(140f) Leveraging Deep Learning for Pharmaceutical Discovery Lead Profiling

As part of small molecule drug discovery, potential lead compounds are studied to ensure that they have a low risk of adverse behavior. Cytochrome P450s (CYP) are a family of enzymes that are an important drug metabolism pathway; drug candidates that inhibit the activity of these enzymes can lead to dangerous drug-drug interactions. Once synthesized, candidate compounds are submitted to high throughput in vitro assays to screen for CYP inhibition. By prioritizing more promising options and measuring the effect of changes to the core structure, the ability to virtually screen compounds for CYP inhibition would save teams time and effort during drug design.

This work describes the development of a model that can predict CYP inhibition from chemical structure using in-house data from BMS. Such quantitative structure activity relationship (QSAR) models have been successfully developed from publicly available data sets, however, the limited amount of public CYP inhibition data hinders their practical utility. As part of an innovative multidisciplinary team at BMS, analytics experts from different functional areas contributed to exploring and analyzing the BMS CYP assay database. Using this much larger set of private data (>1M entries) combined with recent advancements in machine learning algorithms, a model suitable for virtual screening was developed. Of the many modeling approaches that were explored, >85% classification accuracy was achieved using a deep learning approach. Important caveats for handling real-world data as well as a comparison of different modeling methodologies will be presented.