Skip to content

web

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

web

README.md

Sisyphus PBPK Console

A scientific-editorial web interface for the Sisyphus structure-only Cmax core. It is a research console, not a clinical dosing tool. Built from the design handoff ("Direction B — Workspace Console") and wired to the real Sisyphus engine.

Two supported workflows, driven by real model artifacts:

Workflow What it shows
predict SMILES → C(t) curve (real ODE), 4-track meta-learner, body graph, development-residual 90% interval, pipeline log
benchmark N=107 retrospective development scatter + per-track AAFE; explicitly not an independent holdout

Stack

  • Vite + React 18 + TypeScript, hand-rolled SVG charts (no chart library — preserves the design's typographic look).
  • No backend (Phase 1). All numbers are pre-computed by the real engine and read from public/data/console_data.json. See the data layer below.

Develop

cd web
npm install
npm run dev        # http://localhost:5173
npm run build      # type-check + production bundle → dist/
npm run smoke      # headless jsdom render test of both workflows

Data: real engine, two tiers

The console reads public/data/console_data.json — every numeric field is a genuine Sisyphus engine solve / pipeline.predict call, not a mock. It is produced offline by:

# from the repo root, with the locked deps available
/opt/miniconda3/bin/python scripts/gen_console_data.py

This regenerates web/public/data/console_data.json (curated core predictions) and web/public/data/benchmark.json (the N=107 development evidence). Each preset uses one strict prediction; its Cmax contract and matching engine curve are serialized without a second solve or curve rescaling.

The data layer (src/data.ts) talks to the engine through an EngineClient interface:

  • Static tier: SisyphusClient reads the pre-computed JSON. Works on GitHub Pages with zero backend.
  • Live tier: when VITE_API_URL is configured, the same client calls the FastAPI core for arbitrary SMILES.

Curves use the matching ODE single-dose response. The empirical development 90% residual interval (÷×~10.24; 91.1% coverage on the repeatedly used development set) and its calibration caveat are surfaced. Regimen, TDM, DDI, and dose-adjustment workflows are deliberately outside this console.

Deploy (GitHub Pages — sisyphus-pbpk.io/app/)

The console is served at /app/, alongside the Jekyll homepage at /, with no change to the Pages source. The bundle uses base: "./" (path-agnostic), so the build is committed to the repo-root /app/ folder and the existing branch-based Jekyll Pages deployment serves it verbatim. The root CNAME keeps the custom domain.

To update the deployed console:

cd web
npm run build:pages   # build + sync web/dist → ../app
cd .. && git add app && git commit -m "chore(web): rebuild console" && git push

.github/workflows/web-ci.yml build- and smoke-tests web/ on every PR/push so broken bundles never land. (It does not deploy — /app/ is the committed build.)

Layout

src/
  types.ts                # data contracts (mirror gen_console_data.py output)
  data.ts                 # EngineClient + useConsoleData() hook
  pk.ts                   # core Cmax/engine-curve display helpers
  styles.css              # the scientific-editorial design system
  components/
    App.tsx               # rail + nav + run flow + state
    RailInputs.tsx        # contextual per-workflow inputs
    charts.tsx            # concentration and development-scatter charts
    panels.tsx            # Stat, TrackBars, BodyGraph, Pill, Legend, …
    workflows/            # supported PredictView + BenchmarkView only
public/data/              # real-engine JSON (generated)
scripts/smoke.mjs         # headless render smoke test