TOXICOLOGY AND APPLIED PHARMACOLOGY 72, 9 I- 10 1 ( 1984)

Application of Biplot Methods to the Multivariate Analysis

of Toxicological and Pharmacokinetic Data

JANET S. SHY-M• DJESKA,’ J. EDMOND RIVIERE,“~ AND JOHN 0. RAWLINGS”

Laboratory of Pharmacology and Toxicology, School of Veterinary Medicine and Interdisciplinary Toxicology
Program, and Department of Statistics, North Carolina State University, Raleigh, North Carolina 27606

Received April 30, 1983; accepted August I, 1983

Application of Biplot Methods to the Multivariate Analysis of Toxicological and Pharmacokinetic

Data. SHY-M• DJESKA, J. S., RMERE, J. E., AND RAWLINGS, J. 0. ( 1984). Toxicol. Appl. Pharmacol.
72, 91-101. The biplot technique was applied to aminoglycoside renal toxicological and phar-
macokinetic data in beagles. The biplot obtains a two-dimensional approximation to a matrix
and plots row effects and column effectsjointly, depicting relationships among different observed
variables and simultaneously showing the relationship of experimental units as individuals and
as treatment groups to those variables. This graphical representation of the matrix allows inspection
of relationships, trends, clusters, approximate correlations, and variances existing in the data.
Biplots were generated from gentamicin dosage regimen nephrotoxicity data. Six dogs classified
as being intoxicated by established indicators of renal toxicity were a distinct cluster. A cluster
of nonintoxicated dogs was separated into two groups approximating nephrectomized and normal
dogs, thus revealing variables signihcant in separating toxic and nontoxic as well as nephrectomized
and normal dogs. Biplots from pharmacokinetic data were able to separate different renal disease
states on the basis of disease-induced changes in gentamicin pharmacokinetic parameters. In
conclusion, the biplot technique proved to be a very useful tool in exploring this type of data
by revealing clear relationships between nephrotoxicity and physiological and pharmacokinetic
variables and by separating different disease states based on these data.

The biplot is a graphical multivariate statistical the data matrix, allows visual inspection of
tool which displays a two-dimensional ap- patterns existing in the data (e.g., relationships,
proximation to a data matrix of samples and trends, and clusters) as well as approximate
variables and allows for a visual examination correlations and variances (Gabriel, 197 1,
of the structure of the data. A single graph 1972). Although this technique has been pub
can depict relationships among different ob- lished in some fields (Gabriel, 1972, 198 1;
served parameters, (i.e., clinical pathologic Strauss et al., 1979), it has not been extensively
data, histopathologic scores, and pharmaco- utilized in toxicology (Tepper et al., 1982) al-
kinetic constants) and simultaneously illus- though other multivariate techniques have
trate the relationships of experimental units proved to be of value (Gad and Weil, 1982).
as individuals and as treatment groups to each It is the purpose of this paper to demonstrate
other and to these parameters. The technique, the application of the biplot technique to tox-
based on a singular value decomposition of icological and pharmacokinetic data through
the analysis of the nephrotoxic potential and
pharmacokinetics of the aminoglycoside an-
’ Laboratory of Pharmacology and Toxicology.
* Address correspondence and reprint requests to tibiotic gentamicin in dogs with experimen-
Dr. J. E. Biviere. tally induced renal insufficiency. The biplot
’ Department of Statistics. was particularly useful in analyzing a com-

91 0041-008X/84 $3.00
Copyright 0 1984 by Academic Press Inc.
All rights of reproduction 1” any for”? rewn’ed
92 SHY-MODJESKA, RIVIERE, AND RAWLINGS

