target-validation-scorer

Files (7)

• SKILL.md
• checksums.sha256
• commands.sh
• demo_input.json
• environment.yml
• target_validation_scorer.py
• test_target_validation_scorer.py · tests/

SKILL.md (first lines)

# 🎯 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 | Max score | Source | What it measures |
|-----------|-----------|--------|-----------------|
| Disease association | 20 | Open Targets | Genetic and functional evidence linking target to disease |
| Druggability | 20 | ChEMBL + UniProt | Is this target class historically druggable? Known ligands? |
| Chemical matter | 20 | ChEMBL | Do bioactive compounds exist? Best potency? |
| Clinical precedent | 20 | ChEMBL + ClinicalTrials.gov | Have compounds reached clinical trials? |
| Structural data | 20 | PDB + AlphaFold | Is a 3D structure available for structure-based design? |

### 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

### Safety penalties (applied after scoring)

- Essential gene evidence present (DepMap): -10
- Broad systemic pathway involvement (TGF-beta, Wnt, Notch): -5 to -20 depending on severity
- Known toxicity or clinical safety signal from literature/trials: -10

If a target has strong disease relevance but also major systemic safety liability, prefer CONDITIONAL_GO over GO.

Safety penalties reduce the final score but do not change sub-scores.
A target can score 80 on evidence but drop to 65 after safety adjustment.
Safety is treated as a post-hoc penalty rather than a scoring dimension to ensure
that strong biological evidence is not masked by safety concerns, but explicitly adjusted.

### Decision tiers

| Adjusted score | Decision | Meaning |
|----------------|----------|---------|
| 75-100 | GO | Strong evidence across multiple dimensions |

From ClawBio/ClawBio, MIT.

More from ClawBio/ClawBio