Propose the mechanism by which a missense variant causes loss-of-function (LoF), synthesizing evidence from 5 independent layers: AlphaMissense pathogenicity, AlphaFold structural context, ESMC sequence likelihood, SAE feature disruption, and DynaMut2 stability ΔΔG. Distinguishes 'structural stability LoF' (mis-folding) from 'direct functional disruption' (catalytic / binding / PTM site damage). Use for coding missense variants where you need a mechanistic causal model, not just a pathogenicity score.
Propose the mechanism by which a missense variant causes loss-of-function (LoF), synthesizing evidence from 5 independent layers: AlphaMissense pathogenicity, AlphaFold structural context, ESMC sequence likelihood, SAE feature disruption, and DynaMut2 stability ΔΔG. Distinguishes 'structural stability LoF' (mis-folding) from 'direct functional disruption' (catalytic / binding / PTM site damage). Use for coding missense variants where you need a mechanistic causal model, not just a pathogenicity score.
disable-model-invocation
true
Protein LoF Mechanism Synthesis
For a single missense variant, integrate 5 independent computational signals to propose a specific loss-of-function mechanism. Each signal answers a different question:
Signal
Tool
Answers
Pathogenicity
AlphaMissense_get_variant_score
"Is this variant damaging?"
Structural context
alphafold_get_prediction
"Is the mutation in a folded vs disordered region?"
Sequence likelihood
ESM_score_sequence
"Is the substitution evolutionarily plausible?"
Feature disruption
ESM_explain_variant_mechanism (or ESM_score_variant_sae_disruption + ESM_describe_sae_feature for raw control)
"Which biological feature breaks?"
Stability
DynaMut2_predict_stability
"Does the protein still fold correctly?"
When to use this skill
Apply for missense (coding) variants where:
You already have evidence the variant is damaging (or want to find out)
You need to know WHY it's damaging in molecular terms
Downstream work depends on the mechanism: drug rescue strategies need to know what's broken, gene-therapy decisions need to distinguish "fix the protein" vs "replace the protein", clinical reporting wants a mechanism narrative
AlphaMissense_get_variant_score(
uniprot_id="P04637",
position=175,
ref_aa="R",
alt_aa="H",
)
# → returns score 0..1; ≥0.564 is the "likely pathogenic" threshold per the# AlphaMissense paper. Above 0.9 is very confident damaging.
If AlphaMissense says benign (≤0.34), the rest of the analysis is exploratory — most benign variants don't have a clear LoF mechanism.
Step 2: AlphaFold structural context
alphafold_get_prediction(uniprot_id="P04637")
# → returns structure data including per-residue pLDDT
Check pLDDT at the mutation position:
pLDDT > 70 → well-folded region; the variant is in a structured area; structural / functional disruption is meaningful
pLDDT 50-70 → flexible / partially folded; interpretation is ambiguous
pLDDT < 50 → disordered; SAE / stability signals may not be reliable
Step 3: ESMC sequence likelihood
ESM_score_sequence(
sequence=ref_sequence,
model="esmc-600m-2024-12", # or 300m for cheaper
)
# Then for the mutant sequence:
ESM_score_sequence(
sequence=mutant_sequence,
model="esmc-600m-2024-12",
)
Compute ΔlogP = mean_logP(mutant) − mean_logP(reference) at the position window:
ΔlogP < −1 → mutation is evolutionarily implausible (strong signal of functional cost)
ΔlogP ≈ 0 → mutation is conservative or in a tolerant position
ΔlogP > 0 is rare and usually noise
Step 4: SAE feature disruption (the unique signal)
Dominant category among top lost features = the function most likely disrupted.
