How to Use ESMC Feature Interpretation

ESMC Feature Interpretation exposes residue-level patterns represented inside Biohub's ESMC-6B protein language model. A sparse autoencoder transforms the model's dense internal representation into a codebook of 16,384 learned features, with at most 64 active features per residue. Individual features may correspond to motifs, structural regions, binding-site patterns, biophysical properties, or protein-family signals.

On Neurosnap, researchers submit one amino-acid sequence and receive ranked feature summaries together with sparse residue-level activation profiles. Interpretations come from Biohub's separately published, MIT-licensed feature table, including summaries, categories, activation patterns, exemplar families, and reliability thresholds. These descriptions remain model-derived hypotheses rather than experimentally established annotations, so strong activations should be compared with curated databases, known motifs, structures, or experiments.

The service runs ESMC-6B and collects its hidden representation at transformer layer 60. A Top-K sparse autoencoder expands each residue into a 16,384-feature codebook while retaining at most 64 active features. Maximum Activation identifies features with a strong localized response, while Prevalence reports how many residues exceed the selected display threshold and can highlight broader domain-like patterns.

Normalize Features applies Biohub's feature-wise UniRef90 maximum and inverse-document-frequency statistics. This down-weights ubiquitous features and emphasizes rarer signals, making features easier to compare within an exploratory analysis. Number of Features limits the ranked result set, and Activation Threshold removes weak residue-level responses from prevalence calculations and exported activation rows. Biohub's Reliability Threshold is different: it is defined on the raw activation and indicates when the generated biological description is considered reliable. Neurosnap reports both the normalized and raw maximum so these thresholds are not conflated.

A feature ID is meaningful only for the exact SAE model and revision reported with the job. Descriptions are automatically generated hypotheses based on activations in SwissProt. They should guide follow-up analysis, not replace experimental or curated functional annotation.

• Residue-level interpretation: Sparse activations show where learned ESMC patterns occur along the submitted sequence.
• Localized and broad signals: Maximum activation highlights sharp motif-like responses, while prevalence helps identify features spanning larger regions.
• Reproducible model identity: Results record the exact backbone, SAE, normalization setting, threshold, and model revisions used for analysis.