Evaluating generalization in protein-ligand cofolding methods

Deep learning has driven major breakthroughs in protein structure prediction; however, one of the next critical steps forward is accurately predicting how proteins interact with small-molecule ligands, to enable real-world applications such as drug discovery. Recent cofolding methods aim to address this challenge, but evaluating their performance has been inconclusive because of the lack of relevant benchmarking datasets. Here we present a comprehensive evaluation of four leading all-atom cofolding methods using our newly introduced benchmark dataset, Runs N' Poses. Runs N' Poses comprises 2,600 high-resolution protein-ligand systems released after the training cutoff used by these methods. We demonstrate that current cofolding approaches largely memorize ligand poses from their training data, hindering their use for de novo drug design. With this assessment and benchmark dataset, we aim to accelerate progress in the field by allowing for a more realistic assessment of the current state-of-the-art deep learning methods for predicting protein-ligand interactions.