🧬 AlphaFraud

Catch AlphaFold where the fold is wrong: weekly, automatically, on freshly deposited human structures.

🌐 Website

alphafraud.mdeller.com

✉️ Contact

🐙 GitHub

AlphaFraud watches the RCSB PDB for newly deposited human protein structures, matches each one to its blind AlphaFold prediction from the AlphaFold Structure Database (EBI), superposes them, and quantifies exactly where the prediction disagrees with the experiment. AlphaFold2 was trained only on PDB chains released before 2018-04-30, so any human structure deposited after that date is something the model predicted without ever seeing its coordinates: the ideal test set.

Why it matters: the claim that "AlphaFold has solved protein folding" is only testable on structures the model never trained on, and those arrive continuously as the PDB grows. AlphaFraud turns that stream into a standing experiment: every week it surfaces the structures AlphaFold got most wrong, and ranks the "confidently wrong" cases (high pLDDT, low agreement) where the model was sure and still missed. It is useful for: structural biologists auditing AlphaFold reliability, method developers looking for hard cases, and anyone testing whether predicted models are safe to build on.

A live instance runs at alphafraud.mdeller.com, having backfilled the entire post-cutoff archive (96,827 human protein entities, 82,796 analysed) and updating every week. The landing page is a cumulative dashboard across every processed week; the worst offenders so far are a roster of amyloid-forming proteins (transthyretin, β2-microglobulin, islet amyloid polypeptide) that AlphaFold predicts as confidently folded but which crystallise or cryo-EM in a completely different aggregated form.

✨ Features

Capability	What it does
Automated PDB watch	Weekly systemd timer fires after the PDB's Wednesday 00:00 UTC release; resumes from the last processed run
Human, post-cutoff filter	RCSB Search for taxonomy 9606 protein entities deposited after AlphaFold's 2018-04-30 training cutoff
Blind-model matching	Maps each entity to its UniProt accession and AlphaFold DB model, handling fragments, mutants and missing loops via sequence alignment
Sequence-novelty score	Max percent identity to any pre-cutoff PDB chain, so genuinely unseen sequences are flagged (not just easy homologs)
Full metric suite	Global, local, backbone, per-domain and confidence-calibration metrics (see below)
Confidence audit	pLDDT vs actual lDDT calibration, and a PAE-honesty check comparing AlphaFold's self-reported error to the observed error
Two-tier archive backfill	A fast TM-score screen across every structure, running the full metric suite only on the disagreements; makes the whole ~96k-entity archive tractable
Structural imagery	Every worst offender is drawn as a deviation-coloured Cα ribbon (the experimental structure, coloured residue-by-residue by its distance from the AlphaFold model, on an absolute-Ångström scale) on the leaderboard, entry page and weekly highlights; an interactive 3Dmol.js viewer with an optional translucent AlphaFold "ghost" overlay; hover-preview thumbnails on the scatter; and a "divergence" ribbon banner in the header
On-demand Calculate tab	Type any human PDB ID and the full pipeline runs on demand (fetch AlphaFold model, superpose, score, render) to return the auto-example panel, with strict autocomplete (exists + human + has an AlphaFold model) and no LLM
Home & live Statistics	A Home navigation page outlining every tab, and an exhaustive Statistics page grouping all KPIs, coverage, runs, server (CPU/memory/uptime), storage/database and API stats into themed panels
Cumulative dashboard	Default landing page aggregating every processed week: the "fraud quadrant" scatter, metric histograms, a weekly trend, and a browsable per-week / per-structure archive
Branded, mobile-responsive report	Flask app with the signature scatter, heatmaps, per-domain tables and an all-time leaderboard; every plot has a CSV export and a plain-language explanatory panel that defines each metric twice over (a lay sentence and the underlying equation), and every table exports to CSV
One-command deploy	Provisioning, deploy and one-shot release scripts for a DigitalOcean droplet (gunicorn, nginx, certbot TLS)

🏗️ Architecture and workflow

One controller (AlphaFraud.py, subcommands) drives a small Python package; a Flask app serves the results. Two runtimes share one SQLite file: a systemd timer runs the weekly pipeline (writes the database and the ribbon/coords assets), and an always-on gunicorn process serves the site by reading that same database live.

