The Limited Contribution of Structural Information Relative to Language-Model Signal in Data-Limited Cryptic Pocket Prediction
Year: 2026
Type: preprint
Abstract: Cryptic pockets, drug-binding sites absent from a protein’s static structure but exposed through conformational change, expand the scope of druggable targets, yet remain difficult to identify computationally. PocketMiner, a structure-based graph neural network trained on molecular dynamics simulation data, predicts cryptic pocket formation from a single static structure (ROC-AUC 0.87). Recent protein language model (PLM)-based approaches report further improvements over this structure-only baseline (up to 0.93), but do so by simultaneously replacing the structural architecture with a sequence-only model and switching to substantially larger, differently-sourced training data–leaving it unclear whether these gains stem from the languagemodel signal itself or from increased training data scale. We isolate this variable directly: holding architecture family and PocketMiner’s original, small molecular-dynamics-derived training set fixed, we compare three matched variants–geometry-only (baseline), ESM2 embeddings only, and both combined (fused)–under identical training conditions. The geometry-only baseline achieves chance-level test performance (ROC-AUC 0.498), while both the ESM2-only (0.817) and fused (0.835) variants perform far above it, with statistically indistinguishable confidence intervals between the two, indicating that language-model signal alone accounts for nearly all of the achievable performance, and that structural information contributes marginally under these conditions. This result is confirmed on CryptoBench, an independent, larger benchmark (baseline 0.494 vs. fused 0.716, non-overlapping confidence intervals), and holds consistently across every pocket-formation mechanism examined in a stratified analysis. These findings suggest that a substantial share of the improvement reported by prior sequence-based methods is attributable to language-model signal itself rather than training data scale, with implications for data-efficient strategies in structure-based pocket prediction more broadly.
Source: ChemRxiv
Authors: Emmanuel Brown, Phenil Shah
Language: English
Subfield: Computational Theory and Mathematics
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7172327998