Environmental Toxicology
pdahal60
Wk.9ArticleReview-II.pdf
TOXICOLOGICAL SCIENCES 126(1), 114–127 (2012)
doi:10.1093/toxsci/kfr339
Advance Access publication December 13, 2011
Evaluation of Drugs With Specific Organ Toxicities in Organ-Specific Cell Lines
Zhiwu Lin and Yvonne Will1
Compound Safety Prediction, Worldwide Medicinal Chemistry, Pfizer Global R&D, Groton, Connecticut 06340
1To whom correspondence should be addressed. E-mail: [email protected].
Received September 25, 2011; accepted December 7, 2011
Safety attrition of drugs during preclinical development as well
as in late-stage clinical trials continues to be a challenge for the
pharmaceutical industry for patient welfare and financial reasons.
Hepatic, cardiac, and nephrotoxicity remain the main reasons for
compound termination. In recent years, efforts have been made to
identify such liabilities earlier in the drug development process,
through utilization of in silico and cytotoxicity models. Several
publications have aimed to predict specific organ toxicities. For
example, two large-scale evaluations of hepatotoxic compounds
have been conducted. In contrast, only small cardiotoxic and
nephrotoxic compound sets have been evaluated. Here, we
investigated the utility of hepatic-, cardiac-, and kidney-derived
cell lines to (1) accurately predict cytotoxicity and (2) to
accurately predict specific organ toxicities. We tested 273
hepatotoxic, 191 cardiotoxic, and 85 nephrotoxic compounds in
HepG2 (hepatocellular carcinoma), H9c2 (embryonic myocar-
dium), and NRK-52E (kidney proximal tubule) cells for their
cytotoxicity. We found that the majority of compounds, regardless
of their designated organ toxicities, had similar effects in all three
cell lines. Only approximately 5% of compounds showed
differential toxicity responses in the cell lines with no obvious
correlation to the known in vivo organ toxicity. Our results suggest
that from a general screening perspective, different cell lines have
relatively equal value in assessing general cytotoxicity and that
specific organ toxicity cannot be accurately predicted using such
a simple approach. Select organ toxicity potentially results from
compound accumulation in a particular tissue, cell types within
organs, metabolism, and off-target effects. Our analysis, however,
demonstrates that the prediction can be improved significantly
when human Cmax values are incorporated.
Key Words: in vitro; cytotoxicity; liver-cardiac-nephrotoxicity;
Cmax.
Attrition in drug development remains high, with toxicity
being the leading cause at all stages of the drug development
process. This causes a problem for the pharmaceutical industry
for patient welfare and monetary reasons. In a recent review,
Allen et al. (2010) estimated that a 10% improvement in
predicting failure before the initiation of expensive and time-
consuming clinical trials could save upwards of $100 million in
the costs associated with drug development. Efforts have been
made, and continue to be made, to position safety assessment
earlier and earlier in the drug development process.
Testing for potential liabilities has included the areas of drug
metabolism and pharmacokinetic (PK) characteristics, various
safety endpoints including predevelopment safety pharmacol-
ogy, general toxicology, and genetic toxicology, and interroga-
tion of counterscreen data to identify off-target affinities (i.e.,
receptors, ion channels, transporters, kinases, etc.) that pose
a concern (Bass et al., 2009).
Among the many important areas of concern are the
potential for toxicities of the major organ systems. Hepatotox-
icity and cardiotoxicity remain the two major reasons for drug
attrition (Schuster et al., 2005), as well as kidney toxicity
within Pfizer’s own portfolio. Considerable progress has been
toward prediction of organ-specific toxicity, particularly for
hepatotoxicity (O’Brien et al., 2006; O’Connell and Watkins,
2010; Xu et al., 2008), with more limited efforts toward the
prediction of cardiotoxicity (e.g., QT prolongation) and other
organ-specific toxicities (Fermini and Fossa, 2003; Frid and
Matthews, 2010; Inoue et al., 2007; Zhang et al., 2007).
In the future, toxicology will move to a predominantly
predictive science focused upon broad inclusions of target-
specific, mechanism-based biological observations from a pre-
dominantly observational science at the level of disease-specific
models (Andersen et al., 2010; Berg et al., 2011). This implies
that the traditional in vivo–centered toxicity practice will be
shifted more toward in vitro cell-based studies. Consequently,
there has been increased effort in toxicology to develop and
validate new approaches to risk assessment (Huang et al., 2008;
Marchant et al., 2009; McKim, 2010; Nadanaciva et al., 2010;
Xia et al., 2008).
One of the approaches has been measuring cytotoxic drug
effects on cultured cells using either single or multiple
endpoints using plate-based or high-content screening assays.
For example, Xia et al. (2008) tested > 1000 compounds
previously tested in one or more traditional toxicologic assays
by high-throughput screening (HTS) in 13 human and rodent
cell types derived from six common targets of xenobiotic
� The Author 2011. Published by Oxford University Press on behalf of the Society of Toxicology. All rights reserved. For permissions, please email: [email protected]
D ow
nloaded from https://academ
ic.oup.com /toxsci/article/126/1/114/1713735 by guest on 20 N
ovem ber 2022
toxicity (liver, blood, kidney, nerve, lung, and skin). The
authors found that some compounds were cytotoxic to all cell
types at similar concentrations, whereas others exhibited
species- or cell type–specific cytotoxicity.
Another example by O’Brien et al. (2006) used high-content
screening to predict a drug’s potential to cause hepatotoxicity
in HepG2 cells. Of the eight parameters utilized in this assay,
cell number was found to be the most sensitive predictor of
cytotoxicity. The predictivity of the assay further increased
when compounds were tested at 30 times the human maximal
efficacious serum concentration (Ceff) reported. The advantage
of such cell-based assay approaches is that they could
potentially provide mechanistic information in addition to
having high-throughput capability.
To our knowledge, no such comprehensive compound (drug)
set has been evaluated for the prediction of either cardiac or
nephrotoxicity.
Here, we expanded on such studies by performing a large-
scale evaluation of 273 hepatotoxic compounds, 191 cardio-
toxic compounds, 85 nephrotoxic compounds, and 72
compounds with no reported organ toxicity. These compounds
were either commercially available drugs or Pfizer internal
compounds. We tested these compounds for their cytotoxicity
potential in HepG2 cells (hepatocellular carcinoma), H9c2 cells
(embryonic myocardium), and NRK-52E (kidney proximal
tubule) cells. We first investigated their utility for predicting
organ toxicity and whether using a target organ–related cell
line could accurately predict organ toxicity, and secondly,
whether Cmax-considered analysis would increase such pre-
diction. Our results suggest that three organ-specific cell lines
have relatively equal value in assessing general cytotoxicity
and that organ toxicity cannot be accurately predicted using
such a simple approach. Such toxicity potentially results from
compound accumulation in a particular tissue, metabolism, and
off-target effects. However, our analysis demonstrates that the
prediction can be significantly improved when human Cmax
values are incorporated. In fact, projection against 303 and
1003 the human Cmax value increased the predictivity even
further.
MATERIALS AND METHODS
Materials
All tissue culture reagents were purchased from either Sigma (St Louis, MO)
or Invitrogen (Carlsbad, CA). The CellTiter-Glo kit was purchased from Promega
(Madison, WI). The 384-well microplates (Corning 3712) for the luminescence
assays were purchased from Corning (Corning, NY). The 384-well microplates
(Matrix 4312) for compound spotting were purchased from Thermo Fisher
Scientific (Waltham, MA). Compounds were purchased from Sigma and Toronto
Research Chemicals (Toronto, Canada) or were synthesized internally at Pfizer.
Methods
Tissue culture. H9c2, NRK-52E, and HepG2 cells were purchased
from the American Type Culture Collection (Manassas, VA) and cultured
in Dulbecco’s Modified Eagle’s Medium (Invitrogen 11995-065) con-
taining 25mM glucose, 1mM sodium pyruvate, supplemented with 5mM
N-2-hydroxyethyl piperazine-N#-2-ethanesulfonic acid, 10% fetal bovine
serum, and penicillin-streptomycin (50 unit/ml and 50 lg/ml as final
concentrations, respectively). Cells were cultured at 37�C, 5% CO2, and 95%
humidity. H9c2, NRK-52E, and HepG2 cells were harvested in the exponential
growth phase for the experiments and passaged every 3 days up to a maximum
passage of 15. Cells were maintained on 175-cm2 flasks and seeded onto 384-
well plates for individual experiments with a density of 600 cells per well for
HepG2 and NRK-52E and 1000 cells per well for H9c2.
Compounds. Four hundred ninety-seven compounds were utilized in this
study, which included 305 commercially available compounds as well as 192
Pfizer proprietary compounds. Of the 497 compounds, 273 were designated as
primarily hepatotoxic, 191 as primarily cardiotoxic, 85 as primarily
nephrotoxic, and 72 as nontoxic. This classification of toxic and nontoxic
was based on clinical data of hepatotoxicity (Kaplowitz, 2005; Lee, 2003;
Stricker, 1992; Zimmerman, 1999) and for cardiac and nephrotoxicity on
automated queries to retrieve public information on drug-toxicity relationships,
which was manually curated to our specifications.
