More protein-ligand data are needed for AlphaFold-like models to enable drug discovery

Structure-based drug design (SBDD) is an evolving paradigm that leverages protein structural information to improve small molecule therapeutic design. Building on more than 50 years of data curation from the Protein Data Bank, the recent emergence of protein structure prediction models (PSPMs) promises to enable new computationally driven approaches for therapeutic discovery. However, it is critical to assess the limitations of these models using blind challenges and to expand existing datasets to better reflect real-world drug design tasks. Here, we discuss recent efforts to benchmark existing PSPMs and identify their limitations. We offer a hierarchical framework for parsing which tasks the current models perform well, and which tasks remain challenging or unexplored. Finally, we emphasize the need for systematic dataset generation to support the development of frontier models and highlight recent efforts to generate experimental and physics-based datasets for challenging tasks in drug discovery.