Protein language model signal, not structure, drives cryptic pocket prediction
- Architecture and PocketMiner's small MD-derived training set were held fixed while geometry-only, ESM2-embedding-only and fused inputs were compared.
- The comparison was repeated on CryptoBench and across different pocket-formation mechanisms.
- Geometry alone performed at chance, whereas ESM2-based variants did far better with no distinguishable difference between them, suggesting language-model signal accounts for nearly all performance when data are limited.
💡 Points Based on the abstract
Geometry scoring at chance on a pocket task is suspicious, not just informative: with a small MD-derived training set, I'd check for sequence-level leakage or label artifacts before concluding structure is useless. ESM2 may simply be memorizing homologs. If so, test on sequence-clustered splits for your own targets before trusting it. Also, this is an unreviewed preprint.