posite set of data previously reported in in- (yk, Rk), k = 1, 2, 3, 4. The reader is referred to Gabriel
dependent studies (Riviere and Coppoc, (1972) for the details of these computations. The biplot
is especially useful if the first two principal components
1981a,b; Riviere et al., 1981; Riviere, 1982a). account for most of the total variability:
In spite of the complex influence of kidney
disease on the detection of nephrotoxicosis
and on the elimination of drugs from the kid-
ney, a clear relationship between nephrotox-
where n = rank of matrix. “Total variabiity” is measured
icity and both physiological and pharmaco- here as the sum of sum of squares for each variable. Since
kinetic parameters was seen. all variables were standardized, the sum of squares for
each is equal to its degrees of freedom, m - 1, and
C xf = n(m - 1). Row constants, ylr, and column con-
METHODS stants, X~W~, were then plotted jointly as tlri vs t)ri and as
h,w,- vs X2w2j, respectively, for i = 1, 2, . . . m and j
Description of Biplot Analysis = 1,2, * - - n. Column vectors were constructed by draw-
ing lines from the origin (0,O) to their end points (Xl Wrj,
The technique of biplotting uses the singular value de- X,w,). Row vectors were left as points (Uri, ~ri). This two-
composition (SVD) of a matrix to give a graphical display dimensional plot is the best two-dimensional represen-
of its two-dimensional approximation. The approximation tation of the original n-dimensional space defined by the
obtained by using the first and second singular value com- data matrix, i.e., a maximum proportion of the dispersion
ponents of a matrix is the best possible two-dimensional in the original space is retained and the rank two matrix
approximation in the least-squares sense (Gabriel, 1972). represented by this plot is the best rank two approximation,
In our analysis, the m X n matrix Y consisting of n in the least-squares sense, of the original rank n matrix.
variables observed on each of m experimental animals The biplot reveals the structure of the data insofar as it
was standardized by correcting each variable (column) for can be represented in two dimensions. If two dimensions
its mean and dividing by its standard deviation to give are inadequate, more biplots can be used to study the
variables of unit variance. This new matrix X was used third, fourth, etc. dimensions.
asthe input matrix for Gabriel’s biplot analysis to compute For illustrative purposes, an idealized biplot is presented
the first four singular values (X3 and singular vectors in Fig. 1. Each plotted vector represents the projection of

First principal component for rows

“1
0.0

First principal component for columns

FIG. 1. Idealized biplot of data matrix.

 APPLICATION OF BIPLOT TO TOXICOLOGY 93

a variable’s original vector onto this two-space and each culated by dividing the electrolyte clearance by the si-
poini X represents an experimental unit. Since all the multaneously determined CL, and expressed as a per-
vectors are of equal length, m - 1 in n-space, the relative centage. Freewater clearance was calculated by subtracting
lengths of the plotted vectors reflect the closeness with the osmolar clearance from the rate of urine production.
which they fall in this two-space. Had the variables not All clearances used in the analyses were the average of
been standardized before the analysis, the relative lengths two 24-hr determinations, normalized to body weight.
would include differences in variances. Angles between During drug dosage, peak (30 min postdosing) and
vectors represent approximate correlations with small acute trough (immediately prior to next dose) gentamicin con-
angles showing high positive correlations between variables centrations were determined in all dogs on Days 1, 4, I 1.
(e.g., a and b), right angles showing correlations of 0 (e.g., and 14. The daily elimination rate constant (Kel) was
b and c), and large obtuse angles showing high negative then calculated according to the formula:
correlations (e.g., e and b) (- 1 6 r C 1). Clusters of points
would represent experimental units with similar parameter In peak (&ml) - In trough (&ml)
Kel (hr-‘) =
values. Perpendicular projections of the points onto any dosing interval - 0.5 hr
vector approximate the relative values of these. experi-
mental units for that variable. where peak and trough are defined as before. In addition
the ratio of the first and last Kel’s was calculated:

Sample Data Sets Kel( 14)

