structure-prediction

name	structure-prediction
description	Protein structure prediction using Chai-1 or AlphaFold with confidence scoring, quality assessment, and critic evaluation
metadata	{"openclaw":{"requires":{"env":["OPENAI_API_KEY"],"bins":["python3"],"anyBins":["chai","alphafold"]},"primaryEnv":"OPENAI_API_KEY"}}

Structure Prediction — Chai-1 / AlphaFold

Predict 3D protein structures from amino acid sequences using Chai-1 or AlphaFold backends. Includes confidence scoring, quality assessment, and an integrated critic evaluation step that flags low-quality predictions.

Capabilities

• Single-chain folding: Predict monomer structures from amino acid sequences
• Complex prediction: Predict multi-chain protein complexes (co-folding)
• Confidence scoring: pLDDT, pAE, ipTM, and aggregate metrics
• Multiple backends: Support for Chai-1 and AlphaFold prediction engines
• Quality assessment: Automated critic evaluation with weighted scoring (prediction quality, analysis depth, methodology, efficiency)
• Artifact DAG integration: All predictions stored as immutable artifacts with full provenance

Usage

python skills/structure-prediction/scripts/run_prediction.py \
  --sequence "MKWVTFISLLLLFSSAYSRGV..." \
  --backend chai \
  --num-models 5 \
  --output-dir ./prediction_output

Or with a FASTA file:

python skills/structure-prediction/scripts/run_prediction.py \
  --fasta input.fasta \
  --backend alphafold \
  --use-templates \
  --output-dir ./prediction_output

The skill will:

1. Parse the input sequence (raw string or FASTA file)
2. Run structure prediction with the selected backend (Chai-1 or AlphaFold)
3. Score models with confidence metrics (pLDDT, pAE, ipTM)
4. Run critic evaluation to assess prediction quality
5. Store results as artifacts in the Artifact DAG with provenance tracking

Parameters

• sequence: Amino acid sequence (raw string)
• fasta: Path to FASTA file (alternative to sequence)
• backend: Prediction backend — "chai" or "alphafold" (default: "chai")
• num_models: Number of models to generate (default: 5)
• use_templates: Whether to use template structures (default: false)
• device: Compute device, e.g. "cuda:0" or "cpu" (default: "cuda:0")
• output_dir: Output directory for results (default: ./prediction_output)
• artifact_store: Path to artifact store for DAG integration
• sequence_name: Name for the sequence (default: "query")
• constraints: JSON string of folding constraints

Quality Thresholds

• min_confidence: Minimum acceptable pLDDT confidence (default: 60.0)
• high_confidence: High-confidence threshold (default: 80.0)
• min_coverage: Minimum sequence coverage (default: 0.9)
• max_clash_score: Maximum acceptable clash score (default: 10.0)

Critic Evaluation Weights

• prediction_quality: 0.30
• analysis_depth: 0.25
• interpretation_accuracy: 0.20
• methodology: 0.15
• efficiency: 0.10

Output Artifacts

The skill produces artifacts in the Artifact DAG (Layer 3):

• SEQUENCE artifacts: Input sequences with metadata
• PDB_STRUCTURE artifacts: Predicted 3D structures with confidence scores
• SCORE artifacts: Per-model quality metrics (pLDDT, pAE, ipTM)
• ANALYSIS artifacts: Critic evaluation results with improvement suggestions