Drug structure design by deep learning to avoid hepatotoxicity

Abstract

In drug discovery, it is extremely difficult to identify hit compounds. To do this efficiently, there are structure generators that automatically generate the structural formula of a compound with the desired activity. However, the previous methods can only consider one target molecule or chemical property. Recently, structure generators that utilize gene expression profiles, which can consider exhaustive cellular responses, have attracted much attention. However, none of the existing methods considers the safety of the compound. In this study, we propose a method for generating the structure of drug candidate compounds based on gene expression profiles, considering not only the desired target activity but also toxicity. We focused on hepatotoxicity. First, we extracted hepatotoxicity-specific genes by comparing the compound-responsive gene expression profiles of compounds that exhibit and do not exhibit hepatotoxicity. Next, we designed a gene expression profile that counteracts the expression pattern of genes involved in hepatotoxicity included in the drug target perturbation response gene expression profile. Finally, the designed gene expression profile was input into a structure generator. As a case study, we applied our proposed method to a kinase, a drug target for cancers, and generated the structures of new inhibitor candidate compounds. Compounds generated with consideration of toxicity were structurally similar to approved drugs, and they are structurally dissimilar to known toxic compounds. These results suggest the usefulness of the proposed method.