AI-guided competitive docking for virtual screening and compound efficacy prediction

Abstract

Machine learning has transformed how we predict protein structures and interactions, but its full potential in drug discovery is only beginning to be realized. In this study, we demonstrate that advanced deep learning tools --such as AlphaFold3 and Boltz-1/2-- not only predict protein-ligand interactions with high accuracy but can also separate active drug compounds from inactive ones. We present a straightforward strategy called pairwise competitive docking, which ranks drug candidates by directly comparing how well they bind to a proteins target site. When applied to both bacterial and human enzymes, this method produced rankings that closely matched experimental results. We further show how this approach can guide the design of improved antibiotics and speed up the discovery of promising drug candidates from large chemical libraries. Overall, our findings highlight how machine learning can make structure-based drug design faster, more reliable, and more cost-effective.