to match the calibration stand-
This indicates that the incorporation of Sn into the glass
during the manufacturing process results in a leaching ards. Histidine 4 (Table VII) was selected as this repre-
that is, for the most part, independent of the chemical sentative matrix because it is the center point of the
properties of the buffer formulation (pH, and so forth). study DoE, representing the midpoint between the low
As Sn is the only element with prescribed ICH-Q3D lim- and high settings for a factor. It was also selected as the
its, this provides further evidence that regardless of the placebo drug product formulation used to perform a full
buffer formulation used (at least within the broad confines validation of the method according to USP <233>.
of the buffer systems selected for this study), Sn presents Details and acceptance criteria of the validation study
a very low risk of being an element of concern with vari- are presented in Tables VIII and IX, and all acceptance
able leaching that could ultimately impact patient safety. criteria for a general elemental impurities leachable
method were satisfied.
Development and Validation of a General Quantitative
Elemental Impurities Method Applying a general method with a representative buffer
to matrix-match calibration standards presents the chal-
The purpose of the buffer leachable DoE study was to lenge of not being appropriately validated specifically
evaluate only the elemental impurities that could for individual, varying drug product formulations. This

Table VIII
Validation Study Criteria for a General Method for Elemental Impurity Leachables, Using the Histidine 4 Formulation
for Matrix Matching the Calibration Curves and As a Test Sample. The PDE For 100 g/day Parenteral Dosages Is
Adopted as the Method Reporting Limit and the Concentration Around Which the Validation Is Structured
Validation Test Procedure and Acceptance Criteria

Method quantification Procedure: Sample blank and a single 30% of 100 g/day
limit (MQL) PDE sample spike
Criteria: Spike recovery of 70%–150%
Method reporting limit 100 g/day PDE concentration for each element
(MRL)
Linearity of standards Procedure: Blank + (5) (triplicate) standards over the calibration range, including 30%,
50%, 75%, 100%, and 150% of the 100 g/day PDE
Criteria: Mean recoveries of 50%, 100%, and 150% of the 100 g/day PDE are 680%–
120%. Correlation coefficient for mean values ≥0.99
Accuracy Procedure: Sample blank and (3) each of 50%, 100%, and 150% of the 100 g/day PDE
spikes
Criteria: Mean spike recoveries of 70%–150%
Repeatability/Precision Procedure: Six independent samples spiked at the 100 g/day PDE
Criteria: %RSD <20%
Ruggedness/Intermediate Procedure: Six independent samples spiked at the 100 g/day PDE on a second day, using a
precision second instrument or by a second analyst
Criteria: %RSD for combination of ruggedness and repeatability (n = 12) <25%
Specificity Satisfied by the accuracy requirements
PDE, permitted daily exposure; RSD, relative standard deviation.
presents some inherent risk, albeit small. Although a Implications

full validation using a representative placebo (e.g., His-
tidine 4) can be expected to provide a high level of cer- The buffer leachable DoE described in this study
tainty that the method is appropriate for its intended allows for the comprehensive assessment of elemental
use (quantitative determination of elemental impurity impurity leaching in glass vials subjected to 14 differ-
leachables in any parenteral drug product formulation), ent buffer formulations that are representative of those
it is prudent to perform some degree of qualification that may be incorporated into final parenteral drug
the first time that a new drug product is subjected to products. The data indicated that, although the majority
this general method. Such qualification is referred to as of elemental impurities did not leach into the solutions
‘bridging’ and serves as additional support that the ana- at concentrations exceeding 10 ng/mL, some elements
lytical method can be applied to the new formulation. such as Al and Sn did exhibit leaching to various
Bridging is performed by analyzing the formulated degrees. Regardless, the calculated patient exposure
drug product control sample (not exposed to glass), using the maximum leached concentrations and poten-
followed by three spike-and-recovery studies at 50%, tial maximum daily dose showed that the elemental
100%, and 150% of each element’s PDE for a 100 g/day impurities were considerably below the permitted lim-
dose. Once recovery acceptance criteria are satisfied, the its according to ICH-Q3D and presented no risk to
method is deemed successfully bridged and subse- patient safety. The results of this study could be incor-
quently acceptable for use for elemental impurity leach- porated into a risk assessment to avoid further testing.
ables in that specific drug product. Data from a bridging
spike-and-recovery study using a parenteral drug prod- The analytical ICP-MS methodology (sample prepara-
uct in development using the general method are shown tion, matrix-matched calibration standards, instrument
in Table X and demonstrate that spike recovery accep- parameters, and so forth) refined to perform this study
tance criteria of 70%–150% as per USP <233> are was verified as an appropriate general method for the
satisfied. assessment of elemental impurity leachables in any
Vol. 77, No. 3, May--June 2023 207

Table IX Table X
Analyte Limits (mg/g), Based on a 100 g/day Dose Results of the Bridging Study for a Drug Product
Formulation Analyzed According to the General
Reporting Minimum Quantifiable
Elemental Impurities Leachable Method. The
Analyte Limit (mg/g) Limit (mg/g)
Bridging Study Acceptance Criteria Is for an
Li 2.5 0.75 Average Recovery of Triplicate Spiked Samples at
V 0.1 0.03 Various Concentrations (50%, 100%, and 150% of
Cr 11 3.30 the 100 g/day PDE) of 70%–150%. For Simplicity,
Ni 0.2 0.06 Only Single Isotopes for Each Elemental Impurities
Co 0.05 0.15 Are Shown. Results Demonstrate the Method Is
Cu 3 0.90 Successfully Applied to This Drug Product
As 0.15 0.45 Formulation
Se 0.8 0.24
Element Recovery (%)
Mo 15 4.5
100% of
Rh 0.1 0.03 50% of Target Target 150% of Target
Ru 0.1 0.03 7
Li 94 94 94
Pd 0.1 0.03 51
V 104 97 99
Ag 0.1 0.03 53
Cr 102 97 102
Cd 0.02 0.006 59
Co 101 95 96
Sn 6 1.80 60
Ni 112 125 100
Sb 0.9 0.27 65
Cu 106 100 102
Ba 7 2.1 75
As 102 97 98
Ir 0.1 0.03 78
Se 101 98 99
Pt 0.1 0.03 98
Mo 117 111 116
Au 1 0.3 103
Rh 102 95 101
Hg 0.03 0.009 104
Ru 102 97 99
Tl 0.08 0.024 108
Pd 104 97 101
Pb 0.05 0.15 109
Ag 101 100 99
111
Cd 109 101 107
120
Sn 102 94 98
parenteral drug product in development, subject to a 123
Sb 106 97 100
simple bridging study. The availability of a singular, 138
Ba 103 97 100
validated ICP-MS method that assesses all ICH ele- 193
Ir 105 104 107
mental impurities allows for the generation of data at 196
Pt 97 97 99
various time points in long-term stability studies that 197
Au 97 98 99
are both comparable within a study and between differ- 202
Hg 90 99 103
ent drug products. The data from such drug product- 205
Tl 101 101 103
specific elemental impurities leaching studies are 208
Pb 103 99 101
appropriate for use in an ICH-Q3D elemental impur-
ities risk assessment to satisfy the CCS contribution of
the component option.
Furthermore, the study identified the degree of elemen-
tal impurity leaching from glass vials under fixed envi-
Conclusion ronmental controls using multiple buffer systems that
are common in parenteral drug formulation systems.
The preceding study presented a new, general method Results indicated that for the vials investigated during
that can determine the extent of elemental impurities this study, elemental impurity leachables pose no risk
leachables in biological parenteral drug formulations. to patient safety, regardless of the buffer formulation
The method developed and validated was successfully tested. This method has potential further applications
applied to a number of biologics under development. in comparing different vial types and sizes, as well as a

variety of different buffer formulations and excipient Accumulation in Pharmaceutical Drug Products
components. and Related Solutions. PDA J. Pharm. Sci. Tech-
nol. 2011, 65 (2), 166–176.
Acknowledgements
8. Jenke, D. R. Evaluation of Model Solvent Sys-
The authors would like to acknowledge the following tems for Assessing the Accumulation of Container
Bristol Myers Squibb colleagues: Derrick Parfitt, Shilpa Extractables in Drug Formulations. Int J Pharm
Patel, Olga Yee, Dana DeSantis, Daniel Paniagua, and 2001, 224 (1–2), 51–60.
Celia Wang for their technical contributions to this
paper. 9. Zdravkovic, S. A. Comparison of the Solubilization
Properties of Polysorbate 80 and Isopropanol/Water
Conflict of Interest Declaration Solvent Systems for Organic Compounds Extracted
from Three Pharmaceutical Packaging Configura-
The authors declare that they have no competing interests. tions. Eur. J. Pharm. Sci. 2016, 93, 475–483.
References 10. Jenke, D.; Odufu, A.; Poss, M. The Effect of

Solvent Polarity on the Accumulation of Leach-
1. U.S. Food and Drug Administration, Guidance for ables from Pharmaceutical Product Containers.
Industry: Container Closure Systems for Packag- Eur. J. Pharm. Sci. 2006, 27 (2–3), 133–142.
ing Human Drugs and Biologics. Chemistry, Man-
ufacturing, and Control Documentation. Center 11. Shelby, J. E. Introduction to Glass Science and
for Biologics Evaluation and Research. U.S. Technology, 2nd ed.; Royal Society of Chemistry:
Department of Health and Human Services: Rock- Cambridge, U.K., 2005.
ville, MD, 1999.
12. White, W. B.; Theory of Corrosion of Glass and
2. U.S. Pharmacopeial Convention, General Chapter Ceramics. In Corrosion of Glass, Ceramics and
<232> Elemental Impurities—Limits. In USP 41— Ceramic Superconductors: Principles, Testing,
NF 36, USP: Rockville, MD, 2018; pp 6147–6151. Characterization and Applications; Clark, D. E.;
Zoitos, B. K. Eds.; Noyes Publications: Park
3. U.S. Pharmacopeial Convention, General Chapter Ridge, NJ, 1992; pp 2–28.
<233> Elemental Impurities—Procedures. In
USP 41—NF 36, USP: Rockville, MD, 2018; pp 13. Schaut, R. A.; Porter Weeks, W. Historical Review
6151–6455. of Glasses Used for Parenteral Packaging. PDA J.
Pharm. Sci. Technol. 2017, 71 (4), 279–296.
4. U.S. Food and Drug Administration, Guidance for
Industry: Q3D Elemental Impurities. Center for Bio- 14. Swift, R.; Glass Containers for Parenteral Products. In
logics Evaluation and Research. U.S. Department of Pharmaceutical Dosage Forms—Parenteral Medica-
Health and Human Services: Rockville, MD, 2020. tions, 3rd ed; Nema, S.; Ludwig, J. D., Eds.; Informa
Healthcare: New York, 2012; Vol. 1, pp 287–304.
5. International Conference for Harmonization .Q3D
Elemental Impurities Guidance for Industry, ICH: 15. Kang, J.; Lin, X.; Penera, J. Rapid Formulation
Geneva, 2015. Development for Monoclonal Antibodies. BioPro-
cess Int. 2016, 14 (4), 40–47.
6. Zdravkovic, S. A. Comparison of a Lyophilized
Drug Product to Other Solid and Liquid Media for 16. Van Hoecke, K.; Catry, C.; Vanhaecke, F. Determi-
the Extraction of Elastomeric Oligomers from a nation of Elemental Impurities in Leachate Solutions
Butyl Rubber Stopper. PDA J. Pharm. Sci. Tech- from Syringes Using Sector Field ICP-Mass Spec-
nol. 2017, 71 (6), 488–501. trometry. J Pharm. Biomed. Anal. 2013, 77, 139–144.
7. Jenke, D. A General Assessment of the Physio- 17. Solomon, P.; Nelson, J. Profiling Extractable and
chemical Factors That Influence Leachables Leachable Inorganic Impurities in Ophthalmic
Vol. 77, No. 3, May--June 2023 209

Drug Containers by ICP-MS. Pharm. Dev. Tech- Leachables from Saline-Containing IV Bags Can
nol. 2018, 23 (3), 247–254. Alter Therapeutic Protein Properties. Pharm. Res.
2010, 27 (11), 2402–2413.
18. Dorival-Garcia, N.; Carillo, S.; Ta, C.; Roberts,
D.; Comstock, K.; Lofthouse, S.; Ciceri, E.; 20. Menzel, R.; Pahl, I.; Loewe, T.; Hauk, A. Compar-
D’Silva, K.; Kierans, G.; Kaisermayer, C.; Lin- ative Extractables Study of Autoclavable Polyether-
deberg, A.; Bones, J. Large-Scale Assessment of sulfone Filter Cartridges for Sterile Filtration. PDA J.
Extractables and Leachables in Single-Use Bags Pharm. Sci. Technol. 2018, 72 (3), 298–316.
for Biomanufacturing. Anal. Chem. 2018, 90 (15),
9006–9015. 21. Schaut, R. A.; Peanasky, J. S.; DeMartino,
S. E.; Schiefelbein, S. L. A New Glass Option for
19. Chang, J. Y.; Xiao, N. J.; Zhu, M.; Zhang, J.; Parenteral Packaging. PDA J. Pharm. Sci. Tech-
Hoff, E.; Russell, S.; Katta, V.; Shire, S. J. nol. 2014, 68 (5), 527–534.

TECHNOLOGY/APPLICATION
Rapid Sterility Test Systems in the Pharmaceutical Industry:

Applying a Structured Approach to Their Evaluation,
Validation and Global Implementation
SVEN DEUTSCHMANN1, MOUSUMI PAUL2, MARJA CLAASSEN-WILLEMSE3, JONAS VAN DEN BERG4,
PIETA IJZERMAN-BOON3, VIVIANE GRUNERT DA FONSECA5, ELLEN BRUNBECH6, LYNN JOHNSON7,
CHRIS KNUTSEN8, LUCILE PLOURDE9, JOANNY SALVAS10, PHILIP VILLARI2, and LISA WYSOCKI11
1
Roche Diagnostics GmbH, Penzberg, Germany;2Merck & Co., Inc., Kenilworth, NJ;3MSD, Oss, The Netherlands;4Roche
Diagnostics GmbH, Mannheim, Germany;5Roche Diagnostics GmbH, Penzberg, Germany;6Novo Nordisk A/S, Bagsvaerd,
Denmark;7Takeda, Lexington, MA;8Bristol-Myers Squibb, New Brunswick, NJ;9Sanofi Pasteur, Marcy-l’Etoile,
France;10Pfizer, Kirkland, QC, Canada; and11GlaxoSmithKline, Collegeville, PA © PDA, Inc. 2023
ABSTRACT: The current compendial sterility test has a 14-day incubation time and is often the time-limiting step in the
Assess and Release Process of pharmaceutical products. There is an ever-increasing number of technologies available
on the market that have benefits in addition to faster Time to Result, such as standardization and automation of readout
(eliminating analyst subjectivity) and improved data integrity (including eliminating the need for contemporaneous ver-
ification of the result by another analyst). Regulators have been encouraging the pharmaceutical industry to adopt these
innovative systems; however, it has taken a considerable time before receiving the first approvals from various health
authorities (including both the European Medicines Agency and Food and Drug Administration) for the use of an alter-
native and rapid sterility test for the release of sterile drug product lots. This article describes a systematic 9-step
approach to the evaluation, equipment qualification, validation, and deployment of alternative sterility tests that can be
applied by pharmaceutical companies wanting to take advantage of the numerous benefits of alternative sterility tests.
Two case studies are presented to illustrate the validation and implementation approach, including statistical methods.
Although most of the steps toward implementation are aligned, the validation and transfer have been approached differ-
ently for each of the case studies because of differences in the chosen technology as well as independent company inter-
nal decisions to comply with validation guidelines. However, both case studies show successful implementation of an
alternative sterility test for sterile drug products with an 50% reduced incubation time.
KEYWORDS: Alternative and rapid microbiological methods, Alternative sterility test, Method validation, Microbio-
logical method transfer, Regulatory filing, Statistical methodology.
Introduction (4) demonstrated the existence of certain pathogenic micro-

organisms using swan-neck bottles containing broth (direct
Compendial sterility testing (1–3) is performed using the inoculation technique), which he examined for turbidity.
technique of either membrane filtration or direct inoculation Although some parts of the method have changed over
of culture media with the sample to be tested. Two separate time, like the membrane filtration, rinsing of the membrane,
media must be incubated for 14 days at two different tem- length of incubation, and media types, the mechanism of
peratures. Results are based on visual observation and ex- detection (the human eye) has not changed. Second analyst
amination of the media for turbidity. The basis of this verification has been added to reduce interpretation errors,
method has been in existence for generations; Louis Pasteur doubling the efforts to read the results. Recent advances in
alternative and rapid sterility testing circumvent the inher-
*
Corresponding Author: BioPhorum Operations ent subjectivity of the compendial tests while delivering
Group Limited, 5 Westbrook Ct., Sharrowvale Rd., faster time to detection, in this article referred to as Time to
Sheffield, S11 8YZ, United Kingdom. E-mail: Result (TTR), than the compendial method.
margit.franz-riethdorf@biophorum.com
doi: 10.5731/pdajpst.2021.012672 This article provides a structured approach to the evalua-
tion, validation, and implementation of alternative and
Vol. 77, No. 3, May--June 2023 211

