Bioinformatics, 33(17), 2017, 2753–2755

doi: 10.1093/bioinformatics/btx285
Advance Access Publication Date: 4 May 2017
Applications Note

Sequence analysis

modlAMP: Python for antimicrobial peptides

Alex T. Müller, Gisela Gabernet, Jan A. Hiss and Gisbert Schneider*

Downloaded from https://academic.oup.com/bioinformatics/article/33/17/2753/3796392 by guest on 15 January 2021

Department of Chemistry and Applied Biosciences, Swiss Federal Institute of Technology (ETH), CH-8093 Zurich,
Switzerland
*To whom correspondence should be addressed.
Associate Editor: Alfonso Valencia
Received on December 30, 2016; revised on April 10, 2017; editorial decision on April 22, 2017; accepted on April 22, 2017

Abstract
Summary: We have implemented the molecular design laboratory’s antimicrobial peptides pack-
age (modlAMP), a Python-based software package for the design, classification and visual repre-
sentation of peptide data. modlAMP offers functions for molecular descriptor calculation and the
retrieval of amino acid sequences from public or local sequence databases, and provides instant
access to precompiled datasets for machine learning. The package also contains methods for the
analysis and representation of circular dichroism spectra.
Availability and Implementation: The modlAMP Python package is available under the BSD license
from URL http://doi.org/10.5905/ethz-1007-72 or via pip from the Python Package Index (PyPI).
Contact: gisbert.schneider@pharma.ethz.ch
Supplementary information: Supplementary data are available at Bioinformatics online.

1 Introduction der Walt et al., 2011) arrays and pandas (McKinney, 2010) data
frames, where possible. We implemented unit testing to ensure high
The interest in membranolytic antimicrobial peptides (AMPs) has code quality. The package comes with detailed online documenta-
constantly increased over the last decade (Fjell et al., 2012). tion (URL https://pythonhosted.org/modlamp), including elaborate
Research foci have shifted from isolating natural AMPs towards the examples, that demonstrate the use of the various data classes and
computer-assisted design of synthetic analogues and mimetics with analysis methods. A sample script showcases a machine learning
improved properties (Jenssen et al., 2008; Juretic et al., 2017). workflow for classifying AMPs versus other peptides.
Several successful examples of computationally de novo generated
AMPs have been reported (Maccari et al., 2013; Müller et al.,
2016), together with new online AMP prediction tools (Waghu 2 Package description
et al., 2014; Wang et al., 2016). However, to this point one had to The modlAMP package currently consists of nine modules:
connect descriptor calculation, activity prediction and analysis tools
through custom scripts, which requires skills in different program- 1. modlamp.descriptors – molecular descriptor calculations
ming languages and environments. We here present modlAMP, 2. modlamp.sequences – in silico sequence design
a Python package to ease the discovery and design of novel synthetic 3. modlamp.database – queries to peptide databases
AMPs via the amalgamation of sequence generation, descriptor cal- 4. modlamp.datasets – precompiled classification datasets
culation, machine learning and data analysis into a single program- 5. modlamp.plot – visualization tools
ming environment. modlAMP provides functions for calculating a 6. modlamp.ml – machine learning models and functions
variety of different molecular properties and amino acid residue- 7. modlamp.wetlab – interpretation of experimental data
based peptide descriptors. Furthermore, it enables the in silico 8. modlamp.analysis – comparison of sequence libraries
generation of bespoke peptide libraries with desired properties. The 9. modlamp.core – helper functions and parent classes
package design follows a modular, object-oriented architecture. The
functions used by most of the methods are located in the core mod- 2.1 Descriptor calculation
ule and accessed by other modules through local import. We deliber- The two main classes provided by the descriptors module are
ately kept the number of objects small, and relied on numpy (van GlobalDescriptor and PeptideDescriptor. The available

C The Author 2017. Published by Oxford University Press. All rights reserved. For Permissions, please e-mail: journals.permissions@oup.com
V 2753
2754 A.T.Müller et al.