-z Kel of Day 14 of treatment
Kel( I) Kel of Day I of treatment
Nephrotoxicity. Twenty-four adult female beagles were
Following the final assessment of renal function, dogs
subtotally nephrectomized (3/4 or 7/8) and assigned to
were euthanized and necropsied. Renal tissue samples
one of four gentamicin dosage regimens so that mean
were collected and prepared for light microscopy: histo-
pretreatment serum creatinine 2 weeks postsurgery was
logical sections were examined to quantib the degree of
as balanced as possible among treatments (Table 1). (Ri-
interstitial nephritis, tubular necrosis, tubular dilatation.
viere et al., 1982). All four treatment groups contained
and tubular regeneration by assigning values from 0
approximately equal numbers of high and low serum cre-
through 5 for severity of the condition. A histopathologic
atinine. A fifth group of six dogs, with intact kidneys, was
score (0 to 20) was defined for each beagle as the sum of
assigned to the standard dosage regimen.
these individual scores. Established indicators of genta-
Renal function was assessedin all dogs by the following
micin toxicity were considered to be increased SCR, SUN,
methods: serum creatinine (SCR), serum urea nitrogen
CINa, and CIK, decreased CL, and low KeI ratio (Schentag,
(SUN), 24-hr endogenous creatinine clearance (CLJ, urea
1982; Cronin et al.. 1980; Brinker et al., 1981; Lufi ct
and free water clearances (Cl,,, Cl&, and fractional
al., 1978).
sodium and potassium clearances (CINa, Clx).
The first biplot (Fig. 2) was based on the 29 dogs having
Renal clearances of all substances were calculated ac-
complete data for all variables: histopathologic score, Kel
cording to the formula:
ratio, and a transformation on the magnitude of change
Clearance = i&V/P, (ml/min) of the 7 physiological parameters from pre- to postdrug
administration. The transformation of the data giving the
where Ux = urinary concentration of compound X, P, largest separation among dosage regimens, i.e., the trans-
= serum concentration of compound X, and V = urine formation with the highest F ratio for treatments from
flow (ml/min). Fractional electrolyte clearances were cal- an analysis of variance of a randomized incomplete block
design, was chosen for each variable: log (post Cb,/pre
TABLE 1 C4A pre Cl,,-post Ck, log (post &/pre CIK), log (post
SCR/pre SCR), post SUN/pre SUN, preCI,,-postcl,,.
EXPERIMENTAL DESIGN and log (post CI&pre Cl&.
The second biplot (Fig. 3) was based on the same 29
Dosageregimen n beagles and 9 variables but with the 7 physiological pa-
rameters recorded as direction of change from pre to post
DW Interval” Nephrec- (+ 1 if post-pre > 0, - 1 if post-pre < 0, and 0 if post
Treatment h/kg) (hr) tom&d Normal
= pre). Both data sets were standardized before analysis.
Correlations among the 9 variables were computed for
Standard 3 8 6 6
Fixed interval 1.5 or 1.0 8 6 - both data setsby Proc CORR in Statistical Analysis System
Fixed dose 3 16 or 24 6 - (SAS, 1979), and clusters among the 29 dogs were com-
Control none none 6 - puted by Proc CLUSTER in SAS to confirm relationships
suggested in the biplots.
a Time between successivedosesduring the 14day treatment Pharmacokinetics. A second data set consisting of 8
period. pharmacokinetic parameters(ClB, @, Vs,, Vd(-), V,, Klz,
94 SHY-MODJESKA, RIVJERE, AND RAWLINGS

First principal component for animals

-0.40 -0.20 0.0 0.20 0.40

.

0.60- cIK*
- 0.40

0.40 - 0.20
Kek/Kel i

0.0

-0.40-

- -0.40
-0.60-

I
I I 1 1 1 I r I +
-0.60 -0.40 0.0 0.40 0.60

First principal component for physiologic 6

pharmacologic parameters

FIG. 2. Biplot of transformed physiologic data, Kel ratio, and histopathologic index. R& = 76%. A, Toxic
dogs, 0, nontoxic nephrectomized dogs, l , nontoxic dogs with intact kidneys, *, transformations of variables.

K2,, Kel) and Cl, for 23 beagles was also analyzed by the The parameters obtained by nonlinear regression analysis
biplot technique. These data were obtained from analysis were then fit to a two-compartment open pharmacokinetic
of the serum gentamicin decay curve after a single iv dose model by methods previously described (Riviere and Cop
of drug was administered; samples were collected for 6 pot, 1981a; Riviere ef al.. 1983). Six of the dogs which
hr and analyzed for gentamicin by radioimmunoassay. were nephrectomized (7/8,3/4) had received the standard

First prlnclpal component for anhndr

.4a

1
-0.60 0.0 0.60

First principal component for physiologic L

pharmacokinetic parameters

FIG.3. Biplot of direction of change of physiologic data, Kel ratio, and histopathologic index. R$, = 69%.
A, Toxic dogs, 0, represents nontoxic nephrectomized dogs, 0, nontoxic dogs with intact kidneys.
 APPLICATION OF BIPLOT TO TOXICOLOGY 95