rapid sterility test systems. This structured approach has The general framework has already been applied to auto-
been developed collaboratively by experts challenged mated colony counting systems (6). A paper on biofluor-
with the task of implementing alternative and rapid ste- escent particle counters using this framework is planned
rility test systems in their respective pharmaceutical for mid-2024, and a paper on adventitious agents is
companies. Two comprehensive case studies are pro- planned for publication in late 2021/early 2022 [7].
vided; whereas Steps 1 to 6 are aligned, Steps 7 and 8
will provide details of the differences including study In the present article, the general framework is applied
design and statistical methods. The case studies cover to alternative sterility testing.
the following technologies:
The 9 steps described in this framework and applied
1. a respiration-based alternative sterility test, and here are as follows:
2. an adenosine triphosphate (ATP) bioluminescence- Step 1: Identify the operational/business need

based alternative sterility test in liquid growth media.
Step 2: Define the application
The case studies are based on experience in two com-
panies and were developed independently from each Step 3: Assess the requirements
other. Both case studies are discussed extensively, and
recommendations are provided based on the collective Step 4: Compare options and technologies—landscap-
experience among the authors and contributors and the ing and candidate(s) selection
feedback of competent authorities. In the case studies, a
range of products were assessed, including vaccines, a Step 5: Develop a business case: technical, quality, and
biologic therapeutic protein drug product, and a biologic business evaluation and justification
therapeutic monoclonal antibody, all of which are distrib-
uted to global markets. The specific application pertains Step 6: Perform proof-of-concept studies/feasibility
to the commercial release of the medicinal product/drug studies/prevalidation studies
product registered as a sterile, parenteral product in ac-
cordance with the regulations specific to the distribution Step 7: Validate at pilot or primary site
markets. Note that the respiration-based alternative steril-
ity test as described in Case Study 1 has been approved Step 8: Deploy global/company-wide qualification of
by health authorities for in-process testing, but not yet for additional laboratories
release testing of final drug product.
Step 9: Define regulatory filings and implementation
Neither case study is a recommendation for any spe- strategy
cific vendor or technology.
All or some of these steps may be taken by a company
Approach to Alternative and Rapid Microbiological as necessary to fulfill the needs of that firm. It must be
Methods Adoption: Applying a 9-Step Framework noted that this may not be the only approach, but it is
based upon the input from members of the industry
There are variations in the approaches taken by compa- who are active in this process.
nies to implement any Alternative and Rapid Microbio-
logical Method (ARMM), such as alternative sterility Step 1: Identify the Operational/Business Need
testing or automated colony counting. The process can
be broken up into 9 steps, from initial evaluation up to Some of the business needs are more readily apparent
final implementation of the method. This process has than others. For example, the compendial sterility test
been described by a BioPhorum collaborative team in a requires at least 14 days of incubation and is consistently
previous publication (5) that provides a general frame- a limiting factor in final product release, so a faster TTR
work for the evaluation, validation, and implementa- is an obvious driver. Input from stakeholders helps to
tion of alternative and rapid microbiological methods, define the business needs specific to the product being
addressing the questions and considerations relevant in tested. At both companies, stakeholders included analyt-
this context. ical leads, sterility test analysts, quality data integrity

groups, quality assurance product leads, and regulatory Step 2: Define the Application
leads for the product franchise, subject matter experts in
sterility testing, and others. For an alternative and rapid Alternative and rapid sterility testing can be used in
sterility method, the following business needs were any situation where the compendial test is currently
identified: being used. These include:
1. Data Integrity: 1. Testing of drug substance (if applicable) and drug

product as a release test
Automated readout
2. Testing for absence of microbial contamination in
Reduced subjectivity low bioburden biologics as an in-process control
Reduced human errors 3. Stability – sterility of bulk and final container
21 Code of Federal Regulations (CFR), part 11, 4. Culture media in-release testing
compliance, including a more detailed and com-
prehensive audit trail 5. Processes that require post-incubation subculture
2. Accelerated TTR: Step 3: Assess the Requirements
Early notification of positive results will expe- The determination of company-specific requirements
dite root cause investigations and prevent long can only be a snapshot of the current state of technol-
shut down of the filling line. ogy available for alternative sterility testing at the time
of a study.
Cycle time reduction to support batch release,
relevant especially for products with a limited In this article, four technologies were selected for com-
shelf life parison. These alternate technologies are characterized
by the same similarities to the compendial method, that
Reduction of inventory, resulting in significant is, growth-based microbial detection, except for solid-
cost savings phase cytometry. The principle of detection is different
with these technologies:
Freeing up of incubator space, resulting in addi-
tional cost savings if you can downsize or avoid Respiration (CO2 detection): A fully automated ste-
upsizing of incubator capacity rility test system utilizing colorimetric detection of
carbon dioxide production for the detection of mi-
3. Automation and Efficiency: crobial growth. The system incubates, agitates, and
monitors cultures continuously (8).
High throughput, efficient, cost-effective solution
needed in comparison with compendial sterility ATP bioluminescence: Two technologies using an
testing ATP-based bioluminescence reaction. For both meth-
ods, a microbial enrichment in culture broths (9) or on
Desire to move toward a paperless lab agar plates (10) is mandatory after the filtration of the
sample.
4. Quality:
Solid-phase cytometry: A test system that employs a
Capability for real time monitoring of the steril- combination of direct fluorescent labeling and solid-
ity test (if applicable) phase laser scanning cytometry to rapidly enumerate
viable microorganisms without any prior growth
See also Data Integrity preceding amplification (11).
Vol. 77, No. 3, May--June 2023 213

Step 4: Compare Options and Technologies - technology investment, does the alternate method
Landscaping and Candidate(s) Selection add or reduce lab costs?
After identifying possible technologies for an alterna- 5. Do automated methods prevent human error by the
tive sterility test that theoretically fulfills the user standardization of operations (e.g., compared with
requirements, the different technologies still have to be the compendial method, the visual observation to
compared with each other for a final decision to be assess contamination can be replaced by an auto-
reached. mated data interpretation that eliminates contempo-
raneous verification)?
With the relevant information gathered for the compet-
ing options, the technologies can be readily compared Step 5: Develop a Business Case: Technical, Quality,
attribute by attribute (Table I). Most importantly, any and Business Evaluation and Justification
technology that does not meet mandatory requirements
should not be selected, which may narrow the options For many of these technologies, there is an upfront cost
down to a single candidate. In some cases, however, to pursuing the alternative and rapid microbiological
there are multiple competing viable options. In these method, which is later balanced by the Return on
situations, the application-specific requirements (Table Investment (ROI). Upfront costs for capital, investment
II) can be examined more thoroughly. Sites can then in method development and validation, data lifecycle
perform analysis using a local process, in which the management, generation of documents required to per-
various requirements are assessed—either formally or form the alternative microbiological method, change
informally—to help reach a decision. Users may control procedures, and so forth, may discourage com-
decide, for example, that TTR is the most important panies, so it is essential to convey the ROI addressing
requirement, whereas decreasing complexity of the both the qualitative and quantitative aspects. A strong
method is secondary. The importance of all the require- business case can help the technology gain approval
ments should be evaluated. In terms of the assessment and sponsorship for capital investment. Typically,
or the selection process, it is imperative to consider these assessments are processed through several levels
questions about the entire process from end-to-end, of- of management approval as well as various teams that
ten with input from analysts who would routinely use evaluate the value proposition and impact to the busi-
the method. Some considerations include: ness (e.g., regulatory, quality control [QC], quality
assurance, safety, legal, and supply chain departments).
1. Are the attributes of the technology inferior, equiva-
lent, or superior to those of the compendial method? Presenting a strong business case requires researching
the current state of the product or site and a comparison
2. Is the product matrix filterable or does it have growth with the benefits that the selected equipment/technol-
inhibitory properties causing false-negative results or ogy delivers. Examples can include a comparison of
interference with the readout technology causing the workflow with the alternative method versus the
false-positive results? With filterable products, the workflow with the compendial method. For the respira-
primary choice would be for the technologies that tion-based method, it involves direct inoculation of the
follow the principle of membrane filtration as recom- media bottles with the product, loading the bottles into
mended by the pharmacopoeias (1–3) to allow rins- the incubator, and reporting the results, which is auto-
ing of the membrane to remove antimicrobial mated. Compared with the compendial method—which
components or properties interfering with the readout involves filtering, rinsing, adding growth media, incu-
technology (use of the membrane filtration technique bation, performing interim reads, and performing a
vs. the direct inoculation technique). final read with second-person verification—one can
illustrate that the rapid method has less room for
3. What volume of product will be tested? Some tech- human error. On the other side, the respiration-based
nologies may require multiple replicates to accom- method has some disadvantages, for example, a third
modate the entire volume. medium must be added for the detection of slow-grow-
ing aerotolerant bacteria when targeting 50% of the
4. For a rapid sterility test, what is the minimum time current incubation time; a specific workaround is
saving needed to create value? To balance with required for detection of molds, as it is not possible to

Table I
Comparison of Sterility Method Platform Candidates Using Mandatory Core Requirements
Mandatory Core Requirements

Required Solid-Phase
Characteristic Attribute Compendial Respiration ATP Bioluminescence Cytometry
On Solid Growth
In Liquid Growth Media Media
Safety Equivalent or Current methods System safety was System safety was Detection tower Detection platform
(Ergonomics) superior to are in line with assessed for pinch assessed for pinch points, camera does not is a closed system

existing company SOPs and points, electronic electronic hazards, and meet definition without access to
method legal (e.g., OSHA hazards, and spills. spills. Safety was found to of a laser and the laser, no
in the US) Safety was found to be be at an acceptable level was not assessed identified safety risk
requirements at an acceptable level with no additional controls as a safety risk
with no additional
controls
Equivalent to existing Equivalent to existing Equivalent to Equivalent to
method method existing method existing method
CFR 21 part 11 Yes N/A Yes Yes Yes Yes
Compliant
Software
Data Integrity Addresses Current methods Current software on Current software on Automated data Automated data
data integrity use manually system and on the system is 21 CFR part 11 capture and capture and storage,
concerns entered results and workstation, computer capable with storage, exporting to
(audit trail, contemporaneous is 21 CFR part 11 configuration. Automated exporting to common database
storage of verification is capable with data capture and storage, common programs available
data, access needed configuration exporting to common database
control, etc.) database programs programs
available, audit trail is available, audit
default trail is default
Access and password Customizable security Access and Access and
controls, audit trail in levels for access and password password controls,
place password controls in place controls, audit audit trail in place
trail in place
Reputable, Current methods Reputable and Reputable and established
established. make use of established company company
215
Table I
216
(continued)

Required Solid-Phase
Characteristic Attribute Compendial Respiration ATP Bioluminescence Cytometry
On Solid Growth
In Liquid Growth Media Media
Vendor Robust and consumables that Reputable and Reputable and
Reputation/ reliable are widely established established
Status equipment. available company company
LIMS Proven, may Current methods Yes, may require some Yes, may require some IT Yes, may Yes, may require
Connectivity require some require manual IT and vendor and vendor interaction and require some IT some IT and vendor
IT and vendor transcription of interaction and customization and vendor interaction and
interaction results into the customization interaction and customization
and LIMS. No device customization
customization integration
possible.
Regulatory Accepted as a Accepted by all Accepted by Health Accepted by Health CBER has Accepted by Health
Authority viable method Health Authorities Authorities, approvals Authorities, approvals confirmed the Authorities,
Acceptance from major markets, from major markets for technology to be approvals from
most of the world some applications acceptable as an major markets,
alternative most of the world
sterility method. for some
Other Health application
Authorities have
approved the
technology
Established in Validated and Widely used; it is a Validated and in GMP Validated and in GMP Validated and in Validated and in
Biopharma / in GMP regulatory operation in industry operation in industry GMP operation GMP operation in
Extent of Use operation in requirement in industry industry
in the Industry industry, or
clear path
established
Technology Proven N/A. Proven technology with Proven technology with Proven Proven technology
Platform / technology. some development some minimal technology with with minimal
Minimal work for end user some minimal development work

ATP, adenosine triphosphate; CBER, Center for Biologics Evaluation and Research; GMP, good manufacturing practice; IT, information technology; LIMS,
perform non-agitating incubation, which is required to
improve detection of, for example, A. brasiliensis; or
chemical solution)
complex matrix as
Solid-Phase
buffer, water or
Cytometry
the technology is limited to the direct inoculation tech-
Laboratory Information Management System; N/A, not applicable; OSHA, Occupational Safety and Health Administration; SOP, standard operating procedure.
(only for not
nology, which can have other disadvantages. Where
for end user

possible, quantify these advantages and/or disadvan-
tages. Some examples include: the respiration-based or
the broth ATP-bioluminescence alternative sterility tests
allow release of product 6–7 days (for example) faster,
which translates to 6–7 fewer days of batch holding time;
On Solid Growth
the equipment/technology eliminates second-person verifi-

development
work for end
cation of samples, which translates to a specific number of
Media
Full-Time Equivalent (FTE) savings over the course of the

year. In this way, ROI can be perceived as positive, result-
user
ATP Bioluminescence
ing in more readily acceptance by sponsors.
Because ROI may not be solely a financial gain but a gain

development work for end
In Liquid Growth Media
in different aspects like the ones listed previously, a holis-

tic business case utilizing a ranking system (similar to a
risk assessment) that evaluates finance, efficiency gains,
user/industry feedback, and user requirements (Column
“Category”) could help identify the right instrument for
your application (Table III). In such an approach, you
should determine which categories (finance, efficiency

user
gains, and so forth) are the most important ones within

your company. This will help prioritize that category and
allow a higher weighted score to highlight its importance
(Column “Weighting”). Weighted scores are arbitrary val-
Respiration
ues and can be determined at the discretion of a cross-func-

tional team. Every company’s values are different; hence,
the holistic business cases will likely differ between each
company even for the same instrument. As with any risk
assessment, parameters (Column “Parameters”) need to be
identified within each category as topics of evaluation and
a high, medium, and low score should be assigned to each
parameter with the appropriate definitions (Column “Rank-
ing”—“High”, “Medium”, and “Low”). The document
Compendial
owner or cross-functional team may develop as many pa-

rameters as needed to produce a thorough evaluation.
Once a detailed system is in place, it could be used to rank
the instrument under evaluation, using the example shown
in Table III.
user (i.e., plug-
Example Calculation:
development
work for end
Required
Attribute
and-play.)
Score = [(Sum of parameter scores) Finance Weight]
+ [(Sum of parameter scores) Efficiency Gains

Weight]
Characteristic
Maturity of the
Technology
(continued)
+ [(Sum of parameter scores) User/Industry Feed-

back Weight]
Table I
+ [(Sum of parameter scores) URS Weight]
Vol. 77, No. 3, May--June 2023 217

Table II
218
Comparison of Sterility Method Platform Candidates Using Application-Specific Requirements
Application-Specific Requirements
Characteristic Desired Attribute Compendial Respiration ATP Bioluminescence Solid-Phase Cytometry
In Liquid Growth On Solid Growth
Media Media
Monitoring and Continuous No, it is a Continuous No, it is an end-point No, it is an end- No, it is an end-point
Reporting monitoring, manual system monitoring with assay point assay assay
reporting of failure read out every
immediately 10 minutes and
reporting of
positives
immediately
Proprietary No – multiple No Yes Yes Yes Yes
Instrument vendors
Proprietary No – multiple No Yes Partial Yes Yes
Consumables vendors (growth media are
compendial media but
reagents for readout
are proprietary)
Reference Readily available No Yes Yes N/A Yes
Standards (just negative
controls, training
tools)
Detection - Detects Ph. Eur./ USP-required Yes – issues with Detects Ph. Eur./USP Yes, but noted Yes, new kit available
Specificity USP challenge microorganisms detecting mold. challenge issues for some to better detect spores
microorganisms, a P. acnes isolation microorganisms, a bacteria (i.e., and fungi
broad panel of requires iNST broad panel of isolates Methylo-
isolates especially bottles especially from past bacteriaceae)
from past events and events and slow-
slow-growing growing
microorganisms microorganisms
Sensitivity - Equivalent or better Able to recover Yes Yes Yes Yes
Limit of than compendial <100 CFUs
Detection (LOD)
63 min 41 min 30–60 min 30 mins 120 mins

Table II
(continued)
Media Media
Sample Fewer
Preparation Time manipulations
Inoculation Filtration preferred, Membrane Direct Alternative sterility Membrane Membrane filtration

