target-prioritization

Target Prioritization

A multi-source drug-target due-diligence pipeline for ranked gene lists.

When this skill triggers

The user has a list of candidate genes (typically from a DE / DEG / scRNA-seq analysis) and wants a per-gene dossier across multiple evidence dimensions plus a composite re-ranking. The DE statistical rank is just the entry point; the final priority is informed by protein biology, genetics, druggability, and research maturity.

Common input shapes:

• A CSV with a gene column (DE output like expression_table_pass_either_1s.csv)
• A plain-text gene list (one symbol per line)
• A list of symbols inline in the user's message

Output

Four files inside <output_dir>/:

1. targets_report.md — one section per gene, sorted by composite score, with a short LLM-written rationale and recommended next step
2. targets_report.html — self-contained interactive report (sortable + searchable summary table, tier-filterable per-gene cards, score-component bars, UniProt links). Built from the .md and .csv as the final step, after Claude has filled the rationale slots and executive summary.
3. targets_summary.csv — flat table for sorting/filtering in Excel/pandas
4. raw_data/<source>.json — raw API responses (audit trail, reusable across future re-scorings)

Pipeline

input gene list
   │
   ▼
scripts/orchestrate.py
   │
   ├─► fetch_uniprot.py        → protein localization, surface, MHC, coding
   ├─► fetch_opentargets.py    → tractability, approved drugs, associated

How to invoke

• --input accepts a CSV (with --gene-col, default gene), a .txt/.tsv, or any file where the first column has gene symbols. Skips header if first cell is gene/symbol/case-insensitive.
• --top limits the dossier to the top N input genes (default 50) — input order is preserved up to that cut, then composite-score re-ranks within.

orchestrate.py runs the five fetchers in parallel (Python threads, since all calls are I/O-bound). Each writes a self-contained JSON to <output_dir>/raw_data/<source>.json. Then aggregate.py merges them, computes the composite score using weights.yaml, writes targets_summary.csv, and emits a targets_report.md skeleton with one section per gene — the rationale and risks fields are left blank for Claude to fill.

Composite score

Weights live in weights.yaml and can be overridden per-run with --weights. Defaults aim for "find druggable, genetically supported targets with clean therapeutic window and expression in the cell of interest":

ChEMBL contributes dossier columns (chembl_target_id, chembl_best_pchembl, chembl_best_ic50_nm, chembl_top_compounds) but no score component — its job is to surface concrete tool compounds for the "Suggested next step" slot.

Each component is normalized to [0, 1]. The composite is therefore roughly in [-w7, sum(w1..w6)] and is min-max rescaled before reporting. Read weights.yaml for the current defaults.

Writing the rationale

After aggregate.py produces targets_report.md with blank rationale slots, Claude reads the per-gene dossier rows and writes a 2-3 sentence rationale per gene. Use the template in prompts/rationale_template.md — it specifies the structure (one line on the most compelling evidence, one line on the main risk, one line on the suggested next experimental step).

For the top 5–10 genes by composite score, also write a short executive summary at the top of the report. Keep it factual and grounded in the dossier data; do not hallucinate beyond what the JSONs contain.

Building the HTML report (final step)

Once targets_report.md has its rationale slots and executive summary filled in, run build_html_report.py to produce a self-contained interactive HTML report:

This is always the last step. It reads targets_summary.csv and targets_report.md from --report-dir and writes targets_report.html in the same directory. The output is a single file with no external dependencies: sortable summary table at top, live search + tier-filter buttons, one card per gene with chips (surface / secreted / MHC / focus-disease / approved-drug), full dossier grid, and horizontal bars for each score component. Open it directly in a browser; share it as-is.

If rationale slots are still blank when this runs, the per-gene cards will show "not yet written" in those spots — useful for previewing the layout, but the user should be told to fill rationales before sharing.

Data source notes

All free, no API key needed. Rate limits handled in fetchers:

• UniProt REST — 100 req/sec, batched via accession query
• OpenTargets GraphQL — generous, single endpoint; provides disease genetics signal via integrated associatedDiseases
• PubMed E-utilities — 3 req/sec without key; fetchers respect this
• Human Protein Atlas — search_download.php for symbol→ENSG, then per-ENSG /<ENSG>.json; no rate limit documented, fetcher sleeps 0.15s/gene
• DepMap CRISPR essentiality — fetched via target.depMapEssentiality inside the OpenTargets call (no separate endpoint)
• gnomAD constraint — fetched via target.geneticConstraint inside the OpenTargets call (avoids gnomAD's WAF on direct API access)
• ChEMBL REST — target/search.json then activity.json; ~5 req/sec friendly, fetcher sleeps 0.2s/gene

For deeper API details and field mappings, see references/api_endpoints.md.

Retargeting the focus disease + cell context

The skill ships with an autoimmunity / T-cell default but is intentionally disease-agnostic. Three edits switch the focus:

• scripts/fetch_opentargets.py and scripts/aggregate.py — change FOCUS_DISEASE_TERMS to the lowercased substrings that should mark a drug or disease association as "in-scope" (e.g. ("cancer", "carcinoma", "lymphoma") for oncology; ("alzheimer", "parkinson", "huntington", "als") for neurodegeneration; ("diabetes", "obesity", "fatty liver", "nash") for metabolic disease).
• scripts/aggregate.py — change FOCUS_CELL_TYPES to the HPA single-cell type names that should drive cell_context_score. Must match HPA's exact strings (case-sensitive); see comment block above the tuple for examples per domain.
• scripts/fetch_pubmed.py — adjust the focus_disease and cell_context queries in CONTEXTS (these power the PubMed counts in the dossier).

No other code changes are needed; the CSV column names already use the neutral focus_disease_* / cell_context prefixes.

When NOT to use this skill

• Single-gene look-ups (overkill — just ask Claude to web-search)
• Non-human genes (most APIs are human-only; fetchers will silently return empty)
• Pure literature review without target ambition — use scholar-deep-research or literature-review instead

Iteration tips

The pipeline is designed to be re-runnable cheaply:

• Raw JSON cache means re-scoring with different weights.yaml is a one-second aggregate.py rerun
• To add a new evidence source, add scripts/fetch_<source>.py that writes raw_data/<source>.json with the same {gene: {fields}} shape, then add a corresponding term in aggregate.py::compute_composite_score.