dosage regimen and were part of the previously described eters, Kel ratio, and histopathologic scores had
experiment (see data above). Seventeen additional dogs, a rank 2 goodness of fit of 76% (I?$) = 76%)
not part of the previous experiment, were also utilized:
(Fig. 2). High positive correlations are seen
10 were considered to be normal (Riviere and Coppoc,
198 1a; Riviere, 1982a), 4 had glomerulonephritis (Riviere between Kel and Q,, SUN and SCR, and
et al., 198 I), and 3 had catheters surgically implanted in SCR and histopathologic scores (p significantly
the left lateral ventricle of the brain for determination of different from 0 at p < 0.05). The correlation
gentamicin concentrations in the cerebral spinal fluid between Clk and Cln,o and between CIN, and
(CSF) (Riviere and Coppoc, 1981b).
Because five beagles had no Cl, data, this parameter Cl,,, is not significantly different from 0. Cor-
was eliminated from the first biplot on this data set, leaving relations approaching - I exist between SCR
just pharmacokinetic data to be analyzed for all 23 dogs and CL,, histopathologic score and Kel, and
(Fig. 4). between Clk and CIN, (Table 2). Vectors falling
The subset of 18 dogs containing information on all 9 on the right-left axis (SUN, SCR, histopath-
variables consisted of 8 normal, 6 nephrectomized, and
4 glomerulonephritis beagles. Biplots were constructed for ologic score, kel ratio, and C&J define a pri-
the normal dogs; normal and nephrectomized dogs; and mar-y axis and form a highly correlated set of
normal, nephrectomized. and glomerulonephritis dogs variables. Vectors to the right of center (Kel
with a subset of the 9 variables: C&, Vss, Vc, and Cl, ratio and Cl,,) and those to the left of center
(Fig. 5). These pharmacokinetic variables were selected (SUN, SCR, and histopathologic score) are
because they are the model independent, primary phar-
macokinetic, parameters. highly correlated among themselves but are
The four data sets used to generate these biplots were negatively correlated to each other. They ap-
standardized before analysis, and correlations among the pear to separate the dogs into two clusters. A
four sets of variables were again computed with Proc cluster analysis (Proc CORR) grouped the
CORR in SAS for verification of correlations shown by same six dogs into a distinct cluster and the
the biplots.
remaining 23 dogs into another cluster. The
group of six dogs had physiological and phar-
RESULTS
macokinetic alterations consistent with those
Nephrotoxicity previously associated with aminoglycoside
nephrotoxicosis and were therefore classified
The biplot analysis of transformations of as “toxic” dogs; similarly, the 23 dogs were
magnitude of change of physiological param- classified as “nontoxic” (Tables 3 and 4). The

First prmcipal component for animals

-0.50 -0.25 0.00 0.25 0.50
I I 5 I 1
d
4.0-

First principal Componenl for pharmacokmehc parameters

FIG. 4. Biplot of pharmacokinetic data. Rc2)

* - 75%. 0, Normal dogs, A, nephrectomized dogs. 0, cerebral
spinal fluid dogs, X, glomerulonephritis dogs.
 F~rsl prmcipal component for animals

c
2.f3- 0.60

l.! 5-
040
1.C1-

0.f j- 0.20

0.c )-
0

-0.: i-
0

-1.c )-
-0.40
-1.5 F
E
‘u -2.c k 0 -0.60
$ -2.0 -1.0 0.0 1.0 2.0
I I I I I I
x -0.50 -0.25 0.0 o 0.25 0.50
-2 3.C )- 0.60
.-i? 0
2 2.0I-
lki
E
c2 1.c
CL

5 0.c
F
2 -1.c -0.20
zl
E
x -2.a -0.40
z
.a
2 -3.c -0.60
'C -3.0
I -1.0
I 1 .o
I 3.0
P I I I I I I I
2 4.0 I.40 -0.20 0.0 0.20 0.4 0.50
8
fz 3.0 C

2.0 0 x 0.30
x vs
1.0 V
X 0.10
0.0 C 0.00
-0.10
-1.0
0
-2.0 0 -0.30
0
0 O
T CICR
-0.50
-4.0

-4.0 -3.0 -2.0 -1.0 0.0 1.0 2.0 3.0 4.0

first principal component for pharmacokinetic parameters

FIG. 5. (A) Biplot of pharmacokinetic data for normal dogs. R,,* - 83%. 0, Normal dogs. (B) Biplot of
pharmacokinetic data for normal and nephrectomized dogs. R& = 93%. 0, Normal dogs, A, nephrectomized
dogs. (C) Biplot of pharmacokinetic data for normal, nephrectomized, and glomerulonephritis dogs.
R$) = 93%. 0, Normal dogs, A, nephrectomized dogs, X, glomerulonephritis dogs.
96
 APPLICATION OF BIPLOT TO TOXICOLOGY 97

TABLE 2
CORRELATION COEFFICIENTS OF TRANSFORMED PHYSIOLOGICAL AND
PHARMACOKINETIC DATA AND HISTOPATHOL~GIC INDEX

SCR” CI, SUN Cl,, Cl,, clK Histo’ Kel

SCR 1.00 -0.76’ 0.79’ 0.29 -0.44’ -0.70’ 0.51’ 0.82’ -0.76’
CL, 1.00 -0.52’ -0.3 1 0.54’ 0.61’ -0.47’ -0.78’ 0.63‘
SUN 1.00 0.20 -0.32 -0.69’ 0.55’ 0.54” -0.54’
Cl”, 1.00 -0.44’ 0.02 -0.36’ 0.32 -0.24
Cl”,, 1.00 0.16 0.07 -0.59’ 0.53’
Ch, 1.oo -0.75’ -0.50’ 0.45 i
Cl, 1.oo 0.42’ 0.31
Histo b 1.oo -0.83’
Kel 1.00

’ All variables transformed as defined under Methods.

