The Toxicologist - Society of Toxicology

containing substituents and/or hydrophobic ring systems as being significantly correlated
with PL activity. Model features and scaffolds were organized using a hierarchical
scaffold tree to illustrate their relatedness and their ability to predict PL independently.
Closer examination of subsets of related scaffolds show the effects of
various molecular substitution patterns on the correlation with PL activity, as well
as underscore the need to be judicious in selecting molecular features that best predict
the training set without overly compromising the model’s domain of applicability
by defining the local structure environment too narrowly.
137 IN SILICO PREDICTIVE MODEL FOR DRUG-
INDUCED PHOSPHOLIPIDOSIS USING BIOEPISTEME
SOFTWARE.
S. S. Choi, L. G. Valerio and N. Sadrieh. U.S. FDA, Silver Spring, MD.
Drug-induced Phospholipidosis (DIPL) is a recognized finding in pharmaceutical
drug development. DIPL is characterized by accumulation of drugs and phospholipids
in lysosomes. Pathologically, DIPL manifests foamy macrophages or cytoplasmic
vacuoles in various tissues of both animals and humans. These pathologic
findings can be confirmed by the appearance of lamellar inclusion bodies by electron
microscopy. CADs (cationic amphiphilic drugs) are known to induce PL and
share common structural features of containing hydrophobic ring structure and hydrophilic
amine portion. This has led FDA to investigate the structural similarities
and differences for active and inactive compounds. FDA has been developing a PL
database that currently contains over 700 PL positive and negative drugs. Using this
database and various computational programs, FDA has developed predictive
QSAR models for DIPL. In this study, new In silico models were generated and validated
using BioEpisteme, a toxicity screening and QSAR model development program,
developed by the Prous Institute for Biomedical Research. BioEpisteme facilitates
creation and application of In silico toxicity predictions based on molecular
descriptors. In addition, it provides predictions on mechanism of action (MOA)
(mechanism of action) for 432 different pharmacological targets. Preliminary performance
statistics of a new DIPL QSAR model using a genetic algorithm of
BioEpisteme show 87% specificity, 73% sensitivity, and 82% accuracy. These results
were compared with other available FDA QSAR models. In addition, the
MOA model was investigated to predict possible pharmaceutical MOA targets of
the DIPL data set. These highly performing In silico predictive models will help to
predict DIPL early in drug development and will subsequently serve to support regulatory
decisions.
138 PREDICTION OF DRUG-INDUCED LIVER INJURY IN
HUMANS WITH TOXICOGENOMICS.
M. Zhang, Q. Shi, M. Chen and W. Tong. NCTR, U.S. FDA, Jefferson, AR.
Drug induced liver injury (DILI) is a leading cause of liver failure and the withdrawal
of drugs from the market. However, the mechanism of DILI is still not well
understood, and the prediction of the DILI potential for specific drugs remains a
challenge. Over the past few years, large-scale microarray experiments testing DILI
have been performed on animal models and have provided a good opportunity to
understand DILI on the gene expression level. Combining our recently developed
liver toxicity knowledge base- benchmark database (LTKB-DB), which groups
drugs by DILI potential based on their drug labels, we used the microarray data
from a previous toxicogenomic study to distinguish the DILI potentials of drugs by
analyzing gene expression profiles of rats treated by those drugs. We focused on
eighteen drugs with different DILI potentials and achieved a classification accuracy
of over 80% using the leave one compound out cross validation process, which not
only provided in silico evidence for drug categories in LTKB-DB, but also suggested
the feasibility of predicting drug induced liver injury using microarray expression
data. Additionally, the effects of the ALT/TBL level in the treated samples
on the DILI potentials of the drugs were also discussed.
139 PREDICTIVE VALUE OF CHEMICAL AND
TOXICOGENOMIC DESCRIPTORS FOR DRUG-
INDUCED HEPATOTOXICITY.
Y. Low 1, 5 , T. Uehara 1, 2 , Y. Minowa 2 , H. Yamada 2 , Y. Ohno 3 , T. Urushidani 2, 4 ,
A. Sedykh 5 , D. Fourches 5 , H. Zhu 5 , I. Rusyn 1 and A. Tropsha 5 . 1 Environmental
Sciences & Engineering, University of North Carolina, Chapel Hill, NC,
2 Toxicogenomics Informatics Project, National Institute of Biomedical Innovation,
Osaka, Japan, 3 National Institute of Health Sciences, Tokyo, Japan, 4 Doshisha
Women’s College of Liberal Arts, Kyoto, Japan and 5 Eshelman School of Pharmacy,
University of North Carolina, Chapel Hill, NC.
Quantitative Structure-Activity Relationship (QSAR) modeling and toxicogenomics
have been extensively explored independently as predictive toxicology tools.