Therapeutic exposure levels (Cmax) were obtained from a combination of
literature searches and commercially available databases (PubMed, Physicians’
Desk Reference, Prous, and Pharmapendium). The therapeutically active
average plasma maximum concentration (Cmax) values were collated upon
single-dose administration at commonly recommended therapeutic doses. In
cases where multiple recommended doses were available, the average total
Cmax corresponding to a single administration at the median dose was used.
A compound was declared safe if the above-mentioned tools did not indicate
human organ toxicity. For Pfizer internal compounds, toxicity was found
preclinically below the projected human efficacious plasma drug concentration.
All compounds were reconstituted in 100% dimethyl sulfoxide (DMSO) as
30mM stock solutions. All compounds were soluble in DMSO at 30mM.
Compound plates were stored at room temperature in inert nitrogen boxes for
a maximum of 3 months. Test compounds were prepared using 11 doses in
a twofold dilution scheme ranging from 30 to 0.03mM. One microliter of each
dilution was spotted onto an assay plate for use in the assay within 24 h. One
microliter of 1003 compound spotted in compound plates was mixed with
99 ll of high glucose–containing growth media to obtain 13 compound
dilutions for the immediate treatment of cells. DMSO titrations were performed
at 5, 4, 3, 2, 1, 0.5, and 0.1%. DMSO concentrations � 2% did not affect
viability (data not shown). The final concentration of DMSO in the culture
media was 1%. The final drug concentrations were 300–0.3lM. Each
compound was tested in triplicate within a run, and three independent runs
(different days) were performed for each of the three cell lines.
Cytotoxicity (ATP depletion) assay. ATP content was measured as
follows: 72 h after compound treatment, cells in 30 ll media were cooled to
room temperature and mixed with an equal volume of assay reagents (CellTiter-
Glo, Promega). Cells and reagents were mixed by shaking the plates on
a shaker (VWR, Microplate shaker) at 280 revolutions per minute for 5 min.
After an additional 15 min of incubation at room temperature, the plates were
measured for their luminescent light intensity using an EnVision reader (2103
Multilabel PerkinElmer).
Data analysis. Pfizer internal data analysis software (SIGHTS), GraphPad
Prism, and Microsoft Excel were used for the data calculation and analysis.
Raw data were first analyzed using SIGHTS for assay quality evaluation. Only
plates that generated Z# value (Sui and Wu, 2007) > 0.5 were used for The half
maximal inhibitory concentration (IC50) calculation. IC50 values were calculated
using a nonlinear regression curve fit with four parameters. The maximum
(DMSO control) and minimum values (total cell kill) were fixed. Each IC50
value was generated from the data of three identically treated cell plates for each
of the three independent experiments. Means ± SDs were calculated from three
subsequent runs. Statistical analysis (ANOVA/Bonferroni posttest) was
performed using GraphPad Prism (GraphPad Software Inc., La Jolla, CA).
IN VITRO PREDICTION OF ORGAN TOXICITY 115
D ow
nloaded from https://academ
ic.oup.com /toxsci/article/126/1/114/1713735 by guest on 20 N
ovem ber 2022
RESULTS
Cell Line Performance With Reference Compounds
To evaluate assay performance, a set of six compounds
spanning the IC50 range from 1 to 300lM was run in triplicate
on three consecutive days. Mean and SD are shown in Table 1.
For all compounds, the mean ± 3SD was less than threefold
different between the mean �3SD and the mean þ 3SD.
Therefore, only IC50 values with > threefold difference were
considered statistically significant.
Cell Line Sensitivities to Compound Treatment
Three cell lines of different target tissue origin, namely,
HepG2 (human hepatocellular carcinoma), H9c2 (rat embry-
onic myocardium), and NRK-52E (human kidney proximal
tubule), were utilized to investigate the performance of each
cell line toward the detection of general (any type of organ
toxicity) cytotoxicity. Cells were exposed to compounds in
a dose-response curve for 72 h as described in detail in the
‘‘Materials and Methods’’ section, and cell viability was
evaluated using an ATP depletion assay. The compound test
consisted of 497 compounds, reported to cause hepatotoxicity,
cardiotoxicity, nephrotoxicity, or no toxicity in animals and/or
humans. This compound set included commercially available
compounds as well as Pfizer proprietary compounds.
Figure 1 shows that, of the 497 compounds, 13% were
detected within a 10lM IC50 cutoff range using H9c2 cells, 9%
using NRK-52E cells, and 10% using HepG2 cells. For the
cumulative up to a 25lM IC50 cutoff range, we detected 24%
using H9c2 cells, 17% using NRK-52E cells, and 20% using
HepG2 cells. This analysis was continued for IC50 cutoff
TABLE 1
Cell Line Performance With Reference Compounds
Drugs Clozapine Mibefradil Diphenhydramine Dexamethasone Nefazodone Troglitazone
Cell H9c2 H9c2 H9c2 H9c2 H9c2 H9c2
Day 1 10.5 ± 0.8 5.0 ± 0.2 300.0 ± 0.0 300.0 ± 0.0 15.0 ± 1.9 117.0 ± 19.3
Day 2 9.9 ± 1.3 4.6 ± 0.1 300.0 ± 0.0 300.0 ± 0.0 18.5 ± 2.2 89.3 ± 12.8
Day 3 12.7 ± 0.8 5.1 ± 0.2 300.0 ± 0.0 300.0 ± 0.0 16.2 ± 2.5 128.0 ± 21.5
Mean 11.0 4.9 300.0 300.0 16.6 111.4
SD 1.4 0.2 0.0 0.0 1.7 19.9
Mean ± 3SD 6.7–15.4 4.2–5.7 300.0 300.0 11.2–21.9 51.6–171.3
Cell NRK-52E NRK-52E NRK-52E NRK-52E NRK-52E NRK-52E
Day 1 70.1 ± 5.2 9.57 ± 8.4 300.0 ± 0.0 300.0 ± 0.0 40.8 ± 1.2 115.0 ± 3
Day 2 62.8 ± 1.7 11.1 ± 1.0 300.0 ± 0.0 300.0 ± 0.0 38.4 ± 2.2 94.9 ± 3.1
Day 3 55.6 ± 2.4 9.16 ± 0.2 300.0 ± 0.0 300.0 ± 0.0 32.4 ± 15.1 76.0 ± 2.8
Mean 62.8 9.9 300.0 300.0 37.2 95.3
SD 7.2 1.0 0.0 0.0 4.3 19.5
Mean ± 3SD 41.1–84.6 6.9–13.0 300.0 300.0 24.2–50.2 36.8–153.8
Cell HepG2 HepG2 HepG2 HepG2 HepG2 HepG2
Day 1 37.5 ± 3.2 5.6 ± 0.4 300.0 ± 0.0 300.0 ± 0.0 21.0 ± 2.1 119.0 ± 19.6
Day 2 35.9 ± 3.5 5.1 ± 1.4 300.0 ± 0.0 300.0 ± 0.0 21.2 ± 2.8 149.0 ± 24.6
Day 3 27.1 ± 3.2 4.8 ± 0.2 300.0 ± 0.0 300.0 ± 0.0 18.5 ± 2.3 95.6 ± 16.6
Mean 33.5 5.2 300.0 300.0 20.2 121.2
SD 5.6 0.4 0.0 0.0 1.5 26.7
Mean ± 3SD 16.7–50.3 3.9–6.5 300.0 300.0 15.7–24.7 40.9–201.5
Notes. HepG2, H9c2, and NRK-52E cells were treated with six reference compounds for 72 h. Cellular ATP levels were measured and applied for IC50
calculation. Compounds that failed to generate IC50 within the maximum concentration 300lM were expressed as IC50 ¼ 300lM. Data are mean and SD, n ¼ 3.
IC50 (µM) range H9c2 NRK-52E HepG2
0.29-<10 13% 9% 10%
0.29-<25 24% 17% 20%
0.29-<75 39% 32% 38%
0.29-<100 41% 37% 42%
0.29-<300 56% 54% 58%
0 .2
9 -<
1 0
0 .2
9 -<
2 5
0 .2
9 -<
7 5
0 .2
9 -<
1 0 0
0 .2
9 -<
3 0 0
0
10
20
30
40
50
60
70
IC50 (µM) cut-off range
% to
ta l c
om po
un ds
FIG. 1. Cell line sensitivities to compound treatment. H9c2 (d), NRK-
52E(n), and HepG2 (:) cells were treated with 497 compounds for 72 h.
Cellular ATP levels were measured, and IC50 values were calculated (n ¼ 3).
Cell sensitivities were calculated as percentage of total compounds, which were
detected within specific IC50 cutoff ranges.
116 LIN AND WILL
D ow
nloaded from https://academ
ic.oup.com /toxsci/article/126/1/114/1713735 by guest on 20 N
ovem ber 2022
ranges up to 75 and 100lM (see table in Fig. 1). The total
number of detectable compounds between 0.29 and 300lM
IC50 values were 56% for H9c2, 54% for NRK-52E, and 58%
for HepG2 cells. One-way ANOVA (and nonparametric)
analysis was performed using the three datasets, and no
significant difference was found between the three cell lines for
any of the IC50 cutoff ranges.