Module	Responsibility
`pdb.py`	RCSB Search (new human protein entities) + Data GraphQL (UniProt mapping, dates, method)
`afdb.py`	AlphaFold DB fetch: the blind model + PAE, choosing the fragment covering the entity's span
`novelty.py`	max % identity to any pre-cutoff chain (so genuinely unseen sequences are flagged)
`compare.py`	the full metric suite on the aligned span (TM, lDDT, GDT, RMSD, per-residue deviation, PAE honesty)
`ribbon.py`	Cα-deviation SVG + 3D-viewer coords + the AlphaFold "ghost" (vendored, offline)
`annotate.py` / `analysis.py`	CATH / SCOP / citation enrichment and the `/analysis` snapshot (enrichment, clustering, PCA)
`examples_auto.py`	the deterministic, no-LLM panel builder (match or failure framing)
`calc_index.py` / `calculate.py`	the Calculate tab's strict autocomplete index + on-demand single-entity compute
`db.py` / `report.py` / `webapp.py`	SQLite persistence, Plotly figure JSON, and the Flask routes

The weekly pipeline (AlphaFraud.py run, one entity at a time):

for entity in RCSB.search(human, protein, deposited_after = 2018-04-30):
    if entity in db: continue
    meta = RCSB.graphql(entity)                       # UniProt, dates, method, chains
    if not meta.single_uniprot: skip("antibody / fusion / chimera"); continue
    af = AlphaFoldDB.model(meta.uniprot)              # the blind prediction (fragment covering the span)
    if af is None:               skip("no AlphaFold model"); continue
    exp, model = download(entity), download(af)
    m        = compare(exp, model)                    # TM, lDDT, GDT, RMSD, per-residue dev, PAE honesty
    novelty  = max_identity_to_pre_cutoff(meta.sequence)
    fraud    = mean_i( pLDDT_i/100 * clamp(dev_i, 0, 15Å) / 15 )   # confidence-weighted error
    confidently_wrong = (mean_pLDDT > 70) and (TM < 0.5)          # sure, and still wrong
    ribbon.render(exp, model)                         # deviation SVG + coords + AF ghost
    db.store(entity, m, novelty, fraud, confidently_wrong)

analysis.rebuild()            # /analysis enrichment snapshot
rebuild_weekly_examples()     # auto 'example of the week' + weekly archive
calc_index.refresh(metas)     # keep the Calculate autocomplete index current

The always-on web app (gunicorn) reads the same SQLite and renders cached Plotly JSON:

GET /            -> cumulative "fraud quadrant" dashboard (figures cached, short TTL)
GET /entry/<id>  -> per-structure heatmaps, calibration, per-domain metrics
GET /examples    -> positive control + auto example-of-the-week + curated deep-dives
GET /calculate   -> type a human PDB id  ->  (reuse if known | compute on demand)  ->  panel

The Calculate tab reuses the very same process_entity() as the pipeline, run in an isolated subprocess so a heavy computation never blocks a web worker, and stored apart so it never pollutes the science:

POST /calculate/run { pdb }:
    entity = resolve_human_entity(pdb)                # strict index, else a live RCSB check
    if not qualifies:  return reason                 # not human / no single UniProt / no AF model
    if entity already computed:  return "ready"      # cached -> instant
    mark(entity, "computing"); spawn AlphaFraud.py calculate <entity>   # same pipeline, subprocess
    return "computing"                               # the browser polls, then injects the panel

🎯 Metric suite

Every matched structure is scored on the aligned residue span (all pure-Python except CATH, which is a lookup):

Family	Metrics
Global fold agreement	TM-score (4 normalisations), Cα / backbone / all-atom / core RMSD, GDT_TS, GDT_HA, MaxSub, structural overlap
Local, superposition-free	lDDT (global and per-residue), contact-map Jaccard / precision / recall, distance-matrix difference, CAD-score
Backbone and secondary structure	Q3 secondary-structure agreement, mean Δφ / Δψ, radius-of-gyration difference
Domains	Per-CATH-domain and per-PAE-cluster-domain TM and RMSD (exposes inter-domain hinge errors a global fit hides)
Confidence calibration	pLDDT vs lDDT correlation, PAE vs observed-error correlation, overconfident-pair fraction
Composite	FRAUD score (confidence-weighted error) and a confidently-wrong flag (mean pLDDT > 70 yet TM-score < 0.5)

📋 Requirements

Python 3.11 or newer
The packages in requirements.txt (requests, tenacity, biotite, tmtools, numpy, scipy, pandas, plotly, flask, gunicorn, python-dotenv), all pip-installable with no compiled toolchain
Outbound HTTPS to search.rcsb.org, data.rcsb.org, files.rcsb.org and alphafoldebi.ac.uk
For deployment: a Linux droplet with root/SSH access, a domain, and DNS pointed at it

🔧 Installation

git clone https://github.com/bellcheddar/AlphaFraud.git
cd AlphaFraud
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python3 AlphaFraud.py init          # create dirs, DB schema, vendor plotly.min.js

🚀 Usage

# Preview discovery for a window without writing anything
python3 AlphaFraud.py run --since 2024-01-01 --until 2024-01-10 --limit 6 --dry-run