’ Histopathologic score.
’ r significantly different from 0 at p < 0.05.

top-bottom axis (Clx , Cl,,, CIHZO, and Cl,,) = 69% (Fig. 3). The same six dogs still grouped
is only partially correlated to the right-left together in the biplot and the cluster analyses
axis. Correlations among these variables are and again were classified as toxic, while most
near 0 and - 1. The nontoxic dogs are begin- of the non-nephrectomized dogs fell in another
ning to be separated into groups approxi- cluster, and the remaining nephrectomized
mating those that were nephrectomized and (nontoxic) dogs fell in a third cluster. (When
those six that were not. R& is low, the biplot procedure can be ex-
Direction of change of physiological pa- tended to include more dimensions.) High
rameters, histopathologic score, and Kel ratio, positive correlations exist between SUN and
when subjected to a biplot analysis, gave a Clx, SUN and CIN,, Clx and CIN,, and Kel
wider separation among toxic, nontoxic, ne- and C&,. Negative water clearance has a cor-
phrectomized, and normal dogs and an R$, relation not significantly different from 0 with
SUN, ClK, and CINa; while both Kel and Cl,,
TABLE 3 have correlations near - 1 with histopathologic
score. Again, the right-left axis (Kel ratio, Cl,,
PHusro~ocrc, PHARMACOKINETIC, AND HIST~-
PATHOLOGIC PARAMETERS IN TOXIC Dots (n = 6)
SCR, and histopathologic score) is made up
of two sets of variables highly correlated
Parameter Pre Post among themselves and negatively correlated
to each other and apparently separate neph-
CL 0.67 (0.11)” 0.44 (0.14) rotoxic and nontoxic dogs. The top-bottom
SCR 1.87 (0.25) 4.27 (1.23)
SUN 40.83 (7.83) 72.17 (19.77) axis (C1n20, Cl,,, Clx, CIN,, and SUN) is
Cl?& 0.95 (0.10) 1.51 (0.31) made up of two sets of variables highly cor-
ClK 27.98 (4.86) 61.32 (22.97) related among themselves and uncorrelated
Kel ratio NAb 0.48 (0.07) to each other and separates normal from ne-
Histopathologic phrectomized dogs (Fig. 3). This biplot was
index NA 10.50 (0.56)
used to help generate a “nephrotoxicity index”
a Mean (SE). by giving an indication of variables that sig-
b NA, Not appropriate. nificantly separate toxic and nontoxic dogs.
98 SHY-MODJESKA, RIVIERE, AND RAWLINGS

TABLE 4
PHYSIOLOGIC,PHARMACOKINETIC,ANDHISTOPATHOLOGICPARAMETERSFORNONTOX~CDOGS(~ = 23)

Nephrectomized dogs Normal dogs (intact Nephrectomized dogs

receiving gentamicin kidneys) receiving receiving no gentamicin
(n = 11) gentamicin (n = 6) (n = 6)

Parameter Pre Post Pre Post Pre Post

CL, 0.93 (0.07)” 1.11 (0.10) 2.09 (0.23) 2.57 (0.23) 0.66 (0.11) 1.05 (0.17)
SCR 1.50 (0.14) 1.46 (0.15) 0.60 (0.03) 0.68 (0.08) 1.72 (0.22) 1.48 (0.17)
SUN 32.91 (2.96) 3 1.09 (4.66) 13.50 (0.92) 18.00 (2.08) 34.17 (5.28) 30.83 (4.04)
cLNa 0.93 (0.12) 0.64 (0.10) 0.3 1 (0.09) 0.46 (0.10) 1.10 (0.26) 0.70 (0.14)
CL, 29.67 (4.91) 20.42 (4.66) 5.68 (0.76) 12.28 (2.35) 35.13 (8.14) 17.12 (4.36)
Kel ratio NAb 0.85 (0.04) NA 0.95 (0.11) NA NA
Histopathologic
index NA 3.09 (0.44) NA 2.17 (0.60) NA 0.67 (0.33)

’ Mean (SE).
b NA, Not appropriate.