Cell Line Specificities to Compound Treatment
Following the cell line sensitivity analysis, we next
investigated if the cell lines detected the same compounds or
if there were significant differences between them. To address
this question, we first looked at each individual compound IC50
distribution between two cell lines. Figures 2A–C show the
scatter plots comparing compound IC50 values between cell
lines. We observed that most compounds were populated along
the diagonal line, which indicates that those compounds share
similar IC50 values between paired cell lines. Statistical
analysis revealed that indeed no statistical difference was
found to exist between any of the cell lines with respect to
compounds detected. The correlation coefficients were 0.90
between H9c2 and NRK-52E cells, 0.91 between H9c2 and
HepG2 cells, and 0.89 between NRK-52E and HepG2 cells.
However, we did find that 23 compounds displayed IC50
values that differed by more than threefold from cell line to cell
line. The compounds and their respective IC50 values can be
found in Table 2. For example, axitinib was much more potent
in H9c2 and NRK-52E cells than in HepG2 cells, whereas
azathioprine was less cytotoxic to H9c2 cells but moreso to
NRK-52E and HepG2 cells.
Specificity and Sensitivity Assessment for General Cytotoxicity
The ultimate goal of cytotoxicity screening is to predict
compound toxicity in either preclinical species or humans.
Toward this end, we compiled 212 commercial compounds
with toxicity information available from the literature and 72
compounds with no reported toxicity. Compounds were
categorized as toxic or nontoxic as described in detail in the
‘‘Materials and Methods’’ section. Compound annotation can
be found in Supplementary tables A–D. Our analysis focused
on evaluating the relationship between compound in vitro cytotoxicity (IC50 value) and compound toxicity in human. To
do so, we ranked the compounds based on their cytotoxicity
(IC50) values from lowest to highest IC50 value. Next, for each
IC50 value, the sensitivity (defined as the fraction of correctly
predicted positives to all true positives in the clinic) and the
specificity (defined as the fraction of correctly predicted
negatives to all true negatives in the clinic) were calculated
(Saah and Hoover, 1997).
For the predictive organ toxicity analysis, only commercial
compounds with literature-reported human Cmax information
were included. No Pfizer internal compounds were included,
due to the lack of human Cmax information. This reduced the
compound set to 109 hepatotoxicants, 62 cardiotoxicants, 41
nephrotoxicants, and 72 compounds not known to cause
toxicity in humans. All these compounds were commercially
available compounds.
Figure 3 (black lines) shows that all three cell lines had
similar sensitivity and specificity values, and no statistically
significant difference was found. Overall, at 80% sensitivity,
less than 40% specificity was achieved. Therefore, we
expanded our analysis and investigated if consideration of
human Cmax value would change the sensitivity and specificity
for the individual cell lines. For this exercise, all cytotoxicity
IC50 values were divided by their compound’s Cmax values
(lM), and these values are defined as Cmax-normalized IC50
values. Just like before, compounds were ranked based on this
normalized IC50 values, and sensitivity and specificity were
calculated. Figure 3 (colored lines) shows that sensitivity and
specificity looked different from the values displayed in
Figure 3 (black lines). However, again, the differences between
the cell lines were not statistically significant. However, the
overall predictivity was significantly higher for each of the cell
lines when Cmax was taken into consideration (p < 0.0001). At
80% sensitivity, specificity was 68% with H9c2, 59% with
HepG2, and 55% with NRK-52E.
Figure 3 demonstrates an increase in sensitivity for each cell
line for the Cmax-normalized IC50 values, and this difference
was found to be statistically significant.
Can We Predict Specific Organ Toxicities?
Thus far, we have shown equal predictivity toward toxicity
for each of the cell lines and that Cmax incorporation increased
sensitivity significantly. Here, we extended our focus toward
the prediction of a particular organ toxicity. In other words, we
wanted to know if HepG2 cells were most suitable for detecting
hepatotoxicants, H9c2 cells for detecting cardiotoxicants, and
NRK-52E for detecting nephrotoxicants. For the predictive
organ toxicity analysis, only commercial compounds with
literature-reported human Cmax information were included. No
Pfizer internal compounds were included, due to the lack of
human Cmax information. This reduced the compound set to
109 hepatotoxicants, 62 cardiotoxicants, 41 nephrotoxicants,
and 72 compounds not known to cause toxicity in humans.
Supplementary tables A–D show the individual compounds
sorted into the following categories: hepatotoxicants, cardio-
toxicants, nephrotoxicants, and nontoxic compounds with the
measured IC50 values. Compound annotation and Cmax values
were derived as stated in the ‘‘Materials and Methods’’ section
and were of our best knowledge at this time. This does not
imply that compounds could potentially change categories over
time as more safety evaluation becomes available. We applied
the same approach as above to answer the following questions:
(1) Does the usage of an organ-specific cell line provide an
advantage over the use of non–organ-specific cell lines? (2)
IN VITRO PREDICTION OF ORGAN TOXICITY 117
D ow
nloaded from https://academ
ic.oup.com /toxsci/article/126/1/114/1713735 by guest on 20 N
ovem ber 2022
0
50
100
150
200
250
300
0 50 100 150 200 250 300
HepG2 IC50 (µM)
H 9c
2 IC
50 (µ
M )
HepG2 IC50 (µM)
0
50
100
150
200
250
300
0 50 100 150 200 250 300
N R
K- 52
E IC
50 (µ
M )
H9c2 IC50 (µM)
0
50
100
150
200
250
300
0 50 100 150 200 250 300
N R
K- 52
E IC
50 (µ
M )
A
B
C
FIG. 2. Individual compound IC50 distribution maps between two cell lines. (A) H9c2 versus HepG2, (B) NRK-52E versus HepG2, and (C) NRK-52E versus
H9c2 cells. IC50 values for each compound were the mean value of three independent experiments.
118 LIN AND WILL
D ow
nloaded from https://academ
ic.oup.com /toxsci/article/126/1/114/1713735 by guest on 20 N
ovem ber 2022
Does a Cmax-considered analysis strengthen the organ toxicity
prediction? and (3) Does the organ toxicity prediction differ
from the previously reported general toxicity prediction?
Does the Use of HepG2 Cells Increase the Predictivity for Hepatotoxicity?
For this analysis, 109 compounds known to cause
hepatotoxicity and with reported human Cmax values and 72
compounds not known to cause toxicity, which also had
reported Cmax values, were utilized. Similar to the results in
Figure 3, the curves were generated for either the IC50 value
only or the Cmax-normalized IC50 values (Fig. 4A). As for the
general toxicity prediction, the Cmax-normalized values im-
proved the predictivity significantly (p < 0.001). Next, we
compared the prediction for hepatotoxicity to that of the total
(general) cytotoxicity (Fig. 4A). The IC50-based analysis
showed no statistically significant difference between the
hepatotoxicity prediction and the general toxicity prediction
(Fig. 4A, black lines). The same was true for the Cmax-based
analysis (Fig. 4A, red lines), which showed an increase in
predictivity over non–Cmax-normalized prediction but no
difference between Cmax-normalized hepatic and general
toxicity prediction. Next, the HepG2 Cmax-based hepatotoxicity
prediction was compared with the Cmax-based hepatotoxicity
prediction of the other two cells, namely, NRK-52E and H9c2
(Fig. 4B). No statistical difference was found between the three
cell lines for the prediction of hepatotoxicity. The detailed data
can be found in Table 3. At 90% specificity, HepG2 cells
display 45% sensitivity. The positive predictive value (PPV;
fraction of toxic compounds correctly predicted) was 0.88,
which means that 88% of the time a compound was called out
as toxic correctly. The negative predictive value (NPV; fraction
of nontoxic compounds correctly predicted as nontoxic) was
0.52, which means that 52% of the time a negative compound
will be truly negative. However, 48% of the time, the
compound could still be toxic (FNR, false-negative rate;
percentage of toxic compounds not correctly predicted). The
false-positive rate (FPR; percentage of nontoxic compounds
predicted to be toxic) for this category was 13%. The Cmax
cutoff value was 34lM, which means that in order to achieve
a 90% specificity no compounds with Cmax value > 34 could be
considered. Values are also reported for 95 sand 100%
specificity. H9c2 cells and NRK-52E cells performed very
similar to HepG2 cells (Table 3). At a specificity of 90%, H9c2
cells had a PPV of 0.87 and NRK-52E of 0.86. The FPR were
TABLE 2
Cell Type–Specific Organ Toxicants
Drugs Toxicity class
H9c2
IC50 (lM) SD
NRK-52E
IC50 (lM) SD
HepG2
IC50 (lM) SD
IC50 ratio
H9c2/NRK/HepG2
Axitinib Cardiac 13.4 1.7 30.4 2.7 115.8 19.4 1.0/2.3/8.6
Azathioprine Hepatic and nephro 101.8 14.9 21.2 3.1 45.8 24.7 4.8/1.0/2.1
Chloroquine phosphate Hepatic and cardiac 176.9 112 144.7 21.1 31.5 4.1 5.6/4.5/1.0
Cisapride Cardiac 38.9 13.7 208.3 90.5 55 5 1.0/5.3/1.4
Clarithromycin Hepatic 17.9 2.1 101.2 40.2 85.3 7.3 1.0/5.6/4.7
Cyanocobalamin Nontoxic 232.3 61.2 36 9.4 53.3 2 6.4/1.0/1.4
Deferoxamine mesylate Hepatic 131 30.6 15.5 5.7 43 24.2 8.4/1.0/2.7
Erlotinib (Tarceva) Nontoxic 54 11.1 300 0 259.3 70.4 1.0/5.5/4.8
Floxuridine Hepatic 296.7 5.8 1.8 0.2 300 0 164/1.0/166
Flutamide Hepatic 70.9 2.1 120.4 26 19 2.9 3.7/6.3/1.0
Ketotifen Nontoxic 8.6 2.2 139 11.5 65.7 3.2 1.0/16/7.6
Methotrexate Hepatic 300 0 300 0 0.3 0 1000/1000/1.0
6-Mercaptopurine Hepatic 268.7 47.5 17.6 4.2 36.8 19.8 15/1.0/2.0
Quinine Hepatic and cardiac 231 115 267.7 43.1 44.5 2.8 5.1/6.0/1.0
Ribavirin Nontoxic 300 0 37.6 13.3 116 36.7 7.9/1.0/3.0
0 20 40 60 80 100
0
20
40
60
80
100
Sensitivity (%)
S p
e c if ic
it y ( %
)
FIG. 3. Specificity and sensitivity assessment for general cytotoxicity.
