Recent advances in machine learning predictions of protein-ligand binding affinities

Accurately predicting protein-ligand binding affinities is a central task in rational drug design, as it directly influences hit discovery, lead optimization, and compound prioritization. Traditional approaches often suffer from limited accuracy, high computational cost, or dependence on heuristic scoring functions. Recent advances in machine learning (ML) have introduced new paradigms for the binding affinity prediction. In this review, we survey the latest developments in ML-based predictions of protein-ligand binding affinities across various directions, including structure-based approaches that leverage three-dimensional conformational data, ligand-based models that utilize mathematical approaches that employ topological invariants, and hybrid or alternative frameworks addressing diverse prediction scenarios. We highlight representative algorithms ranging from traditional supervised learning to deep learning architectures. Additionally, we discuss the current challenges faced in this domain. Finally, we outline emerging trends and potential future directions, which are poised to further enhance the accuracy and applicability of binding affinity prediction in drug discovery pipelines.