Pharmacokinetics Positive correlations exist among Vc, Vss,

V++ and CIB, between K,z and & , and
Considering pharmacokinetic data similarly Kel and p. & and K2, have correlations not
(Fig. 4), one sees normal dogs near the center significantly different from 0 with both Kel
of the biplot, implying little change in their and /3 and have negative correlations with Cle ,
parameters. Nephrectomized dogs fall to the Vd(area)y VSS, and VC (Table 5).
left, with low Kel and Cla values. The CSF The biplot of the eight normal dogs with
dogs are within the cluster of normal dogs respect to CL,, Vc, Cln, and VSSreveals no
suggesting that the catheterization did not af- clusters or relationships, reflecting random
fect the kinetics of gentamicin. The glomer- variation among the dogs (Fig. 5A). Inclusion
ulonephritis dogs tend to fall to the right of of nephrectomized dogs gives an entirely dif-
the normal dogs with slightly higher Vc, Vss, ferent biplot. Normal dogs cluster together
Vdc-), and Cln values (R& was 75%). midway up the four vectors while nephrec-

TABLE 5
COWLATIONCOEFFKIENTSOFPHARMACOKINETICDATA

B ClB VSS Vd(arca) vc 42 &I Kel

B 1.00 0.70” 0.15 0.19 0.32 -0.22 -0.18 0.74”

c&3 1.oo 0.76“ 0.82* 0.78” -0.32 -0.35 0.61’
VSS 1.00 0.94’ 0.86” -0.15 -0.16 0.13
V d(Ma) 1.00 0.81” -0.22 -0.31 0.30
VC 1.oo -0.44” -0.40” 0.02
Kt2 1.00 0.90” 0.13
K21 1.00 -0.02
Kel 1.00

Br significantly different from 0 at p < 0.05.

 APPLICATION OF BIPLOT TO TOXICOLOGY 99

tomized dogs cluster along the vectors in the however, such a table of clinical data does not
opposite direction (Fig. 5B). Addition of the also show the relationships among toxicologic
four glomerulonephritis dogs reveals another variables. The biplot shows both kinds of in-
relationship. Although there is more variability formation simultaneously. This attribute is
among the glomerulonephritis dogs, they tend especially powerful in the situation above
to have higher kinetic parameters and lower where variables indicative of nephrotoxicosis
CL, readings than the normal dogs. Again, the (i.e., CL,, SUN) may be confounded with
nephrectomized dogs have lower values for variables indicative of the underlying renal
all four parameters when compared to both insufficiency. The biplot was successful in sep-
glomerulonephritis and normal dogs. Rank 2 arating these two states of renal insufficiency
goodness of fit for those three biplots was 83, because additional variables specific for each
93, and 93%, respectively. Vector lengths in- of the conditions were utilized. It is of interest
crease (Fig. 5) because degrees of freedom to note that the high correlation between Cl,,
(m - 1) increased as more dogs were included and the Kel ratio implies that changes in drug
in the analysis. elimination over time in this study are pri-
marily due to the drug-induced decrease in
glomerular filtration rate. In this situation the
DISCUSSION compound under study, i.e. gentamicin, serves
as an excellent marker of glomerular f&ration.
This study has shown that the biplot is a The biplot can also discriminate between
useful tool in exploratory analysis of the struc- the effects of different disease states on drug
ture of toxicological and pharmacokinetic disposition in animals. For example, normal
data. Not only does it provide graphic ap- dogs and nephrectomized dogs differed in all
proximations to complex data sets, it also gives pharmacokinetic variables measured and
insight into relationships between toxicologic therefore appeared as two separate groups in
and pharmacokinetic variables as well as be- the biplot (Figs. 4 and 5B). Although nephrec-
tween individual treatment responses. This tomized and glomerulonephritis dogs both
technique replaces the need for multiple two- have renal insufficiency characterized by a low
dimensional scatter plots or tables of corre- glomerular filtration rate estimated by low
lation coefficients; however, in all cases pre- CL,, differences become apparent in the biplot.
sented in this study, the relationships between Glomerulonephritis dogs have higher Vss, I’, ,
variables demonstrated on the biplot were and C1ii compared to nephrectomized dogs
confirmed by independent correlation and which have markedly reduced CIB and some-
cluster analyses. what contracted Vssand Vc compared to nor-
A biplot analysis can determine if clusters mal. In healthy animals aminoglycoside CIB
of samples exist as a result of specific disease is directly correlated to CL or to other mea-
states (e.g. nephrectomy) and further, what sures of glomerular filtration rate (Chiu et al.,
variables separate the clusters. A nephrotox- 1976; Pechere and Dugal, 1979; Schentag,
icity index consisting of variables separating 1980; Riviere, 1982b). The biplot clearly sug-
toxic and nontoxic dogs was developed by gests an uncoupling of this normal association
employing the biplot in this manner. The pri- between Cla and CL,, an event discussed in
mary parameters separating toxic dogs, i.e., the original studies (Riviere et al., 198 1; Ri-
SCR, CL,, Kel ratio, and histopathologic in- viere, 1982a). This uncoupling can be appre-
dex, are similar to those reported by other ciated from the clustering evident in the biplots
investigators for aminoglycoside toxic ne- (Figs. 4 and 5C) or from the increased angle
phropathy (Schentag, 1982; Cronin et al., between the Cl* and Cl, vectors when glo-
1980; Brinker et al., 1981; Lufi et al., 1978); merulonephritis dogs are added to the input
100 SHY-MODJESKA, RIVIERE, AND RAWLINGS

