Protein ligand structure prediction: From empirical to deep learning approaches

Protein-ligand structure prediction methods, aiming to predict the three-dimensional complex structure and binding energy of a compound and target protein, are essential in many structure-based drug discovery pipelines, including virtual screening and lead optimization. Traditional empirical approaches use explicit scoring functions and conformational search techniques to predict protein-ligand structures and binding affinities. With the recent advent of deep learning (DL) methods, DL-based models learn both the scoring function and conformational sampling by approximating the underlying data distribution from training data. In this review, we first discuss the key components of both empirical and DL-based structure prediction methods to provide a unified view. We categorize these computational methods into two main groups based on whether a template protein structure is required, and briefly overview the important methods in each category. Finally, we discuss the major challenges and opportunities, focusing on the future development of DL-based approaches.