Profile-QSAR (pQSAR) is a 2D screening method that allows obtaining unprecedentedly accurate predictions of IC50 values for targets of interest covering about 50% of the estimated druggable genome. Calculating predictions for compounds is a two-step process. The first step is to calculate IC50 values using previously built Random Forest (RF) models for each assay/target independently of all other assays/targets. We call these values our basis set. During the second step, the basis set for each compound is used in a linear equation to obtain predicted IC50 for the desired assay/target. The linear equation is generated by Partial Least Squares (PLS) during model training. In order to better support drug discovery projects, we wanted to evaluate how frequently we need to update pQSAR models. For virtual screening of the historical archive that is usually performed at the lead generation phase, we developed a unique “novelty” test set to account for these models’ wide domain of application. For these archive screens, models did not need frequent retraining. However, in prospective applications, normally applicable in the lead optimization stage, totally novel compounds that lay outside of the scope of the models’ domain could be made. A chronological held-out test, developed to evaluate models’ degradation over time, indicated a need for frequent model retraining. Calculating RF predictions using KNIME was always the rate limiting step in obtaining pQSAR predictions for millions of compounds in the company database. In order to speed up virtual screening of the company library we pre-calculated our basis set and stored it in flat files. The process took months, while the rest of the steps could be completed in less than 3 hours. The pQSAR workflow was re-engineered with a focus on distributed computing to eliminate the bottleneck basis set calculations. This allowed us to speed up predictions for novel compounds and to update models more often. It also enabled predictions on large virtual combinatorial libraries.

Predictions are tough, especially about the future: Assessing model quality, need for rapid automated model updates, and distributed computing

253rd National Meeting (2017)

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