Step 5: DynaMut2 stability (ΔΔG)
DynaMut2 needs a PDB structure. Two options:
Option A — use a PDB ID if available for this protein:
# Look up PDB cross-references from UniProt entry
UniProt_get_entry_by_accession(accession="P04637")
# → check `uniProtKBCrossReferences` for entries with database == "PDB"# → pick a structure that covers the mutation position
DynaMut2_predict_stability(
pdb_id="2FEJ", # example TP53 DNA-binding domain crystal structure
chain="A",
mutation="R175H",
)
# → returns ddG in kcal/mol
Option B — if no experimental PDB covers the position, use the AlphaFold model (output of Step 2). DynaMut2 accepts AlphaFold PDBs the same way.
Step 6: Synthesis — decide the LoF mechanism category
Apply the upstream variant_lof_mechanism decision rule:
Signal pattern
Inferred mechanism
ddG > +1 kcal/mol AND ΔlogP < 0
Structural stability LoF — mutation destabilizes the fold; protein may misfold / be degraded. Drug rescue strategy: pharmacological chaperones, refolding agents.
ddG ≈ 0 (in [-0.5, +1]) AND SAE features lost are catalytic
Direct catalytic LoF — protein folds normally but the active site is broken. Strategy: substrate analog / cofactor supplementation.
Missense only. Indels, nonsense, splice variants need other workflows.
Single-isoform. Uses UniProt canonical. Variants in non-canonical isoforms may not apply.
SAE labels are inferred, not curated.ESM_describe_sae_feature labels are best-effort aggregations from a 10-protein panel; some features stay "uncategorized" with high activation values — flag this as low-confidence interpretation.
DynaMut2 needs PDB. If no PDB covers the mutation position and AlphaFold confidence is low (pLDDT < 50), the stability signal is unreliable.
Long-range allosteric effects not captured. SAE window is ±8 residues. Some mutations break protein function via distant effects (e.g., dimerization interface modulation) — SAE alone won't see this.
Synthesis decision rule is heuristic. The 6-category mapping in Step 6 — it's a reasonable starting point, not a clinical gold standard. For high-stakes interpretation, the rule should be treated as a hypothesis to be tested experimentally.
Non-commercial license on SAE outputs (per EvolutionaryScale Cambrian License).
Optional: ThermoMPNN instead of DynaMut2 (advanced users with GPU)
DynaMut2 is the recommended default — it's already wired into TU via a hosted academic API (BioSig lab, UQ Australia) and requires zero extra setup. But if you need the higher accuracy of the newer ThermoMPNN model (Dieckhaus et al., PNAS 2024), here are your options:
Local install (you have GPU + want full control)
git clone https://github.com/Kuhlman-Lab/ThermoMPNN.git
cd ThermoMPNN
# Install conda env from environment.yaml — note the GitHub README warns# the .yaml may install CPU-only PyTorch by mistake; verify GPU PyTorch
conda env create -f environment.yaml
conda activate ThermoMPNN
# The checkpoint thermoMPNN_default.pt ships in models/
python custom_inference.py --pdb <your.pdb> --mutation <e.g. R175H>
Requirements: NVIDIA GPU with CUDA 11.8, ~15 min one-time setup.
License: MIT (no commercial restrictions).
Citation: Dieckhaus et al. (2024). Transfer learning to leverage larger datasets for improved prediction of protein stability changes. PNAS 121(6):e2314853121. doi:10.1073/pnas.2314853121.
Hosted SaaS (no GPU, but registration + API key required)
Several commercial platforms host ThermoMPNN with free tiers (specific quota varies and is not always documented):
Tamarind Bio — REST API at app.tamarind.bio/api/, x-api-key auth
These are commercial SaaS — "free tier" usually means "try a few calls then pay". Not as clean as DynaMut2's purely academic endpoint, which is why TU defaults to DynaMut2.
When ThermoMPNN actually matters over DynaMut2
For LoF mechanism classification (this skill's use case), the binary distinction ddG > +1 vs ddG ≈ 0 is what drives the synthesis. Both models give this signal correctly for clear-cut cases. ThermoMPNN's edge over DynaMut2:
More accurate ΔΔG near the threshold (e.g., distinguishing +0.5 from +1.2 reliably)