Method if not possible then filtration method inoculation test can be performed filtration method method
consider direct or the direct method using either the or the direct
inoculation inoculation membrane filtration inoculation
method method or the direct method
inoculation method
Maximum Sufficient number N/A Expandable 162 samples per run Limited capacity Can read only one
Samples at a based on user capacity. with 50 tests/day sample at a time. Max.
Time preferences Range is 240 per throughput is 20 tests/
cart, can have 8 day
carts per 1
controller
Time To Result ≤10 days 14–21 days 7–8-day 6–7-day incubation, 5–7-day 5 hours
(TTR) depending on if a incubation then 40 minutes incubation, then
subculture step is automated readout on 15 minutes
needed instrument automated
readout on
instrument
Results Automatic, Manual, interim Automatic Automated data Automated data Automated data
Reporting integrated with reads at 3 and readout. Reports interpretation, capture, interpretation, interpretation, capture,
LIMS 7 days, final read generated in and storage capture, and and storage.
at 14 days OBSERVA are storage Additional step not
LIMS capable integrated if needed:
verification of positive
result by microscope
observation
Non-Destructive Non-destructive – Non-destructive Non-destructive – ATP readout uses only ATP readout is Non-destructive
viable cells for – viable cells for viable cells for part of the sample destructive
219
identification identification identification material and is
Table II
220
(continued)
Media Media
destructive. The
remaining part of the
original sterility test
containers can be used
for identification
User Minimal, closed Manipulations in Minimal, closed Initial manipulations Some manual Some manual stages:
Manipulations system an isolator/ system, bottles in an isolator/ stages: filtration, filtration, reagent
cleanroom inoculated in an cleanroom. Manual reagent addition, addition, and
isolator sample preparation in and membrane membrane transfer
BSC. Manipulation of transfer
ATP-readout minimal
Complexity/ Simple – user On-the-job Vendor-provided Vendor- provided Simple, vendor- Vendor-provided
Training in training for a couple training training needed training needed on provided training training for a couple of
Addition to of days on software and operation of the days
Fluency in equipment equipment and on
Aseptic Skills operation software for a couple
of days
Equipment Reliable and Pumps used for Reliable. Reliable. Consumables are Reliable.
Reliability and supported by vendor filtration are Company has Company has supported. Company has
Lifecycle from a life cycle robust reputation for reputation for robust Current reputation for robust
Management management robust and and reliable equipment technology will and reliable equipment
perspective reliable not be supported (second-generation
equipment by the vendor in available since 2013)
the future, but the
next-generation
platform is in
development.
Level of Fewer potential Manual system Fewer human Yes. Potential Filtration has few No automation—
Automation human error sources prone to human errors because of contamination point potential errors filtration, all staining
error barcoding of during readout steps, and final readout

Table II
(continued)
Media Media
compared with samples, manipulations, but this (scan) are manual. The
existing. automatic read of has no impact on the initial staining step is
samples, and outcome with critical, the other/

system alarms appropriate safety subsequent steps are
when there is a measures, e.g., less critical
positive result sampling under
laminar flow
Sample Volume Compendial Guided by Same Same requirements as Less than Same requirements as
per Test compendia requirements as compendia compendia compendia
compendia
Preventative ≤ every 6 months Annual Annual Annual Annual Annual
Maintenance calibration of
pumps
ATP, adenosine triphosphate; BSC, biological safety cabinet; LIMS, Laboratory Information Management System; LOD, limit of detection; N/A, not applicable.
221
Table III
Example for Decision-Making Matrix (to Be Completed by the End User)
Category Weightinga Parameters Ranking

High (10) Medium (5) Low (3)
Finance (ROI) 5 Cost per Test Cost of each test is Cost of each test is Cost of each test is
less than that of equivalent to that of more than that of
the traditional the traditional the traditional
method method method
<input <additional <additional <additional
additional definition high> definition medium> definition low>
parameters>
Efficiency 10 Rapid Results Shorter TTR than TTR equivalent to Longer TTR than
Gains the that of that of the traditional that of the
traditional method method traditional method
parameters>
User/Industry 10 Health Alternative Alternative sterility Alternative
Feedback Agency sterility test is test is accepted by sterility test is not
feedback accepted all the major Health yet accepted by
around the globe Authorities any Health
Authority
parameters>
User 5 Number of Alternative Alternative sterility Alternative
Requirements manual steps sterility test test and traditional sterility test
requires fewer method require the requires more
manual steps than same number of manual steps than
the traditional manual steps the traditional
method method
parameters>
a
A company may use another weighted score. Typical scores are 3, 5, and 10. A typical score of 3 is applied for less
important values, a score of 5 for values of medium importance, and a score of 10 for most important values.
ROI, return on investment; TTR, time to result.
Additional assessments supporting the business case can with the product(s) that will be tested and with the needs
include literature searches on the equivalence of the instru- of the laboratory. Supplier documentation, publications,
ment, discussions with vendors, and benchmarking with and communication in conferences are used to help the
external industry laboratories that have experience with the microbiologist initiate the project and evaluate the advan-
technology. Therefore, before making any investment in a tages and the disadvantages in terms of costs, training, user
new technology, microbiology laboratories must look for manipulation, level of automation, operational lead time,
the best one and/or the best adapted of all methods already data integrity, and so forth Whenever available, regulatory
available. There are some alternative and rapid microbio- intelligence is another important element to convince inter-
logical methods that are applicable to the sterility test; nal stakeholders to invest in alternative microbiological
however, the technology chosen should be compatible methods.

Step 6: Perform Proof-of-Concept Studies/Feasibility to include commonly found environmental microorgan-

Studies/Pre-Validation Studies isms that were isolated, subcultured, identified, and stored
as facility isolates. In addition to facility isolates, microor-
Proof-of-concept studies (sometimes also referred to as ganisms that pose a challenge to the method should be
“feasibility studies” or “prevalidation studies”) serve to included, for example, slow-growing, stressed, or other-
investigate the feasibility of the technology for the use as wise difficult-to-detect microorganisms. One of the species
an alternative and rapid sterility test. These studies are im- that may be important to be recovered in the alternative
portant to identify potential issues with the equipment or sterility test is Cutibacterium acnes (formerly Propionibac-
the method and to find mitigations before any validation terium acnes), because it is a well-known slow-growing
activities. Vendors of new technologies often provide pro- microorganism.
tocols that can be used to perform the method. A gap anal-
ysis should be performed to demonstrate that the vendor’s
The alternative and rapid sterility method with the new
method is sufficient to meet the companies’ requirements
TTR should be statistically comparable (equivalent or
for the alternative sterility test. Further method develop-
better, that is, Noninferior) to the compendial sterility
ment might be required to customize the method for its
intended use. The extent of the development work in this test. The prevalidation studies should ensure that non-
step will depend on the maturity of the technology, data inferiority can successfully be demonstrated in the pri-
available from the vendor, and benchmarking information mary validation. Parameters that could be assessed in
from the industry. All activities of the proof-of-concept the prevalidation studies are the limit of detection
studies should be carefully planned, as they form the basis (LOD) and specificity with a limited number of experi-
for the validation strategy in Step 7. ments. Subsequently, a limited robustness study could
be performed, investigating the capability of the alter-
One of the drivers for alternative and rapid sterility testing native rapid method to remain unaffected by small but
methods is analytical lead time reduction in comparison to deliberate variations in method parameters and provid-
the compendial sterility test (at least 14 days according to ing an indication of its reliability during normal usage.
the pharmacopoeias [1–3]). For this reason, the first stage Alternatively, robustness studies performed by the ven-
of proof-of-concept studies is to determine the TTR of dor can be leveraged.
the alternative sterility test method. According to the infor-
mation available from published data or from the supplier, Another aspect of a proof-of-concept study is the prod-
the time is often approximately known; however, it should uct-interference testing. Often this is done with a pilot
be verified in the facilities/companies with in-house iso- product or a limited number of products from the portfo-
lates. When the alternative and rapid sterility test method is lio to test whether the product sample is compatible with
growth-based, the choice of the media and the incubation
the method. The product-interference testing should test
conditions should be evaluated. Media and incubation con-
for inhibitory effects by performing spiking studies with
ditions may need to be optimized to meet the targeted
respect to recovery or TTR in comparison with the alter-
TTR, for example, an additional medium might be needed
native sterility test without product. Additionally, any
for the detection of slow-growing aerotolerant bacteria
using the respiration-based alternative sterility test. When product-interfering properties on the assay or the detec-
using growth conditions or media that differ from the tion itself should be evaluated, for example, quenching of
requirements described in the pharmacopoeias, the proof- reagents used for the assay or an increase in background
of-concept studies should ensure that the new method can signal caused by the product. Product-interference testing
detect all potential microbial contaminants (pharmacopoeial is an important preparation for the product validation and
reference microorganisms, in-house-isolates, and slow- the method suitability for every product.
growing and/or stressed microorganisms).
During the prevalidation phase, operational attributes
For proof-of-concept studies, relevant microorganisms of the new method, such as ergonomics, method com-
should be used, for example, standard reference strains as plexity, training needs, and risks, are finally assessed.
described in the compendia that are used routinely for In addition, the ability to meet internal equipment qual-
growth promotion testing of media and the method suit- ification standards and compliance requirements such
ability test (MST) (see pharmacopoeial requirements [1– as data integrity are assessed.
3)]) that is Staphylococcus aureus, Bacillus subtilis, Pseu-
domonas aeruginosa, Clostridium sporogenes, Candida Although some guidelines (16–18) mention or suggest
albicans, and Aspergillus brasiliensis. It is also important the use of stressed microorganisms to challenge the
Vol. 77, No. 3, May--June 2023 223

Step 7: Validate at Pilot or Primary Site

Table IV
Method Validation Parameters for Qualitative
Validation of a new alternative microbiological method
Methods
comprises two parts: equipment qualification and method
Validation USP Ph. Eur. PDA validation. Equipment qualification ensures that the
Parameter <1223> 5.1.6 TR33 equipment functions properly and meets quality and regu-
Accuracy X latory requirements (e.g., data integrity requirements). It
Specificity X X X ensures that the equipment is configured in its final state
to avoid changes that could affect functionality and
Limit of Detection X X X
results. At the primary site, a divisional equipment quali-
Robustness X X X
fication is performed. Using information gained during
Repeatability X
this qualification, the primary site develops a standard
Ruggedness X X
package (e.g., data integrity assessment, equipment quali-
fication, maintenance, operational standard operating pro-
cedure (SOP), test method, and so forth) containing
validation of ARMMs, what constitutes a stressed orga- global documents that can be used to qualify the equip-
nism is not clearly defined. The idea behind the use of ment at secondary sites.
stressed microorganisms is that the microorganisms are
subject to different challenging conditions (e.g., heat, pH- The method validation can be separated into two main
conditions, or starvation) resulting in microorganisms activities:
with suboptimal growth capabilities and/or a prolonged
lag phase (12). Thus, microorganisms recovered from 1. the primary method validation (usually without prod-
production environments differ from microorganisms uct) and
kept in culture conditions with respect to readiness for
growth. When performing studies with microorganisms 2. the validation for the intended use (with product),
from culture collections, usually the frozen cells are composed of
thawed, subcultured, and a fresh overnight colony is used
for inoculation of samples resulting in immediate prolif- product validation (if needed – e.g., for methods
eration without long lag phases. This could make the use using direct inoculation) and/or
of these cultures less challenging for certain alternative
sterility test method validation purposes and lead to short method suitability (‘suitability testing’)
TTR values and the potential underestimation of micro-
bial contamination in real situations during manufactur- The preceding terms can be open to interpretation, so
ing processes.
specific explanations are provided following.
There are some methods described to generate stressed

The main purpose of method validation is to demon-
microorganisms, for example, subculturing on a poor
strate statistically noninferior performance (equivalent
medium like Reasoner’s 2 A Agar (R2A) or applying and/or better performance) compared with that of the
heat (12, 13). However, most methods do not work for compendial method with respect to one or more valida-
all microorganisms, and species-specific conditions to tion parameters. Method validation parameters are
introduce stress must be developed. determined before initiation of testing and are based on
the guidance about method validation of alternative
For both primary method validation approaches described microbiological methods in the pharmacopeial chapters
in the next step, cryopreserved stocks that contain freeze- Ph. Eur. 5.1.6 (16) and USP <1223> (17), as well as in
dried microorganisms that can be used directly without the Parenteral Drug Association (PDA) Technical
subculturing were selected. Such cells are consequently Report No. 33 (18). Table IV demonstrates all relevant
different from freshly prepared cells with regard to their parameters for a qualitative method such as the sterility
readiness for growth. Exposure to freezing temperatures test. The parameters LOD, specificity, and robustness are
can lead to metabolic responses, to the formation of ice mentioned in all three guidelines. According to Ph. Eur.
crystals, and DNA and protein conformational changes 5.1.6 (16), validation of the accuracy of the alternative
that are quantifiable responses considered as “metabolic method with respect to the compendial method can be
stress” (14, 15). used instead of validation of the LOD. The intent is to

Figure 1
Validation strategy and global deployment of an alternative sterility test based on (a) Case Study 1 and (b)
Case Study 2.
demonstrate the sensitivity of the method. For growth- of Technical Requirements for Pharmaceuticals for
based methods, the LOD is defined as the detectable Human Use (ICH) Q2A(R1) (19) and the PDA Technical
number of microorganisms present in the sample before Report No. 33 (18) define ruggedness as intermediate pre-
the growth phase. The parameters repeatability and rug- cision. Precision, including repeatability, is described in
gedness are mentioned in Table I of USP <1223> (17). the text of USP <1223> (17) only in relation to quantita-
Ruggedness may be based on data from the technology tive tests; from a statistical perspective, this makes sense.
supplier. Although ruggedness can be interpreted in dif- As recognized by Ph. Eur. 5.1.6 (16), precision is not ap-
ferent ways, the International Council for Harmonisation plicable to qualitative tests.
Vol. 77, No. 3, May--June 2023 225

Table V
Summary of Statistical Validation Approaches for Accuracy (non-Inferiority) and LOD
Case Study 1 Case Study 2

Study Number of microorganisms 6 12
design Spike levels (CFU/sample) 1 / 2 / 4 CFU 0.1 / 1 / 10 CFU
Experimental design (paired/ Unpaired Paired
unpaired; number of sample
replicates per microorganism &
method)
40 / 40 / 40 sample replicates 10 / 28 / 10 sample replicates
Accuracy Comparison measure and Ratio of (organism) detection Difference of (sample)
corresponding non-inferiority probabilities (independent of detection probabilities (for the
criterion CFU levels) 3 CFU levels)
0.7 0.2
Statistical methodology Parametric estimation per Descriptive per organism and
organism and pooled (Poisson pooled (0.1/10 CFU).
model).
Parametric non-inferiority test Non-parametric non-
for organisms pooled inferiority test for organisms
(independent of CFU levels) pooled (1 CFU)
LOD Definition Lowest number of Lowest concentration for
microorganisms x per sample which the sample is detected
that is detected with 95% positive with 10% probability
probability
Estimation (no statistical Binomial detection model per Logistic regression model for
hypothesis test) organism (independent of organisms pooled (using all 3
CFU levels) CFU levels)
LOD, limit of detection.
Although referred to as equivalency in Ph. Eur. 5.1.6 including the spike levels and the number of replicates
(16), USP <1223> (17) recommends demonstrating per spike level used, should reflect the statistical proce-
noninferiority in the sense that the alternative sterility dure that will be used to evaluate the data. The statistical
test method is (statistically) equivalent or better than procedure should be described and justified in the valida-
the compendial method. In addition, equivalence is not tion documentation.
really a validation parameter on its own, but rather an
evaluation method that should be applied to one or Noninferiority compared with the compendial method can
more of the validation parameters or a corresponding be demonstrated using different validation strategies. In
measure to evaluate the sensitivity of the method (e.g., case the rapid sterility test will be used at multiple sites
accuracy or LOD). To show equivalence on the valida- for multiple products, it is often efficient to perform the
tion parameters, a spiking study is most efficient for primary method validation (including a noninferiority
the method validation of an alternative sterility test, study) without product and to perform an additional non-
because parallel testing during routine production for a inferiority study with product (product validation). Oth-
predefined number of samples or a predefined period of erwise, the primary validation can be performed directly
time (Ph. Eur. 5.1.6 [16]) would result in comparing with product, and there is no need for an additional prod-
(mostly) negative results. Equivalence (in terms of uct validation. Also, USP <1223> (17) states that, once
noninferiority) can be studied using various statistical equivalency (i.e., Noninferiority) for an ARMM to the
procedures, as demonstrated in Case Studies 1 and 2 compendial method is demonstrated with one product,
(see following). The design of the validation experiment, it is not necessary to repeat such a study for every new

