Profile-QSAR (pQSAR) is a massively-multitask, 2-level, stacked model trained on 13,000 bioactivity and ADMET assays. For level 1, 13,000 conventional single-assay random-forest regression (RFR) models are trained on the available pIC50s for each individual assay. Compounds are described by chemical substructure fingerprints. For level-2, new, separate, PLS regression models are trained for each assay, now using the profile of predicted pIC50s from level-1 models as compound descriptors. The improvement over the single-assay models is dramatic. 9,000 (72%) of the assays give successful pQSAR models vs. 8% of single-assay RFR models. The median correlation with experiment for pQSAR models is r2=0.53, comparable to 4-concentration IC50 experiments, vs. only r2=0.05 for the single-assay models. Every month, we retrain the 13,000 pQSAR models, calculating and storing the 66 billion predicted pIC50s for the 5.5 million Novartis compounds. pQSAR has contributed to hundreds of projects over the last 15 years in many capacities: virtual screening, hit-list triaging, hit expansion, MoA prediction, Off-target safety profiling, polyphamacology, virtual counter-screens, biological profile-based scaffold hopping, etc. This presentation will describe 4 recent extension of this already very mature method: * Federated pQSAR for safe collaboration among competing companies * Conformal uncertainty intervals * pQSAR in generative chemistry * Prediction of gene expression profiles

Recent extensions of Profile-QSAR: beyond successful IC50 models for 9,000 assays

Fall 2022 abstracts, posters, and presentations will launch in mid-September.

ACS Fall 2022

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