Table 1. Amino acid property scales available for descriptor calculation through Moreau-Broto type correlation with the
PeptideDescriptor class

Code-IDa Description Reference

AASI Amino acid selectivity index Juretic et al. (2009)

Argos Argos hydrophobicity scale Argos et al. (2005)
Bulkiness Side chain bulkiness Zimmerman et al. (1968)
Charge_Phys Residue charge (pH 7) Cock et al. (2009)
Charge_Acid Residue charge (pH < 6; H ¼ þ1) Cock et al. (2009)
Eisenberg Eisenberg consensus scale Eisenberg et al. (1982)
Ez Energy of lipid bilayer insertion Senes et al. (2007)
Flexibility Side chain flexibility Bhaskaran and Ponnuswamy (1988)
GRAVY Hydrophobicity Kyte and Doolittle (1982)

Downloaded from https://academic.oup.com/bioinformatics/article/33/17/2753/3796392 by guest on 15 January 2021

Hopp-Woods Hydrophobicity Hopp and Woods (1981)
ISA–ECI Isotropic surface area–electronic charge index Collantes and Dunn (1995)
Janin Hydrophobicity Cornette et al. (1987)
KyteDoolittle Hydrophobicity Kyte and Doolittle (1982)
Levitt_alpha a-helical propensity Levitt (1978)
MSS Side chain topological shape and size Raychaudhury et al. (1999)
MSW Principal components of steric and 3D residue properties Zaliani and Gancia (1999)
pepCATS Binary pharmacophoric features Koch et al. (2013)
Polarity Amino acid polarity Zimmerman et al. (1968)
PPCALI Principal components of selected side chain properties Koch et al. (2013)
Refractivity Relative refractivity values McMeekin et al. (1962)
t_scale Principal components of GRID derived values Cocchi and Johansson (1993)
TM_tend Transmembrane propensity Zhao and London (2006)
z3 Original three-dimensional Z-scale Hellberg et al. (1987)
z5 Extended five-dimensional Z-scale Sandberg et al. (1998)

Optionally, users can use their own, locally saved amino acid property scales.
a
Code-ID refers to the scalename input option of the PeptideDescriptor class.

property scales and the corresponding scalename option codes for GLFDIVKKVVGALGSL GLFDIVKKVVGALGSL

PeptideDescriptor instantiation are listed in Table 1.

Holistic, one-dimensional peptide representations, e.g. total charge,
molecular weight, hydrophobic ratio, or aromaticity, are calculated in Eisenberg value

the GlobalDescriptor class. The PeptideDescriptor class han-

dles property-based descriptors computed by Moreau-Broto correl-
ation functions with variable sliding sequence windows (Broto et al.,
1984). Amino acid sequences can be imported as individual residue
window size: 5
strings, a list of strings, or in FASTA format, e.g. as follows:
Sequence position
o from modlamp.descriptors import PeptideDescriptor
o desc ¼ PeptideDescriptor(‘GLFDIVKKVVGALGSL’, Fig. 1. Examples of helical wheel (left) and hydrophobicity (right) plots gener-
‘pepCATS’) ated with the helical_wheel and plot_profile functions
o desc.calculate_crosscorr(window¼7)
o desc.descriptor (Fig. 1). In addition, GlobalAnalysis from the analysis module
array([[0.6875, 0.46666667, 0.42857143, . . .]]) provides a graphical overview of the properties of given sequence
libraries (Supplementary Fig. S1).

