The AlphaFold companion — inspect, validate, and compare AlphaFold predictions, entirely in your browser.

Inspect any AlphaFold model (confidence, disorder, PAE, domains). Validate it against the real experimental structure (RMSD, TM-score, GDT-TS, lDDT). See where it was confidently wrong.

---

OpenFoldUI is a companion to AlphaFold. Point it at any protein and it will:

• Inspect the AlphaFold model on its own — pLDDT confidence, predicted disordered regions, PAE, and domains — for any of the 200M+ proteins, no experimental structure required.
• Compare it against the best experimental structure of the same protein — Kabsch superposition, RMSD · TM-score · GDT-TS · lDDT, and the scientific payload: the correlation between AlphaFold's confidence (pLDDT) and its actual per-residue error.
• Fold a raw sequence (ESMFold, or your own AlphaFold/ColabFold endpoint) and feed it straight into a comparison.

Everything runs in your browser. There is no backend.

The headline question it answers: where was the model confident but wrong — high pLDDT yet far from the real structure.

Quick start

npm install
npm run dev          # http://localhost:5173

Type a protein name or UniProt accession (e.g. p53 or P04637), click an example, upload your own model + reference, or fold a sequence. See the metrics, the calibration charts, the per-residue tracks, and a 3D overlay you can recolour.

Run the engine headlessly on real data (no browser):

npx vite-node scripts/validate.ts                # p53 (P04637) vs 2OCJ
npx vite-node scripts/validate.ts P24941 6q4g    # any UNIPROT [PDB]

What it does

Core comparison

• AlphaFold-DB lookup, best-experimental-structure selection (X-ray > coverage > resolution), SIFTS numbering done right (per-atom UniProt numbers from PDBe updated mmCIF).
• From-scratch engine: Kabsch superposition, RMSD · TM-score · GDT-TS · lDDT, per-residue deviation, and Spearman(pLDDT, error) — validated against canonical TM-align (WASM) to ~0.001 TM-score.
• Mol* 3D overlay (deviation / pLDDT colouring, bound-ligand display, PNG screenshot).

Deeper science

• lDDT (superposition-free) with the correct pLDDT-vs-lDDT calibration plot.
• PAE domain decomposition + per-domain RMSD (catches "domains right, orientation wrong").
• Distance-difference matrix (contact-map comparison) for topological errors.
• Secondary-structure breakdown, binding-site impact, B-factor vs deviation, divergent regions, workspace-wide confidence calibration.
• UniProt feature track (domains, sites, modifications, variants); multi-state comparison with a conformational-landscape readout ("which state did AlphaFold predict, and how flexible is the protein?").
• Multi-chain / assembly analysis: per-chain RMSD/TM/lDDT across an oligomer's copies and assembly-interface deviation vs the core.
• Predictor face-off: fold the same sequence with ESMFold and benchmark AlphaFold vs ESMFold vs experiment, metric by metric.
• Dataset analytics: score a whole list of accessions at once — aggregate pLDDT distribution, disorder, ranked table, CSV export.

Bring your own data

• Upload .pdb/.cif models + references; align by author numbering, UniProt, or sequence (Needleman–Wunsch).
• Fold a sequence with a persistent queue — ESMFold by default, or point it at your own AlphaFold/ColabFold endpoint.
• Batch mode over many IDs with live distributions; CSV/column import.

A real workspace

• IndexedDB history with favorites, tags, projects, notes, annotations; a dashboard with stats, sparklines, and a tag/project filter.
• Command palette (⌘K), keyboard shortcuts, onboarding + interactive guided tour (learn the metrics on a live p53 comparison), undo, jump-back, dark mode, installable PWA.
• Example gallery: curated, one-click, shareable comparisons (every result reproduces from its URL — no account).
• Export everything: Excel/CSV, per-comparison reports (HTML→PDF), a paper bundle (Markdown + SVG figures + BibTeX), replication logs, JSON backup, and a full .zip.

Architecture

Pure, headless engine at the core; everything else is a thin layer around it.

Layer	Path	Notes
Engine	src/engine/	Pure math + parsing. No DOM, no network, no Mol*. Fully unit-tested.
API	src/api/	One typed module per source (UniProt, PDBe, RCSB, AlphaFold) + pipeline.ts.
Viewer	src/viewer/	Lazy, error-boundaried Mol* wrapper.
Charts	src/charts/	Observable Plot.
Workspace	src/workspace/	IndexedDB persistence, export, reports, paper bundle.
Batch / Fold / Search	src/batch/, src/fold/, src/search/	Bounded-concurrency pools; ESMFold; Foldseek.
UI shell	src/ui/	Sidebar, top bar, command palette, toasts, onboarding.

SPEC.md is the source of truth · CLAUDE.md the summary · BUILD_PLAN.md the status · docs/PHASE0.md the de-risking + real-data validation.

Tech stack

React · TypeScript (strict) · Vite · Vitest · Mol* · Observable Plot · ml-matrix · idb · tmalign-wasm · ESMFold/Foldseek (best-effort) — no backend, no tracking, no accounts.

Testing

npm test          # 206 unit tests (engine, api, workspace, batch, fold, search)
npm run typecheck # tsc --noEmit
npm run build     # production build (Mol* + tmalign are lazy chunks)

The from-scratch engine is verified two ways: known-answer unit tests for every metric, and a real-data run (p53/2OCJ → RMSD 0.51 Å, TM 0.99, GDT 0.99) that agrees with TM-align.

Commands

Command	What it does
npm run dev	Dev server — the full app
npm test	Vitest suite (network-free)
npm run typecheck	tsc --noEmit
npm run build	Production build to dist/
npx vite-node scripts/validate.ts [UNIPROT] [PDB]	Headless real-data validation

License

MIT

---

Made by Ben Vaccaro