ESM-C Mutation Scoring
v3.4.3Score amino acid substitutions with masked protein language models.
ESM-C Mutation Scoring ranks single amino acid substitutions by how strongly a protein language model prefers each replacement to the original token. It masks one position at a time, keeping the rest of the submitted sequence unchanged.
The webserver runs Biohub's MIT-licensed ESMC-300M, ESMC-600M and ESMC-6B checkpoints on ProteinIQ's compute. Every run returns complete mutation scores, residue entropy, ranked substitutions and plots. ESM-C provides embeddings, hidden states and unmasked logits.
Runs cost 23 credits per minute with ESMC-300M or ESMC-600M, and 42 credits per minute with ESMC-6B. The minimum reservation equals one minute at the selected rate, not a minimum final charge. Completed runs are charged in proportion to measured runtime, rounded up to a whole credit; unused reserved credits are returned. Set a spending limit before submission.
| Input | Accepted values and limits |
|---|---|
Protein | Raw single-letter sequences, FASTA text, or sequences fetched from RCSB. |
| Uploaded file | .fasta, .fa, .fas or .txt; up to 10 MiB per file. |
| Sequence count | Up to 50 sequences per job. |
| Sequence length | Up to 2,046 sequence tokens per sequence and 20,000 tokens across the job. |
Supported tokens are the 20 standard amino acids plus X, B, U, Z, O, . and - gaps, and | chain breaks. Whitespace is removed and letters are converted to uppercase. Positions retain ambiguous amino acids, gaps and chain breaks; unsupported characters are rejected.
| Parameter | Type | Default | Description |
|---|---|---|---|
Model variant | enum | ESMC-300M | ESMC-300M, ESMC-600M or ESMC-6B; the 6B model uses a larger GPU. |
Each sequence produces all seven files below. Filenames include the sequence index and label to distinguish repeated labels. A downloadable run.log records the outcome of the whole job. In the arrays, L is the sequence-token count.
| File | Contents |
|---|---|
*_masked_scores.npz | Full (L, 64) masked logits and LLRs, entropy, negative fractions, sequence tokens, vocabulary, positions, and model identity. |
*_positions.csv | Every position, original token, entropy in bits, and negative fraction. |
*_substitutions.csv | Complete ranked substitutions with LLRs in natural-log units. |
*_mutation_heatmap.svg | The tutorial’s mutation heatmap, including wild-type markers. |
*_entropy.svg | The tutorial’s entropy plot. |
*_negative_fraction.svg | The tutorial’s negative-fraction plot. |
*_mutation_provenance.json | Exact source, model, sequence, runtime, settings, and interpretation. |
The Mutation map, Ranked substitutions, Data, Files, and Logs views retain every sequence. Ambiguous amino acids, gaps, and chain-break tokens keep their input positions; their LLR reference is the submitted token.
For position i, ESM-C sees the original sequence with only that position masked. Scores use the model output at token i + 1, accounting for the leading start token. All other positions retain their original sequence.
ln p(substitution | masked sequence) - ln p(wild type | masked sequence). Positive values mean the model prefers the substitution to wild type; negative values mean it prefers wild type. Wild-type scores are zero. These are sequence preferences, not measured changes in activity or stability.-sum(p * log2(p + 1e-9)), in bits. Probabilities are normalized across all 64 vocabulary outputs. The 20 standard amino acids are selected only after normalization for the heatmap.The heatmap uses the Biohub tutorial’s amino-acid order: descending isoelectric point. Red indicates negative LLR, white zero, and blue positive LLR, on a symmetric scale. Click a cell or use the arrow keys to inspect its score. The entropy chart and exact selected-position values remain available below the map.
These are single substitutions in the original sequence context. Adding their scores does not model epistasis or predict interactions between mutations. No joint multi-mutant prediction is performed. Unmasked logits from ESM-C embeddings are not mutation scores.
Mutation scoring runs on ProteinIQ’s compute using Biohub’s MIT-licensed code and checkpoints. It adapts the tutorial’s API calls to the documented self-hosted model and executes the tutorial’s entropy and LLR functions. The provenance download records the source commit, checkpoint revision, precision, settings, equations, and indexing.
Analyze protein sequences and generate embeddings.
Predict non-germline residues and generate embeddings for paired or unpaired antibody sequences.
Antibody language model for scoring, classification, embeddings and masked-residue filling
Predict protein aggregation nucleation propensity from amino acid sequences.
Generate protein embeddings with ESM-C.
Predict protein solubility from sequence with feature-level and windowed profile outputs.
Identify co-evolving residue sectors in protein families using Statistical Coupling Analysis.
Restore residues and generate antibody-specific representations or likelihood scores.
Predict intrinsically disordered regions and per-residue disorder probabilities from protein sequences.
Score single and multi-site protein mutations with evolutionary and inverse-folding profiles.