# Process a window for real (omit --since to resume from the last run)
python3 AlphaFraud.py run --since 2024-01-01 --limit 50

# Backfill a historical range (full metrics on everything)
python3 AlphaFraud.py backfill --from 2023-01-01 --to 2023-12-31

# Two-tier backfill: fast TM-score screen on everything, full metrics only on the
# disagreements (TM below --tm-threshold). Far cheaper on time, disk and API load.
python3 AlphaFraud.py backfill --from 2023-01-01 --to 2023-12-31 --two-tier --tm-threshold 0.7

# Two-tier backfill of the ENTIRE post-cutoff archive (~96k entities, monthly chunks,
# resumable). Screens all, fully analyses only the tail AlphaFold got wrong.
python3 AlphaFraud.py backfill --all --two-tier

# --workers N adds a download thread pool (network-bound work parallelises well). Single-
# threaded is the reliable default for long unattended runs; workers give a partial speedup
# but are more sensitive to the droplet's core count.
python3 AlphaFraud.py backfill --all --two-tier --workers 6

# Database summary + worst-offenders leaderboard
python3 AlphaFraud.py status

# Local dev web server (production uses gunicorn)
python3 AlphaFraud.py serve --port 8000

Command	Purpose
`init`	Create directories, the SQLite schema, and vendor `plotly.min.js`
`run`	Discover, match, score and store one deposit window (`--since`, `--until`, `--limit`, `--dry-run`)
`backfill`	Process a historical window (`--from`, `--to`); `--two-tier` screens by TM-score and fully analyses only disagreements below `--tm-threshold`; `--all` covers the entire post-cutoff archive in resumable monthly chunks
`status`	Print a database summary and the current leaderboard
`serve`	Run the Flask app locally (`--port`, `--debug`)

📊 Output

The web app serves live from SQLite. Every plot carries a plain-language explanatory panel (what the axes mean, the cutoffs, a take-home summary, and a per-metric breakdown giving both a lay explanation and the equation, e.g. TM = (1/L) Σᵢ 1/(1 + (dᵢ/d₀)²)) plus a CSV export; every table exports to CSV too.

The site is live at alphafraud.mdeller.com.

Route	Shows
`/`	Home — a concise navigation page: the project's purpose, headline KPIs, and a colour-coded card per tab with a "what you'll learn" line
`/all`	Cumulative "All structures" dashboard across every processed week: KPIs, the pLDDT-vs-TM fraud-quadrant scatter, metric histograms, the weekly trend, and a top-ranked table. A pulldown jumps to any single week
`/week/<label>`	One release week, same layout scoped to that batch
`/entry/<id>`	Per-structure detail: per-residue error tracks, calibration scatter, distance-matrix and PAE-vs-observed heatmaps, per-domain breakdown, and every metric
`/leaderboard`	The all-time worst AlphaFold failures across every processed week
`/analysis`	Structural deep dive over the worst offenders: fold/family enrichment (Wilson-CI + Fisher), sequence-similarity clustering, failure-mode PCA, theme flags, per-superfamily blind-spot scorecards, conformational-heterogeneity detection, a "new this week" ribbon, and RCSB + DOI links on every structure
`/examples`	Curated storytelling layer: a positive control (an excellent AlphaFold match), an auto-generated "example of the week" plus a full weekly-release archive, and hand-authored deep-dive panels with interactive viewers (see below)
`/calculate`	Type any human PDB ID and get the same panel computed on demand: strict autocomplete (exists + human + has an AlphaFold model), full pipeline in an isolated subprocess, no LLM (see below)
`/archive`	Every PDB release week since the 2018 cutoff, at weekly granularity, each linking to that week's structures and confidently-wrong count
`/stats`	An exhaustive live Statistics page: themed panels for catch/quality, coverage & schedule, runs & compute, server (CPU/memory/uptime), storage & database, data sources & APIs, and traffic
`/api/week/<label>`, `/api/leaderboard`, `/api/entry/<id>`	JSON for external tools

🔬 Examples

The Examples tab turns raw catches into readable stories and deliberately shows both sides of AlphaFold's record, not just the failures:

