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 discovery’s search for active chemical matter. In typical virtual screens, however, 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 studies reporting new scoring methods have not validated their models prospectively within the same study. Here, we report a 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. In retrospective benchmarks, our classifier shows outstanding performance relative to other scoring functions. In a prospective context, nearly all candidate inhibitors from a screen against acetylcholinesterase show detectable activity; beyond this, 10 of 23 compounds have IC50 better than 50 μM. Without any medicinal chemistry optimization, the most potent hit has IC50 280 nM, corresponding to Ki 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.

Selecting machine-learning scoring functions for structure-based virtual screening

“…This shows again that optimal models can be anticipated without any need for PM decoys. Table 1 compiles several freely-available off-the-shelf generic ML-based SFs that could be used on the test set: NNScore [44,45], Δ vina RF 20 [46], Δ vina XGB [47], Convolutional NN [26], RF-Score-VS [27] or vScreenML [48]. For any employed SF, it is necessary to find out which protein-ligand complexes where used for training it and remove them from the test set if also found there.…”

“…In addition to evaluating the SFs on the entire test set, their performance on certain test subsets would be very informative. For instance, on the subset of complexes bound by molecules dissimilar to any training set molecule, as SFs performing well here should discover a higher proportion of novel compounds [5,8,48]. Another example is the subset of complexes with most potent actives/binders (e.g.…”

Nanoprinted high-neuron-density optical linear perceptrons performing near-infrared inference on a CMOS chip

“…As a machine-learning-based classification device, the decryptors presented in this work will always show a certain false match rate—a challenge inherent to the field of machine learning classification 38 – 40 . A number of techniques have therefore been developed to decrease the false match rate in a given classification setting, which can be equally applied to the decryptors presented in this work.…”

Challenges and advances in clinical applications of mesenchymal stromal cells

“…AI can find new molecular compounds and emerging drug targets much faster than traditional methods, thus speeding up the progress of drug development [ 184 , 185 ]. At the same time, AI can more accurately predict the follow-up experimental results of new drugs, so as to improve the accuracy at each stage of drug development [ 186 ]. Computer-aided drug design techniques are thus revolutionizing MSCs therapies.…”