Conventional single-target regression models for virtual screening (VS) appear to have excellent predictive power when tested on random held-out test sets. However, project teams intentionally order compounds highly dissimilar to the training sets. When tested on realistically novel compounds, single-target models are unable to predict IC50. Deep learning “Profile-QSAR 1.0”, developed several years ago to expand the domain of application, trained models for individual kinases on 5 million IC50s covering 500K compounds from 728 historical kinase assays. Accuracy for novel compounds was improved, and it was widely applied in VS, but the models were still not accurate enough for broad selectivity calculations across large targets collections. “Profile-QSAR 2.0” has greatly improved accuracy, now comparable to high-throughput 4-concentation IC50 assays even for archive compounds with almost no similarity to the training set. This unprecedented accuracy allows unparalleled estimates of potency, ligand efficiency, lipid efficiency, and selectivity for very sophisticated VS. It also now enables broad profiling applications such as virtual toxicity-panels, mode-of-action discovery, and promiscuity assessment. The method has been extended to many protein families, as well as many idiosyncratic targets outside those protein families. Key points: ● Random held-out test sets do not predict real-life VS performance. ● Single-target models fail on realistically novel compound sets. ● “Profile-QSAR 2.0” kinase models are comparable to 4-pt experimental IC50 even for novel compounds. ● Accurate IC50 predictions enable very sophisticated virtual screens. ● Extended to GPCRs, proteases, cyp 450s, and many non-kinase adenosine binding proteins. ● Beyond VS and hit-list triaging, applications now included selectivity prediction, mode-of-action discovery, safety profiling, promiscuity assessment, etc.

Deep learning profile-QSAR 2.0 IC50 predictions as accurate as 4-pt IC50s: Applications to polypharmacology

253rd National Meeting (2017)

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