A positive control. The page opens with the one thing the rest of the site does not: an excellent match. Human Artemis (DCLRE1C), a genuinely novel post-cutoff nuclease (only 29% identity to anything AlphaFold trained on), was predicted blind to TM-score 0.996, lDDT 0.985 and 0.45 Å Cα-RMSD at a justified mean pLDDT of 95, and AlphaFraud scores it FRAUD 0.02. It is the proof that the pipeline certifies good predictions, not just hunts for bad ones.
An auto "example of the week". After every weekly run the pipeline picks that release's most instructive confidently-wrong catch and writes its panel deterministically from the database (no LLM), badged and dated.
A complete weekly archive. Every PDB release week since the 2018 cutoff is listed: weeks with a catch expand to a full panel with its Cα-deviation ribbon, and the quiet weeks (where AlphaFold was never overconfident) are recorded too, collapsed to the most recent by default.
Curated deep-dives. Hand-authored panels across the failure spectrum: ordinary single-chain globular folds AlphaFold gets subtly wrong (right fold, wrong conformation or assembly — the PP2A scaffold, STIM1, Aha1, WWP2) and the dramatic whole-fold catastrophes (TMEM106B, alpha-synuclein, transthyretin, factor X). Each panel carries live metrics, an interactive 3Dmol.js viewer (deviation colouring plus the AlphaFold "ghost") and a structure / mechanism / disease write-up.
Linked from the scatter. The curated examples are overlaid on the home fraud-quadrant plot as pulsing red (failure) and green (accurate) markers, each labelled with its protein name and clickable straight through to its panel.

Each panel is summarised below — click any title to open its live, interactive version. Metrics: TM-score and lDDT (1.0 = identical), Cα-RMSD (0 Å = identical), pLDDT (AlphaFold's own confidence, 0–100), FRAUD (confidence-weighted error, higher = worse), novelty (100% = wholly unseen before AlphaFold's cutoff).

🟢 Protein Artemis (DCLRE1C) · `6TT5_1`

Field	Detail
Take-home	A novel post-cutoff nuclease AlphaFold predicted blind almost perfectly — the control proving AlphaFraud certifies good predictions, not only failures
AlphaFold vs experimental	_{Experimental structure coloured by Cα deviation from AlphaFold (blue = agree, red = diverges) — click for the interactive 3D viewer with the AlphaFold "ghost" overlay}
Link to live example	Open the interactive panel →
KPIs	TM 0.996 · lDDT 0.985 · pLDDT 95.2 · Cα-RMSD 0.45 Å · FRAUD 0.02 · novelty 71%
Stats	11 depositions, 0 confidently wrong · SS-Q3 92.5% · pLDDT↔lDDT +0.56 (calibrated)
Structure	Metallo-β-lactamase + β-CASP nuclease domain with a Zn active site; the disordered regulatory C-terminus sits (correctly) outside the crystallised fold
Why AlphaFold gets it right	A single, conserved globular domain with a deep MSA and one native conformation — AlphaFold's ideal regime; confidence is earned (PAE-overconfident 0.00)
Biology & disease	NHEJ endonuclease that opens V(D)J hairpins with DNA-PKcs; loss of function causes radiosensitive SCID / Omenn; pursued as a radiosensitiser drug target
Key fact	29% identity to anything pre-cutoff yet GDT_TS 99.2 — genuine generalisation, not memorisation

🔴 PP2A scaffold subunit A / PR65 (PPP2R1A) · `8UWB_3`

Field	Detail
Take-home	Near-maximal confidence (pLDDT 97) on a flexible 15-HEAT solenoid whose curvature is wrong for 99% of residues
AlphaFold vs experimental	_{Experimental structure coloured by Cα deviation from AlphaFold (blue = agree, red = diverges) — click for the interactive 3D viewer with the AlphaFold "ghost" overlay}
Link to live example	Open the interactive panel →
KPIs	TM 0.293 · lDDT 0.311 · pLDDT 96.8 · Cα-RMSD 15.0 Å · FRAUD 0.77 · novelty 0%
Stats	26 depositions, 1 confidently wrong (this one) · pLDDT↔lDDT −0.13 (confidence carries no signal)
Structure	15 tandem HEAT repeats forming a horseshoe α-solenoid — a spring, not a fixed ruler
Why AlphaFold gets it wrong	Local helices are right, but tiny per-repeat hinge errors compound across 15 repeats into a grossly wrong global shape — a curvature failure, not a fold failure
Biology & disease	Sole scaffold of the PP2A tumour-suppressor phosphatase; cancer hotspots (P179R, R183, S256F…) cluster in the B56-binding repeats
Key fact	The source paper needed an experimental trimer structure to guide the prediction