In this study, we employed a rigorous QSAR modeling workflow to predict the he-
patotoxicity for 127 well-known drugs in the Toxicogenomics Informatics Project
using both approaches. Predictor variables included chemical descriptors as well as
biological descriptors represented by the short-term gene expression profile for rat
liver (24h after dosing). The prediction target, hepatotoxicity, was defined by
histopathology and serum chemistry results after 28 days of treatment. We first developed
conventional QSAR models using a comprehensive set of chemical descriptors
and several classification methods. Using only chemical descriptors, external
predictivity expressed as Correct Classification Rate (CCR) and evaluated by
5-fold external cross-validation was as high as 59%. In parallel, models were built
using only gene expression data. Optimal models comprising of 85 genes had CCR
of 76%. Finally, hybrid models combining both chemical descriptors and gene expression
data were also developed but their predictivity was under 76%. Although
the accuracy of the QSAR models involving only chemical descriptors was limited
due to the high chemical diversity of the small dataset, using both chemical and biological
descriptors enriched the interpretation of the modeling results: the 85 predictor
genes were mechanistically relevant, involving hepatotoxicity-related pathways,
and structural alerts were identified among hepatotoxicants. This study shows
that concurrent exploration of chemical features and early treatment-induced
changes in gene expression affords predictive and interpretable models of subacute
hepatotoxicity.
140 CLASSIFICATION OF HUMAN CYTOCHROME 3A4
LIGANDS BY MEANS OF MOLECULAR DOCKING.
Y. Tie 1 , B. T. McPhail 1 , H. Hong 2 , R. D. Beger 2 , B. A. Fowler 1 and E.
Demchuk 1 . 1 Division of Toxicology & Environmental Medicine, ATSDR/CDC,
Atlanta, GA and 2 Division of Systems Biology, NCTR/U.S. FDA, Jefferson, AR.
An interagency collaboration of ATSDR/CDC and NCTR/FDA was developed to
model adverse drug-drug interactions mediated by CYP3A4 and 2D6. The modeling
of CYP3A4 was carried out using several computational methods, including
molecular docking. Molecular docking is a technique frequently used to estimate
the binding interactions between ligands and a protein. Initially, 121 drugs were
classified as strong and weak binders of CYP3A4. Docking studies were then performed
by eHiTS, FRED, Glide, and SYBYL. Docking scores were used to develop
multiple logistic regression (LR) models. The model with the least number of parameters
(specificity of 100%, sensitivity of 42.4%, and overall accuracy of 84.3%
at a LR probability cutoff of 0.5) was selected for final modeling and cross-validation
(CV). A 10-fold CV was used. The average accuracy at the CV stage was
83.8% and 79.3% for the training and CV sets, respectively. These results were
compared to those of two other models: SDAR (spectrum-data-activity relationship)
and a DF-SAR (decision forest structure-activity relationship). Although the
accuracy of classification was comparable among the modeling methods, the number
of parameters, and consequently the robustness of the models, was different.
While only 7 parameters were used in the LR docking, 27 descriptors were used in
SDAR, and even a greater number in the DF-SAR: 5 trees ranging from 5 to 17 descriptors.
All three models were used to predict which ligands in a test set of 120
were strong or weak binders. In summary, molecular docking provides a unique
way of CYP3A4 ligand classification, which is complementary to traditional
S(D)AR modeling. While S(D)AR techniques are focused at intra-molecular properties,
molecular docking takes account inter-molecular interactions. The three
modeling methods combined provide an a priori assessment of hepatic drug-drug
interactions and could serve as an adjunct to the conventional FDA drug approval
process and to the public health review process of chemical safety by ATSDR.
141 BUILDING STRUCTURE FEATURE-BASED MODELS
FOR PREDICTING ISOFORM-SPECIFIC HUMAN
CYTOCHROME P450 (HCYP 3A4, 2D6 AND 2C9)
INHIBITION ASSAY RESULTS IN TOXCAST.
P. Volarath 1 , B. Bienfait 2 , J. Gasteiger 2 , K. Houck 1 and A. Richard 1 . 1 NCCT,
U.S. EPA, Research Triangle Park, NC and 2 Molecular Networks GmbH, Erlangen,
Germany.
EPA’s ToxCast project is using high-throughput screening (HTS) to profile and prioritize
chemicals for further testing. ToxCast Phase I evaluated 309 unique chemicals,
the majority pesticide actives, in over 500 HTS assays. These included 3
human cytochrome P450 (hCYP3A4, hCYP2D6, hCYP2C9) inhibition assays
that are routinely used to evaluate drug candidates for potential drug-drug interactions.
isoCYP is a structure-activity relationship (SAR) prediction model built on a
training set of hCYP isoform-specific metabolism data that uses a decision-tree approach
to predict which of the these 3 CYP isoforms is preferentially involved in
drug metabolism. The published isoCYP models were derived from a training set of
146 drugs for which the particular single metabolizing CYP isoform is known in
each case (Terfloth et al., 2007, J.Chem. Inf. Model., 47:1688-1701). ToxCast results
for the 3 hCYP inhibition assays differ significantly from the isoCYP model
SOT 2011 ANNUAL MEETING 29