HepG2 (d), H9c2 (:), and NRK-52E (¤). Black lines represent the IC50-based
toxicity prediction, whereas the colored lines present the Cmax-normalized IC50
toxicity prediction. Sensitivity (X) was calculated as percentage of total
toxicants detected within a specific IC50 or Cmax-normalized IC50 range
(toxicants detected divided by total toxicants). Specificity (Y) was calculated as
percentage of total nontoxicants that were not detected within a specific IC50 or
Cmax-normalized IC50 ranges (nontoxicants without being detected divided by
total nontoxicants).
IN VITRO PREDICTION OF ORGAN TOXICITY 119
D ow
nloaded from https://academ
ic.oup.com /toxsci/article/126/1/114/1713735 by guest on 20 N
ovem ber 2022
13 and 14% for H9c2 and NRK-52E, respectively. The NPV
was also similar to that of HepG2 cells. It was 0.51 and 0.49 for
H9c2 and NRK-52E, respectively. The Cmax cutoff values are
33 and 31lM for H9c2 and NRK-52E, respectively. All three
cell lines showed a rather large FNR (48–51%) at a FPR below
14%. In other words, many compounds, which were toxic to
humans, remained undetected in any of the cell lines (Table 3).
Does the Use of H9c2 Cells Improve the Predictivity for Cardiotoxicity?
For this analysis, 62 compounds known to cause cardiotox-
icity and with reported human Cmax values and 72 compounds
not known to cause toxicity, which also had reported Cmax
values, were utilized in the prediction analysis. The analysis
was carried out in the same way as for the hepatotoxicants.
0 20 40 60 80 100
0
20
40
60
80
100
Sensitivity (%)
S p e c if ic
it y ( %
)
0 20 40 60 80 100
0
20
40
60
80
100
Sensitivity (%)
S p
e c
if ic
it y
( %
)
0 20 40 60 80 100
0
20
40
60
80
100
Sensitivity (%)
S p
e c if ic
it y ( %
)
A
C
0 20 40 60 80 100
0
20
40
60
80
100
Sensitivity (%)
S p
e c if ic
it y ( %
)
0 20 40 60 80 100
0
20
40
60
80
100
Sensitivity (%)
S p
e c
if ic
it y
( %
)
0 20 40 60 80 100
0
20
40
60
80
100
Sensitivity (%)
S p
e c if ic
it y ( %
)
B
D
FE
FIG. 4. Organ-specific toxicity prediction. (A) Hepatotoxicity versus general toxicity prediction using HepG2 cells, (C) cardiotoxicity versus general toxicity
prediction using H9c2 cells, and (E) nephrotoxicity versus general toxicity prediction using NRK-52E cells. The black lines represent the IC50 prediction, whereas
the red lines represent the Cmax-normalized IC50 prediction. In addition, the solid circles represent general toxicity prediction and solid triangles organ-specific
toxicity prediction. (B) Hepatotoxicity prediction with different cell lines. (D) Cardiotoxicity prediction with different cell lines and (F) nephrotoxicity prediction
with different cell lines. The black line represents HepG2 cells, the red line H9c2 cells, and the blue line the NRK-52E cells. Sensitivity (X) and specificity (Y)
were as described for Figure 3.
120 LIN AND WILL
D ow
nloaded from https://academ
ic.oup.com /toxsci/article/126/1/114/1713735 by guest on 20 N
ovem ber 2022
Similar to the results shown in Figure 4A for hepatotoxicants,
the curves were generated for either the IC50 value only or the
Cmax-normalized IC50 values (Fig. 4C). Cmax normalization
significantly improved the overall predictivity for cardiotoxicity
(p< 0.004). Next, we compared the prediction for cardiotoxicity
to that of the total (general) cytotoxicity prediction (Fig. 4C).
The IC50-based analysis (black lines) showed no statistically
significant difference between the cardiotoxicity prediction and
the general toxicity prediction (Fig. 4C). The same was true for
the Cmax3 based analysis (red lines), which showed an increase
in predictivity over non–Cmax-normalized prediction but no
difference between Cmax-normalized cardiac or general toxicity
prediction. Next, the Cmax-based H9c2 cardiotoxicity prediction
was compared with the Cmax-based cardiotoxicity prediction of
the other two cells, namely, NRK-52E and HepG2. Figure 4D
illustrates the data. No statistical difference was found between
the three cell lines for the prediction of cardiotoxicity. The
detailed data are found in Table 3. At 90% specificity, H9c2
cells display 43.5% sensitivity. The PPV was 0.79 (79%). The
NPV was 0.66. The FPR for this category was 21%. The Cmax
cutoff value was 33lM. Values are also reported for 95 and
100% specificity. HepG2 cells and NRK-52E cells were very
similar to the H9c2 cells (Table 3). At a specificity of 90%,
HepG2 cells had a PPV of 0.77 and NRK-52E of 0.75. The FPR
was 23 and 25% for HepG2 and NRK-52E, respectively. The
NPV was also similar to that of H9c2 cells. It was 0.64 and 0.63
for HepG2 and NRK-52E, respectively. All three cell lines have
a rather large FNR (34–38%). The Cmax cutoff value for H9c2
cells was 33lM, very similar to HepG2 (34lM) cells and for
NRK-52E cells, 31lM.
Does the Use of NRK-52E Cells Improve the Predictivity for Nephrotoxicity?
For this analysis, 41 compounds known to cause nephrotox-
icity and with reported human Cmax values and 72 compounds
not known to cause toxicity, which also had reported Cmax
values, were utilized in the prediction analysis. The analysis was
carried out in the same way as for the hepatotoxicants and
cardiotoxicants. Similar to the results shown in Figures 4A and
4C for the hepatotoxicants and cardiotoxicants, the curves were
generated for either the IC50 value only or the Cmax-normalized
IC50 values (Fig. 4E). Unlike in Figures 4A and 4C, the
predictivity toward nephrotoxicity (black lines) was much lower
than the predictivity toward general toxicity. However, when
TABLE 3
Toxicity Prediction at 100, 95, and 90% Specificity
Hepatotoxicants HepG2 H9c2 NRK-52E
Specificity (%) 100.0 95.0 90.0 100.0 95.0 90.0 100.0 95.0 90.0
Sensitivity (%) 5.5 14.7 45.0 4.6 17.4 43.1 8.3 16.5 38.5
PPV 1.00 0.80 0.88 1.00 0.83 0.87 1.00 0.82 0.86
FPR 0.00 0.20 0.13 0.00 0.17 0.13 0.00 0.18 0.14
NPV 0.41 0.42 0.52 0.41 0.43 0.51 0.42 0.43 0.49
FNR 0.59 0.58 0.48 0.59 0.57 0.49 0.58 0.57 0.51
Cmax cutoff < 2 < 10 < 34 < 2 < 10 < 33 < 2 < 10 < 31
Cardiotoxicants H9c2 HepG2 NRK-52E
Specificity (%) 100.0 95.0 90.0 100.0 95.0 90.0 100.0 95.0 90.0
Sensitivity (%) 4.8 25.8 43.5 3.2 22.6 38.7 3.2 17.7 33.9
PPV 1.00 0.80 0.79 1.00 0.78 0.77 1.00 0.73 0.75
FPR 0.00 0.20 0.21 0.00 0.22 0.23 0.00 0.27 0.25
NPV 0.56 0.61 0.66 0.55 0.60 0.64 0.55 0.58 0.63
FNR 0.44 0.39 0.34 0.45 0.40 0.36 0.45 0.42 0.38
Cmax cutoff < 2 < 10 < 33 < 2 < 10 < 34 < 2 < 10 < 31
Nephrotoxicants NRK-52E HepG2 H9c2
Specificity (%) 100.0 95.0 90.0 100.0 95.0 90.0 100.0 95.0 90.0
Sensitivity (%) 9.8 31.7 58.5 9.8 34.1 58.5 9.8 31.7 61.0
PPV 1.00 0.76 0.77 1.00 0.78 0.77 1.00 0.76 0.78
FPR 0.00 0.24 0.23 0.00 0.22 0.23 0.00 0.24 0.22
NPV 0.66 0.71 0.79 0.66 0.72 0.79 0.66 0.71 0.80
FNR 0.34 0.29 0.21 0.34 0.28 0.21 0.34 0.29 0.20
Cmax cutoff < 2 < 10 < 31 < 2 < 10 < 34 < 2 < 10 < 33
Notes. The analysis was carried out based on the Cmax-normalized IC50 values. At 100% specificity, the IC50/Cmax ratio was < 2. All compounds with an IC50/
Cmax ratio < 2 were considered assay positive, whereas all others were considered assay negatives. Sensitivity was calculated by dividing the number of correctly
identified toxicants (TP) by the number of total toxicants (TT) in the test set. Specificity was calculated by dividing the number of correctly identified nontoxicants
(TN) by the number of total TN in the test set. The PPV is defined as the ability to correctly identify toxicants, and the NPV is defined as the ability to identify
nontoxic compounds as truly negative. The FPR was calculated as 1 � PPV and the FNR as 1 � NPV.
