An Antigravity 2.0 IDE experiment that builds a real, citation-grounded, scroll-driven explainer of what heat does to human proteins — using Google DeepMind's Science Skills plugin (PDB, AlphaFold, PubMed, Human Protein Atlas, ClinVar, Reactome) to fetch real biological data, and 3Dmol.js to render real protein structures. Not a clinical tool. A lab.
This repo is the small, deliberately-low-stakes home for an idea that was floated during the HeatThreshold Google I/O 2026 hackathon push and ultimately deemed too risky to ship inside a 48-hour scoped submission. The risk wasn't technical — it was epistemic: claiming anything about human biology under heat stress, on a hackathon stage, with no domain reviewer, is a bad idea. Doing it here, framed as an experiment, with disclaimers up front and every claim back-linked to a real source, is a fine idea.
The companion question is: how does Antigravity 2.0 feel as a creative coding environment when the brief is half-scientific? The Science Skills plugin gives the IDE first-class access to a stack of public scientific databases. This repo is the petri dish for finding out what that combination can produce.
A single-page, public scrollytelling explainer at index.html that walks the reader through what happens inside a human body as core temperature climbs from a normal 37 °C up through fever (39 °C), hyperthermia (40 °C), medical emergency (41 °C), cellular catastrophe (42 °C), and beyond. Every chapter centers on a real protein, with its real structure, with real citations, with real expression data, with — where relevant — real clinical variants.
The narrative arc:
| Ch | Title | Anchor |
|---|---|---|
| 0 | The premise | One screen. The hook + the disclaimer. |
| 1 | Meet your cellular thermometer | HSF1, the master heat-shock regulator (PDB structure, expression) |
| 2 | When heat arrives | HSF1 activation — HSP90 abandonment → trimerization → DNA binding |
| 3 | The first line of defense | HSP70 (HSPA1A) — the chaperone that refolds damaged proteins |
| 4 | When proteins melt | A real thermosensitive enzyme — visual denaturation (clearly labeled as visualization, not MD) |
| 5 | Heat shock proteins across the body | Tissue expression heatmap from the Human Protein Atlas |
| 6 | When the genome strains | ClinVar variants in HSPA1A / HSF1 — known clinical significance |
| 7 | The full pathway | Reactome "Cellular response to heat stress" diagram (R-HSA-3371556, candidate) |
| 8 | Bridge to HeatThreshold | A WBGT input → predicted core-temperature progression → chapter checkpoint markers |
The full phased build plan lives in PROJECT.md. The visual & UX design system — including per-chapter prompts the Antigravity Stitch MCP can consume — lives in DESIGN.md. The running bibliography lives in references.md. Project context for Antigravity 2.0 lives in AGENTS.md. The build is being tracked as a developer-relations case study; the framework lives in DEVREL.md.
- Build-time data fetching. A set of Python scripts in
scripts/calls the Science Skills CLIs (PDB, AlphaFold, PubMed, HPA, ClinVar, Reactome) viauv run. Each script writes plain JSON or downloaded structure files into the committeddata/directory. - Static page reads from
data/. No server, no build framework initially. Plain HTML + CSS + ES modules. 3Dmol.js for protein rendering, IntersectionObserver-driven CSS for scroll triggers, Reactome's own SVG export for the pathway diagram. - All data is committed. The page works offline; rebuilding the data is an explicit, separate workflow.
- Every claim is cited. Inline footnote-style links to PubMed, PDB, AlphaFold, ClinVar entries. Sources also collected in references.md.
- No medical or clinical claims. Not advice. Not diagnostic. Not predictive of individual human outcomes. If you are experiencing heat illness, call a doctor, not a JavaScript app.
- No PII, no biometric ingestion, no accounts, no analytics. The page is purely static and educational.
- No physical simulation. The denaturation animations are visualizations of denaturation, not molecular dynamics. This distinction is surfaced in the UI, never buried.
- No HeatThreshold integration without explicit say-so. This is a sibling experiment, not a feature.
Open index.html in a browser. There is no build step for the page itself. To refresh the underlying scientific data, run the fetcher scripts in scripts/ — they require uv, jq, and the Science Skills plugin installed in Antigravity (see PROJECT.md § Phase 0).
Live at https://heat-protein-lab.pages.dev/. The two project-anchored DevRel posts are mirrored at /posts/ and canonical on craigmerry.com.
The site auto-deploys to Cloudflare Pages via .github/workflows/cloudflare-pages.yml on every push to main. One-time operator setup:
- Create a Cloudflare Pages project named
heat-protein-lab. - Generate a Cloudflare API token scoped to Account → Cloudflare Pages → Edit.
- Add two secrets to this repo:
CLOUDFLARE_API_TOKENandCLOUDFLARE_ACCOUNT_ID.
The deploy workflow stages a publishable subset (index.html, src/, data/, LICENSE, README.md, plus _headers/_redirects if present) into ./dist and uploads that. Internal directories (notes/, scripts/, .github/, the design/project/agents docs) stay in the repo but out of the deployed site.
_headers at the repo root configures Content-Type and caching for the deployed site — most importantly, it serves .cif.gz mmCIF files as application/gzip without Content-Encoding: gzip, so the in-browser DecompressionStream decode path in src/main.js keeps working.
MIT for the project code. The fetched data files in data/ carry the licenses of their upstream sources (RCSB PDB, EBI AlphaFold, NCBI PubMed/ClinVar/dbSNP, Human Protein Atlas, Reactome). See references.md for source attribution and PROJECT.md § Data licensing for details.
Craig Merry (craigm26 on GitHub). Personal experiment; PRs welcome, pace will be slow.