Who contributes to whom: following the evidence behind protein language-model predictions of backbone variability

A fine-tuned ESM3-based model, ESMfluc, predicts local backbone variability on ATLAS proteins with AUROC 0.889 and links predictions to evidence at specific sequence positions.

• Why it matters: Sequence models can forecast backbone variability, but attention weights cannot show whether sequence positions support or oppose predicted local conformations.
• What they did: ESMfluc was trained on ATLAS molecular dynamics simulations using ESM2 or ESM3 representations to classify residues sampling one versus multiple Protein Block states, then evaluated on 208 held-out proteins and 985 mdCATH domains.
• The result: On 985 mdCATH domains, mean within-domain Spearman correlation was 0.616 (vs 0.589 for PEGASUS), and evidence decomposition identified recurring positive/negative influence bands; uniform attention reduced multi-PB-supporting bands to fewer than one per protein.