Proceedings - Interdisciplinary Center for Nanotoxicity

136
Conference on Current Trends in Computational Chemistry 2009
Predictive QSAR modeling of animal toxicity endpoints using
a combination of chemical and biological descriptors of
molecules
Alexander Tropsha
Laboratory for Molecular Modeling, Carolina Center for Computational Toxicology
and Center for Environmental Bioinformatics,
University of North Carolina at Chapel Hill, Chapel Hill, NC U.S.A.
Modern “‐omics” technologies produce a wealth of biological assay data for large
numbers of chemicals with known in vivo toxicity profiles. For instance, to develop effective
means for rapid toxicity evaluation of environmental chemicals, the Tox21 partnership between
the National Toxicology Program (NTP), NIH Chemical Genomics Center, and the National
Center for Computational Toxicology at the US EPA are conducting quantitative high‐
throughput screening studies with thousands of chemicals. Concurrently, the ToxCast program
established at the US EPA is addressing the Tox21 goals by using ca. 600 in vitro assays to create
bioactivity profiles for 100s of compounds with known in vivo toxicities measured in ca. 70
animal assays. The analysis of this data requires new computational approaches to link
chemical structures, in vitro responses and in vivo toxicity effects.
We have developed a general purpose predictive QSAR modeling workflow that relies on
effective statistical model validation routines. Traditionally, we have used this workflow for
conventional QSAR modeling studies of in vivo and points utilizing chemical descriptors of
molecules. Recently, we have started to employ both chemical and biological (i.e., in vitro assay
results) descriptors of molecules to develop in vivo chemical toxicity models. We have
developed two distinct methodologies for in vivo toxicity prediction utilizing both chemical and
biological descriptors. In the first approach, we employ biological descriptors in combination
with chemical descriptors to build models. Obviously, this approach requires the knowledge of
biological descriptors to make toxicity assessment for new compounds. Our second modeling
approach employs the explicit relationships between in vitro and in vivo data as part of the two‐
step hierarchical modeling strategy. In the first step, all compounds are partitioned into classes
defined by patterns of in vitro – in vivo relationships (e.g., compounds designated as “active” in
both in vitro and in vivo assays form one class whereas those that have these designations
disagree for in vitro and in vivo assays, form another class). Then, a conventional classification
QSAR model using chemical descriptors only is built to discriminate compounds that belong to
different classes defined by the in vitro – in vivo correlation patterns. In the second step, the
class‐specific conventional QSAR models are built using chemical descriptors only. Thus, this
hierarchical strategy affords external predictions using chemical descriptors only.
We will present the results of applying the above QSAR modeling strategies to several
datasets including ToxCast Phase I data; the ZEBET dataset with in vitro IC50 cytotoxicity values
and in vivo rodent LD50 values for more than 300 chemicals; and others. All studies suggest
that utilizing in vitro assay results as biological descriptors afford prediction accuracy that is
superior to both the conventional QSAR modeling that utilizes chemical descriptors only and
the in vivo effect classifiers based on in vitro biological descriptors only. We will discuss how
our models are being used to prioritize compound selection for the ToxCast Phase II studies.