IN VITRO PREDICTION OF ORGAN TOXICITY 121
D ow
nloaded from https://academ
ic.oup.com /toxsci/article/126/1/114/1713735 by guest on 20 N
ovem ber 2022
normalized to Cmax (red lines), the predictivity now was very
similar to the one for general toxicity.
The detailed data are found in Table 3. At a 90% specificity,
NRK-52E cells display 58% sensitivity. The PPV was 0.77 and
the NPV was 0.79. The FPR for this category was 23% and the
FNR was 21%. Values are also reported for 95 sand 100%
specificity. HepG2 cells and H9c2 cells were very similar to the
NRK-52E cells (Table 3). At a specificity of 90%, H9c2 cells
had a PPV of 0.78 and HepG2 cells of 0.77. The FPR was 22
and 23% for H9c2 and HepG2, respectively. The NPV was also
similar to that of HepG2 cells. It was 0.79 and 0.80 for HepG2
and H9c2, respectively. The Cmax cutoff value was very similar
between all three cell lines. It was 31, 33, and 34lM for NRK-
52E, H9c2, and HepG2 cells, respectively.
Does the Application of Multiple of Cmax Projection Improve Predictivity Toward a Particular Organ Toxicity?
Xu et al. (2008) showed that for hepatotoxicity, predictivity
improved if the safety margin was calculated using a 1003
Cmax. Here, we explored how the predictivity would change if
we calculated the safety margin over the 13, 303, and 1003
Cmax values. Table 4 shows the results. Indeed, the analysis
indicates that 1003 Cmax increases the sensitivity for hepato-
toxicants significantly. Whereas the sensitivity at 13 Cmax was
4.6%, it increased to 39.4% at 303 Cmax and increased to 67.9%
at the 1003 Cmax. This increase in sensitivity, however, was
achieved at the cost of losing specificity. Whereas the
specificity at 13 Cmax was 100%, it decreased to 90.3% at
303 Cmax and further decreased to 75% at the 1003 Cmax. The
same was also observed for the cardiotoxicity and nephrotox-
icity prediction (Table 4). In both cases, sensitivity increased
significantly from 13 Cmax to 303 Cmax and further at 1003
Cmax, but the specificity was reduced by close to 20%. In
order to achieve relatively good specificity (90%), our data
imply that 303 Cmax is the best cutoff for the prediction
of hepatotoxicants, cardiotoxicants, and nephrotoxicants
(Table 4).
DISCUSSION
Late-stage drug attrition of drugs accounts for a substantial
amount of total development costs. Consequently, there is a strong
need to improve predictive toxicology through the development
and deployment of assays that accurately forecast toxicity of
compounds as early as possible in the drug development process.
Hepatic and cardiac toxicity have been the main reasons for
late-stage drug attrition. Those, as well as nephrotoxicity, remain
Pfizer’s own preclinical and clinical toxicity problems. In recent
years, efforts have been made to identify such liabilities earlier in
the drug development process, through utilization of in silico and in vitro toxicity models. For example, pharmaceutical
companies for many years have routinely utilized the Ames
mutagenicity assay to assess carcinogenic potential of new
chemical entities (NCEs). Recently, several assays have been
developed and applied to predict possible cardiovascular toxicity
in humans, such as QT prolongation, which have been linked to
torsade de pointes (Antzelevitch, 2007; Bauman et al., 1984).
Although in vitro assays have been utilized for many years,
the data have not been applied consistently to the decision
making process in early drug development because in vitro toxicity data have not proven to be a reliable predictor of
in vivo toxicity in either animals or humans (McKim, 2010).
Recently, Benbow et al. (2010) reported that safety findings in
rodent in vivo studies could be approximated using cytotoxicity
values. Greene et al. (2010) expanded on these observations,
by demonstrating that compounds with IC50 values < 50lM
were five times more likely to give rise to one or more adverse
finding at Cmax < 10lM in a rodent in vivo study.
Here, we took a stepwise approach to understand assay
performance of each of the three cell lines, their ability to
predict general organ toxicity, as well as a particular organ
toxicity, such as cardiac, hepatic, or nephrotoxicity.
The ATP assay was used throughout our studies as
a surrogate for cell viability. The assay is HTS compatible,
has excellent reproducibility and sensitivity, and is free from
fluorescent interference. However, the assay cannot differentiate
TABLE 4
Toxicity Prediction at 13 Cmax, 303 Cmax, and 1003 Cmax Cutoff
Hepatotoxicity prediction with HepG2 Cardiotoxicity prediction with H9c2 Nephrotoxicity prediction with NRK-52E
Cmax TI < 13 < 303 < 1003 < 13 < 303 < 1003 < 13 < 303 < 1003
Sensitivity (%) 4.6 39.4 67.9 3.2 41.9 54.8 9.8 58.5 73.2
Specificity (%) 100.0 90.3 75.0 100.0 88.9 80.6 100.0 88.9 77.8
PPV 1.00 0.86 0.80 1.00 0.76 0.71 1.00 0.75 0.65
FPR 0.00 0.14 0.20 0.00 0.24 0.29 0.00 0.25 0.35
NPV 0.41 0.50 0.61 0.55 0.64 0.67 0.66 0.79 0.84
FNR 0.59 0.50 0.39 0.45 0.36 0.33 0.34 0.21 0.16
Toxicants 109 109 109 62 62 62 41 41 41
Nontoxic compounds 72 72 72 72 72 72 72 72 72
Notes. At a 303 Cmax cutoff, all compounds were defined as assay positives if the 303 Cmax value exceeded their IC50 values and as assay negative if the Cmax
value was lower than their IC50 values. Sensitivity, specificity, PPV, NPV, FPR, and NPR were calculated in the same way as in Table 3. TI ¼ therapeutic index.
122 LIN AND WILL
D ow
nloaded from https://academ
ic.oup.com /toxsci/article/126/1/114/1713735 by guest on 20 N
ovem ber 2022
between cytotoxic and cytostatic compounds (Kangas et al., 1984; Petty et al., 1995). Other assays such as 3-(4,5-
dimethythiazol-2-yl)-2,5-diphenyl tetrazolium bromide (MTT)
or lactate dehydrogenase (LDH) release could potentially also
be used and should yield similar results (Miret et al., 2006;
Mueller et al., 2004). Our assay validation demonstrated
consistent means and low SDs over three consecutive days of
testing. From these data, we established that only compounds
with IC50 values more than threefold difference could be
considered significantly different.
Next, we engaged in a large-scale evaluation of compounds
that were hepatotoxic, cardiotoxic, or nephrotoxic or had no
reported organ toxicity. We screened these compounds in all
three organ-specific cell lines. We found that the cell lines
exhibited a similar sensitivity to the entire set of compounds.
In fact, only 23 compounds differed in their IC50 value by
more than threefold between the three cell lines. Interestingly,
the higher sensitivity found in a particular cell line for any of
these 23 compounds was not always indicative of its proposed
primary organ toxicity. For example, clarithromycin, known to
have cardiac liabilities, was much more potent in H9c2 cells
(17.9lM) than in HepG2 (85.3lM) and NRK-52E cells
(101.2lM), whereas floxuridine, a potent hepatotoxicant, gave
no response in HepG2 cells (300lM) but was very potent in
NRK-52E cells (1.8lM).
