Addressing the Bottlenecks in Machine Learning and Physics-based Approaches to Structure-based Virtual Screening

Researchers have long sought to harness the power of computers to assist in drug discovery. Structure-based virtual screening (SBVS) uses the 3D structure of a target binding site to quickly score protein-ligand binding affinities, enabling the screening of large chemical libraries. Although this technique has been used for decades, its success has mostly been limited to targets with many known binders. The two predominant SBVS paradigms, machine learning (ML) and physics-based approaches, both face significant challenges with novel targets. ML models struggle to generalize beyond their training data, leading to unreliable predictions for unseen targets. Physics-based methods, while more theoretically grounded, suffer from an inherent tradeoff between speed and accuracy; absolute binding free energy (ABFE) calculations, the most reliable approach, remain computationally prohibitive for large-scale screening. I have addressed these limitations by improving both ML and physics-based approaches. I first enhanced the generalization capabilities of ML models by incorporating a broader range of training data. I additionally developed a novel algorithm for proposing 3D ligand binding poses using a physics-inspired ML approach. I also developed a new metric for assessing the performance of virtual screening models in more realistic scenarios. I also accelerated ABFE calculations through a novel application of spatial data structures, improving efficiency without sacrificing accuracy. Finally, I developed a hybrid algorithm that integrates ML and physics-based methodologies, leveraging the strengths of both to improve SBVS performance on novel targets.