Score point mutations for functional impact with a protein language model

Hand Claude a wild-type protein sequence and a list of substitutions; get back a ranked table of zero-shot fitness scores — the masked-marginal log-likelihood ratio each mutation receives from a protein language model — with no deep-mutational-scanning data and no task-specific training.

   
Problem class Data analysis
Subject areas Integrative Structural and Computational Biology, Molecular and Cellular Biology
Evidence level Proposed
Complexity One skill or MCP
Availability Fully open
Compute Workstation with GPU

Problem

You have a protein and a set of single-residue substitutions — clinical VUS to triage, an enzyme you want to engineer, a panel of designed mutants to pre-screen before ordering DNA. You want a defensible, data-free ranking of which mutations are likely tolerated and which are likely deleterious, before committing wet-lab budget. Supervised predictors need labelled deep-mutational-scanning (DMS) data you don’t have for this protein. A protein language model trained only on natural sequences solves this zero-shot: the log-probability it assigns to the mutant residue relative to the wild-type residue is a strong proxy for evolutionary tolerance, and the masked-marginal version of that score is the field-standard heuristic. Solved looks like: paste a FASTA sequence and a list like A123G, R45K, ..., get a CSV ranked by score with a plain-language tolerated/deleterious call per mutation.

  1. Install the ESM skill (K-Dense scientific-agent-skills):

    npx skills add K-Dense-AI/scientific-agent-skills

    Enable the esm skill when prompted. For the input-fetch step (optional) also enable the gget skill so Claude can pull the canonical UniProt sequence by accession instead of you pasting it.

  2. Provide the sequence and the mutation list. If you only have an accession, ask Claude to fetch the canonical sequence first (gget → UniProt), then confirm the residue numbering matches your mutation list (1-indexed, position 1 = the first residue of the canonical isoform).

  3. Prompt for masked-marginal scoring. A minimal version:

    Use the esm skill to score these point mutations on the protein
below with the masked-marginal log-likelihood-ratio scheme
(Meier et al. 2021), using ESM C (or ESM-2 if available locally).

For each mutation pos/wt/mut:
1. Run a forward pass with that position MASKED.
2. Read the per-position log-probabilities (logits -> log-softmax)
   over the 20 amino acids at the masked position.
3. Score = log P(mut) - log P(wt) at that position.
4. A more-negative score = more likely deleterious; near-zero or
   positive = likely tolerated.

WT sequence (FASTA):
>my_protein
MK...   <paste full canonical sequence>

Mutations: A123G, R45K, D88N, ...

Return a CSV sorted ascending by score with columns:
mutation, position, wt, mut, score, call
where call = "deleterious" if score < -<threshold> else "tolerated".
Pick the threshold from the score distribution (e.g. the 25th
percentile) and state the value you used.

  4. Use the score distribution, not an absolute cutoff. Zero-shot scores are relative within one protein; calibrate the tolerated/deleterious split from the spread of your own mutation set (or anchor it on a few known-benign and known-pathogenic mutations if you have them). State the threshold explicitly.

  5. For deep scans, switch to wt-marginal. If you are scanning every position to all 20 amino acids (a full single-mutation landscape), ask for the wt-marginal scheme instead — one forward pass on the unmasked sequence scores the whole landscape, at a small accuracy cost versus per-position masking.

Why this assembly

Rung 2. One skill (ESM, optionally a second read-only skill for the sequence fetch) computes the whole ranking; the masked-marginal score is a single model forward pass per mutation. Claude Code alone (rung 1) cannot do this — it has no protein-language-model weights and will confabulate scores. A rung-3 toolbelt or a rung-4 autonomous system buys nothing for a one-shot scoring table. The escalation that would justify rung 3 is closing a design-build-test-learn loop (score → synthesize → assay → retrain a fitness head on the results); for a standalone zero-shot ranking, the single skill is enough.

Availability

Fully open. The ESM and gget skills are MIT-licensed OSS. ESM open weights are downloaded from EvolutionaryScale / Hugging Face under their model license (free for research; check the EvolutionaryScale Community License for ESM3/ESM C terms before commercial use). UniProt sequences are public. The cloud Forge/Biohub API path requires a free ESM_API_KEY; the local-weights path needs no account.

Compute requirements

GPU workstation recommended. Running ESM C or ESM-2 (650M) locally for masked-marginal scoring is comfortable on a single GPU with ~8–16 GB VRAM; each masked forward pass is sub-second for a typical <500-residue protein, so a list of dozens of mutations finishes in well under a minute. Scoring a full single-mutation landscape (sequence length × 19) is heavier — prefer the wt-marginal one-pass scheme there, or batch on the GPU. CPU-only is possible for small proteins and short lists but slow. The largest models (ESM-2 3B/15B) want ≥24 GB VRAM; the Forge API offloads compute entirely if local GPU is unavailable.

Evidence

Proposed. No documented attempt of this exact K-Dense esm-skill-driven scoring assembly is known. The underlying method is well-validated at the component level: Meier et al. (NeurIPS 2021) introduced ESM-1v and showed zero-shot masked-marginal log-likelihood-ratio scoring matches or beats supervised state-of-the-art on variant-effect prediction with a single forward pass, no labelled data (bioRxiv 2021.07.09.450648). The masked-marginal heuristic is the canonical zero-shot scorer in the ProteinGym benchmark across hundreds of DMS assays (Notin et al., NeurIPS 2023). The scheme is in active 2025 use: Zhang et al. (Nat. Commun. 2025) seed an automated biofoundry directed-evolution loop with ESM-2 zero-shot predictions of 96 variants, reaching up to 2.4-fold activity gains in four rounds over 10 days (doi:10.1038/s41467-025-56751-8), and ESM-Scan packages the same scoring to guide amino-acid substitutions (bioRxiv 2023.12.12.571273). What is not independently benchmarked is the convenience layer — Claude driving the K-Dense skill to assemble the ranked CSV.

Alternatives considered

  • Interpret a clinical variant (BioMCP, rung 2). Reach for that instead when the variant is already catalogued — it reads ClinVar significance, gnomAD frequency, and curated predictor calls directly. The ESM recipe is the complement: it scores novel or uncharacterized substitutions that have no database entry, which is exactly where database lookups return nothing.
  • Supervised DMS models. If you already have a deep-mutational-scanning dataset for this protein, a supervised model trained on it will beat zero-shot. Use zero-shot when you have no labels — its whole point is needing none.
  • Structure-aware scoring (e.g. graph + PLM ensembles). Recent work shows combining ESM embeddings with structure graphs improves generalization to highly diverged sequences (Ash et al., bioRxiv 2025). That is a rung-3 toolbelt (needs a structure source plus a custom model) — escalate only if pure-sequence scores generalize poorly on your protein family.

See also

Sources

Tried this recipe?

Share feedback — what worked, what didn’t, what you’d change. The form opens with this recipe pre-selected and a link back to this page.