Machine learning classification can reduce false positives in structure-based virtual screening

Abstract

With the recent explosion in the size of libraries available for screening, virtual screening is positioned to assume a more prominent role in early drug discoverys search for active chemical matter. Modern virtual screening methods are still, however, plagued with high false positive rates: typically, only about 12% of the top-scoring compounds actually show activity when tested in biochemical assays. We argue that most scoring functions used for this task have been developed with insufficient thoughtfulness into the datasets on which they are trained and tested, leading to overly simplistic models and/or overtraining. These problems are compounded in the literature because none of the studies reporting new scoring methods have validated their model prospectively within the same study. Here, we report a new strategy for building a training dataset (D-COID) that aims to generate highly-compelling decoy complexes that are individually matched to available active complexes. Using this dataset, we train a general-purpose classifier for virtual screening (vScreenML) that is built on the XGBoost framework of gradient-boosted decision trees. In retrospective benchmarks, our new classifier shows outstanding performance relative to other scoring functions. We additionally evaluate the classifier in a prospective context, by screening for new acetylcholinesterase inhibitors. Remarkably, we find that nearly all compounds selected by vScreenML show detectable activity at 50 {micro}M, with 10 of 23 providing greater than 50% inhibition at this concentration. Without any medicinal chemistry optimization, the most potent hit from this initial screen has an IC50 of 280 nM, corresponding to a Ki value of 173 nM. These results support using the D-COID strategy for training classifiers in other computational biology tasks, and for vScreenML in virtual screening campaigns against other protein targets. Both D-COID and vScreenML are freely distributed to facilitate such efforts.