Assessing the Robustness and Scalability of Machine Learning Methods to Accelerate Ultralarge High-Throughput Docking Campaigns

Structure-based virtual screening methods are, nowadays, one of the key pillars of computational drug discovery. In recent years, high-throughput docking campaigns aided by machine learning (ML)-based protocols have emerged as a way to accelerate the identification of top-scoring molecules within ultralarge chemical molecule libraries. However, studies validating these ML approaches used one or two targets and/or small molecule libraries. Herein, we extended the validation of ML protocols at retrieving virtual hits in an accelerated fashion by using two standard publicly available ∼100M molecule libraries and also a comprehensive benchmark set involving molecular docking scores of a 10M molecule library in 10 diverse protein targets with two docking programs, PLANTS and AutoDock Vina. In the 10M benchmark set, we have shown that, on average, more than 60 and 70% of the top 10k and top 1k molecules, respectively, can be retrieved while reducing the number of docking evaluations by more than 97%, indicating a robust performance of the ML protocol. With larger molecule libraries, we have shown that a proportional increase in the training set size enhances the performance of the ML model at retrieving virtual hits. In summary, our results support the use of ML methods to retrieve top-scoring molecules for chemical libraries containing hundreds of millions or even billions of molecules, where the role of ML models becomes even more critical as brute-force exploration of such chemical libraries through molecular docking is inaccessible in reasonable time frames.