🔴 Stromal interaction molecule 1 / CC1 (STIM1) · `6YEL_1`

Field	Detail
Take-home	A bistable three-helix Ca²⁺ switch; AlphaFold committed to one conformer, the structure is the other
AlphaFold vs experimental	_{Experimental structure coloured by Cα deviation from AlphaFold (blue = agree, red = diverges) — click for the interactive 3D viewer with the AlphaFold "ghost" overlay}
Link to live example	Open the interactive panel →
KPIs	TM 0.389 · lDDT 0.341 · pLDDT 78.9 · Cα-RMSD 35.8 Å · FRAUD 0.75 · novelty 2%
Stats	1 deposition, confidently wrong · SS-Q3 69.6% (secondary structure right, packing wrong)
Structure	CC1 is a compact antiparallel three-helix bundle that toggles between clamped (resting) and extended (active) states
Why AlphaFold gets it wrong	Right fold class, wrong tertiary packing — a single-answer predictor cannot represent a two-state switch
Biology & disease	Gates store-operated Ca²⁺ entry via Orai1; gain-of-function → Stormorken / tubular aggregate myopathy; loss → CRAC-channelopathy
Key fact	35.8 Å RMSD with correct secondary structure = a conformer error, not a misfold

🔴 Activator of Hsp90 ATPase 1 / Aha1 (AHSA1) · `7DME_1`

Field	Detail
Take-home	Two well-folded domains bolted together at the wrong angle — a flexible-linker ensemble collapsed to one guess
AlphaFold vs experimental	_{Experimental structure coloured by Cα deviation from AlphaFold (blue = agree, red = diverges) — click for the interactive 3D viewer with the AlphaFold "ghost" overlay}
Link to live example	Open the interactive panel →
KPIs	TM 0.432 · lDDT 0.748 · pLDDT 83.3 · Cα-RMSD 23.4 Å · FRAUD 0.82 · novelty 0%
Stats	4 depositions, 1 confidently wrong · pLDDT↔lDDT +0.33 (high confidence hides the joint)
Structure	N- and C-terminal domains on a partly-ordered linker; each folds correctly in isolation
Why AlphaFold gets it wrong	Domains are right (lDDT 0.75) but their relative orientation is wrong (TM 0.43) — a conformational-ensemble failure, not a misfold
Biology & disease	Strongest stimulator of the Hsp90 ATPase; drives ΔF508-CFTR maturation (cystic-fibrosis target) and pathological tau
Key fact	The pieces are right; only the assembly instruction is wrong

🔴 NEDD4-like E3 ligase WWP2 / WW module (WWP2) · `6RSS_1`

Field	Detail
Take-home	A tiny two-Trp β-sheet AlphaFold folds fine, but the deposited chain's inter-module geometry is flexible — and, tellingly, the model was honestly unsure
AlphaFold vs experimental	_{Experimental structure coloured by Cα deviation from AlphaFold (blue = agree, red = diverges) — click for the interactive 3D viewer with the AlphaFold "ghost" overlay}
Link to live example	Open the interactive panel →
KPIs	TM 0.46 · lDDT 0.503 · pLDDT 74.5 · Cα-RMSD 21.9 Å · FRAUD 0.71 · novelty 8%
Stats	7 depositions, 1 confidently wrong · pLDDT↔lDDT +0.15 (lowest confidence of the set)
Structure	NEDD4-family HECT E3 ligase; the WW domain is a ~38-residue three-stranded antiparallel β-sheet that reads PPxY motifs
Why AlphaFold gets it wrong	Uncertain relative orientation between small folded modules — and its low pLDDT half-admits the uncertainty
Biology & disease	Ubiquitinates SMAD7, PTEN, TP53, OCT4; isoform balance shifts drive EMT and metastasis
Key fact	The "honest" failure of the set — low confidence tracking real uncertainty

🔴 Transmembrane protein 106B (TMEM106B) · `7U14_1`

Field	Detail
Take-home	AlphaFold built a tidy globular β-sandwich at pLDDT 94; the real structure is a novel brain amyloid — every ordered residue wrong
AlphaFold vs experimental	_{Experimental structure coloured by Cα deviation from AlphaFold (blue = agree, red = diverges) — click for the interactive 3D viewer with the AlphaFold "ghost" overlay}
Link to live example	Open the interactive panel →
KPIs	TM 0.157 · lDDT 0.536 · pLDDT 94.4 · Cα-RMSD 27.1 Å · FRAUD 0.87 · novelty 100%
Stats	29 depositions, 27 confidently wrong · pLDDT↔lDDT 0.03 (confidence uninformative)
Structure	Native luminal domain is a fibronectin-III β-sandwich; the deposited form is a five-layered cross-β amyloid core
Why AlphaFold gets it wrong	Amyloid is a self-templating polymer state absent from training, and the sequence is 100% novel — nothing steers it off its globular prior
Biology & disease	Lysosomal protein; top GWAS hit for FTLD-TDP; its C-terminal domain forms fibrils across ageing and neurodegeneration
Key fact	100% sequence-novel — no pre-2018 fibril template existed