product. Hence, a separate product validation study is Primary Method Validation: For the primary method
not always needed. Method suitability, however, validation of the rapid sterility test, the parameters specific-
should be demonstrated for each new product and at ity and accuracy/LOD were evaluated in separate studies
each new testing site as described in USP <71> (2) (referred to as the Specificity study and the LOD study),
and Ph. Eur. 2.6.1 (1). Figure 1 demonstrates the strat- both in comparison with the compendial sterility test with
egy for validation of a new test method at the primary 14 days of incubation and in the absence of product.
site using the examples of the two case studies.
Fifteen different microorganisms were used for the method
Microorganisms for Validation: The microorganisms validation. The LOD study was challenged with six cryo-
used for the method validation should comprise at least the preserved reference strains recommended for testing in
standard reference strains that are used for growth promo- USP <71> (2) and Ph. Eur. 2.6.1 (1). The Specificity study
tion testing of sterility test media according to USP <71> was challenged with nine environmental isolates including
(2) and Ph. Eur. 2.6.1 (1). Additionally, several environ- slow growers. Cryopreserved reference strains were chosen
mental isolates should be selected, for example, predomi- for the LOD study as precise spiking was required. Frozen
nantly facility isolates, isolates from a previous sterility cultures of environmental isolates were used for the Speci-
failure, slow growers (additional stress conditions may be ficity study that aimed to recover a large panel of relevant
applied to prolong the lag phase), spore formers, microor- microorganisms including slow growers.
ganisms with aerobic and anaerobic metabolism, microor-
ganisms coming from potentially different sources (i.e., Noninferior accuracy of the rapid sterility test versus
human, environment, and water) as well as from different the compendial method was determined in the LOD
groups (bacteria, molds, and yeasts). The guidelines do not study. To ensure that both methods would test samples
specify how many microorganisms should be used for with on average the same spike levels, one big, spiked
method validation nor for which study. stock solution was created, which was split into differ-
ent volumes to create dilutions from which all test sam-
The case studies following show two different approaches ples were taken. In order to reduce the risk that due to
toward method validation, especially with regard to the imprecise spiking, the true spike level would be too far
design of experiments and statistical methods applied. Ta- away from the optimal spike level of around 2 Colony
ble V provides an overview of the differences between Forming Units (CFUs)/sample (20), the LOD study
the statistical approaches for accuracy (including the non- was performed for each microorganism using three tar-
inferiority assessment) and LOD. Both approaches to get concentrations: 1, 2, and 4 CFU/sample, each with
method validation were approved by health authorities 40 replicates per method (independent group design);
with the data presented. Adaptations to these validation total of 240 tests per microorganism (120 per method),
approaches are possible; in fact, another BioPhorum each providing a positive or negative test result.
member has also gained approval for one of the technolo-
gies with slight adaptations to the method validation. For the analysis, a binomial detection model (20–22) was
assumed, in which each single microorganism has a prob-
Case Study 1: Validation of a Respiration-Based ability h, the detection proportion (depending on the type
Alternative Sterility Test of microorganism and the method), to be detected by the
method. As a result, for a sample with x microorganisms
Description of the Rapid Sterility Test: The method in it, the (binomial) probability that a sample becomes
uses direct inoculation of the sample to be tested and three positive, is
types of media: one for detection of aerobic microorgan- x
px ¼ 1 ð1 hÞ :
isms and two for anaerobic microorganisms. One of the an-
aerobic media was added for the detection of slow-
growing microorganisms such as Cutibacterium acnes. Note that this model assumes no false positives (px ¼ 0
Incubation takes 8 days in total, where the anaerobic media when x ¼ 0), which is addressed under specificity by
are incubated at 30˚C–35˚C and the aerobic media at 20˚ testing blank samples. However, the exact number of
C–25˚C. The aerobic media use a delayed entry into the microorganisms in a sample is not known, only an esti-
sterility test system to allow for nonagitating germination mate of the average number of microorganisms k per
of mold spores (1-day static incubation in a traditional in- test sample may be available. Assuming, additionally
cubator, followed by 7 days in the system). to the binomial detection model, that the number of
Vol. 77, No. 3, May--June 2023 227

microorganisms in a sample follows a Poisson distribu- value is commonly used for validation of mycoplasmas
tion with mean k, the expected proportion of positive or other nucleic acid amplification techniques (Ph. Eur.
samples (the positive rate), equals 2.6.7 [26], 2.6.21 [27]). LODs were reported per method
and microorganism, under the assumption that the esti-
p ¼ 1 expðhkÞ:
mate for the average spike level k was correct, because h
and, therefore, the LOD cannot be estimated from the for-
Accuracy (Noninferiority Testing). To address noninfer-
mula for p without knowledge of the spike level k.
iority1 (d) of the detection capabilities of the rapid sterility
Because of the uncertainty of assumed spike levels, one
test method compared with that of the compendial method,
should be careful with comparing LOD results between
the accuracy or recovery for each type of microorganism
studies. No formal acceptance criterion was used, because
was defined as the ratio of microorganism detection pro-
the detection capabilities of the two methods were al-
portions hR =hC of the rapid and the compendial method.
ready formally compared using a noninferiority criterion
Estimates of this ratio were calculated per type of microor-
under accuracy.
ganism across all spike levels. Note that the detection pro- Specificity. The specificity of the method is defined as its
portions and therefore also the ratio are independent of the ability to detect a range of challenge microorganisms spe-
spike levels. For the estimation of the ratio, it is important cific to the technology (Ph. Eur. 5.1.6 [16], USP <1223>
that both methods test samples from the same spike levels, [17]). In the Specificity study, nine relevant and worst-case
which may be accomplished by deriving all samples from environmental isolates were tested (chosen based on an
the same stock solution. An overall accuracy estimate was assessment of predominant species recovered from differ-
also calculated across all microorganisms and spiking lev- ent locations across sites) using the same inoculum solu-
els (23), provided that the assumption of homogeneity of tion for both the rapid sterility test and the compendial
the accuracy across microorganisms was not rejected. In sterility test method at concentrations of 10–100 CFU/me-
order to fulfil noninferiority, the corresponding two-sided dium bottle and three replicates per microorganism and
90% confidence interval (CI) (or equivalently, the one- test method. Detection of the panel of microorganisms by
sided 95% lower confidence limit) of the ratio of organism the rapid sterility test with an incubation time of 8 days is
detection proportions should be entirely above the nonin- compared with that of the compendial sterility method
feriority margin, which was set at 0.70 (or 70%). This with an incubation time of 14 days, and the number of pos-
means that the null hypothesis H0 in the following test itive samples should be equal to or higher than for the
problem is rejected at a significance level of 5%: compendial test method. Note that one would generally
expect all three replicates per microorganism and test
H0 : hR =hC 0:70 versus H1 : hR =hC > 0:70. method to be positive, but if this is not the case, it is con-
sidered sufficient if the rapid sterility test has at least as
The noninferiority margin of 0.70 was motivated by an many positives as the compendial test. Positive tests from
acceptance criterion of ‘at least 70% recovery’ sug- the rapid method were identified to confirm the inoculated
gested for microbial tests in PDA Technical Report No. isolates.
33, Chapter 5.3.1 (18) and in previous versions of Ph.
Eur. 5.1.6 (24) and USP <1227> (25). Note that apply- The specificity of the method is also its ability to detect
ing this noninferiority margin to the ratio of organism only the required microorganisms (Ph. Eur.5.1.6 [16]).
detection proportions is more stringent than when it Hence, false positivity was evaluated by inoculating 60
would be applied to the ratio of positive samples at a replicates per media type with sterile dilution fluid and
given spike level (22). testing them with the rapid sterility test. If all 60 sam-
Limit of Detection. The LOD was defined as the small- ples would be negative, then the estimate for the true
est number of microorganisms x in a sample, such that false-positive rate would be 0.0%, and we could claim
the sample becomes positive according to the formula with 95% confidence that the true value would be
for px with at least 95% probability, that is below 5%, because the one-sided 95% CI would be just
below 5.0%.
LOD = [ln(0.05)/ln(1h)], Robustness. The capacity of the method to remain
where the brackets [·] indicate the nearest integer above unaffected by small but deliberate variations in method
the calculated value represented by the dot. A 95% cutoff parameters has been well studied by the supplier. Addi-
tionally, a study was performed to investigate the effect
1
Please note that in USP <1223> this is mentioned under LOD, but Ph. of the stand-time of inoculated media bottles on sensi-
Eur. 5.1.6 suggests Accuracy instead.
tivity. The stand-time is the time between adding the

test sample to the media bottle and incubation in the facility isolates. Based on internally predefined accep-
system for automated reading. The results were input tance criteria, the test is considered suitable when the
for the routine procedure. bottles show a positive read-out within the 8-day incu-
Repeatability/Ruggedness. Repeatability and rugged- bation period, and the TTR of the bottles with product
ness (or intermediate precision) were not considered is smaller than two times the TTR of the bottles with-
relevant for Case Study 1, because the rapid sterility out product sample (control). A product control for
test is a qualitative test for which precision does not each media type is incubated to check for false
apply (see also the text about Table IV). positivity.
Product Validation: Because the primary method vali- Case Study 2: Validation of an ATP Bioluminescence–
dation is performed without product, a product valida- Based Alternative Sterility Test in Liquid Growth Media
tion is performed to confirm that the product matrix
under the given test conditions does not affect the vali- Description of the Rapid Sterility Test: The method
dation parameters as a result of interference. This prod- uses membrane filtration2 of the sample to be tested.
uct validation bridges the gap between the extensive Each sample is split over two membranes, which subse-
primary method validation without product and the quently are rinsed with an appropriate solution to
limited method suitability with product. Note that prod- remove the sample. Fluid thioglycolate medium (F) is
uct in this case may refer not only to final drug product added to one membrane, tryptic soy broth (TSB) is
but also to other matrices like process bulk. The prod- added to the other membrane. The first part of the
uct validation seeks to demonstrate noninferiority on method is identical to the procedure described in USP
the overall accuracy, similar to the primary method <71> (2), Ph. Eur. 2.6.1 (1), and JP 4.06 (3).
validation, but at an internally predefined noninferior-
ity margin of 50% of the rapid sterility test with versus The samples in TSB are incubated at 20˚C–25˚C, the
without the product, rather than comparing versus the samples in F are incubated at 30˚C–35˚C for a total
compendial method. This noninferiority margin of 50% incubation time of not less than 7 days. After incuba-
was applied to the overall accuracy of the rapid sterility tion, all samples must be agitated to ensure a represen-
test in the presence of product compared with without tative dispersion of microorganisms in the growth
product, that is, the ratio of probabilities to detect a sin- medium. A small aliquot is withdrawn from the sam-
gle organism, which is more stringent than it would be ples and inserted into a luminometer, which automati-
when applied to the ratio of positive samples at a given cally adds reagents to detect ATP via bioluminescence.
spike level (22). For this purpose, the product (usually
from one batch) was added to 20 aerobic and/or anaero- Primary Method Validation: The product-independent
bic media bottles per microorganism. Another 20 primary method validation evaluated the validation pa-
media bottles were left without product, and all 40 rameters specificity, LOD, ruggedness, robustness, and
media bottles were subsequently inoculated with 2 repeatability considering different sets of data.
CFU of each of the six standard reference strains that
were used in the primary method validation. All 40 rep- The most comprehensive data set addressed the specific-
licates (20 with and 20 without product) were incu- ity of the rapid method and compared the performance
bated for 8 days according to the rapid sterility test of the rapid sterility test and the compendial sterility test
method. with respect to LOD and accuracy (i.e., sensitivity in
terms of detection probabilities). The comparison of
Method Suitability: Method suitability testing is per- detection probabilities via a statistical noninferiority hy-
formed by the testing laboratory to demonstrate that pothesis test allowed the conclusion of statistically sig-
the new method is compatible with the product or nificant noninferiority regarding the performance of the
actual material used during routine testing. Suitability rapid method in comparison with that of the compendial
testing of the rapid sterility method is performed on method. Twelve different microorganisms (pharmaco-
three batches and is very similar to the testing peial reference strains, in-house isolates, and stressed
described in Ph. Eur. 2.6.1 (1) and USP <71> (2). microorganisms) were used for this data set. Three
Media bottles are inoculated with and without product
2
and spiked with <100 CFU of the six standard refer- Theoretically, the rapid sterility method can also be combined with the
direct inoculation method.
ence strains, C. acnes, and two predominant local
Vol. 77, No. 3, May--June 2023 229

inoculation levels for each microorganism (nominal lev- The inferiority null-hypothesis was to be rejected if the
els of 10 CFU, 1 CFU, and 0.1 CFU) were prepared. The lower one-sided 95% nonparametric Bootstrap confi-
inoculation levels were chosen because the LOD was dence interval (28) for the difference in detection prob-
expected to be close to 1 CFU. At levels 0.1 CFU and 10 abilities was entirely above D.
CFU, respectively, most of the test results were expected Limit of Detection. As the lowest concentration of
to be concordant (negative by both methods or positive microorganisms in a test sample that can be detected,
by both methods). Per microorganism, 10 replicates the LOD was internally defined as the lowest concen-
were prepared for the samples with 10 CFU and 0.1 tration (18) at which the detection probability is at least
CFU, and 28 replicates were prepared for the samples 10% (0.1). A concentration at the LOD level should
with 1 CFU. In total, 48 test runs were performed for have a detection probability clearly above 0; a proba-
each microorganism. bility of at least 10% is considered meaningful for all
practical purposes.
Rinsing fluid was spiked with the microorganisms and
subsequently filtered through the sterility test con- The pooled data (three inoculum levels of all microorgan-
tainer. Sterile growth media was added to each con- isms) were considered for the estimation of LODs of the
tainer. The containers were then incubated for in total two sterility test methods. The LOD was not meant to be
14 days. After 7 days (6 4 h), the samples were exam- specific to a particular type of microorganism.
ined using the rapid method and after 14 days (6 4 h),
the same samples were examined with the compendial A nominal (binary) logistic regression model (28) was
method. Statistically, the data correspond to a paired fitted to the pooled data of the alternative sterility test
sample design. The data served to study specificity, method and of the compendial method, respectively.
LOD, and accuracy (noninferiority).
Accuracy (Noninferiority Testing and Descriptive Analysis). The actual inoculum levels (verified via an inoculation
Different approaches were adopted for each inoculation count control procedure) rather than the nominal levels
level to compare the performance of the two sterility (0.1, 1, and 10 CFU) were used in the analyses, as the
methods. A descriptive analysis was carried out for the actual levels are assumed to be closer to the microbiologi-
inoculum levels 0.1 CFU and 10 CFU (10 sample repli- cal reality.
cates per microorganism, 120 in total) by reporting the
difference of sample detection proportions (i.e., esti- The following nominal logistic regression models were
mates of the probabilities detecting a sample to be posi- fitted separately to the rapid and compendial results:

tive) and also the ratio of detection counts, that is, the p
ln ¼ intercept þ slope log10 ðinoculum levelÞ
number of positive results with the alternative method 1p
divided by the number of positive results with the com-
pendial method. where p is the probability of a positive result.
With respect to the inoculum level of 1 CFU with 28 sample The limits of detection of the alternative sterility test
replicates per microorganism (336 in total), a noninferiority and the compendial method were estimated (in a con-
hypothesis test was performed considering the following servative way) via the corresponding upper one-sided
test problem regarding sample detection probabilities: 95% confidence limit of the inverse-predicted inocu-
lum level at a detection probability of 10% (=0.1).
H0: P(Rapid=positive) P(Compendial=positive)
≤ D = 0.2 The estimated detection limits of the rapid method and
of the compendial method were reported and compared
versus relative to each other. Ideally, the detection limit of the
rapid method should be lower than or equal to the
H1: P(Rapid=positive) P(Compendial=positive) detection limit of the compendial method. This require-
> D = 0.2 ment was not considered a mandatory acceptance crite-
rion, as the LOD verification was already addressed by
The noninferiority margin of D = 0.2 for the compari- the noninferiority test presented previously demonstrating
son of qualitative methods is a recommendation in USP noninferior detection probabilities of the alternative steril-
<1223> (17). ity method in comparison with the compendial sterility

method at the inoculum level of 1 CFU (following recom- Repeatability. Additional data were generated by test-
mendations of USP <1223>). ing four microorganisms with an inoculation level of
Specificity. After the incubation period, the positive not more than 100 CFU per test sample (actual inocula-
sterility test containers for each set of microorganisms tion level between 25 and 50 CFU) with the rapid
were subcultured; colony morphology and Gram stain- method. The procedure was applied repeatedly to mul-
ing was performed. DNA sequencing was performed tiple samplings of the same suspension of microorgan-
on one plate for each of the test microorganisms. The isms. One test container for each organism was
acceptance criterion for specificity was met when each prepared (4 in total). From each test container, 10 repli-
microorganism of the panel was successfully detected cates were taken and measured (40 measurements in
with the alternative sterility test method and was cor- total). All samples were positive in the ATP readout;
rectly identified. therefore, the acceptance criterion for repeatability was
Ruggedness. Additional data were generated by test-
met.
ing two microorganisms at an inoculation level of not
more than 100 CFU per test sample (actual inoculation
level between 25 and 45 CFU) with the rapid method Product Validation: In Case Study 2, the term product
under a variety of test conditions. The validation pa- validation relates to a verification of some of the char-
rameter ruggedness (also referred to as intermediate acteristics of the primary validation in the presence of
precision) is an evaluation of the intrinsic resistance of actual product. Because this case study uses the mem-
the method to the influences arising from operational brane filtration technique, the product validation is
or environmental variables. Several variables were intended to be performed only for a limited number of
introduced, like different operators, instruments, lots of products. Additionally, the FDA requested a product
reagents, and days. Three test samples for each orga- validation; for all other health authorities, the method
nism and each day (12 samples in total) were prepared. suitability test was sufficient. The product validation
The individual tests were measured in different combi- was designed to compare the LODs of the alternative
nations (48 measurements in total). All samples were sterility method and the compendial sterility method in
positive in the ATP readout; therefore, the acceptance the presence of drug product with a limited number of
criterion for ruggedness was met. microorganisms and replicates. The LODs of the alter-
Robustness. Additional data were generated by testing native and compendial method were estimated and
two microorganisms at an inoculation level of not more compared using a nominal logistic regression model as
than 100 CFU per test sample (actual inoculation level described in the primary validation. The study demon-
between 15 and 45 CFU) with the rapid method. The strated that the LOD of the alternative method is not in-
validation parameter robustness is an evaluation of the ferior to that of the compendial method. Equivalence
method’s capacity to remain unaffected by small but of both methods is verified through descriptive analysis
deliberate variations in method parameters. Variations (report of the read-outs analogous to the approach in
introduced were related to the reconstitution volume, primary validation). In addition, specificity in the pres-
sample volume, storage conditions of reagents, varia- ence of drug product of the alternative method was
tions of incubation temperature and incubation time, as verified by identification of the spiked microorganisms.
well as different types of the sterility test containers.
Three sterility test containers for each organism and
variation (21 in total) were prepared. All test samples Method Suitability: Method suitability testing is per-
taken from the different test containers were measured formed by the testing laboratory to demonstrate that
under different test conditions (102 measurements in the new method is compatible with the product or
total). During the robustness study, the incubation tem- actual material used during routine testing. The test for
perature was identified as a critical parameter for the growth-inhibiting properties of the product is identical
detection of C. acnes. Growth of C. acnes was delayed to the testing described in Ph. Eur. 2.6.1 (1), USP
at low temperatures of around 30˚C; therefore, it is im- <71> (2), and JP 4.06 (3) and is performed on three
portant to tightly control the incubation temperature for batches. Samples with and without product are spiked
routine testing. With the tight temperature control with <100 CFU of the six reference strains and two
applied, all samples were positive in the ATP readout; local in-house isolates. After incubation, the growth of
therefore, the acceptance criterion for robustness was microorganisms must be visually comparable for the
met. samples with and without product.
Vol. 77, No. 3, May--June 2023 231

