D2Screen: Embedding Pretrained Representation Learning Model and Molecular Docking for Virtual Screening

Virtual screening and molecular docking are widely used computational strategies for drug discovery. Recently, deep learning has emerged as a powerful approach that leverages known pharmacological data to identify compounds with similar bioactive profiles. Here, we developed a new end-to-end pipeline called "D2Screen" (Deep learning and Docking-based Screening) that combines both deep learning and molecular docking. D2Screen exhibits higher accuracy compared with individual deep learning or molecular docking prediction with improved BedROC and EF1% metrics. In a case study on developing antidrug resistant inhibitors against SARS-CoV-2 M^pro, we employed D2Screen to successfully discover a group of noncovalent inhibitors with the most potent compound having the IC₅₀ of 5.9 μM. This inhibitor showed significantly less susceptibility to known drug-resistant M^pro mutations T21I-E166V with a 7.6-fold reduction versus an over 1000-fold reduction of the clinically used drug Nirmatrelvir for COVID-19 treatment.