🔴 Alpha-synuclein (SNCA) · `9A1Q_1`

Field	Detail
Take-home	A protein with no single fold; AlphaFold's confidence is actively anti-correlated with accuracy
AlphaFold vs experimental	_{Experimental structure coloured by Cα deviation from AlphaFold (blue = agree, red = diverges) — click for the interactive 3D viewer with the AlphaFold "ghost" overlay}
Link to live example	Open the interactive panel →
KPIs	TM 0.138 · lDDT 0.177 · pLDDT 75.2 · Cα-RMSD 36.3 Å · FRAUD 0.71 · novelty 0%
Stats	210 depositions, 206 confidently wrong — the archive's most-repeated failure · pLDDT↔lDDT −0.29
Structure	140-residue intrinsically disordered monomer; the deposited forms are cryo-EM cross-β fibrils
Why AlphaFold gets it wrong	No native fold to predict, and polymorph-specific packing is not in the sequence — where it is more confident it is measurably more wrong
Biology & disease	Presynaptic protein; fibrils define Parkinson's, DLB and multiple system atrophy; PD mutations A30P/E46K/H50Q/G51D/A53T
Key fact	210 depositions, 206 confidently wrong — the most-deposited catch on the site

🔴 Transthyretin (TTR) · `9BZS_1`

Field	Detail
Take-home	pLDDT 98 on the native tetramer fold; the deposited cardiac amyloid fibril scores TM 0.21 — one of the starkest confidently-wrong cases
AlphaFold vs experimental	_{Experimental structure coloured by Cα deviation from AlphaFold (blue = agree, red = diverges) — click for the interactive 3D viewer with the AlphaFold "ghost" overlay}
Link to live example	Open the interactive panel →
KPIs	TM 0.214 · lDDT 0.489 · pLDDT 98.0 · Cα-RMSD 22.8 Å · FRAUD 0.95 · novelty 0%
Stats	160 depositions, 34 confidently wrong · pLDDT↔lDDT −0.16
Structure	Native 127-aa β-sandwich homotetramer with thyroxine sites; the amyloid form is a cross-β fibril from cleaved segments
Why AlphaFold gets it wrong	A wrong-state failure: pLDDT measures confidence in the known native fold, not whether a deposition is that state; TTR is 0% novel, squarely in training
Biology & disease	Thyroxine/retinol transporter; ATTR amyloidosis (wild-type cardiac + hereditary V30M); uniquely druggable (tafamidis, patisiran)
Key fact	FRAUD 0.95 — near the very top of the whole site

🔴 Coagulation factor X (F10) · `6Q9F_2`

Field	Detail
Take-home	Not a refold but a wrong-context capture — a 39-residue EGF1 fragment threaded through a partner enzyme's active site
AlphaFold vs experimental	_{Experimental structure coloured by Cα deviation from AlphaFold (blue = agree, red = diverges) — click for the interactive 3D viewer with the AlphaFold "ghost" overlay}
Link to live example	Open the interactive panel →
KPIs	TM 0.127 · lDDT 0.468 · pLDDT 93.5 · Cα-RMSD 6.6 Å · FRAUD 0.37 · novelty 0%
Stats	47 depositions, 26 confidently wrong · pLDDT↔lDDT −0.59
Structure	Modular zymogen (Gla–EGF1–EGF2–protease); the worst deposition is a linear EGF1 peptide in AspH's channel, not a folded module
Why AlphaFold gets it wrong	A context failure, not aggregation — the RMSD is modest (6.6 Å) but TM collapses because a linear fragment has no compact fold to superpose
Biology & disease	Central coagulation zymogen; activated factor Xa is the target of rivaroxaban, apixaban and edoxaban
Key fact	The non-amyloid contrast — the error is missing biological context (fragment / complex / PTM), not misfolding

🧪 Calculate

The Calculate tab runs the whole pipeline on demand for any human PDB ID you type and returns the same panel format as the auto examples — KPIs, the Cα-deviation ribbon, the interactive 3D viewer and a deterministic write-up — with no AI-generated text.

Strict autocomplete. The input only completes to entries that (1) exist, (2) are human, and (3) have an AlphaFold model. That comes from a precomputed index — seeded for free from the corpus (the ~83k entities already carrying a confirmed AlphaFold model), back-filled with a one-off pre-cutoff human sweep, and refreshed automatically by every weekly run. Press Tab to complete.
Any human structure, blind or not. Post-cutoff entries are genuine blind tests; pre-cutoff entries compute too, with a note that AlphaFold may have seen them in training (so it is not a blind prediction).
Clear reasons on rejection. If an entry doesn't qualify you are told exactly why: not a human protein structure, no single UniProt mapping (antibody / fusion / chimera / synthetic), or no AlphaFold model for the sequence.
Right or wrong, honestly. A good prediction renders the green "AlphaFold gets it right" panel (like the positive control); a confidently-wrong one renders the failure panel — the same deterministic builder (examples_auto.build) decides which from the metrics.
Isolated and cached. On-demand results are stored under a separate run and status, so they never enter the leaderboard, stats, analysis or archive; the heavy compute runs in an isolated subprocess; and a repeat request for the same structure is instant.