Our work compares to Xia et al. (2008) who tested 1408
compounds in 13 human and rodent cell lines. Cells were either
transformed or primary. The authors report that some
compounds were cytotoxic to all cells at similar concentrations,
whereas others exhibited species- or cell type–specific cytotox-
icity. The authors concluded that this approach might be
valuable for prioritizing compounds for further toxicology
evaluation and to identify compounds with a particular
mechanism and/or action. In our study using 497 commercial
and Pfizer proprietary compounds, we observed the same
phenomenon: cell response was compound specific but only in
rare cases cell type dependent. Our study is also in concert with
a study conducted by Schoonen et al. (2005) who tested 100
compounds in HepG2 cells and HeLa cells and 60 compounds
in ECC-1 and CHO cells. The authors concluded that in general,
all tested compounds gave similar dose and toxicity profiles in
all cell lines. Only three compounds differed.
Next, we expanded our studies by testing for each cell line
for its ability to detect a particular organ toxicity. We had
hypothesized that HepG2 cells would be superior in detecting
hepatotoxicants, whereas H9c2 and NRK-52E would be
superior for cardiac and nephrotoxicants, respectively. This
was not the case. All cell lines were equally suitable for
detecting any of the organ toxicities, which is in accord with
our first study that showed equal predictivity to all toxicants.
So how does our study differ from others that aimed to predict
a specific organ toxicity? It is commonly believed that in order
to predict specific organ toxicity, one must measure a suitable
endpoint in an organ-specific cell line. For example, hepato-
toxicity is generally investigated using rat or human hepato-
cytes (Schutte et al., 2011), whereas cardiotoxicity is assessed
using rat embryonic cardiomyocytes (Fu et al., 2010). The
endpoints utilized for toxicity evaluation range from the
cytotoxic endpoints, such as LDH, MTT, and ATP, often run
in plate-based application, to more sophisticated mechanistic
endpoints, such as reduced glutathione (GSH) content, reactive
oxygen species formation, mitochondrial function, apoptosis,
endoplasmatic reticulum stress, and DNA damage utilizing
newer and more sophisticated imaging applications, such as
high-content imaging or flow cytometry.
These endpoints, however, can potentially contribute to
a variety, if not the majority, of organ toxicities. For example,
it has been well established that GSH depletion can lead
to hepatotoxicity as well as to nephro- and cardiotoxicity
(El-Shitany et al., 2008; Fu et al., 2010; Li et al., 2011).
Mitochondrial dysfunction and apoptosis have been implicated
in hepatotoxicity, nephrotoxicity, and cardiotoxicity (Dykens
and Will, 2007; Dykens et al., 2007).
All cells have the basic machinery for cell survival and
replication (if immortalized) and cell injury pathways that can
be activated after xenobiotic insult. Therefore, we wanted to
understand how much impact the choice of cell line would
truly have in predicting the organ toxicity outcome. Searching
the literature, we also found that when researchers examined
a particular organ toxicity, they rarely included compounds that
caused different organ toxicity than the one of interest. In other
words, nobody studies the cardiotoxic effects of doxorubicin
using rat hepatocytes. For example, O’Brien et al. (2006) and
Xu et al. (2008) have conducted large-scale in vitro evaluations
of hepatotoxic compounds with the aim of predicting drug-
induced liver injury in vivo. The first study, O’Brien et al. (2006), used HepG2 (human hepatoma) cells and a high-content
imaging approach. HepG2 cells are a simple, readily accessible,
and almost unlimited source of cells from a human liver.
However, a major limitation is their reduced drug-metabolizing
capability in comparison to primary hepatocytes, which can
contribute to the differences in cytotoxicity found in both cellular
models (Westerink and Schoonen, 2007a,b). O’Brien et al. (2006) tested 243 drugs and chemicals, which were grouped as
follows: (1) severely hepatotoxic drugs, (2) moderately hepato-
toxic drugs, (3) nontoxic drugs, and (4) drugs toxic to other
organs. In addition, a variety of toxic and nontoxic chemicals
were also tested. Cells were incubated for 72 h with the highest
drug concentrations being 303 Cmax (maximum achieved
plasma concentration of a particular drug) or 100lM (if Cmax
was not available) in a dose-response curve. IC50 values were
generated for cell number, nuclear area, calcium content, and
mitochondrial membrane potential. The sensitivity for this assay
was reported to be 85% with a specificity of 98%. However,
compounds known to cause other organ toxicities were included
in the testing, but not in the prediction for hepatotoxicity.
The second large-scale study was conducted by Xu et al. (2008). This study tested over 300 drugs and chemicals, many
IN VITRO PREDICTION OF ORGAN TOXICITY 123
D ow
nloaded from https://academ
ic.oup.com /toxsci/article/126/1/114/1713735 by guest on 20 N
ovem ber 2022
of which caused liver injury in humans. The study utilized
primary human hepatocytes maintained in a matrigel-overlayed
culture. Hepatocytes were treated with each drug at a concen-
tration equal to 1003 Cmax (reported for humans) on day 3 and
subjected to high-content imaging of intracellular glutathione,
mitochondrial damage, and oxidative stress after 24 h.
A compound was considered true positive if any of the
endpoints (or multiple endpoints) were positive at the 1003
Cmax testing concentration. The assay had a true positive rate
between 50 and 60% at the testing dose of 1003 Cmax and
a FPR of 0–5%. How did Xu derive the scaling factor of 100?
For Xu to identify an idiosyncratic hepatotoxic drug with
a serious hepatotoxic event happening in less than 1 in 1000
patients, he needed to predict a population’s outlier response
rather than its mean response. Xu used a scaling factor of
sixfold to account for population Cmax variability from the
average therapeutic Cmax that was collated from the literature to
account for patient genetic (such as metabolic enzymes and
transporters) and epigenetic factors (such as age and preexist-
ing disease), which affect drug clearances. Additionally,
a sixfold uncertainty factor was used to account for higher
drug exposure to the liver via liver portal vein for an orally
dosed drug and a final threefold uncertainty factor to account
for drug-drug or drug-diet interactions due to increased usage
of polypharmacy and naturaceuticals, and potential for in-
creased drug exposure upon multiple days of dosing compared
with the single-dose Cmax values, resulting in a combined 100-
fold of the Cmax values (Xu et al., 2008). Our study shows that
the prediction of hepatotoxicity at 303 Cmax was given as
a slightly lower sensitivity than Xu observed, but a much
higher specificity. This difference might be due to the fact that
Xu also tested compounds at much higher concentrations than
300lM. Our results are more in concert with O’Brien et al. (2006), who also found that 303 Cmax resulted in the best
prediction. However, again, compounds known to cause other
organ toxicities were included into the testing, but not into the
prediction for hepatotoxicity.
In contrast, much smaller compound sets have been
evaluated for cardiac and nephrotoxicity (Inoue et al., 2007;
Wu et al., 2009; Zhang et al., 2007). Zhang et al. (2007)
evaluated three hepatotoxicants and three nephrotoxicants in the
human hepatoma cell line Bel-7402 and the human renal tubular
epithelial cell line HK-2. The authors measured cell viability,
mitochondrial membrane potential, and the neutral red assay.
The authors found good correlation for 2/3 hepatotoxicants and
1/3 nephrotoxicants when using the neutral red assay. Metabolic
activation by S9 increased the toxicity for 2/6 compounds. The
authors conclude that in vitro assays could potentially predict
organ toxicity, but if no correlate existed to the observed organ
toxicity, then absorption, distribution, metabolism, and excre-
tion properties most likely would have to be considered.
Inoue et al. (2007) used neonatal rat heart cells and adult rat
hepatocytes to predict cardio- and hepatotoxicity. They tested
four anthracyclines and five hepatotoxicants in their system.
All anthracyclines displayed an IC50 value < 2lM at 24 h in
the neonatal cardiac cells, whereas the IC50 values in the rat
hepatocytes were between 10 and 20lM. Of the five
hepatotoxicants, one compound, namely, acetaminophen, was
more toxic in the rat neonatal cells than in the rat hepatocytes.
For the other four compounds, IC50 values were less than
twofold different in the two cell lines. The authors suggest that
taking the ratio of one cell line over the other would be
predictive of possible organ toxicity. Given the less than
twofold difference for the hepatotoxicants, this approach might
not be valid with a larger compound set and certainly was not
the case with our large compound set with the exception of the
said 23 compounds.
To our surprise, the prediction of cardiotoxicity and
nephrotoxicity was also increased at 303 Cmax. Our findings
are supported by a recent expert opinion review by Weiss
(2011) who examined drug uptake and metabolism in the heart.
He reports a heart-to-plasma concentration of 20 for idarubicin
and 40 for doxorubicin 24 h after injection into rabbits.
Minocycline, which is an antibiotic with cardioprotective
effects, was shown to accumulate in rat hearts by 24-fold. In
addition, Nagai (2006) reported a selective accumulation of
aminoglycosides in the renal cortex, causing toxicity. The same
is true for cephaloridine (Koren, 1989).
Our study expands on that observation by testing for
possible organ toxicity in organ-specific cell lines and
including Cmax projection. Our results suggest that the
compound-induced cytotoxicity is of a more general nature
and the mechanisms are present in all cell lines. Based on our
data, any of these three cell lines (possibly others as well) could
be used as a first pass to examine possible cytotoxicity.
