Skill: Predict Protein Heat Stability

Use When

Use this skill when users need to assess the thermal stability of a protein or compare variant heat stability:

• "Is this protein thermostable?"
• "What is the predicted Tm of this enzyme?"
• "How does this mutation affect heat stability?"
• "Which of these variants is most thermostable?"
• "Will this enzyme be stable at 60°C?"
• "Estimate the melting temperature of this sequence"
• "Compare the thermostability of these two protein variants"

Inputs

• Required:
  • Protein sequence (--sequence or --fasta)
• Optional:
  • Structure file (.pdb/.cif/.mmcif)
  • Mutation list (A123V format, one per line or comma-separated)
  • Organism optimal growth temperature (OGT) for context
  • Target operating temperature (e.g., for industrial applications)
  • Functional constraints (active-site, cofactor positions) for risk-aware output

Workflow

Choose the smallest workflow that still supports decision-grade confidence.

Path A: Sequence-Only Assessment

Use when no structure is available.

Script:

• scripts/predict_heat_stability.py

Capabilities:

• sequence validation (alphabet, length, ambiguous residues)
• composition-based features: GRAVY index, charged residue ratio (Arg+Lys+Asp+Glu), aromatic content (Trp+Tyr+Phe), instability index (Guruprasad et al.)
• salt bridge potential (Arg+Lys vs Asp+Glu charge balance)
• predicted thermophilicity class from feature thresholds (calibrated on known thermophile/mesophile datasets)
• confidence band with calibration caveats
• homolog-guided OGT estimate (if WebSearch available)

Example:

python3 scripts/predict_heat_stability.py \
  --fasta wt.fasta \
  --output stability_report.json

Path B: Structure-Aware Prediction

Use when a structure is available or can be generated (AlphaFold/ColabFold).

Steps:

1. Validate structure (chain selection, residue completeness, numbering match to FASTA).
2. Repair structure with FoldX RepairPDB before scoring.
3. Run FoldX Stability to estimate ΔG of folding.
4. Optionally run HoTMuSiC or MAESTRO for Tm-scale estimates.
5. Cross-validate with sequence-based features from Path A.
6. Report consensus with per-tool table.

Practical guidance:

• Fix chain and protonation state before comparing WT vs. mutant runs.
• Run WT as a control to anchor relative ΔΔG calculations.
• Do not treat FoldX ΔG absolute values as Tm; use Tm-oriented tools for direct Tm estimates.

Path C: Variant ΔTm Prediction

Use when comparing WT vs. specific mutants for thermostability rank-ordering.

Steps:

1. Validate all mutation notation (1-based, WT residue must match sequence).
2. If structure available, build mutant models with FoldX BuildModel.
3. Score each variant with at least two independent tools.
4. Rank by consensus ΔΔG/ΔTm evidence.
5. Flag variants that overlap functional constraints if provided.

Example (with structure):

python3 scripts/predict_heat_stability.py \
  --fasta wt.fasta \
  --structure model.pdb \
  --mutations A123V,G45S,T210I \
  --active-site 57,102,195 \
  --output variant_stability_report.json

Output Contract

Always include:

1. Predicted thermostability class (hyperthermophilic/thermophilic/mesophilic/thermolabile) with confidence band
2. Key supporting features table (GRAVY, charged ratio, aromatic content, instability index)
3. Predicted Tm range if structure-based tools were run (with ±error caveat)
4. Per-variant ΔΔG/ΔTm table (if mutations provided)
5. Functional risk flags (if constraints provided)
6. Interpretation limits and escalation notes

Interpretation Bands

Thermostability Class (Approximate Tm Ranges)

These bands are approximate; protein family and domain architecture affect calibration.

Sequence Feature Thresholds (Heuristic)

• GRAVY index > 0: net hydrophobic (not a direct thermostability predictor, context-dependent)
• Instability index < 40: likely stable in vitro; ≥ 40: potentially unstable
• Charged residue ratio (RKED/all) > 0.15 and Arg/(Arg+Lys) > 0.5: thermophile-associated pattern
• Aromatic content (WYF) > 0.06: elevated aromatic packing, common in thermophiles

ΔΔG Effect Sizes (Typical, Tool-Dependent)

• |ΔΔG| > 1.0 kcal/mol: strong effect candidate
• |ΔΔG| 0.3–1.0 kcal/mol: moderate effect
• |ΔΔG| < 0.3 kcal/mol: near-neutral/uncertain

Always verify sign convention per tool (negative = stabilizing in FoldX; varies in others).

References

• references/heat-stability-prediction-playbook.md

Install

From repository root:

# Claude Code
./scripts/install_skill.sh predict-protein-heat-stability --tool claude

# Codex (default)
./scripts/install_skill.sh predict-protein-heat-stability

# Custom destination
./scripts/install_skill.sh predict-protein-heat-stability --dest ~/.my-tool/skills

Limits

• Sequence-only predictions are low-resolution; structural analysis is strongly preferred for final decisions.
• Tm predictions carry ±5–10°C typical error for most tools; treat as triage evidence, not ground truth.
• Heuristic composition features are calibrated on soluble globular proteins; multi-domain, membrane, and intrinsically disordered proteins require special handling.
• Tool calibration may not transfer across distant protein families.
• Stabilizing variants are systematically under-predicted by current DDG tools; treat stabilizing predictions as high-uncertainty until orthogonally validated.
• All predictions require experimental validation (DSF, DSC, thermal shift assay, Tm determination).
• Common failure cases:
  • mutation indexing mismatches between sequence and structure numbering
  • model-quality artifacts from low-confidence structure regions
  • incompatible DDG sign conventions when comparing outputs across tools