Risk model: Table 1, each risk group's clinical and genomic makeup

[alisman]

Request #4 from the analyst's review of the riskmodel view: show the clinical and genomic makeup of each risk group — stage in the low-risk group, whether a group is enriched or depleted for KEAP1 or TP53 — with a test across groups. Built on the risk-model work merged in #37.

Preview

Open this on the PR's preview deployment (or any instance with msk_chord_2024): the gallery's NSCLC model. Click table 1 in the card's toolbar; add a gene such as SMARCA2 (not every panel covers it); try Add to document.

/lab?tab=analysis&doc=eyJkc2wiOiIyLjYiLCJzdHVkeSI6Im1za19jaG9yZF8yMDI0IiwiY29ob3J0cyI6eyJOU0NMQyI6eyJhbGwiOlt7ImtpbmQiOiJjbGluaWNhbCIsImF0dHJpYnV0ZSI6IkNBTkNFUl9UWVBFIiwidmFsdWVzIjpbIk5vbi1TbWFsbCBDZWxsIEx1bmcgQ2FuY2VyIl0sInBlciI6InNhbXBsZSJ9XX19LCJ2aWV3cyI6W3sia2luZCI6InJpc2ttb2RlbCIsImNvaG9ydHMiOlsiTlNDTEMiXSwicnVucyI6NX1dfQ

What changes for a reader

• Table 1 — every risk group side by side, on the patients the curves draw (with validate, the validation half):
  • a categorical attribute's patients per value (the patient's value, or their first sample's; case and whitespace folded) with a χ² test;
  • a numeric one's median [IQR] with Kruskal–Wallis;
  • a gene's mutated patients over those whose mutation panels cover it — a patient whose panels do not is "not profiled", never unmutated, and the count is shown — with a χ² test;
  • Benjamini–Hochberg q across rows.
• Rows that differ by construction are marked. The model's genes (kept in any run) and covariates are "in the model", the endpoint's own columns "outcome": they keep a p-value but get no q-value and stay out of the correction, where their guaranteed tiny p-values would shrink every other row's q. The model's stratum is marked "stratum" and stays in — the score has no stratum term.
• Defaults: the model's stratum and covariates, then the study's attributes in priority order, leaving out survival columns and any that have no value, one value, or more than 12 categories among these patients (each left out is listed with why); the first 8. Genes: the 15 the model kept most often. The dialog adds and drops attributes and genes (≤ 20 each).
• describe: {attributes?, genes?} — new view option. "Add to document" writes the table's resolved lists into the view, so a link reproduces it and the figure's caption (MCP, exports) states Table 1 in words. Without it the table is the dialog's alone.
• Toolbar: three sections on one line — data (genes, stratify by, covariates, clock from), model (settings · groups · table 1), view. Penalty, runs and "validate on held-out half" move into a settings menu (the sequence and course views' idiom); the groups button reads "groups", its current cut in its title.

Commit by commit

1. Risk model: share the modelled-patient filter, and keep each group's members — refactor, no behaviour change: keepRiskRows (who is modelled, every exclusion counted) out of fitRiskmodel; groupRiskmodel also returns each group's members as positions. The riskmodel result hashes identically before and after.
2. Risk model: Table 1, each risk group's clinical and genomic makeup — the engine. Table 1 re-reads the fit's own rows query with its attributes added, keeps them through keepRiskRows, and refuses unless they match the fit's patients one by one (time, entry, event), so each column is the patients its curve drew. Counting is pure (lib/views/riskmodel-table1.ts); tests reuse lib/stats.ts (chiSquareTest, kruskalWallis, quantilesFromHistogram, benjaminiHochberg). describe joins RISK_PRESENTATION_KEYS, so a Table 1 change never refits. Method record (steps + a script recomputing every p, q and median from the counts), verification wording (no new statistic), DSL reference + docs/dsl.md, the gallery example, two planted bugs in check-mutants.
3. Risk model: the figure's caption states Table 1 — and a stress document (20 attributes, 20 genes, five groups) for check:figures.
4. Risk model: a Table 1 dialog, with rows to add and "Add to document"
5. Risk model: "groups" and "table 1" side by side in the model zone
6. Risk model: the fitting settings in a settings menu, three toolbar sections

How it was verified

• Gate (scripts/gate.sh): build, lint, unit tests, Python suite, check:methods, docs, figures, e2e — green on this branch rebased onto master.
• Denominator against ClickHouse, independently: msk_chord_2024 NSCLC, validate, runs 3 — a hand-written query over the validation half gives 3,867 modelled patients, SMARCA2 22/1,093, ZFHX3 185/3,710, KEAP1 535/3,867; Table 1's per-group counts sum to exactly these.
• Statistics: the method script (scipy chi2_contingency(correction=False), kruskal, numpy percentiles, statsmodels fdr_bh over the same family) reproduces every number on the gallery example (21) and on ad-hoc runs with defaults, named lists, strata/covariates and validation (29–32 each). It caught a duplicate-gene bug before commit.
• Unit tests (test/riskmodel-table1.test.ts): the denominator (panel lacking the gene, WES, a group nobody covers), roles and the BH family, folding, numbers, defaults, set-aside rows, and the refusal when rows do not match the fit. Planting either mutant by hand fails them.
• e2e: one button, aligned; groups side by side; model genes marked with no q; adding a gene with the request held covers the table and leaves the controls usable, then is one request; "Add to document" puts describe in the URL and Table 1 in the caption; reopening is served from the view's result with no request; the settings menu shows the document's runs.

Review notes

• Defaults can include near-constant attributes. On an NSCLC cohort, CANCER_TYPE and DIAGNOSIS_DESCRIPTION are kept (~99% one value) because a sample-level attribute takes each patient's first sample, which for 36 patients is not the lung sample. A dominance cut-off would drop them; left as is pending a decision.
• No DOC_VERSION bump: describe is an optional view setting, as cuts and labelAbove were in Risk model: OncoCast's 0–10 score, volcano, custom cuts, and a cached fit #37; an older engine shows no Table 1 rather than a wrong number.
• Not this PR's, noticed on the way: the volcano's gene list runs names into their percentages ("DNMT3A80%") and the figure lint misses it; the gallery lists the risk-model example twice (already so on master — this PR adds Table 1 to the first).

🤖 Generated with Claude Code

https://claude.ai/code/session_01UCpBeF9VMrjXyddFKUfiSX

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
cohort-lab	Ready	Preview	Oct 19, 2026 1:28am UTC