Our goal is to build virtual screening (VS) models with accuracy comparable to expensive and time consuming experimental high-throughput 4-point IC50 screens. Although apparently excellent when tested on random held-out sets, when tested on realistically novel compound sets, the median correlation for conventional random forest regression models was only R2=0.12 for our 728 kinase assays, and only 20% reached R2=0.3, our recommended minimum for virtual screening. “Profile-QSAR” (Gen 1) deep learning models were developed several years ago to expand chemical coverage. They train kinase family-based models for individual kinases using 5 million IC50s covering 500K compounds from the 728 historical kinase assays. On realistically novel test sets, the median correlation was 0.24 and 40% of assays reached 0.3, but still fell short of our goal. Our recently improved “Gen 2” Profile-QSAR algorithm now succeeds on 80% of the 728 kinase assays with median correlation of R2=0.55, as accurate as a 4-point medium-throughput IC50 assay. This accuracy allows good estimates of potency, ligand efficiency, lipid efficiency, and selectivity, which allows for selection of very sophisticated screening sets. The method has also been applied to additional protein families covering about half of the estimated druggable genome. Key points: ● Random held-out test sets do not predict real-life VS performance. ● Conventional individual-target models fail on realistically novel compounds. ● “Gen 2” Kinase-family Profile-QSAR models are comparable to 4-pt experimental IC50 for novel compounds. ● IC50 prediction allows for very sophisticated screening set design. ● Also for GPCRs, proteases, cyp 450s, and many non-kinase adenosine binding proteins. ● Applications have included virtual screening, hit-list triaging, mode-of-action, safety profiling, HTS “rescue”, complex HTS ... ● Success with ChEMBL assays

Profile-QSAR Gen 2: Deep learning kinase IC50 predictions for novel compounds as accurate as 4-pt IC50s

252nd National Meeting (2016)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.