Profile-QSAR achieves unprecedented accuracy and broad domain of application by augmenting a few hundred IC50s for a target of interest with millions of additional IC50s from 100,000s of compounds tested across many hundreds of historical assays from the same protein family. The accuracy and domain of application have now been dramatically further improved by replacing the original Bayesian formulation with a random forest formulation. Highlights of this presentation include - A new random forest-based algorithm yields unprecedented accuracy and domain of application, - A new challenging “Novelty” test set for evaluating virtual screening models, which mirrors the great diversity of actual historical compound selections, - A demonstration that linear PLS models can extrapolate to novel chemistry better than non-linear random forests, - A surprising demonstration that excluding the seemingly most relevant training data can prevent overfitting and greatly improve extrapolation, and - A comparison and analysis of models built on internal and public data sets.

Profile-QSAR generation 2: Perfection, the enemy of the good?

251st National Meeting (2016)

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