2.2 Sequence generator

2.4 Machine learning
The sequence module implements ten classes for in silico sequence
modlAMP provides standard functions for machine learning and
generation (Supplementary Table S1), namely (i) random sequences,
model selection via a pipeline of data scaling, parameter grid search
(ii) sequences with a presumed amphipathic helical structure,
and model cross-validation for both support vector machine (Cortes
(iii) kinked amphipathic helices, (iv) amphipathic helices with a defined
and Vapnik, 1995) and random forest classifiers (Breiman, 2001).
hydrophobic arc, (v) sequences with a linear hydrophobicity gradient,
For example, the function ml.train_best_model performs a
(vi) centrosymmetric sequences, (vii) sequences incorporating a pos-
parameter grid search on the selected model and training dataset. As
sible heparin binding domain, (viii) sequences generated from frequent
the name implies, it returns the best performing model based on the
AMP n-grams, (ix) sequences with the residue probability of known
Matthews correlation coefficient (Matthews, 1975) obtained by
helical anticancer peptides and (x) mixed peptide libraries.
cross-validation. The functions ml.score_cv and ml.score_-
testset evaluate the performance of existing classifiers by per-
2.3 Visualization forming cross-validation or calculating the test set error,
The plot module contains several functions for data visualization respectively. The function ml.predict retrieves the pseudo-prob-
from the matplotlib package (URL https://matplotlib.org) ability of custom-generated peptides to belong to the different
modlAMP 2755

classes, and thereby informs the user about the model’s estimated Cornette,J.L. et al. (1987) Hydrophobicity scales and computational tech-
uncertainty and applicability domain. modlAMP relies on the sci- niques for detecting amphipathic structures in proteins. J. Mol. Biol., 195,
kit-learn package (Pedregosa et al., 2011), providing thoroughly 659–685.
Cortes,C. and Vapnik,V. (1995) Support-vector networks. Mach. Learn., 20,
tested state-of-the-art implementations of machine learning and
273–297.
data preprocessing methods in Python.
Eisenberg,D. et al. (1982) The helical hydrophobic moment: a measure of the
amphiphilicity of a helix. Nature, 299, 371–374.
Fjell,C.D. et al. (2012) Designing antimicrobial peptides: form follows func-
2.5 Circular dichroism spectral analysis
tion. Nat. Rev. Drug Discov., 11, 37–51.
Secondary structure dynamics may be a major feature determining
Hellberg,S. et al. (1987) Peptide quantitative structure-activity relationships,
antimicrobial activity of certain classes of AMPs. Initial laboratory
a multivariate approach. J. Med. Chem., 30, 1126–1135.
experiments usually include circular dichroism (CD) spectroscopy of Hopp,T.P. and Woods,K.R. (1981) Prediction of protein antigenic determin-
peptides in different solvents. modlAMP contains the wetlab mod- ants from amino acid sequences. Proc. Natl. Acad. Sci., 78, 3824–3828.
ule for the analysis of CD data (Supplementary Fig. S2) and signal Jenssen,H. et al. (2008) QSAR modeling and computer-aided design of antimi-

Downloaded from https://academic.oup.com/bioinformatics/article/33/17/2753/3796392 by guest on 15 January 2021