However, a prediction toward organ toxicity cannot be made
using this simple approach. The toxicity of compounds used in
humans especially via oral application is much more
complicated and dependent not only on the compound dose
ingested but also on the compound metabolism. Therefore,
compound PK information is essentially important in com-
pound efficacy and toxicity and will affect in vitro toxicity
prediction. Unfortunately, PK information is not available at
the early stage of compound development. Therefore, retro-
spective study on market compounds using in vitro cytotoxicity
assays becomes meaningful to reveal compound safety space
between in vitro toxic concentration and human toxic dose as
well as compound Cmax values. Our study showed at an
accuracy of 90% the toxic compound prediction can reach up
to 50% when Cmax data (303 Cmax cutoff) can be considered in
contrast to only 20% in the absence of such information. In the
absence of projected efficacious drug concentration or available
Cmax data, the data can at most be used to rank-order
compounds. However, that assumes that the efficacy is equal.
If binding assay Ki values or cell-based assay data are
available, one can potentially compare the cytotoxicity IC50
data as a ratio to those values. Unfortunately, the prediction of
Ceff or Cmax is much more difficult in the preclinical arena.
124 LIN AND WILL
D ow
nloaded from https://academ
ic.oup.com /toxsci/article/126/1/114/1713735 by guest on 20 N
ovem ber 2022
Compounds with very similar profile can become very different
when projected efficacious drug concentration or Cmax values
are finally projected or available. Below are a few examples
that illustrate this.
For example, the antidepressants nefazodone and trazodone
have been associated with liver toxicity, whereas no toxicity
has been reported for buspirone (DeSanty and Amabile, 2007).
In fact, nefazodone has been withdrawn from most markets. At
first glance, it appears that both trazodone (IC50 ¼ 220.3lM)
and buspirone (IC50 ¼ 253lM), are fairly nontoxic in the ATP
depletion assay, whereas nefazodone (IC50 ¼ 24.7) was much
more toxic. However, inclusion of Cmax shows that trazodone
only has a marginal better safety margin (IC50/Cmax) than
nefazodone (43.6 vs. 26.7, respectively), in contrast to
buspirone, which has a very high safety margin (19,523).
Another example is the thiazolidinones troglitazone, piogli-
tazone, and rosiglitazone. Troglitazone was withdrawn from
the market due to idiosyncratic hepatotoxicity. Rosiglitazone
and pioglitazone carry a black box warning for cardiotoxicity
as a possible class effect, and recently, rosiglitazone was
withdrawn due to increased cardiac liabilities. The ATP
depletion assay revealed that all three compounds were fairly
nontoxic (IC50 > 100lM). However, Cmax inclusion reduced
the safety margin for troglitazone to 20 in contrast to
pioglitazone and rosiglitazone, where the safety margin was
> 200 for pioglitazone but only 8 for rosiglitazone, which as
mentioned was recently withdrawn.
Our dataset included a variety of statins such as simvastatin,
atorvastatin, fluvastatin, and lovastatin. All statins have been
associated with some form of mild to moderate muscle toxicity
and in rare cases hepatotoxicity. All statins were rather potent in
the ATP depletion assay (IC50 < 30lM). However, Cmax-based
calculation gave very different safety margins for the statins.
Whereas lovastatin and simvastatin achieved a high safety index
due to their low Cmax values, atorvastatin and fluvastatin had
a much lower safety margin. This might be due to their selected
uptake by The monocarboxylate transporter (MCT4) transporter.
One could speculate that fluvastatin and atorvastatin are much
more potent substrates for MCT4 than, e.g., lovastatin or
fluvastatin. No reports in the literature have examined this to date.
We also examined two anthracyclines. Whereas doxorubicin
has been associated with cardiotoxicity, reports on daunorubicin
are more frequent. Whereas the Cmax-based safety margin is 0 for
doxorubicin, the safety margin for daunorubicin is 10-fold higher.
It appears that the safety margin might be adjusted to the
therapeutic area. For example, an oncology drug might advance
with a much lower safety margin than an antidiabetic drug,
which will be administered for an extended period of time.
Retrospective analysis of commercial drugs, and more impor-
tantly, the commercial drugs and attrited compounds developed
by the pharmaceutical industry, should provide guidance toward
establishing safety margins. In addition, the accurate prediction
of human Ceff is of paramount importance, but often not
achieved until in the clinic.
The above examples demonstrate the value of Cmax-
considered toxicity prediction. However, the experimental
design needs to be taken into some consideration. For example,
when applying 303 Cmax prediction, test compounds with
Cmax values � 10lM have to be tested at 300lM to be
considered negative. Compounds that exceed 10lM Cmax
values cannot be accurately predicted to be negative. The ideal
test compound concentration would be at the 303 Cmax
concentration, although it is not pragmatic in many cases. As to
the nontoxic control compound selection, the Cmax value of
those compounds is of paramount importance. Our experience
shows that in the 303 Cmax prediction analysis, compounds
with Cmax value < 3lM serve better as negative controls in the
data analysis. Compounds with Cmax values > 10lM
automatically become false positive because their 303 Cmax
exceeds the maximum test concentration.
In summary, we have shown that a variety of cell lines can
be used for assessment of cytotoxicity of NCEs. In rare
occasions, differences can be encountered which can be due to
metabolism or transporter differences. There is a strong need
for improvement of the large percentage of remaining false-
negative compounds. Here, we made the assumption that organ
toxicity is a result of cytotoxic response. Many cardiotoxicities
could be ion channel related and would remain undetected for
the most part. For example, the hERG (the human Ether-à-go-go Related Gene) channel is not expressed in the majority of cell
lines. Liver toxicity can be due to reactive metabolite formation
as well as inhibition of the bile salt efflux transporter. We neither
used metabolically competent cells nor do we have bile salt efflux
pump (BSEP) expression in our cell systems. In addition, liver
toxicity due to allergic reactions cannot be modeled here as it
would require coculture with a variety of cell types that can
mount an immune response. In addition, some drugs might need
extended exposure beyond 72 h. Some nephrotoxicants require
megalin expression to exhibit toxicity, such as tobramycin,
amikacin, carboplatin, cisplatin, gentamycin, and netilmicin.
Newer approaches, such as coculture models, high-content
screening of mechanistic endpoints, reporter assays, and zebra-
fish might be able to further improve the false-negative space.
The availability of the projected efficacious drug concentration
or Cmax is essential to truly rank-order compounds, as we have
shown using selected examples.
SUPPLEMENTARY DATA
Supplementary data are available online at http://toxsci.
oxfordjournals.org/.
ACKNOWLEDGMENTS
The authors would like to thank Mrs Rachel Swiss, MS, for
her excellent technical assistance with cell culture and the ATP
depletion assay, Drs Ahmed Enayetallah and Xiangyun Wang
IN VITRO PREDICTION OF ORGAN TOXICITY 125
D ow
nloaded from https://academ
ic.oup.com /toxsci/article/126/1/114/1713735 by guest on 20 N
ovem ber 2022
for the computational support for the drug-toxicity information
collection, and Dr Michael Aleo for discussions of the
manuscript. The data presented here were in part presented at
the 50th Anniversary Annual Meeting & ToxExpo during 6–12
March 2011. Abstract 1671.
REFERENCES
Allen, J., Jeffrey, P., Williams, R., and Ratcliffe, A. J. (2010). Approaches to
assessing drug safety in the discovery phase. Drugs Future 35, 67–75.
Andersen, M. E., Al-Zoughool, M., Croteau, M., Westphal, M., and
Krewski, D. (2010). The future of toxicity testing. J. Toxicol. Environ.
Health Part B Crit. Rev. 13, 163–196.
Antzelevitch, C. (2007). Ionic, molecular, and cellular bases of QT-interval
prolongation and torsade de pointes. Europace 9(Suppl. 4), iv4–iv15.
Bass, A. S., Cartwright, M. E., Mahon, C., Morrison, R., Snyder, R.,
McNamara, P., Bradley, P., Zhou, Y. Y., and Hunter, J. (2009). Exploratory
drug safety: A discovery strategy to reduce attrition in development.
J. Pharmacol. Toxicol. Methods 60, 69–78.
Bauman, J. L., Bauernfeind, R. A., Hoff, J. V., Strasberg, B., Swiryn, S., and
Rosen, K. M. (1984). Torsade de pointes due to quinidine: Observations in
31 patients. Am. Heart J. 107, 425–430.
Benbow, J. W., Aubrecht, J., Banker, M. J., Nettleton, D., and Aleo, M. D.
(2010). Predicting safety toleration of pharmaceutical chemical leads:
Cytotoxicity correlations to exploratory toxicity studies. Toxicol. Lett. 197,
175–182.
Berg, N., De Wever, B., Fuchs, H. W., Gaca, M., Krul, C., and Roggen, E. L.
(2011). Toxicology in the 21st century—Working our way towards
a visionary reality. Toxicol. In Vitro 25, 874–881.
DeSanty, K. P., and Amabile, C. M. (2007). Antidepressant-induced liver
injury. Ann. Pharmacother. 41, 1201–1211.