In addition, the suitability test for the alternative Areas to assess in the quality risk assessment include
method was adapted to include product interference method, facility, equipment/analyst qualification, and
testing. It contains two additional criteria that are not material and shipment. Any risks identified must be
required for the compendial test. First, in addition to mitigated before or during the method transfer. A
visual growth, the samples used to test the growth-in- global alternative sterility test deployment strategy
hibiting properties must also result in a positive signal designed for multiple sites and products is summarized
using the alternative readout. With this test, false nega- in Figure 1, using the examples of the two case studies.
tives can be excluded, which can be caused for exam- The method transfer from the primary site to a second-
ple by product-interfering properties on the ATP ary site evaluates and demonstrates that the method
bioluminescence reagents. Second, false-positive test- performance of the secondary site is equivalent to the
ing is used to demonstrate that the product does not performance of the primary site that performed the
method validation. Globally, method transfer protocols
create an increased background noise or interfering sig-
are harmonized across sites to enable standardized
nals. For this test, a sample with product but without
implementation of the alternative sterility test. The
addition of any microorganisms is prepared. The sam-
method suitability is always the final step before the
ple must result in a negative signal using the alternative
test is applied routinely.
readout.
Case Study 1: Method Transfer Approach for the
Step 8: Deploy Global/Company-Wide Qualification of
Respiration-Based Alternative Sterility Test: Method
Additional Laboratories
transfer is performed without product, allowing the
secondary site to use the method for multiple products
The global deployment of an alternative sterility test
if product validation has demonstrated that the matrix
should be considered early on in the planning of the vali-
does not interfere with the validated method and
dation strategy. By validating the alternative method at
method suitability was demonstrated (see Step 7). The
one site (referred to as the primary site) without product,
method transfer experiments are performed at the pri-
the method validation can be leveraged for a larger num-
ber of products, which allows for more flexibility. After mary and the secondary site using the six standard ref-
equipment qualification and primary method validation, erence strains inoculated into the appropriate media
the primary site facilitates readiness of additional labora- bottles at 2 CFU/medium bottle and 80 replicates
tories (referred to as secondary sites) to receive the each. These numbers are sufficient to achieve a power
validated method. Secondary sites should obtain the nec- of more than 90% under a variety of assumptions.
essary equipment and follow the global package described Results of both the primary and secondary sites must
in Step 7 to qualify the equipment. demonstrate equivalence to confirm qualification of the
secondary site to leverage the method validation data
An important element of the global deployment pack- of the primary site. Equivalence criteria of 70%–130%
age—besides the data integrity-assessment and the were applied to the overall accuracy, that is, the ratio
equipment qualification protocols—is the global test of probabilities to detect a single organism, of the
method required to perform the alternative sterility test receiving site relative to the transferring site. The lower
in routine. The global test method will be generated by margin was chosen similar to the noninferiority margin
the primary site, and this test method should be used of 70% used during primary validation. For the upper
(ideally) by the secondary sites. Under certain circum- equivalence margin, the reciprocal of 70%, that is,
stances, a site needs an additional local test method, for 143%, would also be sensible.
example, a translated version of the global test method
to include additional local requirements. In this case, it
Case Study 2: Method Transfer Approach for ATP
must be ensured via change control processes that the
Bioluminescence–Based Alternative Sterility Test in
local test method does not deviate from the require-
Liquid Growth Media: The method transfer strategy is
ments described in the global test method. Any revision
of either the global test method or local test method(s) designed as a partial revalidation that verifies the speci-
must follow change control procedures. ficity, LOD, and noninferiority to the compendial
method in the presence of product. The product to be
After the equipment is qualified, a risk assessment used in the method transfer at the secondary site is one
should be performed and documented to assess any of the products manufactured at that site. The experi-
gaps between the primary and secondary sites before mental design of the method transfer is based on the
the method transfer for facility readiness assessment. design used for the product validation. The transfer

experiments are only carried out by the secondary site data is recommended and may need to be included in
including site-representative in-house isolates. The the filing.
rounded LOD of the secondary site and the primary
site are compared. The estimated LOD of the alterna- It is easier to implement an alternative sterility test
tive sterility test method at the secondary site (in the method for new products than for currently marketed
presence of product) must be lower than or equal to the products. Existing products will require a postapproval
estimated LOD at the primary site (without the pres- filing (analytical variation) with the health authorities
ence of product), both LODs rounded to the nearest creating a longer lead time for receiving approval from
unit. Similarly, the estimated LOD of the alternative different markets, which might result in a dual-testing
scenario where both methods—the compendial sterility
sterility test method and that of the compendial sterility
test as well as the alternative sterility test—must be
test method at the secondary site (in the presence of
maintained and applied for product release to fulfill dif-
product) are compared, and the ratio of detection
ferent local requirements and to comply with different
counts are reported. After successful completion of the filing strategies. Filing these methods in the earlier
method transfer, the secondary site is qualified to per- stages of a product’s development can allow a company
form method suitability tests and subsequently release to leverage its existing regulatory structure to facilitate
tests for new products. acceptance of the alternative microbiological method as
a replacement for the compendial sterility method.
Figure 1 summarizes the validation and global deploy-
ment process that was used for the technology in Case Regulatory approaches for the implementation of an alter-
Study 1 (Figure 1a) and Case Study 2 (Figure 1b). It native sterility test method should be managed at the
should be noted that the validation and global deploy- product level with appropriate change controls guiding
ment process demonstrated here were specifically the regulatory pathway. If possible and accepted by the
designed to fit the technologies used by the two compa- national competent authorities, both the alternative
nies for Case Study 1 (direct inoculation method) and 2 method and compendial method would usually be filed
(membrane filtration method). The first technology has a together to alleviate any risk to alternative microbiologi-
higher risk of interference by the product matrix owing cal method acceptance in a certain country/market. Either
to the direct inoculation. Different validation approaches method can then be used once approval is received. An
and deployment strategies might be suitable. implementation plan should reflect use of one method,
not both, in routine testing. The alternative sterility test
Step 9: Define Regulatory Filings and Implementation method can then be designated as the primary method to
Strategy be used for routine testing in the QC lab. Some compa-
nies, however, may choose not to file both alternative and
It is recommended to initiate discussions with health compendial methods based on their supply chain configu-
authorities in the early stages of alternative microbio- ration or because of lack of availability of a sterility test
logical test method implementation planning to ensure isolator. Supply chain strategy often will dictate the
alignment with the prospective country’s regulatory implementation approach for a certain product.
requirements for microbial testing of drug substance,
intermediates, and final drug product. It is encouraged Acknowledgements
to discuss the approach and rationale of the proposed
alternative microbiological method to cultivate an Many improvements were suggested by many subject matter
effective pathway for implementation. For example, a experts who reviewed the manuscript. Any remaining errors
thorough evaluation of the proposed method is are our own and should not tarnish the reputations of these
required for compliance with local requirements reviewers. This article has been written by subject matter
where product will be filed. As a result, companies experts from eight companies, collaborating in the
should obtain guidance from their regulatory teams on Alternative and Rapid Microbiological Methods team of the
the possibility of engaging in multiple project-specific BioPhorum Operations Group. More information can be
interactions with the corresponding health authorities. found at www.biophorum.com.
More stringent acceptance criteria may be required
based on the feedback from health authority/national
competent authorities interactions. Consultation with Conflict of Interest Declaration
the equipment manufacturers of the alternative micro-
biological methods regarding the available validation The authors declare that they have no competing interests.
Vol. 77, No. 3, May--June 2023 233

Disclaimer 7. Montero Julian, F. A.; Corrado, A.; Garcia, E. T.;

Barbirato, C.; Goodier, R.; McMutrie, D.; Jørgen-
This document represents a consensus view and, as sen, A.; McAlister, M.; Deutschmann, S.; McCar-
such, it does not fully represent the internal policies of thy, K.; Bonnevay, T.; Dormeyer, M.; Brewer, M.,
the contributing companies. Neither BioPhorum nor A structured approach for the evaluation, valida-
any of the contributing companies accept any liability tion and implementation of NAT-based myco-
to any person arising from their use of this document. plasma detection methods, BioPhorum, 2023.
This article has been developed by a collaborative 8. Jimenez, L.; Rana, N.; Amalraj, J.; Walker, K.;
group of subject matter experts from eight biopharma- Travers, K. Validation of the BacT/ALERT(R)
ceutical companies. It represents a current consensus 3D System for Rapid Sterility Testing of Biophar-
view also based on the feedback from health agencies. maceutical Samples. PDA J. Pharm. Sci. Technol.
Users of the systems are of course free to deviate from 2012, 66 (1), 38–54.
the described validation and implementation approach.
9. Bugno, A.; Almodovar, A. A. B.; Saes, D. P. S.;
References Awasthi, R.; Ghisleni, D. D. M.; de Souza
Braga, M.; de Oliveira, W. A.; Dua, K.; de Jesus
1. Council of Europe, Sterility, Chapter 2.6.1. In Eu- Andreoli Pinto, T. Evaluation of an Amplified
ropean Pharmacopoeia (Ph. Eur.), 10th Edition, ATP Bioluminescence Method for Rapid Sterility
Council of Europe: Strasbourg, France, 2022. Testing of Large Volume Parenteral. J. Pharm.
Innovation 2019, 14 (2), 152–158.
2. U.S. Pharmacopeial Convention, General Chapter
<71> Sterility Tests. In USP 42—NF 37, USP: 10. Parveen, S.; Kaur, S.; David, S. A. W.; Kenney,
Rockville, MD, 2019. J. L.; McCormick, W. M.; Gupta, R. K. Evaluation
of Growth Based Rapid Microbiological Methods for
3. Ministry of Health Labour and Welfare of Japan, Sterility Testing of Vaccines and Other Biological
General Tests, Processes and Apparatus–4.06 Ste- Products. Vaccine 2011, 29 (45), 8012–8023.
rility Test. In Japanese Pharmacopoeia, 18th Edi-
tion (English Version), MHLW: Tokyo, 2021. 11. Smith, R.; Von Tress, M.; Tubb, C.; Vanhaecke,
E. Evaluation of the ScanRDI as a Rapid Alterna-
4. Pasteur, L. (translated from French) [On the tive to the Pharmacopoeial Sterility Test Method:
Extension of the Germ Theory to the Etiology of Comparison of the Limits of Detection. PDA J.
Certain Common Diseases]; Ernst, H.C. (trans). C. Pharm. Sci. Technol. 2010, 64 (4), 356–363.
R. Acad. Sci. 1880; XC, 1033–1044.
12. Sage, A. Stressed Microbial Populations: Real-
5. Alternative and Rapid Micro Methods (ARMM): A World Challenge for Validation of New Applica-
Framework for the Evaluation, Validation and Imple- tions. In Environmental Monitoring: A Comprehen-
mentation of Alternative and Rapid Microbiological sive Handbook; Moldenhauer, J., Ed.; PDA:
Testing Methods. BioPhorum Website. 2020. https:// Bethesda, MD, 2015, pp 325–340.
www.biophorum.com/download/alternative-and-rapid-
micro-methods-armm-a-framework-for-the-evaluation- 13. Gray, J. C.; Staerk, A.; Berchtold, M.; Hecker,
validation-and-implementation-of-alternative-and-rapid- W.; Neuhaus, G.; Wirth, A. Growth-Promoting
microbiological-testing-methods/. Properties of Different Nutrient Media Evaluated
with Stressed and Unstressed Microorganisms: Pre-
6. Deutschmann, S.; Carpenter, B.; Duignan, C.; study for the Validation of a Rapid Sterility Test.
Knutsen, C.; Salvas, J.; Wysocki, L.; Plourde, PDA J. Pharm. Sci. Technol. 2010, 64 (3), 249–263.
L.; Johnson, L.; Eder, W. A Systematic
Approach for the Evaluation, Validation and
14. Shivaji, S.; Prakash, J. S. S. How Do Bacteria
Implementation of Automated Colony Counting
Sense and Respond to Low Temperature? Arch.
Systems. PDA J. Pharm. Sci. Technol. 2022,
Microbiol. 2010, 192 (2), 85–95.
76 (6), 509–526.

15. Paulussen, C.; Hallsworth, J. E.; Álvarez-Pérez, 22. IJzerman-Boon, P. C.; Manju, M. A.; van den
S.; Nierman, W. C.; Hamill, P. G.; Blain, D.; Heuvel, E. R. Non-Inferiority Testing for Qualitative
Rediers, H.; Lievens, B. Ecology of Aspergillo- Microbiological Methods: Assessing and Improving
sis: Insights into the Pathogenic Potency of Asper- the Approach in USP 1223. J. Biopharm. Stat. 2022,
gillus fumigatus and Some Other Aspergillus 32 (6), 915–941.
Species. Microb. Biotechnol. 2017, 10 (2), 296–
322. 23. Emampour, M.; IJzerman-Boon, P. C.; Manju,
M. A.; van den Heuvel, E. R. Optimal Spiking
16. Council of Europe, Alternative Methods for the Experiment for Noninferiority of Qualitative Micro-
Control of Microbiological Quality, Chapter 5.1.6. biological Methods on Accuracy with Multiple
In European Pharmacopoeia (Ph. Eur.), 9th Edi- Microorganisms. Stat. Biopharm. Res. 2023, 15 (1),
tion, Council of Europe: Strasbourg, France, 2017. 198–213.
17. U.S. Pharmacopeial Convention, General Chapter 24. Council of Europe, Alternative Methods for the
<1223> Validation of Alternative Microbiologi- Control of Microbiological Quality, Chapter 5.1.6.
cal Methods. In USP 42—NF 37, USP: Rockville, In European Pharmacopoeia (Ph. Eur.), 6th Edition,
MD, 2019. Council of Europe: Strasbourg, France, 2008.
18. Parenteral Drug Association Inc. Technical Report 25. U.S. Pharmacopeial Convention, General Chapter
No. 33: Evaluation, Validation and Implementa- <1227> Validation of Microbial Recovery from
tion of Alternative and Rapid Microbiological Pharmacopeial Articles. In USP 39—NF 34, USP:
Methods; Bethesda, MD, 2013. Rockville, MD, 2016.
19. International Conference for Harmonisation, Quality

26. Council of Europe, Mycoplasmas, Chapter 2.6.7. In
Guideline Q2A(R1): Validation of Analytical Proce-
European Pharmacopoeia (Ph. Eur.), 10th Edition,
dures: Text and Methodology. ICH: Geneva, 1995.
Council of Europe: Strasbourg, France, 2022.
20. IJzerman-Boon, P. C.; Van den Heuvel, E. R.
Validation of Qualitative Microbiological Test 27. Council of Europe, Nucleic Acid Amplification
Methods. Pharm. Stat. 2015, 14 (2), 120–128. Techniques, Chapter 2.6.21. In European Pharma-
copoeia (Ph. Eur.), 10th Edition, Council of
21. Van den Heuvel, E. R.; IJzerman-Boon, P. C. A Europe: Strasbourg, France, 2022.
Comparison of Test Statistics for the Recovery of
Rapid Growth-Based Enumeration Tests. Pharm. 28. Good, P. Permutation, Parametric and Bootstrap
Stat. 2013, 12 (5), 291–299. Tests of Hypotheses, 3rd ed.; Springer, 2005.
Vol. 77, No. 3, May--June 2023 235

CASE STUDIES
Multisite Qualification of an Automated Incubator and Colony

Counter for Environmental and Bioburden Applications in
Pharmaceutical Microbiology
HANS JOACHIM ANDERS1, DANIEL MáNNLE1, WILLIAM CARPENTER2, WOLFGANG EDER3, IVANA HECKEL4,
TOBIAS GøTZEN4, CORINNE OECHSLIN4, CEDRIC JOOSSEN5, MARIA EUGENIA GIRIBETS PARRA6,
JASON ROSE7, VAISHALI SHAH8, and DAVID L JONES9,*
1
Novartis Pharma Stein AG, Stein CH4332, Switzerland;2Biogen, Durham, NC;3Roche Diagnostics GmbH, Nonnenwald 2,
82377 Penzberg, Germany;4Lonza, Visp, Valais, Switzerland;5Janssen Pharmaceutica NV, Beerse, Belgium;6Boehringer
Ingelheim Pharma GmbH & Co. KG, Biberach, Germany;7GlaxoSmithKline, Upper Merion, PA;8Kite Pharma, Santa
Monica, CA; and9Rapid Micro Biosystems, Lowell, MA, USA © PDA, Inc. 2023
ABSTRACT: Traditional microbiological techniques have been used for well over a century as the basis for contamina-
tion testing of pharmaceutical products and processes. With more recent focus on faster product release and concerns
around the integrity of the test data, new technologies have been implemented to detect and enumerate organisms faster
and provide paperless processes to minimize data integrity issues. Manual colony counting technologies, where incuba-
tion is performed in a standard incubator, and the plate is manually transferred to the colony counter for a single read at
the end of incubation, have been used for many years to reduce the potential for human error; however, they pose vali-
dation challenges due to poor counting accuracy. Colony counters that automatically perform both the incubation and
enumeration functions (multiple enumeration calculations through the incubation phase) have recently been imple-
mented for quality control (QC) laboratory analytical processes, supporting a cGMP environment. This article summa-
rizes the findings of eight companies demonstrating the qualification of an automated colony counter technology to
perform the majority of microbial tests required for QC, environmental monitoring, and bioburden for in-process, bulk
drug substance, and water system testing. Comparable analytical performance and time to result data generated during
individual studies at all companies allows the system to be qualified and implemented for cGMP processes while reduc-
ing data integrity risks.
KEYWORDS: Automation, Colony Counter, Environmental Monitoring, Bioburden testing, Performance Qualification,
Method Validation, contact plate, Rapid Microbial Methods (RMMs).
Introduction answers the questions posed in the risk analysis. The

key technical risks for colony counters are:
The introduction of automated colony counter tech-
nology in pharmaceutical microbiology is becoming 1. Will the automated colony counter be viewed as an
more common. Significant improvements regarding alternative microbiological test method and subject
data integrity and counting accuracy are the main to full method validation, or merely the automation
drivers for this shift from conventional plate reading of the incubation and reading of a traditional micro-
to automation. biological method and hence subject to a reduced
verification?
When introducing a new technology, a risk analysis
should be performed to ensure that the validation 2. Will the colony counter give higher counts due to its
ability to detect microcolonies from either small,
non-visible colonies or detection of separate colonies
* Corresponding Author: Rapid Micro Biosystems, Low-
that merge and appear to the eye as a single entity
ell, MA, USA; E-mail: djones@rapidmicrobio.com
during CFU enumeration? Higher counts may require
doi: 10.5731/pdajpst.2022.012742
a change of action/alert levels.

