zPocket: Finding Pockets with Frozen Cofolding Features

Identifying protein-ligand binding sites (pockets) is important for understanding protein function and developing drugs. However, no public method has leveraged modern cofolding models for novel pocket prediction; furthermore, this task lacks a suitable benchmark as existing benchmarks are within the training distribution of modern cofolding models. We therefore developed PocketWatch, a curated dataset of biologically relevant ligand pockets, whose test set consists of unseen pockets stratified by sequence similarity, structural similarity, and crypticity to enable evaluation of generalization to novel and challenging pockets. Using this dataset, we develop zPocket, a method that trains pocket-prediction heads on top of frozen features from cofolding models. We show that zPocket heads outperform the competitive baseline P2Rank on our test set across every stratification. Our best zPocket model achieves overall per-pocket AUPRC of 0.745 compared to P2Rank's 0.495. We further show that zPocket can be used to outperform P2Rank on top-N+2 recall of predicted pocket centroids through clustering per-residue predictions or using zPocket to rescore centroids proposed by other models. Together these results suggest that zPocket heads built on cofolding trunk features can greatly advance ligand binding site prediction.