Using Molecular Features of Xenobiotics to Predict Hepatic Gene Expression Response

Despite recent advances in molecular
medicine and rational drug
design, many drugs still fail because toxic effects arise at the cellular
and tissue level. In order to better understand these effects, cellular
assays can generate high-throughput measurements of gene expression
changes induced by small molecules. However, our understanding of
how the chemical features of small molecules influence gene expression
is very limited. Therefore, we investigated the extent to which chemical
features of small molecules can reliably be associated with significant
changes in gene expression. Specifically, we analyzed the gene expression
response of rat liver cells to 170 different drugs and searched for
genes whose expression could be related to chemical features alone.
Surprisingly, we can predict the up-regulation of 87 genes (increased
expression of at least 1.5 times compared to controls). We show an
average cross-validation predictive area under the receiver operating
characteristic curve (AUROC) of 0.7 or greater for each of these 87
genes. We applied our method to an external data set of rat liver
gene expression response to a novel drug and achieved an AUROC of
0.7. We also validated our approach by predicting up-regulation of
Cytochrome P450 1A2 (CYP1A2) in three drugs known to induce CYP1A2
that were not in our data set. Finally, a detailed analysis of the
CYP1A2 predictor allowed us to identify which fragments made significant
contributions to the predictive scores.