No pocket-prediction head needed: cofolding one protein against 16 ligands recovered 176 of 255 annotated sites
Synopsis
The Rowan team tried pocket detection without dedicated training: the same protein was repeatedly cofolded with 16 chemically varied ligands using Boltz-2, and the predicted ligand placements were aligned to a common reference and merged into pocket suggestions. Across 210 PocketWatch proteins and 255 annotated sites, the paper's N+2 metric recovered 176 sites (69.0%), above 138 (54.1%) for standalone zPocket and 158 (62.0%) for the P2Rank+zPocket hybrid. The experiments do not separate the benefit of probe diversity from repeated cofolding itself, and when only the top-ranked suggestion counts, the methods perform similarly.
Interpretation
Pocket detection can proceed without a dedicated pocket-prediction head: cofolding one protein against many ligands and merging the predicted ligand placements yields usable binding-site suggestions. The earlier zPocket preprint from Genesis Molecular AI trained a set of pocket-prediction heads on frozen cofolding features, on the premise that cofolding embeddings already encode binding-site information; that route requires an extra training step. Across 210 PocketWatch proteins and 255 annotated sites under the N+2 metric, the cofolding probes recovered 176 sites (69.0%), versus 138 (54.1%) for standalone zPocket and 158 (62.0%) for the P2Rank+zPocket hybrid.
The method treats ligands as probes: the protein sequence and source MSA stay fixed while 16 ligands spanning different molecular shapes and functional groups are cofolded separately, so the model exposes different accommodation preferences under different ligands. A cofolding model such as Boltz-2 proposes a joint protein–ligand structure even when the ligand may not actually bind; for pocket detection those placements may still carry information about where the protein accommodates binders. The 16 probes were selected by GPT-6 Astra for chemical variety and decent pocket-discovery performance on a subset of PocketWatch proteins; the runs used no target-specific ligand selection, pocket-location hints or experimental templates.
Each prediction is locally aligned on protein residues within 5 Å of ligand heavy atoms to move the probe into a common reference frame, and each suggestion is centered on ligand heavy atoms within 4.5 Å of the protein. Fitting only residues near the ligand keeps distant domain movements from dominating alignment error, and placements with local alignment RMSD above 5 Å are discarded. Remaining suggestions are ranked by the product of ligand ipTM and mean ligand-heavy-atom pLDDT, and suggestions within 4 Å of an already retained higher-ranked center are removed; on the validation set, the anchor and RMSD checks rejected 272 of 2,240 placements.
Perspective
The evaluation covers 210 PocketWatch proteins and 255 annotated sites, 70 from the development stage and 140 added later, with Appendix A reporting the two groups separately; all methods use the same ESMFold2 predicted structure used by zPocket and P2Rank, and no method used an experimental structure as its prediction reference.
The experiments do not isolate the benefit of probe diversity from that of repeated cofolding, and the authors suggest comparing the 16-probe panel with 16 runs of a single probe using different random seeds; PocketWatch labels derive from experimental structures but are not exhaustive, so a prediction that misses them is not necessarily poor, and the labels are mapped onto the predicted structure by local alignment, leaving location uncertainty when local shapes differ.
