cndoin/med-biostat-skill

**English** · 简体中文 · 日本語
Quick start · Analysis methods · Install into your agent · License
What it is
A **cross-agent research skill** for medical and bioinformatics work. Install it into Claude Code, Codex, or WorkBuddy, and when you ask something like *"analyse this GWAS summary statistics"*, *"is this gene causally linked to cataract"*, or *"double-check whether my results are right"*, it loads **automatically** and runs the analysis inside a fixed methodological framework — instead of letting the model improvise code.
The problem it solves is not "can the AI write the code". It is **caliber** (口径) — the definition a number was computed under.
- λGC has two legitimate definitions; mixing them yields two different numbers for the same file.
- How missing values are handled changes every downstream conclusion.
- "Matches the official release" means nothing if it is asserted from "the numbers look close" rather than proven by hashes.
So this skill treats **caliber as a first-class citizen**: a result without a stated caliber is treated as invalid; anything computable is computed by a script; anything verifiable is verified by a hash.
Three principles
| # | Principle | Meaning |
|---|---|---|
| 1 | **Caliber before numbers** | One name under two calibers is two different quantities. A missing caliber is an error, not a default. |
| 2 | **Consistency beats "looks right"** | "Matches the official release" must be proven by hashes and byte counts, never inferred from "close enough". |
| 3 | **Confounding is the main battlefield** | Failures rarely sit in the algorithm. Bias auditing is its own layer, with all twelve items checked one by one. |
Analysis methods
| Group | Count | Covers |
|---|---|---|
| Causal & genetic | 10 | Mendelian randomization (7 estimators + 7 sensitivity checks), drug-target cis-MR chain, colocalization, fine-mapping (SuSiE), PRS, GWAS QC, independent loci, observational causal inference (PSM / IPTW / AIPW), mediation, DiD |
| Omics & pathways | 10 | Differential expression, WGCNA, ORA, GSEA, GSVA / ssGSEA, single-cell pipeline, immune infiltration, pseudotime, spatial transcriptomics, mutational signatures |
| Clinical stats & modelling | 11 | Survival analysis, competing risks (CIF), prediction models (calibration + DCA), C-index / NRI / IDI, regularized regression (LASSO / Cox-LASSO), meta-analysis, publication bias, MICE imputation, multiple-testing correction, mixed effects, power & sample size |
| Data engineering & ophthalmology | 4 | Data cleaning & variable selection, IOL refractive calculation, visual acuity / refractive outcomes, lens gene-set enrichment |
| Preprocessing & cell interaction | 2 | ComBat batch correction, cell–cell communication (ligand–receptor) |
| Extended methods | 4 | Gene–environment interaction (G×E), EWAS methylation, bidirectional / mediation MR, cell-type annotation |
Entry points, required calibers and common pitfalls for each method are in SKILL.md.
Quick start
Requirements
| Item | Requirement |
|---|---|
| Python | **3.9+** (Tier 0 is standard library only) |
| OS | Windows / macOS / Linux |
| Optional stack | numpy + pandas + scipy (Tier 1 only) |
1. Run the self-check (no third-party dependencies)
Tier 0 is pure standard library, so it runs straight after cloning:
2. Install Tier 1 (optional, needed for full capability)
scripts/capabilities.py will then report Tier 1 as ready. **Missing dependencies never produce false failures** — cases that cannot run are reported as SKIP, not FAIL.