3. Will automated vision systems generate false posi- size, colored black on the surface of the media to
tives and cause more action/alert level excursions? improve the signal-to-noise ratio for the detection sys-
tem. During the incubation phase, images of each cas-
4. If a shorter incubation time (TTR) is selected, will the sette are taken at intervals of 4 h, allowing organisms
colony counter miss contamination that would have and debris that are naturally fluorescent under the exci-
been seen with the traditional incubation conditions? tation blue light (465–495 nm) of the imager to be
detected as objects in the green (505–560 nm) spec-
5. Will the technology get regulatory acceptance? trum. Analysis of the behavior of objects over the incu-
bation time by the proprietary algorithms of the vision
These risks are addressed in the article. analysis software allows the Growth DirectV System to
R
distinguish and enumerate the growing colonies from

Any viable (and culturable) microorganism that can be background and debris that do not grow (4). Figure 1
captured on the membrane by sample filtration, spread shows the blue excitation light and green emission
plating, surface contact, or impingement during air mon- from colonies being captured on a charge coupled de-
itoring can be detected and enumerated (1–3). Based on vice (CCD). The imaging method does not harm the
these discussions, data will be presented for verification, cells, and as such is a nondestructive method. The
a position justified by the USP40—NF35 General Infor- microcolonies can grow into visible colonies for use in
mational Chapter <1223> Validation of New Microbio- subsequent microbial identification.
logical Testing Methods and industry practice as found
in the 2013 PDA Technical Report 33 (Revised) Evalua- The Growth Cassette products incorporate standard
tion, Validation and Implementation of Alternative and media varying with the application. For EM two main
Rapid Microbial Methods. The verification approach has media options are available, Tryptone Soy Agar with
also been used in several successful health authority
Lecithin Polysorbate (TSA LP80), or added Histidine
approvals with both European and US regulators.
and Thiosulfate (TSA LP80HT). For product and water
bioburden, TSA, Saboraud Dextrose agar (SDA), and
This study will discuss the advanced imaging system in
R2A media are available, respectively. All media are
the Growth DirectV System as an example of an auto-
R
standard pharmacopeial formulations used in the phar-

mated colony counter, employing data for the three
maceutical industry. The growth promotion is per-
prime applications: environmental monitoring (EM),
formed as a separate study and is not reported in this
bioburden for water, in-process samples from multi-
company implementations. Data for the time to results article. Standard site acceptance criteria would be used
(TTR) determination and method qualification/suitabil- to verify the media as an approved media.
ity obtained by those companies for each application
will also be shown. Regulatory and Compendial Guidance for the
Qualification of Automated Methods
Technical Background of the Advanced Imaging
System Used in the Automated Colony Counter USP 40—NF 35 General Notices 6, Testing Practices
and Procedures, provides guidance for the use of auto-
The Growth DirectV System (Rapid Micro Biosystems
R
mated and alternative test methods. 6.20 Automated
Inc., Lowell, MA, USA) is an automated rapid microbial Procedures states, “Automated and manual procedures
enumeration platform that integrates digital imaging for employing the same basic chemistry are considered
colony counting, robotic cassette handling, incubation, and equivalent.” The statement is equally true for proce-
software control. dures employing the same basic microbiology such as a
plate count and the Growth DirectV System. Further-
R
V
R
The Growth Direct System for rapid microbial enumer- more, 6.30 Alternative and Harmonized Methods and
ation comprises two automated incubators handling up Procedures states that “Alternative methods and/or
to 659 media cassettes. The Growth Cassette products procedures may be used if they have advantages in
are plastic contact plate style cassettes containing stand- terms of accuracy, sensitivity, precision, selectivity, or
ard compendial growth media. adaptability to automation or computerized data reduc-
tion, or in other specialized circumstances. Such alter-
The Growth DirectV test method requires the presence
R
native procedures and methods shall be validated
of a mixed cellulose ester membrane, 0.45-micron pore as described in the USP40/NF35 general chapter Validation
Vol. 77, No. 3, May--June 2023 237

Figure 1
Schematic of the detection of the autofluorescence of microcolonies. Top: Microorganisms fluoresce under blue
light and their location on the membrane captured by the CCD detector. Middle: Images taken at 4-hour inter-
vals and an increase in a fluorescent object size or brightness confirms a growing colony. Bottom: Accuracy of
the system can be shown by comparison with human counts at the end of incubation.
of Compendial Procedures <1225> and must be shown to demonstrate certain method validation requirements as
give equivalent or better results.” The Pharmacopeia are specified in Section 5.0 of the Technical Report. For
being updated to relate to changes in microbial methods. these technologies, at least accuracy and precision
These changes and how the new methods are validated are assessments should be performed, in addition to method
described in USP <1223> Validation of Alternative Micro- suitability and equivalence/comparability studies.” The
biological Methods. For colony counters, USP <1223> view expressed in USP <1223> is fully supported in
states the following: “There are commercially available this industry practice document.
enhancements to growth-based methods that allow colonies
Ph. Eur. 5.1.6, Alternative Methods for the Control of
on solid media to be read more quickly, with substantially
Microbiological Quality, does discuss growth-based
less incubation time, than is possible using only the unaided
methods using the presence of endogenous autofluores-
eye. In the implementation of these enhanced methods for
cent molecules and metabolites such as reduced nico-
the detection of colony growth, only the detection capability
tinamide adenine dinucleotide phosphate (NADPH)
of the method requires verification.” This statement supports and flavoproteins within microorganisms. The revised
the view that the Growth DirectV System is not an alterna-
R
chapter continues to view a technology such as the

tive method requiring method validation. The validity of the Growth DirectV System within the framework of alter-
R
definition as an automated compendial colony counter is native microbiological methods. Taking the risk-based
discussed in a technology review article by Jones and Cun- approach, as recommended in Ph. Eur. 5.1.6, in conjunc-
dell (5). tion with a verification strategy, a simpler approach can
be justified.
Similarly, the PDA Technical Report No. 33 (Revised),
Evaluation, Validation and Implementation of Alterna- Validation Approaches
tive and Rapid Microbial Methods, states the follow-
ing: “Some alternative or rapid technologies may be The analytical testing performed by each site was the
considered automated traditional or compendial micro- same, but depending on company definitions, the name
biological methods, especially when the results are in of the test phase varied, for example, Performance
colony-forming units (CFU). These technologies may Validation/Qualification/Verification. The term PQ
be qualified for their intended use without the need to will be used in this document to encompass all three

terminologies. The same approach will be used for the system and visual counts using standard analysis of
Method Validation/Qualification/Verification. variance methods can be used.
Performance Qualification (PQ)

Method Qualification/Validation/Verification (MQ)
Performance Qualification has a focus on the system
In the MQ phase, tests provide evidence of suitable sys-
and compares the system colony count with the visual
tem and analytical microbial performance. The assess-
colony count on the same cassette, rather than an
ment of precision and accuracy is performed as part of
assessment of microbial analytical performance.
the MQ phase. Studies of the pharmacopeial organisms
Because the basis of the PQ test is to verify the vision should include a mixed organism population of two or
software’s ability to accurately detect and count the more species and at least one environmental organism
range of colony morphologies seen in the QC testing, from the facility, as proposed in TR33.
the media type under the black mixed cellulose mem-
brane is not critical. The same range of colony mor- Precision (Repeatability) shows the variability of the
phologies exists on all media. method when analyzed by the same operator from the
same sample pool. When compared to a control method,
Time to Results (TTR) a statistical test to compare the two methods’ variances
should be used.
The TTR determination can be performed locally or
can be performed at a Center of Microbiological Excel- Accuracy compares the CFU obtained by the colony
lence using laboratory spiked samples of stressed and counter against the standard visual plate count for the
known slow growing microorganisms obtained from same sample.
other sites from a global organization. The global TTR
can then be set to those organisms and the TTR verified Method Suitability (MS)
via on-site validation during a technical transfer pro-
cess. It is possible that a local organism may require a Before routine testing can occur, method suitability crite-
longer incubation so that the site can adapt to its local ria for individual pharmaceutical ingredients and drug
microflora. Water and EM sampling sites with a histor- products must comply with USP <61>, Microbiological
ical recovery trend should be chosen to give sufficient Examination of Nonsterile Products: Microbial Enumera-
counts for detection. tion Tests. Method suitability demonstrates the recovery
of the challenge microorganisms in the presence of prod-
The specification used by all sites to determine the
uct sample preparation.
TTR is 85% of the slowest growing detected CFU num-
ber. This specification arose from the need to define a
simple test that would facilitate a faster indication of a Methods
contamination issue for trending, as well as the need to
recover sufficient organisms to pass the equivalence Performance Qualification (PQ) and Performance
testing while accounting for the inherent imprecision Verification (PV)
of the microbiology test. The USP specification in
force for recovery at the time was >70% from USP For the consistency of the data generated, most phar-
<1227>, which was later harmonized to the current maceutical companies reporting in this article used the
50%–200%. A TTR had to be set to recover >70% of same protocol as defined in the vendors performance
the traditional method accounting for an imprecision in qualification (PQ) documents as part of the modular
the microbiology test of 15% to 30% CV from USP validation. Similar statistical methods were used for
<1223>. The median value of 85% was therefore the analysis of data but are not described in detail in
selected and has proven to be a robust specification. this article. However, most sites employed the Two
Depending on company policies, this number can be One-Sided t-Test (TOST) for noninferiority as defined
altered according to the risk assessment performed. in both USP <1223> and EP 5.1.6. Examples of the
Alternatively, an acceptance criterion that requires no statistical tests using the technology can be found for
significant difference between the Growth DirectV
R
water (6), bioburden (7), and EM (8).
Vol. 77, No. 3, May--June 2023 239

Performance testing used a challenge microorganism CFU numbers can vary depending on the day of the final
stock culture, freshly grown but no more than five pas- visual count, so the final read date needs to be controlled.
sages removed from ATCC or an equivalent source or
reconstituted commercial preparation such as Quanti- The TTR was evaluated from system colony counts
cult or BIOBALLV. The test runs were performed with
R
collected every 4 h. When plotted against incubation
three replicates of each test organism. The smallest time, CFU counts show a sigmoidal colony detection
number of organism types that may be used would be curve (Figure 2). The TTR can be determined from a
three (Escherichia coli, fast growing circular morphol- defined set of library microorganisms or by testing the
ogy, Aspergillus brasiliensis, slower growing irregular sample site and determining the natural growth time
morphology with hyphae, and Bacillus subtilis, fast for the “stressed” organisms present. The time at which
growing irregular morphology) to cover the main col- all detected organisms met an 85% threshold of the vis-
ony morphologies for detection capability by the ual count is the basis of the TTR.
Growth DirectV algorithm software. A company-spe-
R
cific environmental organism could be added at this

Method Qualification (MQ)
phase if the organism demonstrates an unusual mor-
phology. The target inoculum concentration was
Following guidelines from PDA Technical Report No.
between 20 and 100 CFU per 100 lL in 20 mL of sterile
33 (Revised), Evaluation, Validation and Implementa-
buffer or water for bioburden testing or in 50 lL
tion of Alternative and Rapid Microbial Methods, at
directly plated for EM cassettes (lower volume to
least six replicates of each organism were prepared on
obtain good distribution without flooding on the mem-
the Growth DirectV cassette. The same method was
R
brane surface). For molds such as A. brasiliensis, a

smaller inoculum < 20 CFU could be used. A summary used for each of the application types but featured sev-
of the methods used is provided in Table I. eral variations: organism preparation, membrane filtra-
tion for bioburden and water, and the spread plate
technique for EM cassettes. The Growth DirectV cas-
R
After incubation on the Growth DirectV System at the

R
required temperature 20˚C–25˚C, 25˚C–30˚C, 28˚C– settes were incubated and enumerated on the system,
32˚C, or 30˚C–35˚C for 3–7 days, a visual count was then visually counted by the analyst at the end of
performed for the colonies on each Growth Cassette. incubation.
The equivalence of the Growth DirectV system com-
R
pared with the mean of the three analysts’ count was The range of sample points and sample numbers used
performed. in the study is given below:
Time to Results (TTR) For water, sample volumes ranged from 0.1 mL to
200 mL depending on water type with 2 to 25 sample
The TTR is related to the organism type, media type, points yielding 192 to 600 test replicates.
and incubation conditions. Thus, separate studies
should be performed for each sampling site and appli- For active air, sample volumes ranged from 200 to
cation. EM samples were usually taken from the lower 1000 liters from 18 to 83 sample points, yielding 54
environmental grades C or D. For water samples, larger to 215 samples. For contact plates, 6 to 132 sample
volumes than routine can be taken to increase cell num- points yielded 36 to 216 samples.
bers. Product bioburden samples are often uncontami-
nated, so organisms likely to be found in the environment For bioburden testing, the guidelines suggest the Phar-
or product are used to inoculate the sample. For product macopeial organisms, at least one environmental orga-
bioburden, organisms were included that have been nism, and a stressed organism should be tested for each
through a stress treatment present in the process, for test product type. The stressed cells should represent
example, low pH. For EM testing, the samples usually conditions from the process, for example, viral inacti-
contain organisms that are already stressed through disin- vation at low pH (3.0-4.0), as performed by some of
fection, dehydration, or starvation. Sample incubation the companies. Six replicates of each organism were
reflects the maximum duration of the visual method, 3, 5, used for in-process or BDS samples. Calculation of the
or 7 days. Visual counts of the colonies on the plate were mean and standard deviation of each data set allows a
made by three analysts on incubation completion. The statistical test of the variance, for example, Chi-square

Table I
The Media, Test Microorganisms, and Growth DirectV Incubation Temperature and Time at the Different Study
R
Sites
Site and Media Control Microorganism Incubation Temperature and time

Company 1 B. subtilis 30˚C–35˚C for 36 hours
TSA P. aeruginosa
R2A C. albicans
SDA A. brasiliensis
TSA + LP80 R. pickettii
S. aureus
S. epidermidis (in-house)
Penicillium species (in-house)
P. glucanolyticus (in-house)
B. cereus/thuringiensis (in-house)
Mixed culture
Company 2 B. subtilis 22.5˚C–27.5˚C for 44 hours
TSA + LP80 A. brasiliensis
E. coli
E. coli
S. aureus
P. aeruginosa
C. albicans
Mixed culture
In-house isolates
E. coli
Company 5 EM:E. coliB. cepacia EM:Water:
TSA + LP80HT B. subtilis 30˚C–35˚C for 44 hours
A. brasiliensis
R2A Water:
E. coli
Water: 30˚C–35˚C for < 7 days
B. cepacia
Company 6 S.aureus 30˚C–35˚C for 44 hours
TSA + LP80HT C. albicans
S. hominis
M. luteus
C. tuberculostearicum
B. subtilis
A. brasiliensis
E. coli
P. aeruginosa
TSA A. brasiliensis
Vol. 77, No. 3, May--June 2023 241

Table I
(continued)
Site and Media Control Microorganism Incubation Temperature and time

R2A E. coli
TSA + LP80HT
R2A E. coli
Figure 2
TTR curves for 5 pharmacopeial and 7 environmental organisms, namely Staphylococcus epidermidis, Micro-
coccus luteus, Kocuria rhizophila, Bacillus pumilus, Brevundimonas diminuta, Candida tropicalis, and Aspergillus
fumigatus are incubated and imaged every 4 h for 72 h at 30˚C–35˚C. The cumulative percentage of emerging
colonies is shown on the y axis.
distribution. The same sample data used for precision 1. Neutralization of disinfectant residue from surfaces,
are used for accuracy determination.
2. Ability to capture organisms on a surface and pull
Equivalence is included in the method qualification, them away for growth,
which demonstrates results obtained with the test method
agree with the compendial method currently in use. Test 3. Ability to capture organisms in an air flow with-
sample selection is based on relevant species and a suffi- out adverse dehydration resulting from active air
cient number of organisms for statistical significance. sampling.
The incubation time for the new method would be the
assigned TTR or the compendial duration if the data are Performance of these tests may be comprehensive or
to be used to set the TTR. A minimum of 50–100 individ- based on literature data with a minimal test strategy to
ual test samples should be taken and analyzed contempo- verify the published data.
raneously and the resulting CFU counts compared.
The main function of the media in the bioburden and Method Suitability (MS)
water applications is to provide nutrients to allow the
organisms to grow. In all cases, the traditional methods Method suitability followed the same method as that
used for comparison studies had been qualified. Other described for the method qualification, with the excep-
critical functions of the media in the environmental tion that the test sample contained the product. The con-
application include: trol was a suitable buffer (Fluid A, PBS and so forth).