🧮 Coverage and skipped entities

Not every deposited entity can be compared. AlphaFold DB holds a single model per human UniProt sequence (monomers only), so anything that does not map cleanly to one accession, has no model covering its resolved range, or is too small or too large to compare is recorded as skipped rather than forced through.

The completed full-archive backfill (finished 2026-07-18) processed 96,827 post-cutoff human entities: 82,796 analysed (85.5%; 67,314 fast-screened plus 15,482 fully compared), 1,941 confidently wrong (2.3% of analysed, 431 of them sequence-novel), 13,908 skipped (14.4%), and 123 unrecoverable errors (0.1%, broken coordinate-less files). The skips break down as:

Count	Share	Reason	Nature
11,875	85%	No single UniProt mapping — antibodies, chimeras, engineered constructs, fusion proteins, synthetics	Structural fact: no single AlphaFold model exists to compare against
1,186	9%	No AlphaFold model covering the residue range — mostly giant multi-domain proteins (titin-scale) whose resolved fragment falls outside AlphaFold DB's fragmented coverage, plus isoforms it lacks	Coverage gap in AlphaFold DB
464	3%	Too few residues aligned (<10) — tiny peptides or mostly-unresolved chains	Nothing meaningful to compare
381	3%	Structure too large (>40 MB) — the out-of-memory guard for the 3.8 GB droplet	Resource limit (our cap)
2	—	Sequence too short (<3 residues)	Degenerate

The 85% is exactly what the design intends: because AlphaFold DB is monomer-and-single-sequence only, anything that does not map to one human UniProt accession (an antibody Fab, a designed fusion, a chimera) has no counterpart model, so it is flagged and skipped rather than mis-compared.

Skipped entities are terminal by design: retry-errors only re-runs status='error', and the resumable weekly and backfill runs skip anything already recorded. Four of the five buckets are permanent structural properties, so re-running would simply re-skip them. The two worth revisiting are the 381 oversized structures (a memory limit of the small droplet, not a true incomparability, recoverable by raising MAX_STRUCT_BYTES or upsizing the box) and the multi-domain subset of the 1,186 missing-model bucket (recoverable with better fragment selection for AlphaFold models above 2,700 residues).

🛠️ Web deployment

Live at alphafraud.mdeller.com. The report is served by a Flask app (gunicorn behind nginx with a Let's Encrypt certificate) and the weekly pipeline runs from a systemd timer on the same droplet.

# 1. From your Mac: copy .env.example to .env and fill in DROPLET_SSH / DROPLET_PATH / SERVER_NAME
cp .env.example .env

# 2. Push the code to the droplet
bash deploy/deploy.sh

# 3. On the droplet as root: one-time provisioning (packages, user, venv, services, nginx, TLS)
sudo SERVER_NAME=alphafraud.mdeller.com bash /opt/alphafraud/deploy/provision.sh

# 4. (optional) trigger the first pipeline run immediately
sudo systemctl start alphafraud-run.service

After that, deploy/deploy.sh pushes code updates and restarts the web service, and deploy/release.sh "message" does the whole loop in one step (commit, push to GitHub, then deploy). Deploys are safe to run while a backfill is in progress: the web app is a read-only reader, WAL is set once, and DB writes are retry-wrapped, so restarting the web service never disturbs the running pipeline. The weekly timer fires every Wednesday 00:30 UTC.

File	Role
`deploy/provision.sh`	One-time root setup: system packages, service user, venv, systemd units, nginx site, certbot. Also patches HTTP/2 onto certbot's TLS listener afterward (nginx doesn't enable it by default) — idempotent, so it self-heals on a re-run
`deploy/deploy.sh`	Push code from your Mac (rsync over SSH) and restart the web service
`deploy/release.sh`	One-shot: commit, push to GitHub, then deploy to the droplet
`deploy/alphafraud-web.service`	gunicorn web app (always on)
`deploy/alphafraud-run.service` + `.timer`	Weekly pipeline and its Wednesday 00:30 UTC schedule
`deploy/alphafraud-backfill.service`	Supervised one-time full-archive backfill; auto-restarts on crash, survives reboot (`systemctl enable --now alphafraud-backfill`)
`deploy/nginx-alphafraud.conf`	nginx reverse-proxy site (templated with the server name)

🧱 Stack

Python, biotite and tmtools for structure handling and superposition, numpy and scipy for the metrics, Flask and gunicorn for serving, Plotly for the figures, and SQLite for state. No compiled toolchain, no external database.

