drug-research

Files (4)

• SKILL.md
• tool-reference.md · references/
• run.cpython-313.pyc · scripts/__pycache__/
• run.py · scripts/

SKILL.md (first lines)

# Drug Research Strategy

Comprehensive drug investigation using 50+ ToolUniverse tools across chemical databases, clinical trials, adverse events, pharmacogenomics, and literature.

For detailed tool chains, output templates, and validation guidance, see [references/tool-reference.md](references/tool-reference.md).

---

## When to Use

- The user asks about a drug, medication, or therapeutic compound
- The user needs a drug profile, safety assessment, or clinical development overview
- The user requests ADMET evaluation, pharmacogenomics, or regulatory landscape information
- The user provides a compound name, SMILES, or ChEMBL/PubChem identifier for research

---

## Key Principles

1. **Report-first approach** - The agent creates the report file before any data collection, then populates it progressively.
2. **Compound disambiguation first** - The agent resolves identifiers (PubChem CID, ChEMBL ID, DailyMed SetID, PharmGKB ID) before beginning research.
3. **Citation requirements** - Every fact includes inline source attribution with the tool and identifier used.
4. **Evidence grading** - Claims are graded by evidence strength (T1: Phase 3/FDA label, T2: Phase 1-2/large case series, T3: preclinical, T4: computational).
5. **Mandatory completeness** - All 11 report sections must exist, even if marked "data unavailable."
6. **English-first queries** - The agent uses English drug/compound names in tool calls, falling back to original-language terms only if needed. The agent responds in the user's language.

---

## Workflow Overview

```
Step 1:  Create report file ([DRUG]_drug_report.md) with all 11 section headers
Step 2:  Resolve compound identifiers -> Update Section 1 (Identity)
Step 3:  Retrieve FDA label core fields (mechanism, PK, safety, PGx)
Step 4:  Query PubChem / ADMET-AI / DailyMed -> Update Section 2 (Chemistry)
Step 5:  Query FDA Label MOA + ChEMBL + DGIdb -> Update Section 3 (Mechanism & Targets)
Step 6:  Query ADMET-AI tools (fallback: DailyMed PK) -> Update Section 4 (ADMET)
Step 7:  Query ClinicalTrials.gov -> Update Section 5 (Clinical Development)
Step 8:  Query FAERS / DailyMed -> Update Section 6 (Safety)
Step 9:  Query PharmGKB (fallback: DailyMed PGx) -> Update Section 7 (Pharmacogenomics)
Step 10: Query DailyMed / Orange Book -> Update Section 8 (Regulatory)
Step 11: Query PubMed / literature -> Update Section 9 (Literature)
Step 12: Synthesize findings -> Update Executive Summary & Section 10 (Conclusions)
Step 13: Document all sources, run completeness audit -> Update Section 11
```

---

## Report Structure

The agent produces an 11-section report in `[DRUG]_drug_report.md`:

| Section | Content |
|---------|---------|
| Executive Summary | High-level drug profile overview |
| 1. Compound Identity | Database IDs, SMILES, formula, synonyms |
| 2. Chemical Properties | Physicochemical profile, drug-likeness, solubility, salt forms |
| 3. Mechanism & Targets | FDA label MOA, primary targets with UniProt IDs, selectivity |
| 4. ADMET Properties | Absorption, distribution, metabolism, excretion, toxicity |
| 5. Clinical Development | Phase counts, trial landscape, approved/investigational indications, biomarkers |
| 6. Safety Profile | FAERS data, black box warnings, DDIs, drug-food interactions, dose modifications |
| 7. Pharmacogenomics | Pharmacogenes, CPIC/DPWG guidelines, clinical annotations |
| 8. Regulatory & Labeling | Approval status, patents, exclusivity, special populations, timeline |
| 9. Literature & Research | Publication metrics, research themes, real-world evidence |
| 10. Conclusions | Scorecard, strengths, concerns, research gaps, comparative analysis |
| 11. Data Sources | Tool call summary, completeness audit, quality control metrics |

---

## Report Detail Requirements

Each section must be comprehensive and detailed:

- **Tables** for structured data (targets, trials, adverse events)
- **Lists** for features, findings, key points
- **Paragraphs** for narrative synthesis
- **Specific values** including counts, percentages, and confidence levels (not vague terms)
- **Context** explaining what the data means, not just what it is
- **Source attribution** at the end of each data block

---

## Citation Format

The agent attributes every data block to its source:

```markdown
*Source: PubChem via `PubChem_get_compound_properties_by_CID` (CID: 4091)*
```

Section-level source summaries appear at the end of each section:

```markdown
---
**Data Sources for this section:**
- PubChem: `PubChem_get_compound_properties_by_CID` (CID: 4091)
- ChEMBL: `ChEMBL_get_bioactivity_by_chemblid` (CHEMBL1431)
---
```

---

## Critical Rules

- **Avoid `ChEMBL_get_molecule_targets`** — it returns unfiltered, irrelevant results. The agent derives targets from `ChEMBL_search_activities` instead, filtering to pChEMBL >= 6.0.
- **Type normalization** — All IDs (ChEMBL, PubMed, NCT) are converted to strings before API calls.
- **ADMET fallback** — If ADMET-AI tools fail, the agent falls back to FDA label PK sections. Section 4 is never left empty.
- **PharmGKB fallback** — If PharmGKB is unavailable, the agent uses DailyMed PGx + PubMed literature.
- **FAERS limitations** — The agent always includes a data limitations paragraph noting voluntary reporting, causality caveats, and reporting bias.
- **Clinical trial counts** — Section 5.2 shows actual counts by phase/status in table format, not just a list of trials.

---

## Common Use Cases

| Scenario | Focus |
|----------|-------|
| **Approved drug profile** ("Tell me about metformin") | Full 11-section report emphasizing clinical data, FAERS, PGx |
| **Investigational compound** ("What do we know about compound X?") | Preclinical data, mechanism, early trials; safety sections may be sparse |

From lamm-mit/scienceclaw, Apache-2.0.

More from lamm-mit/scienceclaw