Table II
Verification of the Equivalency of the Automated Plate Counter and Traditionally Read Plate Counts
B. subtilis E. coli A. brasiliensis P. aeruginosa S. aureus C. albicans

Site ATCC 6633 ATCC 8739 ATCC 16404 ATCC 9027 ATCC 6538 ATTC 10231
Site 1 Pass Pass Pass Pass Pass Pass
Site 2 Pass Pass Pass
Note: Acceptance Specification: Growth DirectV Colony Count ≥85% of the mean colony count by 3 analysts. Organism
R
colonial shape recorded in PQ to verify vision accuracy by site. Grayed cells indicate the organism was not tested during
company PQ.
Results active air media dehydration and contact recovery

experiments all passed the assigned acceptance criteria.
Performance Qualification (PQ)
Method Suitability (MS)
Results for the PQ phase are shown in Table II. All
sites had acceptable data for the three key organisms. All sites performing MS for in-process and BDS met
Some sites extended the testing to the remaining USP the required recovery acceptance criteria. This demon-
organisms and historical EM samples with similar ac- strates the lack of product interference with the fluores-
ceptable results. cence detection method used.
Time to Results Determination (TTR) Discussion
Findings of the TTR study are shown in Table III. This article summarizes the results from eight global
From the data shown, the TTR for the EM testing companies obtained after qualification and implemen-
ranged from 36 to 76 h using a single incubation format tation of an automated colony counter system including
and with a wide range of incubation temperatures used, automated incubation for routine microbial testing in
22.5˚C to 32.5˚C. Bioburden testing using TSA ranged cGMP manufacturing. The applications cover the main
from 36 to 52 h depending on the company. Water microbiological testing needs of a QC laboratory, EM
gave the longest TTR with use of the low nutrient and bioburden testing for water or product, and follow
media with results from 44 h (rinse water sample) but the verification approach proposed in USP <1223> for
with the majority in the 100 to 116 h range. automated colony counters. The rationale for that
approach was based on the media types and incubation
Method Qualification (MQ) conditions, as specified in the harmonized pharmaco-
peia with the only change being the enumeration of an
All companies passed the acceptance criteria for accu- image of the colonies taken by a CCD camera and
racy, precision, and equivalence for all applications interpreted by software rather than by the human eye.
tested, water, product bioburden, and EM. For EM The visual (human eye) approach has been shown in
equivalence, most companies used both active air and some cases to have wide variability and sensitivity
surface contact samples in the MQ and from a range of (15). This, combined with the enumeration of colonies
room qualities. Where Grade A areas or WFI was used at the end of a 3–7 day incubation period, can lead to
for sample type, there were no false positives found significant errors in the microbial status of a product or
with the testing as all samples gave 0 CFU. For compa- facility through overgrowth due to false negative
nies that performed specific disinfectant neutralization, counts caused by colonial mergers. The BioPhorum
Vol. 77, No. 3, May--June 2023 243

Table III
Determination of TTR. Acceptance Specification, Recovery ≥85% of the Visual Count for the Maximal Incubation
Time
Site System Incubation

and Media Temperature and time Time to Results Comments
Site 1 EM, water and product: 30˚C– Product: 36 hours Many species recovered >85% much
TSA 35˚C for 72 to 120 hours Rinse Water earlier for EM but minimum times will
R2A 44 hours be driven by worst case recovery
SDA EM: 60 hours organism(s)
TSA + LP80
Site 2 EM: 22.5˚C–27.5˚C for EM: 36 hours 100% recovery at 72 hours
TSA + LP80 120 hours Microorganisms included standard
molds and bacteria and in-house molds
and bacteria
Site 3 EM: 30˚C–35˚C for EM 44 hours TTR is defined by ATCC strains, site-
TSA + LP80 120 hours isolates, slow-growers, and heat-
stressed organisms. Verified with real
samples in routine
Site 4 Water: 30˚C–35˚C for Water: 116 hours Representative purified water samples
R2A 144 hoursEM: 28˚C–32˚C EM: 76 hours from site water loop
TSA + LP80 for 120 hoursProduct: 30˚C– Product: 52 hours Real air and surface samples from C/D/
TSA 35˚C for 120 hours CNC areas
pH stressed ATCC strains on phosphate
buffer (without product matrix)
Site 5 EM:25˚C–30˚C(surface and EM: 56 hours Microorganisms for EM TTR:
R2A air monitoring)30˚C–35˚C Water: 116 hours environmental samples from surface,
TSA + LP80HT (personnel monitoring)for air and personnel monitoring
96 hoursWater:30˚C–35˚C Microorganisms for water TTR:
for 164 hours Stenotrophomonas maltophilia (in-house
strain)
Methylobacterium extorquens (in-house
strain)
Water samples of different quality
Site 6 EM: 25˚C–30˚C for EM: 72 hours Site-isolates from air and surface
TSA + LP80HT 120 hours monitoring
Site 7 EM30˚C–35˚C for 168 hours EM: 52 hours Water TTR varies by facility flora.
TSA + LP80HT Water Water: 64 and Bioburden used pH 4.0 stress method
R2A 30˚C–35˚C for 120 hours 108 hours
TSA Product Product: 36 hours
30˚C–35˚C for 120 hours
Site 8 EM30˚C–35˚C for 72 hours EM: 68 hours Microorganisms for WFI water TTR
TSA + LP80 Water 30˚C–35˚C for Water: 100 hours were each “stressed” per JP
R2A 100 hours pharmacopeia:
Pseudomonas protegens (ATCC 17386)
Stenotrophomonas maltophilia (in-
house strain)
Methylobacterium extorquens (ATCC
BAA-2500)

Table III
(continued)
Site System Incubation

and Media Temperature and time Time to Results Comments
Endogenous bacteria were used for
purified water (PUW) and pretreatment
water systems, using historically highest
count sites based on annual trend reports.
EM used typical 5 USP organisms plus 3
in-house organisms.
organization has recently published a reference docu- microflora found in each site may vary (depending on
ment to cover the 9 steps from evaluation to routine global locality and environmental conditions) and the
use of any Automated Colony Counter (9). incubation conditions used may vary as well, it is
unlikely that all sites would see the same TTR for each
To verify the performance of the automated colony application. The sites contributing to this article are
detection, a simple comparison of the visual colony from East and West USA, as well as Europe, with vali-
count versus the colony count obtained by the system dations performed at various times of the year when
can be made on the same test plate. The amount of test- temperature and humidity could affect the microbial
ing required can be minimized as the possible popula- species present. As can be seen from the data presented
tion of shapes and sizes that grow are media agnostic, here, incubation conditions vary with mean incubation
for example, the same colony shapes exist on all media temperatures for EM ranging from 22.5˚C to 32.5˚C.
types depending on organism type, and the vision sys- The different temperatures will each affect growth
tem has been verified at many companies over the last rates and detection times for the organisms present.
10 years. Colony colors are not a factor, as the camera Interestingly, all sites showed good recovery using a
only sees a black and white image of the colony fluo- single incubation temperature rather than the serial
rescence, so the range of variations is significantly less incubation strategy (10–13). With bioburden and water
than with a traditional white light colony counter. As a testing, the incubation temperature was the same for all
result, the three basic shapes are specified for PQ test- sites, 30˚C–35˚C, but TTR times did differ due to the
ing, fast growing discrete E. coli, slower growing irreg- spectrum of microorganisms and local stressors used
ular shape B. subtilis, and the irregular growing hyphae dependent on process. It is noticeable that the TTR for
of A. brasiliensis. As the software accuracy is inde- the water systems is longer than that seen for other
pendent of media type sold by Rapid Micro Biosystems applications. The use of R2A as a low nutrient media
(colony shapes are not unique to media type), it only was standard for all sites but cannot support the growth
needs to be performed once for an application (e.g., rates seen by the rich media used in other applications.
bioburden) and can then be transferred to any other The original article by Reasoner and Geldreich (14)
application to be implemented with no further PQ work suggested that the optimum temperature for water
required. borne bacteria should be 20˚C to 25˚C, so use of 30˚C
to 35˚C may not be optimal; however, the higher tem-
Following confirmation of the accuracy of the technol- perature is specified in the EP. Coupling that informa-
ogy, the QC laboratory must decide on the incubation tion with the adaptation to a very low nutrient environ-
conditions for the required test in terms of temperature ment adds to the slower colony growth rates and TTR,
(single or serial) and time. Most companies have al- even with the earlier microcolony detection. The TTR
ready validated an incubation strategy for their facility should also be set with the shift pattern in use at the fa-
that can be transferred directly to Growth DirectV with
R
cility. If a one-shift pattern is used, there is no benefit
no further system qualification. However, due to the for a sample result to be produced at midnight as no
ability of the technology to detect microcolonies ear- one can react to it. In those cases, a more conservative
lier, many companies opt to determine the TTR for time can be used so that the result also appears during a
their facility and the microflora found there. As the following single shift. With a single shift, the incubation
Vol. 77, No. 3, May--June 2023 245

windows would be 40–48 h, 64–72 h, 88–96 h, and 112– full validation as an alternative microbiological
120 h. method, only verification of the enumeration software.
The technology has been qualified for water, in-process
A significant benefit for all sites implementing the bioburden, and BDS testing as well as for EM applica-
automated colony counter technology centers on the tions in several global pharmaceutical companies,
improvement in data integrity. Recent discussions on using a shortened TTR that is defined by the site-spe-
the so-called “four eyes rule”, in which every plate cific natural flora. The technology has been success-
count should be verified by a second analyst, have not fully implemented for in-process bioburden testing as
led to general adoption. The PDA TR80 Data integrity part of new drug applications to the FDA and EMA.
management system for pharmaceutical laboratories Water and environmental testing have been imple-
discussed the secondary review of microbiological test mented in routine cGMP areas through annual report
plates, noting that: changes.
“Currently a high percentage of the tests con- Conflict of Interest Declaration

ducted in microbiology laboratories are obser-
vational, that is, the results (such as a colony Jason Rose works on behalf of GSK. Any views
count) are viewed and manually recorded on a expressed are his own and he is not speaking on behalf
paper document or in a computer record. Absent of GSK.
an easy, reliable method to verify the recorded
data, some laboratories require microbiologists References
to use second person verification (e.g., supervi-
sor) by physical examination of the test plates. 1. Russell, C.; Parsaei, N.; Jones, D. L. Validation of
the Growth DirectV System for Environmental Mon-
R
Further, the second person verification could be
performed as a discrete step prior to approval of itoring and Define Optimal Incubation Conditions.
the data or combined with the data approval Eur. J. Parenter. Pharm. Sci. 2020, 25 (2), 1–13.
step”
2. Nguyen, K.; Mills, A.; Sage, A.; Jones, D. Val-
The recent update to USP <1117>, Microbiological idation of the Growth DirectV System to Perform
R
Best Laboratory Practices, confirmed the benefit of a Pharmaceutical Water Bioburden Analysis. Eur. J.
second read for the critical sterility test but does not Parenter. Pharm. Sci. 2014, 20 (3), 85–93.
recommend its use for other microbial enumeration
methods. The use of an automated colony counter and 3. Samson, A.; Durusky, A.; Carpenter, B.; Jones,
direct transmission of the counts from plate to labora- D. L.; Intinarelli, T. Validation of the Growth
DirectV System to Perform Pharmaceutical In Pro-
R
tory information management system (LIMS) database
has improved security and negates any need for a sec- cess Bioburden Analysis. IVT Network 2018, Jan
ond reader. Most sites in this study have a LIMS con- 23.
nectivity to the Growth DirectV System and can cover
R
the most common LIMS platforms available. 4. London, R.; Schwedock, J.; Sage, A.; Valley,
H.; Meadows, J.; Waddington, M.; Straus, D.
Conclusions An Automated System for Rapid Non-Destructive
Enumeration of Growing Microbes. PLoS One
The authors believe this review makes a strong case 2010, 5 (1), e8609
that colony counters such as the Growth DirectV Sys-
R
tem qualify as an automated system for the incubation 5. Jones, D.; Cundell, T. Method Verification
and reading of the compendial plate count based on the Requirements for an Advanced Imaging System
conditions stated in USP <1223> Validation of New for Microbial Plate Count Enumeration. PDA J.
Microbiological Testing Methods, and in the industry Pharm. Sci. Technol. 2018, 72 (2), 199–212.
practice document, PDA Technical Report No. 33 (re-
vised) Evaluation, Validation and Implementation of 6. Murphy, T.; Schwedock, J.; Nguyen, K.; Mills,
Alternative and Rapid Microbial Methods. As a com- A.; Jones, D. Evaluation of PDA Technical Report
pendial plate count, the technology does not require No 33. Statistics Testing Recommendations for a

Rapid Microbiological Method Case Study. PDA J. 11. Symonds, I. D.; Martin, D. L.; Davies, M. C. Facil-
Pharm. Sci. Technol. 2015, 69 (4), 526–539. ity-Based Case Study: A Comparison of the Recovery
of Naturally Occurring Species of Bacteria and Fungi
7. van den Heuvel, E.; Jones, D.; Manju, A.; Bros- on Semi-Solid Media When Incubated under Stand-
seau, J.; Parsaei, N. Primary Validation of the ard and Dual Temperature Conditions and Its Impact
Growth DirectV Bioburden System and Media.
R
on Microbial Environmental Monitoring Approach.
Eur. Pharm. Rev. 2020, 25 (5). Eur. J. Parenter. Pharm. Sci. 2016, 21 (1), 7–15.
8. Jones, D. L.; Parsaei, N.; van den Heuvel, E. 12. Sage, A.; Timas, N.; Jones, D. Determining
Primary Validation of the Growth DirectV Envi-
R
Incubation Regime and Time to Results for Auto-
ronmental Monitoring System and Media. IVT mated Rapid Microbiology EM Methods. Eur. J.
Network 2020, 26 (1). Parenter. Pharm. Sci. 2014, 19 (2), 45–55.
9. Deutschmann, S.; Carpenter, B.; Duignan, C.; 13. Moldenhauer, J. Justification of Incubation Con-
Knutsen, C.; Salvas, J.; Wysocki, L.; Plourde, L.; ditions Used for Environmental Monitoring. Am.
Johnson, L.; Eder, W. A Systematic Approach for Pharm. Rev. 2014, 17 (2), 25–29.
the Evaluation, Validation and Implementation of
Automated Colony Counting Systems. PDA J. 14. Reasoner, D. J.; Geldreich, E. E. A New Medium
Pharm. Sci. Technol. 2022, 75 (6), 509–526. https:// for the Enumeration and Subculture of Bacteria
doi.org/10.5731/pdajpst.2021.012646 from Potable Water. Appl. Environ. Microbiol.
1985, 49 (1), 1–7.
10. Gordon, O.; Berchtold, M.; Staerk, A.; Roesti,
D. Comparison of Different Incubation Conditions 15. Paris, A.; Plourde-Awobi, L. From Variable Opera-
for Microbiological Environmental Monitoring. tor Numeration to the Standardized 3P Station Auto-
PDA J. Pharm. Sci. Technol. 2014, 68 (5), 394– mated Colony Counting on Environmental Monitoring
406. Culture Media Plates. La Vague 2020, 66, 27–34.
Vol. 77, No. 3, May--June 2023 247