data matrix (Fig. 5C vs Fig. 5B). This illus- dogs in the biplot (Fig. 4), lending support to
tration suggests a powerful use for the biplot the assumption that the surgical procedure in-
in pharmacokinetic studies, namely, as a volved in implanting ventricular cannulas did
graphical tool to discern the effects of different not alter drug disposition as compared to his-
pathophysiological states on the parameters torical controls. The biplot would be useful
of drug disposition by studying the rotation in toxicologic studies utilizing normal animals
of the parameter vectors induced by the ad- to determine relationships between indepen-
dition of animals with specific disease states dent toxicologic parameters. In addition, by
to the overall data matrix. This information generating sequential biplots composed of
would be especially beneficial if performed in subsets of all measured parameters, the op-
a real-time environment on a video monitor timal subset which discriminates between toxic
where vector displacement could be quantified and nontoxic animals, i.e., the parameters
and highlighted. The above mentioned disease which make up the primary axis, can be se-
states can be compared and contrasted easily lected. The use of these parameters in sub-
from the biplot, conclusions drawn, and out- sequent toxicologic studies should increase the
liers readily noted. Note that when vectors do efficiency and power of the experiments.
not fall into the two-space being plotted, they In summary, the biplot appears to be a use-
are short vectors and the angles between them ful tool in toxicologic and pharmacologic re-
may not be good reflections of their correla- search as a graphical device to explore rela-
tions. tionships between physiologic or pharmaco-
The relationships seen between pharma- logic variables and individual treatment
cokinetic parameters in Fig. 4 are consistent responses to these variables. Biologically sig-
with theoretical assumptions (Gibaldi and nificant variation in the data should be present
Perrier, 1982; Wagner, 1975; O’Flaherty, before the biplot is applied. The biplot is par-
1981; Pechere and Dugal, 1979; Schentag, ticularly useful as an exploratory tool and is
1980, 1982). For example, the K12 and &, not intended for statistical test of hypotheses.
distribution rate constant vectors are perpen- Its use is analogous to that of scattergrams
dicular to the Kel vector indicating no cor- and residual plots in regression analysis
relation. This is implicit to their derivation. (Chattejee and Price, 1977). Associations or
However, note that if only relatively homo- hypotheses suggested by the biplot should be
geneous data are utilized as input (Fig. 5A), verified with independent sets of data to insure
relationships between parameters may not be the results are not unique to that set of data.
very meaningful. In this case, only the cor-
relation structure of homogeneous individuals
within one treatment group is being displayed. ACKNOWLEDGMENTS
Although it is usually evident in a biplot anal-
ysis when only random variation is present, The authors thank Dr. Gordon L. Coppoc, Dr. William
techniques such as the cluster analysis can be W. Carlton, and Dr. Edward J. Hinsman of the School
of Veterinary Medicine, Purdue University in West La-
used to verify such conditions or conclusions.
fayette, Indiana, for their support in the original studies
Biologically significant variation must be from which these data originated. We acknowledge Mr.
present if meaningful interpretations are to be Richard Rogers of the Laboratory of Pharmacology and
made by this technique. Toxicology for assistance in graphical preparation of the
Another application for the biplot is in de- biplots. This work was supported in part by the USDA-
ES No. 12-050300-595. Portions of this work were pre-
termining whether individual animals can be
sented in abstract form at the Annual Meeting of the
classified as normal in terms of more than one American Society of Pharmacology and Experimental
parameter. For instance, the dogs involved in Therapeutics in Philadelphia, Pennsylvania on August I 1,
the CSF experiment cluster with the normal 1983 (Pharmacologist 25, 232, 1983).
 APPLICATION OF BIPLOT TO TOXICOLOGY 101