Dykens, J. A., Marroquin, L. D., and Will, Y. (2007). Strategies to reduce late-
stage compound attrition due to mitochondrial toxicity. Expert Rev. Mol.
Diagn. 7, 161–175.
Dykens, J. A., and Will, Y. (2007). The significance of mitochondrial toxicity
testing in drug development. Drug Discov. Today 12, 777–785.
El-Shitany, N. A., El-Haggar, S., and El-desoky, K. (2008). Silymarin prevents
adriamycin-induced cardiotoxicity and nephrotoxicity in rats. Food Chem.
Toxicol. 46, 2422–2428.
Fermini, B., and Fossa, A. A. (2003). The impact of drug-induced QT interval
prolongation on drug discovery and development. Nat. Rev. Compound Discov. 2, 439–447.
Frid, A. A., and Matthews, E. J. (2010). Prediction of compound-related cardiac
adverse effects in humans–B: Use of QSAR programs for early detection of
compound-induced cardiac toxicities. Regul. Toxicol. Pharmacol. 56,
276–289.
Fu, Z., Guo, J., Jing, L., Li, R., Zhang, T., and Peng, S. (2010). Enhanced
toxicity and ROS generation by doxorubicin in primary cultures of
cardiomyocytes from neonatal metallothionein-I/II null mice. Toxicol. In Vitro 24, 1584–1591.
Greene, N., Aleo, M. D., Louise-May, S., Price, D. A., and Will, Y. (2010).
Using an in vitro cytotoxicity assay to aid in compound selection for in vivo
safety studies. Bioorg. Med. Chem. Lett. 20, 5308–5312.
Huang, R., Southall, N., Cho, M. H., Xia, M., Inglese, J., and Austin, C. P.
(2008). Characterization of diversity in toxicity mechanism using in vitro
cytotoxicity assays in quantitative high throughput screening. Chem. Res.
Toxicol. 21, 659–667.
Inoue, T., Tanaka, K., Mishima, M., and Watanabe, K. (2007). Predictive in
vitro cardiotoxicity and hepatotoxicity screening system using neonatal rat
heart cells and rat hepatocytes. Proceedings of 6th World Congress on
Alternatives & Animal Use in the Life Sciences, 21–25 August 2007.
AATEX 14, 457–462.
Kangas, L., Grönroos, M., and Nieminen, A. L. (1984). Bioluminescence of
cellular ATP: A new method for evaluating cytotoxic agents in vitro. Med.
Biol. 62, 338–343.
Kaplowitz, N. (2005). Idiosyncratic drug hepatotoxicity. Nat. Rev. Drug
Discov. 4, 489–499.
Koren, G. (1989). The nephrotoxic potential of drugs and chemicals.
Pharmacological basis and clinical relevance. Med. Toxicol. Adverse Drug Exp. 4, 59–72.
Lee, W. M. (2003). Drug-induced hepatotoxicity. N. Engl. J. Med. 349,
474–485.
Li, Y., Li, X., Wong, Y. S., Chen, T., Zhang, H., Liu, C., and Zheng, W.
(2011). The reversal of cisplatin-induced nephrotoxicity by selenium
nanoparticles functionalized with 11-mercapto-1-undecanol by inhibition
of ROS-mediated apoptosis. Biomaterials 32, 9068–9076.
Marchant, C. A., Fisk, L., Note, R. R., Pate, M. L., and Suárez, D. (2009). An
expert system approach to the assessment of hepatotoxic potential. Chem. Biodivers. 6, 2107–2114.
McKim, J. M., Jr. (2010). Building a tiered approach to in vitro predictive
toxicity screening: A focus on assays with in vivo relevance. Comb. Chem.
High Throughput Screen. 13, 188–206.
Miret, S., De Grone, E. M., and Klaffke, W. (2006). Comparison of in vitro
assays of cellular toxicity in the human hepatic cell line HepG2. J. Biomol.
Screen. 11, 184–193.
Mueller, H., Kassack, M. U., and Wiese, M. (2004). Comparison of the
usefulness of the MTT, ATP, and calcein assays to predict the potency of
cytotoxic agents in various human cancer cell lines. J. Biomol. Screen. 9,
506–515.
Nadanaciva, S., Dillman, K., Gebhard, D. F., Shrikhande, A., and Will, Y.
(2010). High-content screening for compounds that affect mtDNA-encoded
protein levels in eukaryotic cells. J. Biomol. Screen. 15, 937–948.
Nagai, J. (2006). Molecular mechanisms underlying renal accumulation of
aminoglycoside antibiotics and mechanism-based approach for develop-
ing nonnephrotoxic aminoglycoside therapy. Yakugaku Zasshi 126,
327–335.
O’Brien, P. J., Irwin, W., Diaz, D., Howard-Cofield, E., Krejsa, C. M.,
Slaughter, M. R., Gao, B., Kaludercic, N., Angeline, A., Bernardi, P., et al.
(2006). High concordance of compound-induced human hepatotoxicity with
in vitro cytotoxicity measured in a novel cell-based model using high content
screening. Arch. Toxicol. 80, 580–604.
O’Connell, T. M., and Watkins, P. B. (2010). The application of metabonomics
to predict compound-induced liver injury. Clin. Pharmacol. Ther. 88,
394–399.
Petty, R. D., Sutherland, L. A., Hunter, E. M., and Cree, I. A. (1995).
Comparison of MTT and ATP-based assays for the measurement of viable
cell number. J. Biolumin. Chemilumin. 10, 29–34.
Saah, A. J., and Hoover, D. R. (1997). ‘‘Sensitivity’’ and ‘‘specificity’’
reconsidered: The meaning of these terms in analytical and diagnostic
settings. Ann. Intern. Med. 126, 91–94.
Schoonen, W. G., Westerink, W. M., de Roos, J. A., and Débiton, E. (2005).
Cytotoxic effects of 100 reference compounds on HepG2 and HeLa cells and
of 60 compounds on ECC-1 and CHO cells. I mechanistic assays on
ROS, glutathione depletion and calcein uptake. Toxicol. In Vitro 19,
505–516.
Schuster, D., Laggner, C., and Langer, T. (2005). Why drugs fail—A study on
side effects in new chemical entities. Curr. Pharm. Des. 11, 3545–3559.
Schutte, M., Fox, B., Baradez, M. O., Devonshire, A., Minguez, J.,
Bokhari, M., Przyborski, S., and Marshall, D. (2011). Rat primary
hepatocytes show enhanced performance and sensitivity to acetaminophen
126 LIN AND WILL
D ow
nloaded from https://academ
ic.oup.com /toxsci/article/126/1/114/1713735 by guest on 20 N
ovem ber 2022
during three-dimensional culture on a polystyrene scaffold designed for
routine use. Assay Drug Dev. Technol. 9, 475–486.
Stricker, B. H. C. (1992). Drug-Induced Hepatic Injury. Elsevier, Amsterdam,
The Netherlands.
Sui, Y., and Wu, Z. (2007). Alternative statistical parameter for high-
throughput screening assay quality assessment. J. Biomol. Screen. 12,
229–234.
Weiss, M. (2011). Functional characterization of drug uptake and metabolism
in the heart. Expert Opin. Drug Metab. Toxicol. 7, 1295–1306.
Westerink, W. M., and Schoonen, W. G. (2007a). Cytochrome P450 enzyme
levels in HepG2 cells and cryopreserved primary human hepatocytes and
their induction in HepG2 cells. Toxicol. In Vitro 21, 1581–1591.
Westerink, W. M., and Schoonen, W. G. (2007b). Phase II enzyme levels in
HepG2 cells and cryopreserved primary human hepatocytes and their
induction in HepG2 cells. Toxicol. In Vitro 21, 1592–1602.
Wu, C. F., Chuang, W. P., Li, A. H., and Hsiao, C. H. (2009). Cardiac magnetic
resonance imaging in sunitinib malate-related cardiomyopathy: No late
gadolinium enhancement. J. Chin. Med. Assoc. 72, 323–327.
Xia, M., Huang, R., Witt, K. L., Southall, N., Fostel, J., Cho, M. H., Jadhav, A.,
Smith, C. S., Inglese, J., Portier, C. J., et al. (2008). Compound cytotoxicity
profiling using quantitative high-throughput screening. Environ. Health
Perspect. 116, 284–291.
Xu, J. J., Henstock, P. V., Dunn, M. C., Smith, A. R., Chabot, J. R., and de
Graaf, D. (2008). Cellular imaging predictions of clinical drug-induced liver
injury. Toxicol. Sci. 105, 97–105.
Zhang, L., Mu, X., Fu, J., and Zhou, Z. (2007). In vitro cytotoxicity assay with
selected chemicals using human cells to predict target-organ toxicity of liver
and kidney. Toxicol. In Vitro 21, 734–740.
Zimmerman, H. (1999). Hepatotoxicity: The Adverse Effects of Drugs and Other
Chemicals on the Liver. Lippincott Williams & Wilkins, Philadelphia, PA.
IN VITRO PREDICTION OF ORGAN TOXICITY 127
D ow
nloaded from https://academ
ic.oup.com /toxsci/article/126/1/114/1713735 by guest on 20 N
ovem ber 2022