COMMENTARY
Something for Nothing

JAMES AGALLOCO
Agalloco & Associates © PDA, Inc. 2023
ABSTRACT: This article reviews the evolution of performance standards for aseptic processing as defined by regulatory,
pharmacopoeial, and industry activities. The increasing rigor of the expectations for sterility testing, environmental
monitoring, and process simulation have made for improvements in capability for sterile manufacturing. The document
explains why contemporary requirements are such that satisfying them requires a shift in operating practice from
manned clean rooms to isolation technology.
KEYWORDS: Closed systems, Aseptic processing, Isolation technology, Sterility testing, Environmental monitoring,
Process simulation.
Introduction limitations of the test as definitive proof of the critical at-

tribute of sterility were first discussed in the mid-1900s,
Injectable drug products require the attribute of “steril- and the same circumstances are still pertinent (4–6). The
ity” in which the absence of viable microorganisms U.S. Food and Drug Administration’s (FDA’s) 1987
allows for administration to the patient with minimal risk Guideline on Sterile Drug Products produced by Aseptic
of infection. First used in the late 1880s, they have Processing (AP) provided for a retest and even a second
always incorporated added control measures to protect retest if the test environment or methods was considered
the patient. In the 1930s, the sterility test was introduced contributary to the contamination (7). The execution of a
to check for the absence of microorganisms in sterile comprehensive investigation in the event of a failure was
drug products (1). When classified environments were stressed in making a determination on the sterility of the
first utilized for sterile product manufacture, environ- materials in question. The European Medicines Agency
mental monitoring including both viable and nonviable (EMA) issued its first version of Annex 1 – Sterile Medic-
sampling was introduced as a means for confirming the inal Products including detailed sampling requirements
suitability of the environment (2). In the 1970s, the exe- endeavoring to improve the significance of the results (8).
cution of media fills commenced as a means for confirm- Further constraints were added in the FDA’s 2004 version
ing the overall process capability (3). The expectations of the AP guidance which stated, “An initial positive test
for these monitoring systems have been codified in gui- would be invalid only in an instance in which microbial
dances and regulations and have advanced in concert growth can be unequivocally ascribed to laboratory error.”
with manufacturing technologies. A brief review of their (9). This was possible due to the increasing use of isola-
evolution can prove insightful. tors for sterility testing, which had largely made the false-
positive result a rare occurrence. The 2020 revision draft
Sterility Testing of Annex 1 reinforces sampling requirements for different
sterile manufacturing processes (10).
The original version of the sterility test included the use of
two different media, a dilution scheme, and a 5-day incu- The continued emphasis placed on sterility testing by
bation period. There was also provision for repeat testing regulators is somewhat undercut by some harsh real-
up to two additional times in the event of failure. The ities. This was beautifully articulated by Sutton.
“. . .the test states that it is applicable for meet-

Corresponding Author: Agalloco & Associates, PO
Box 899, Belle Mead, NJ 08502; Telephone: (908) ing the requirements set forth in the monograph,
305-2727; E-mail: jagalloco@aol.com the requirement being that the material meets
doi: 10.5731/pdajpst.2022.012736 the requirements of the test. Note that neither
USP citation requires the finished product to

actually be sterile, only that it meet the require- – One (1) contaminated unit is considered cause
ments of the test for sterility.” (1) for revalidation, following an investigation.
Thus, although the present day harmonized sterility test When filling from 5000 to 10,000 units:
must be conducted with a substantial amount of rigor in
the sampling, testing, and review of results, it is not proof – One (1) contaminated unit should result in
(or disproof) of the very attribute it is named for (11)! an investigation, including consideration of a
Nevertheless, it maintains a position of prominence in reg- repeat media fill.
ulatory thinking and inspectional activities.
– Two (2) contaminated units are considered
cause for revalidation, following investigation.
Environmental Monitoring
When filling more than 10,000 units:

Unlike sterility testing, there is no definitive event that
marked the beginnings of environmental monitoring in
– One (1) contaminated unit should result in
sterile product manufacturing. When the author first
an investigation.
worked in sterile manufacturing in the mid-1970s, via-
ble sampling was conducted of air, surfaces, and per-
– Two (2) contaminated units are considered
sonnel using site-specific limits. The UK Department
cause for revalidation, following investiga-
of Health and Social Security (DHSS) “orange guide”
tion.” (9)
included limits for viable air quality (12). The FDA
described environmental monitoring in its 1987 guid-
As might be expected, EMA’s draft revisions on Sterile
ance but did not define any limits (7). Informally and
Medicinal Products made the expected outcome more
around the same time, FDA Newark District senior in-
restrictive, “The target should be zero growth. Any
vestigator Hank Avallone had expressed a desire for a
contaminated unit should result in a failed process sim-
<3 CFU/test limit for viable samples in aseptic envi-
ulation. . .” (10, 14).
ronments (13). This became a de facto limit for many
of the firms Hank visited. A regulatory limit of <1
CFU/sample was included in the EMA’s 2001 Annex 1 The Elusive Goal—Perfection
and reinforced shortly thereafter by the FDA in 2004
(8, 9). These have been made even more restrictive in The thinking behind each of these is laudable but unat-
the EMA’s more recent draft revisions to Annex 1 (10, tainable—confirmation that there are no microorgan-
14). isms present during aseptic processing. The core belief
is that a sterile product cannot be made in the presence
“It should be noted that for Grade A, the of microorganisms. These expectations ignore the
expected result should be no growth”. objective realities (16):
Process Simulations 1. Sterility testing is incapable of detecting low levels

of contamination given the limited sample size.
The process simulation, or media fill test, resulted from
actions of the World Health Organization (WHO) 2. Environmental monitoring samples a tiny fraction of
including a maximum contamination rate of 0.3% at its the air and surfaces, and there is no direct evidence
introduction (15). The FDA’s 1987 Guidance decrea- linking it to contamination of exposed sterile con-
sed the acceptable rate to 0.1% (7). In 2004, the FDA’s tainers. Increases in environmental monitoring can
expectation became: never serve to confirm absence and actually serve to
increase the risk in contamination ingress.
“Recommended criteria for assessing state of
aseptic line control are as follows: 3. Media fills reflect performance during a limited time
period and to the extent that they are heavily supported
When filling fewer than 5000 units, no conta- by the operator’s aseptic technique they are not directly
minated units should be detected. indicative of past or subsequent performance.
Vol. 77, No. 3, May--June 2023 249

the operator’s role is minimized or even fully eliminated,

the remaining control elements play critical roles in steril-
ity assurance.
The 2022 revision of EMA’s Annex 1—Manufacture

of Sterile Medicinal Products have largely equated
cleanrooms, restricted access barrier systems (RABS),
and isolators (20). Perpetuating the use of manned
cleanrooms is minimally justifiable with the stated
environmental and media fill expectations in the revi-
sion. To further complicate progress toward better
outcomes, the revision establishes additional require-
ments for the more advanced systems beyond those
associated with manned environments. Rather than
advancing the implementation of superior approaches,
the EMA seems to be placing additional obstacles.1
The Means for Improvement
The role of personnel in aseptic processing has

Figure 1
changed substantially over time. At its beginnings, the
operator performed all of the tasks associated with the
Aseptic filling circa 1900.
formulation, sterilization, filling, and sealing of the
product. Figure 1 depicts hand filling of a syringe in
None of these methods, nor any of the emerging rapid
the early 1900s. There were few, if any, additional
methods being considered, are able to establish absence
controls available to protect the patient at the time.
(17). The reality is simple. No matter how restrictive the
Gradually, means for improvement became available:
limits, how frequent the testing, or how large the sample
automated fillers, membrane filters, HEPA filters, bar-
size, there are no means to prove that no microorganisms
riers of various design and sophistication, and so
are present. Underlying all of this, from the origins of asep-
forth. Contemporary aseptic systems incorporate an
tic processing to the present day, is the continued presence
extensive array of systems, machinery, and operating
of the human operator. The evolution of aseptic processing
practices that provide near-perfect protection to the
technologies from the beginning of the last century has
sterile materials being produced. As noted previously,
witnessed a gradual decrease in the extent of operator
this has been accompanied by ever restrictive regula-
intervention and increasing separation between the opera-
tory expectations for the absence of microorganisms.
tor and sterile items (18). This was necessary as the opera- What needs to be recognized is that the protective
tor was recognized early on as both the greatest contributor measures all relate to reduction in the operator’s
and greatest disseminator of microbial contamination. It impact on the aseptic process. The numerous improve-
was only as the operator’s central role in the aseptic pro- ments to aseptic processes have accomplished this by
cess was reduced that process performance increased. That (21):
same performance improvement allowed for tightening
expectations. Increasing the separation of personnel from the criti-
cal environment.
Nevertheless, perfection as projected by sampling and test-
ing with zero contamination limits is not the reality. Sam- Reducing or eliminating direct personnel interaction
pling and testing are not control measures but merely with sterile materials and equipment.
severely limited alarm mechanisms. The industry is gradu-
ally coming to the realization that the means to assure ste-
rility is derived from attention to detail in the design,
1. In 2022, The author and several colleagues wrote to the EMA and
installation, validation, and maintenance of interrelated suggested the imposition of a firm deadline for ending the use of
systems, equipment, and procedures that collectively serve manned cleanrooms for sterile products. The EMA’s reply indicated
that they will not impose such a constraint.
to prevent contamination (19). In those systems in which

Removing personnel entirely from the aseptic ever. Monitoring cannot make aseptic products safer, but
environment. elimination of the operator most certainly can (21). Ab-
sence of microorganisms is unattainable in a manned envi-
Refinements and combinations of the preceding. ronment and approachable in the best closed designs (17).
Realization of that absolute intent will never be demonstra-
The continued advances have made aseptic products ble but can be approached using closed systems. Taking
safer than ever before. The next step is obvious. Fur- the operator completely out of the process by employing
ther progress in performance requires the elimination closed systems can bring the industry closer to the unreal-
of direct personnel involvement in aseptic processes. izable goal of zero contamination. Something for nothing.
Realizing the “zero contamination” goal is possible
only by implementation of advanced aseptic processing Conflict of Interest Declaration
technologies (21). This was originally defined as:
The author declares that they have no competing interests.
“An advanced aseptic process is one in which
direct intervention with open product containers References
or exposed product contact surfaces by opera-
tors wearing conventional cleanroom garments 1. Sutton, S. The Sterility Tests. In Rapid Sterility
Testing; Moldenhauer, J., Ed.; PDA/DHI: Be-
is not required and never permitted.”
thesda, MD, 2011; pp 7–24.
A contemporized version of this original definition that
2. Federal Standard 209: Clean Room and Work Sta-
clarifies its application with the improved aseptic tech-
tion Requirements, Controlled Environment, 1963.
nologies now in use would be:
3. World Health Organization. Technical Report Se-

An advanced aseptic process is one in which
ries No. 530: Annex 4, General Requirements for
direct intervention with open sterile containers
or exposed sterilized product contact surfaces the Sterility of Biological Substances. Part A, Sec-
by operators wearing conventional cleanroom tion 2. WHO: Geneva, 1973.
garments is never permitted following decon-
4. Knudsen, L. F. Sample Size of Parenteral Solu-
tamination of the critical environment
tions for Sterility Testing. J. Am. Pharm. Assoc.,
Isolators (open and closed) and closed RABS (decontami- Sci. Ed. 1949, 38 (6), 332–337.
nated after setup and secured throughout use) satisfy this
5. Bryce, D. M. Tests for the Sterility of Pharmaceuti-
definition. The expanded use of automation and robotics
cal Preparations; the Design and Interpretation of Ste-
in these systems currently underway across the industry
rility Tests. J. Pharm. Pharmacol. 2011, 8 (1), 561–
makes the attainment of an operator-free environment
572.
widely available. Advanced aseptic processing as defined
previously relies heavily on the application of precepts of 6. Bowman, F. W. The Sterility Testing of Pharmaceut-
“closed systems” (22–24). The distinguishing design and icals. J. Pharm. Sci. 1969, 58 (11), 1301–1308.
operational elements underlying “closed systems” remove
the operator from the critical environment to an extent that 7. U.S. Food and Drug Administration, Guideline on
the “absence of” expectations can be realized, something Sterile Drug Products Produced by Aseptic Proc-
unattainable when operator involvement persists. This transi- essing. Center for Biologics Evaluation and
tion increases the importance of equipment design, steriliza- Research, U.S. Department of Health and Human
tion/decontamination, and other elements not subject to the Services: Rockville, MD, 1987.
vagaries of human performance in preventing contamination.
8. European Medicines Agency. Annex 1, Sterile Me-
Conclusion dicinal Products. EMA: London, 2001.
Closed systems are the best available means to achieve 9. U.S. Food and Drug Administration, Guideline on
the safest possible aseptically produced sterile products Sterile Drug Products Produced by Aseptic
Vol. 77, No. 3, May--June 2023 251

Processing. Center for Biologics Evaluation and 17. Madsen, R.; Agalloco, J. Unknown and Unknow-
Research, U.S. Department of Health and Human able. Pharm. Technol. 2019, 2019 (1), 4–9.
Services: Rockville, MD, 2004.
18. Agalloco, J.; Akers, J. The Myth Called Sterility.
10. European Medicines Agency. Annex 1, Sterile Medic- Pharm. Technol. 2010, 34 (3), S44–S45.
inal Products draft revision, EMA: London, 2017.
19. USP <1211> Sterility Assurance. Pharmacopeial
11. USP <71> Sterility Tests. Pharmacopeial Forum Forum 2020, 43 (5), 101–125.
2008, 34 (6), 1418–1425.
20. EMA, Annex 1, Sterile Medicinal Products, 2022.
12. Department of Health and Social Services. Guide
to Good Manufacturing Practices. HMSO, 1983. 21. Akers, J.; Agalloco, J.; Madsen, R. What is Advanced
Aseptic Processing? Pharm. Manuf. 2006, 4 (2), 25–27.
13. Agalloco, J. Personal recollection, 1984.
22. Parenteral Drug Association Inc. Process Simulation
14. European Medicines Agency. Annex 1, Sterile Medic- Testing for Sterile Bulk Pharmaceutical Chemicals.
inal Products draft revision, EMA: London, 2020. PDA Technical Report No. 28, Revised. PDA J.
Pharm. Sci. Technol. 2006, 60 (Suppl S-2), 1–22.
15. World Health Organization. WHO/BS/73.1062: Steril-
ity and Sterility Testing of Pharmaceutical Prepara- 23. Agalloco, J. Closed Systems and Environmental
tions and Biological Substances. WHO: Geneva, 1973. Control: Application of Risk Based Design. Am.
Pharm. Rev. 2004, 7 (4), 26–29.
16. Akers, J.; Agalloco, J. Environmental Monitor-
ing: Myths and Misapplications. PDA J. Pharm. 24. Agalloco, J. Straight Talk on Closed Systems.
Sci. Technol. 2001, 55 (3), 176–184. BioPharm Int. 2020, 20 (5), 32–36.

PDA Vol. 77, Issue 3 - VinaGMP

Uploaded by

Copyright:

Available Formats

PDA Vol. 77, Issue 3 - VinaGMP

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

PDA Vol. 77, Issue 3 - VinaGMP

Uploaded by

Copyright:

Available Formats

vinagmp.

com | Group Facebook: EU GMP in Viet Nam

PDA Journal of Pharmaceutical Science and Technology

Acting Editor-In-Chief PDA, Inc., Leadership

Manuscripts must be submitted online at https://submitjournal.pda.org.

Copyright © PDA, Inc. 1994 – 2023

May–June 2023 Volume 77, No. 3

Published by PDA, Inc.

Artificial Intelligence and Pharmaceutical Production

Vol. 77, No. 3, May--June 2023 145

CPV of the Future: AI-Powered Continued Process

Introduction attributes and process parameters, process capability, man-

146 PDA Journal of Pharmaceutical Science and Technology

Vol. 77, No. 3, May--June 2023 147

148 PDA Journal of Pharmaceutical Science and Technology

Vol. 77, No. 3, May--June 2023 149

150 PDA Journal of Pharmaceutical Science and Technology

remained constant regardless of the hypoxic level. It

0,001 6 0,001 0,012 6 0,002 0,032 6 0,005

cates. In Table I, a summary of the main process pa-

From these results, the next step was to design a culture

hypoxic conditions to reach the higher specific produc-

The experimental setup of the bioprocess is presented

profile through an automatic microburette to maintain a

constant specific growth rate of 0.10 h1. The out-gas

was conducted through a gas analyzer to determine the

1,12 6 0,001 1,12 6 0,01 1,30 6 0,01

lated on-line from these measurements. In addition, a

sensor of volatile compounds was coupled to the fer-

menter to monitor ethanol concentration. In Figure 2B,

Physiological Control Strategy

From the previous bioprocess results obtained, RQ and

to be controlled with the goal to maximize specific pro-

duction rate (qP). Although both parameters can be

more complex and require some estimation techniques

ing agitation power would decrease energy demands

without the use of extra gases to modify gas composi-

Thus, to maintain the RQ in the corresponding set point

value or desired range, the agitation was modified

Vol. 77, No. 3, May--June 2023 151

152 PDA Journal of Pharmaceutical Science and Technology

Vol. 77, No. 3, May--June 2023 153

154 PDA Journal of Pharmaceutical Science and Technology

Vol. 77, No. 3, May--June 2023 155

156 PDA Journal of Pharmaceutical Science and Technology

Vol. 77, No. 3, May--June 2023 157

158 PDA Journal of Pharmaceutical Science and Technology

Conclusions Revolution (50). The digitization, big data, and comput-

Vol. 77, No. 3, May--June 2023 159

160 PDA Journal of Pharmaceutical Science and Technology

Vol. 77, No. 3, May--June 2023 161

References 11. Cos, O.; Ramón, R.; Montesinos, J. L.; Valero, F.

162 PDA Journal of Pharmaceutical Science and Technology

Vol. 77, No. 3, May--June 2023 163

164 PDA Journal of Pharmaceutical Science and Technology

Operations That May Have Impact

Vol. 77, No. 3, May--June 2023 165

Flexible Loading Pattern Approach in Overkill Steam

166 PDA Journal of Pharmaceutical Science and Technology

Vol. 77, No. 3, May--June 2023 167