REFERENCES of cerebrospinal fluid gentamicin in the beagle using

an indwelling cerebral ventricular cannula. Chemo-
therapy 21, 309-312.
BRINKER, K. R., BULGER, R. E., D~BYAN, D. C., STAREY, RIVIERE, J. E., COPPOC, G. L.. HINSMAN, E. J., AND
T. R., SOUTHERN, P. M., HENRICH, W. L., AND CARLTON, W. W. (198 1). Gentamicin pharmacokinetic
CRONIN, R. E. (198 1). Effect of potassium depletion changes in induced acute canine nephrotoxic glomer-
on gentamicin nephrotoxicity. J. Lab. C/in. Med. 98, ulonephritis. Antimicrob. Agents Chemother. 20, 387..
292-301. 392.
CHATTERJEE,S., AND PRICE, B. (1977). Regression Anal- RIVIERE, J. E., COPPOC,G. L., HINSMAN, E. J., CARLTON.
ysis by Example. Wiley, New York. W. W., LANTZ, G. C., AND CARVER, M. P. (1982).
CHIU, P. J. S., BROWN, A., MILLER, S., AND LONG, J. F. Comparative nephrotoxicity of fixed-dose versus fixed-
( 1976). Renal extraction of gentamicin in anesthetized interval reduction of gentamicin dosage in partially ne-
dogs. Antimicrob. Agents Chemother. 10, 277-282. phrectomized dogs. Proc. Intersci. Conf Antimicroh
CRONIN, R. E., BULGER, R. E., SOUTHERN, P. M., AND Agents Chemother. 22, 74.
HENRICH, W. L. (1980). Natural history of aminogly- RIVIERE, J. E. (1982a). Paradoxical increase in amino-.
coside nephrotoxicity in the dog. J. Lab. Clin. Med. glycoside body clearance in renal disease when volume
95,463-474. of distribution increases. J. Pharm. Sci. 71, 720-72 I
GABRIEL, K. R. (1971). The biplot graphic display of RIVIERE, J. E. (198213). Limitations on the physiologic
matrices with application to principal component anal- interpretation of aminogiycoside body clearance derived
ysis. Biometrika 58, 458-467. from pharmacokinetic studies. Res. Commun. Chem.
GABRIEL, K. R. (1972). Analysis of meteorological data Pathol. Pharmacol. 38, 3 1-42.
by means of canonical decomposition and biplots. J. RIVIERE, J. E., COPFQC,G. L., HINSMAN, E. J., CARLTO~,
.4pp/. Meteor. 11, 1071-1077. W. W., AND TRAVER, D. S. (1983). Species dependent
GABRIEL, K. R. (1981). Biplot display of multivariate gentamicin phannacokinetics and nephrotoxicity in the
matrices for inspection of data and diagnosis. In In- young horse. Fundam. Appl. Toxicol. (in press).
terpreting Multivariate Data (V. Bamett, ed.), pp. 147- SAS User’s Guide. Statistical Analysis System Institute
173. Wiley, New York. Inc., Raleigh, N.C., 1979.
GAD, S. C.. AND WEIL, C. S. (1982). Statistics for toxi- SCHENTAG, J. J. (1980). Aminoglycosides. In Applwd
cologists. In Principles and Methods of Toxicology Pharmacokinetics (W. E. Evans, J. J. Schentag, and
(A. W. Hayes, ed.), pp. 273-320, Raven Press, New W. J. Jusko, eds.), pp. 174-209. Applied Therapeutics.
York. Inc., San Francisco.
GIBALDI, M.. AND PERRIER,D. (1982). Pharmacokinetics, SCHENTAG,J. J. ( 1982). AminogJycoside pharmacokinetics
2nd ed. Dekker. New York. as a guide to therapy and toxicology. In The Amino-
LUFT. F. C.. BLOCK, R., SLOAN, R. S., YUM, M. N., glycosides (A. Whelton and H. C. Neu, eds.). pp. 141-
COSTELLO, R., AND MAXWELL, D. R. (1978). Com- 167, Dekker, New York.
parative nephrotoxicity of aminoglycoside antibiotics STRAUSS,J. S., GABRIEL, K. R., KOKES, R. F.. RITZLF.R.
in rats. J. Injtict. Dis. 138, 541-545. B. A. VANORD, A.. AND TARANA. E. (1979). Do psy-
O’FLAHERTY. E. J. (198 I). To.ricants and Drugs: Kinetics chiatric patients fit their diagnoses? Patterns of symp-
and Dynamics. Wiley, New York. tomatology as described with the biplot. J. Nerv. Mental
PECHERE, J. C., AND DUGAL, R. (1979). Clinical phar- Di.r. 167: 105-113.
macokinetics of aminoglycoside antibiotics. Clin. Phar- TEPPER. J. L.. WEISS. B., AND Cox. C. (1982). Micro-
macokin. 4, 170- 199. analysis of ozone depression of motor activity. Tovi<,rJl
RIVIERE. J. E., AND COPPOC,G. L. (198 la). Pharmaco- Appl. Pharmacoi. 64, 3 17-326.
kinetics of gentamicin in the juvenile dog. Amer. J. Vet. WAGNER, J. G. (1975). Fundamentals of Clinical Phar-
Res. 42, 1621-1623. macokinetics. Drug Intelligence Publications, Inc ~
RIVI~RE, J. E.. AND COPPOC.G. L. (198 I b). Determination Hamilton. III.