Skip to content

README.md

Latest commit

 

History

History
124 lines (97 loc) · 5.96 KB

README.md

File metadata and controls

124 lines (97 loc) · 5.96 KB
Raw

NCypher dashboard

The product surface for NCypher: a calibrated, honest, context-specific non-coding regulatory-variant triage engine for paediatric diffuse midline glioma (DMG).

This is the polished, judge-facing React dashboard. It renders the same rich output the score_variant MCP tool returns (three orthogonal axes, a convergence verdict with a calibrated confidence tier, a model-native mechanism, a skeptic check and a go/no-go memo), plus the headline finding and the honest validation.

It is fully self-contained. No backend, no Modal, no network calls at runtime. Every number is bundled from the real result tables into src/data/ncypher.json.

Run it

Requires Node 18+ (built and screenshotted on Node 24, npm 11).

cd dashboard
npm install
npm run dev        # http://localhost:5173  (Vite picks a free port and prints it)

Build the static production bundle:

npm run build      # type-checks with tsc, then bundles with Vite into dist/
npm run preview    # serve the built dist/ locally

npm run build type-checks and bundles with zero errors. The output in dist/ is a static site (HTML + one CSS + one JS + the saliency PNGs) that can be opened from any static host.

The three views

A left-nav switches between three views. Each view is deep-linkable by URL hash.

  1. Variant triage (#triage). Pick a featured variant or type any chr-pos-ref-alt. The rich card shows:

    • the three orthogonal axes (chromatin log2FC and direction, phyloP and the constrained flag, the measured-function axis marked not-available for somatic variants);
    • the convergence verdict as a prominent status pill (GO teal, HOLD amber, NO-GO red) with the calibrated confidence tier and the convergence metrics;
    • the mechanism, either the real ref-vs-alt DeepSHAP saliency logo (for the cached heroes) or the model-native contribution collapse / accessibility-gain visual, plus the named motif only where the PWM disruption is clean;
    • the skeptic (falsification) check;
    • the go/no-go memo: validate first, decisive experiment, kill criterion, therapy angle.

    Featured variants demonstrate every verdict state: NPAS3 (the lead GO), SYN3 / SOX2-OT / FARP1 / GDPD5 (converged GO), TERT (positive control), IGFBP7 (an informative HOLD) and TLE4 (an honest NO-GO). The model-context toggle flips to a breast (out-of-domain) context, where the same variant that is a GO in domain is correctly refused as NO-GO. This is the honesty demo.

    Deep links: #triage?v=chr22-33000876-C-T selects a variant; add &ctx=breast to open in the out-of-domain context.

  2. The finding (#finding). The headline: functional non-coding variants in DMG do not recur at a locus, they converge on a pathway. The axis-decomposition bar chart shows the neurodevelopmental-programme enrichment split by axis (constraint 2.62x, p = 4.8e-13; converged 2.62x, p = 0.029; chromatin alone 1.13x, p = 0.36, n.s.), the quantitative argument that convergence beats a single scorer. NPAS3 leads (three patients, three enhancers, a validated tumour suppressor) with its honest caveats stated, plus the neurodevelopmental-programme to therapy-axis mapping.

  3. Validation (#validation). The caQTL context-specificity bars (progenitor 7.5x / AUROC 0.689, neuron 4.8x, any 3.7x, mismatched PsychENCODE 1.0x, n.s.) that prove the right-cell-context claim, alongside the honest MPRA-negative note (accessibility versus reporter-activity modality gap) and the orthogonality argument.

A persistent honesty-guardrails footer (the bright-line "do not say" list) sits under every view.

Design constraints

  • Light theme only. White / near-white surfaces, dark ink, teal accent #0E9E8A.
  • DNA base colours are fixed: A #2E9E43, C #2F6FE0, G #C67F12, T #DA4A42.
  • Verdict pills: GO teal, HOLD amber, NO-GO red.
  • British English, no em dashes in the UI copy.
  • No external fonts, scripts, styles or images at runtime. Self-contained.

Data provenance

All numbers are extracted from the real result tables (read-only) into src/data/ncypher.json by a small generator script, so nothing is hand-typed and nothing drifts. Sources:

  • data/dmg/sweep_result.tsv (10,869 OPC-regulatory somatic non-coding SNVs, 152 DMG patients; 753 chromatin high-impact, 1,583 constrained, 164 two-axis converged);
  • data/figures/mining/hero_candidates.tsv (DeepSHAP contribution collapse, disrupted motif);
  • data/mpra/caqtl_validation.tsv (the caQTL context-specificity numbers);
  • docs/plan/data-mining-findings.md, docs/audit/validation-result.md (the finding and the honest framing);
  • skill/references/* (thresholds, guardrails, therapy map);
  • the real DeepSHAP saliency PNGs (data/figures/hero/*, data/figures/real/TERT_*), copied into public/saliency/.

Authored narrative (mechanism, skeptic, memo and therapy copy) lives in src/data/content.ts, grounded line by line in the guardrails and the finding document. Numbers stay in the JSON; prose stays in the TS. The two never disagree.

Stack

Vite 6, React 18, TypeScript (strict), Tailwind CSS 3. Charts and the sequence-logo visuals are hand-rolled SVG or the real model PNGs, so the bundle pulls in no chart or component library and stays self-contained.

Layout

dashboard/
  index.html                 self-contained shell, inline SVG favicon
  src/
    main.tsx, App.tsx         hash-routed shell + guardrails footer
    index.css                 Tailwind + light-theme base
    types.ts                  data + content types
    lib/format.ts             number / p-value / variant-id helpers, palette
    data/
      ncypher.json            generated numeric bundle (read from the result TSVs)
      content.ts              authored hero narratives, therapy axes, guardrails
    components/               Sidebar, StatusPill, AxisTriPanel, Mechanism, HBarChart, ui
    views/                    VariantTriage, TheFinding, Validation
  public/saliency/            real DeepSHAP saliency PNGs