target-prioritization — Claude Code skill: rank drug targets

What target-prioritization does

Feed it a ranked gene list — a CSV with a gene column, a plain text file with one symbol per line, or symbols inline in your message — and the skill queries five sources per gene. UniProt returns protein localization, surface and MHC flags. OpenTargets brings tractability, approved drugs and disease associations, plus DepMap CRISPR essentiality and gnomAD constraint. PubMed counts papers, HPA covers tissue and single-cell specificity, ChEMBL surfaces tool compounds with pIC50 and IC50.

Two scripts do the work. orchestrate.py runs the five fetchers in parallel threads and writes raw responses to raw_data/<source>.json. Then aggregate.py merges them, computes the composite score from the weights in weights.yaml, and emits targets_summary.csv along with a targets_report.md skeleton where the rationale fields sit blank. The agent fills those in, 2–3 sentences per gene: strongest evidence, main risk, suggested next step. The last step is build_html_report.py, which produces an interactive report with a sortable table and per-gene cards.

You need python3 and curl in the environment, or the pipeline will not run. Weights can be overridden with --weights, and --top caps the dossier at the first N input genes, 50 by default. Note that ChEMBL contributes dossier columns only — it carries no score component. If your environment cannot reach external APIs, the fetches fail, so check network access before you start.