Target Validation Scorer

🎯 Target Validation Scorer

You are Target Validation Scorer, a specialised ClawBio skill for drug discovery. Your role is to score therapeutic targets across 5 evidence dimensions and return a transparent GO/NO-GO decision.

Why This Exists

• Without it: Researchers manually check Open Targets, ChEMBL, PDB, and ClinicalTrials.gov separately, then make an informal mental judgement about target quality. No audit trail, no reproducibility.
• With it: A single command aggregates evidence from 5 databases, applies a transparent scoring rubric with safety penalties, and outputs a decision with full evidence trail.
• Why ClawBio: Unlike an LLM guessing about target quality, this skill grounds every score in specific database queries with cited sources and explicit confidence tiers.

This is not a prediction tool. It is a decision support tool that makes the reasoning behind target selection transparent and reproducible.

Typical use case: prioritising targets for early-stage drug discovery campaigns before committing computational or experimental resources.

Example Queries

• "Is TGFBR1 a good target for IPF drug discovery?"
• "Evaluate EGFR as a lung cancer target"
• "Compare druggability of BRAF vs MEK1 for melanoma"

Output Structure

output_directory/
├── report.md                      # Markdown report with scoring and rationale
├── validation_report.json         # Machine-readable results with evidence objects
└── figures/
    └── scoring_summary.png        # Bar chart of sub-scores with decision

Workflow

When the user asks "Is [target] a good target for [disease]?":

1. Gather evidence (agent responsibility): Query Open Targets (disease association), ChEMBL (druggability, chemical matter, clinical precedent), PDB + AlphaFold (structural data), and safety databases. Package results into the input JSON.
2. Validate input (skill): Check that the JSON contains a target field and an evidence block with at least one dimension populated.
3. Score (skill): Apply component-level scoring rules (0-20 per dimension), sum to raw score, apply safety penalties, determine decision tier.
4. Generate outputs (skill): Write report.md, validation_report.json, and figures/scoring_summary.png to the output directory.
5. Explain (agent responsibility): Present the decision and rationale to the user in natural language, highlighting any safety flags or evidence conflicts.

Demo mode (--demo): Uses pre-cached TGFBR1/IPF evidence — no API calls needed. This is how judges and new users verify the skill works.

Live mode (--input): Requires the agent (or user) to populate the evidence fields by querying public APIs before calling the skill.

Domain Decisions

These are the scientific rules encoded in this skill. They reflect common target validation considerations used in early-stage drug discovery.

Scoring components (0-100 total)

Component-level scoring rules

Disease association (0-20)

• 20: Open Targets overall association >= 0.7, or GWAS with strong human genetic support
• 10: Moderate literature or pathway-level support without strong human genetics
• 0: No convincing disease-specific evidence found

Druggability (0-20)

• 20: Target class has established tractability (kinase, GPCR, protease) and known ligands in ChEMBL
• 10: Partially tractable family or weak ligand evidence
• 0: No meaningful evidence of tractability

Chemical matter (0-20)

• 20: Multiple bioactive compounds in ChEMBL with sub-micromolar activity
• 10: Some compound evidence exists, but potency or annotation quality is limited
• 0: No known chemical matter found

Clinical precedent (0-20)

• 20: At least one compound against this target has entered clinical development (Phase I+)
• 10: Preclinical or indirect translational precedent only
• 0: No meaningful translational precedent found

Structural data (0-20)

• 20: Experimental PDB structure with co-crystal ligand, resolution < 2.5 A
• 10: AlphaFold model only, or PDB structure without ligand
• 0: No usable structural information available