🧭 Caveats

"Deposited after the cutoff" plus a novelty score is a strong proxy, not proof AlphaFold never saw a homolog: the novelty column makes this explicit so you can filter to the defensible subset.
AlphaFold DB models are monomers, so complex and interface accuracy are not tested in this version.
Engineered constructs, antibodies and fusion proteins that do not map to a single UniProt accession are flagged and skipped, not compared.

✅ To Do

📝 Licence

Released under the MIT Licence (see LICENSE).

👤 Author

Marc C. Deller, D.Phil.
Structural biologist & drug discovery scientist

🌐

marcdeller.com

✉️

marc@marcdeller.com

🐙

github.com/bellcheddar/AlphaFraud

Name		Name	Last commit message	Last commit date
Latest commit History 114 Commits
alphafraud		alphafraud
deploy		deploy
mockups		mockups
.env.example		.env.example
.gitignore		.gitignore
AlphaFraud.py		AlphaFraud.py
FAMILY_SPECTRUM.md		FAMILY_SPECTRUM.md
LICENSE		LICENSE
README.md		README.md
SKIPPED_ENTITIES.md		SKIPPED_ENTITIES.md
requirements.txt		requirements.txt
wsgi.py		wsgi.py

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

🧬 AlphaFraud

✨ Features

🏗️ Architecture and workflow

🎯 Metric suite

📋 Requirements

🔧 Installation

🚀 Usage

📊 Output

🔬 Examples

🟢 Protein Artemis (DCLRE1C) · `6TT5_1`

🔴 PP2A scaffold subunit A / PR65 (PPP2R1A) · `8UWB_3`

🔴 Stromal interaction molecule 1 / CC1 (STIM1) · `6YEL_1`

🔴 Activator of Hsp90 ATPase 1 / Aha1 (AHSA1) · `7DME_1`

🔴 NEDD4-like E3 ligase WWP2 / WW module (WWP2) · `6RSS_1`

🔴 Transmembrane protein 106B (TMEM106B) · `7U14_1`

🔴 Alpha-synuclein (SNCA) · `9A1Q_1`

🔴 Transthyretin (TTR) · `9BZS_1`

🔴 Coagulation factor X (F10) · `6Q9F_2`

🧪 Calculate

🧮 Coverage and skipped entities

🛠️ Web deployment

🧱 Stack

🧭 Caveats

✅ To Do

📝 Licence

👤 Author

About

Releases

Packages

Contributors

Languages

Folders and files

Latest commit

History

Repository files navigation

🧬 AlphaFraud

✨ Features

🏗️ Architecture and workflow

🎯 Metric suite

📋 Requirements

🔧 Installation

🚀 Usage

📊 Output

🔬 Examples

🟢 Protein Artemis (DCLRE1C) · 6TT5_1

🔴 PP2A scaffold subunit A / PR65 (PPP2R1A) · 8UWB_3

🔴 Stromal interaction molecule 1 / CC1 (STIM1) · 6YEL_1

🔴 Activator of Hsp90 ATPase 1 / Aha1 (AHSA1) · 7DME_1

🔴 NEDD4-like E3 ligase WWP2 / WW module (WWP2) · 6RSS_1

🔴 Transmembrane protein 106B (TMEM106B) · 7U14_1

🔴 Alpha-synuclein (SNCA) · 9A1Q_1

🔴 Transthyretin (TTR) · 9BZS_1

🔴 Coagulation factor X (F10) · 6Q9F_2

🧪 Calculate

🧮 Coverage and skipped entities

🛠️ Web deployment

🧱 Stack

🧭 Caveats

✅ To Do

📝 Licence

👤 Author

About

Topics

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

🟢 Protein Artemis (DCLRE1C) · `6TT5_1`

🔴 PP2A scaffold subunit A / PR65 (PPP2R1A) · `8UWB_3`

🔴 Stromal interaction molecule 1 / CC1 (STIM1) · `6YEL_1`

🔴 Activator of Hsp90 ATPase 1 / Aha1 (AHSA1) · `7DME_1`

🔴 NEDD4-like E3 ligase WWP2 / WW module (WWP2) · `6RSS_1`

🔴 Transmembrane protein 106B (TMEM106B) · `7U14_1`

🔴 Alpha-synuclein (SNCA) · `9A1Q_1`

🔴 Transthyretin (TTR) · `9BZS_1`

🔴 Coagulation factor X (F10) · `6Q9F_2`