transformation to mean residue ellipticity. crobial peptides. J. Pept. Sci., 14, 110–114.
Juretic,D. et al. (2009) Computational design of highly selective antimicrobial
peptides. J. Chem. Inf. Model., 49, 2873–2882.
Juretic,D. et al. (2017) Tools for designing amphipathic helical antimicrobial
3 Conclusions
peptides. Methods Mol. Biol., 1548, 23–34.
The modlAMP package provides an application programming inter- Koch,C.P. et al. (2013) Scrutinizing MHC-I binding peptides and their limits
face that efficiently facilitates the handling of large sets of peptide se- of variation. PLoS Comput. Biol., 9, e1003088.
quences. It gives access to a full pipeline of in silico methods for Kyte,J. and Doolittle,R.F. (1982) A simple method for displaying the hydro-
peptide analysis and design, ranging from molecular descriptor cal- pathic character of a protein. J. Mol. Biol., 157, 105–132.
Levitt,M. (1978) Conformational preferences of amino acids in globular pro-
culation to machine learning. The software is provided as open
teins. Biochemistry, 17, 4277–4285.
source under the BSD-3 license (URL https://opensource.org/li
Maccari,G. et al. (2013) Antimicrobial peptides design by evolutionary multi-
censes/BSD-3-Clause) from the Python Package Index (URL https://
objective optimization. PLoS Comput. Biol., 9, e1003212.
pypi.python.org/pypi/modlamp) and can be installed through pip Matthews,B.W. (1975) Comparison of the predicted and observed secondary
(pip install modlamp). Full documentation of the package, structure of T4 phage lysozyme. Biochim. Biophys. Acta, 405, 442–451.
including use cases and sample applications, together with a detailed McKinney,W. (2010) Data structures for statistical computing in Python.
explanation of all classes and functions, is available from URL Proc. 9th Python Sci. Conf., 1697900, 51–56.
https://pythonhosted.org/modlamp/. McMeekin,T.L. et al. (1962) Refractive indices of proteins in relation to amino
acid composition and specific volume. Biochem. Biophys. Res. Commun., 7,
151–156.
Müller,A.T. et al. (2016) Sparse Neural Network Models of Antimicrobial
Acknowledgements Peptide-Activity Relationships. Mol. Inf., 35, 606–614.
The authors thank Arndt R. Finkelmann for technical support. Pedregosa,F. et al. (2011) Scikit-learn: machine learning in Python. J. Mach.
Learn. Res., 12, 2825–2830.
Raychaudhury,C. et al. (1999) Topological shape and size of peptides: identifi-
cation of potential allele specific helper T cell antigenic sites. J. Chem. Inf.
Funding
Comput. Sci., 39, 248–254.
This research was financially supported by the Swiss National Science Sandberg,M. et al. (1998) New chemical descriptors relevant for the design of
Foundation (grants no. 200021_157190, CRSII2_160699). biologically active peptides. A multivariate characterization of 87 amino
acids. J. Med. Chem., 41, 2481–2491.
Conflict of Interest: G. S. is a co-founder of inSili.com LLC, Zurich.
Senes,A. et al. (2007) Ez, a depth-dependent potential for assessing the ener-
gies of insertion of amino acid side-chains into membranes: derivation and
applications to determining the orientation of transmembrane and inter-
References
facial helices. J. Mol. Biol., 366, 436–448.
Argos,P. et al. (2005) Structural prediction of membrane-bound proteins. Eur. Waghu,F.H. et al. (2014) CAMP: Collection of sequences and structures of
J. Biochem., 128, 565–575. antimicrobial peptides. Nucleic Acids Res., 42, 1154–1158.
Bhaskaran,R. and Ponnuswamy,P.K. (1988) Positional flexibilities of amino van der Walt,S. et al. (2011) The NumPy array: a structure for efficient numer-
acid residues in globular proteins. Int. J. Pept. Protein Res., 32, 241–255. ical computation. Comput. Sci. Eng., 13, 22–30.
Breiman,L. (2001) Random forests. Mach. Learn., 45, 5–32. Wang,G. et al. (2016) APD3: the antimicrobial peptide database as a tool for
Broto,P. et al. (1984) Molecular structures: perception, autocorrelation de- research and education. Nucleic Acids Res., 44, D1087–D1093.
scriptor and SAR studies: system of atomic contributions for the calculation Zaliani,A. and Gancia,E. (1999) MS-WHIM scores for amino acids: a new
of the n-octanol/water partition coefficients. Eur. J. Med. Chem., 19, 71–78. 3D-description for peptide QSAR and QSPR studies. J. Chem. Inf. Comput.
Cocchi,M. and Johansson,E. (1993) Amino acids characterization by GRID Sci., 39, 525–533.
and multivariate data analysis. Quant. Struct. Act. Relation., 12, 1–8. Zhao,G. and London,E. (2006) An amino acid ‘transmembrane tendency’
Cock,P.J.A. et al. (2009) Biopython: freely available Python tools for com- scale that approaches the theoretical limit to accuracy for prediction of
putational molecular biology and bioinformatics. Bioinformatics, 25, transmembrane helices: relationship to biological hydrophobicity. Protein
1422–1423. Sci., 15, 1987–2001.
Collantes,E.R. and Dunn,W.J. (1995) Amino acid side chain descriptors for Zimmerman,J.M. et al. (1968) The characterization of amino acid sequences
quantitative structure-activity relationship studies of peptide analogs. in proteins by statistical methods. J. Theor. Biol., 21, 170–201.
J. Med. Chem., 38, 2705–2713.