Transporter Variants

A static website for browsing calibrated pathogenicity predictions for all 349,056 ClinVar variants in 783 human transporter, ion-channel, ABC-transporter and ATPase genes.

Site: https://joevondahab.github.io/transporter-variants/

Each variant carries an ACMG evidence strength derived from a likelihood ratio, validated on gene-disjoint halves of the held-out labelled set. Gene pages show a sortable, filterable variant table and a protein map: the chain drawn across the membrane from UniProt topology, plus the AlphaFold structure, both coloured by the regional average of the predictions.

Two tracks, two licences

Track	Contents	Licence
native	Our model alone, every variant, no third-party predictions	CC BY 4.0
routed	Our model with AlphaMissense where AlphaMissense covers the variant, our model elsewhere	CC BY-NC-SA 4.0, inherited from AlphaMissense

The tracks are kept separate throughout: separate files in docs/data/native/ and docs/data/routed/, separate downloads in docs/downloads/. Do not merge them.

Using the data correctly

• Read evidence, not score. A 0.5 cut-off over-calls, because the training set was 45% pathogenic.
• Rows with in_training TRUE were seen during model fitting. Exclude them from any accuracy estimate.
• On the routed track, read v2_plus only where has_alphamissense is TRUE.
• mechanism_pred is exploratory: a hypothesis generator, not a per-variant call.

Layout

docs/                 the site, served by GitHub Pages
  index.html          landing page
  gene.html           gene page (?g=SYMBOL, optional &t=routed, &q=, &pos=)
  assets/             CSS, JavaScript, vendored 3Dmol.js 2.5.4 (BSD-3-Clause)
  data/index.json     gene index, landing-page numbers, shared dictionaries
  data/native/        one JSON file per gene, native track
  data/routed/        one JSON file per gene, routed track
  data/protein/       one JSON file per gene: sequence and UniProt features
  data/lookup/        rsID and ClinVar VariationID to gene, in 100 shards each
  downloads/          the release files, unchanged, with their licences and checksums
tools/build_data.py   rebuilds docs/data and docs/downloads from the release

There is no build step for the site itself and no server: everything is static files.

Rebuilding the data

python -I tools/build_data.py --release ~/projects/TripleVar/outputs/transporters/release

The script reads the two release files, fetches UniProt features once (cached in tools/cache/, which is not committed), checks the release checksums, and writes docs/data/. AlphaFold models are not stored here; gene pages fetch them from AlphaFold DB in the browser.

Licences

• Site code (HTML, CSS, JavaScript, tools/): MIT, see LICENSE.
• Native track and the native-track data files under docs/data/: CC BY 4.0.
• Routed track and docs/data/routed/: CC BY-NC-SA 4.0.
• UniProtKB annotations and AlphaFold DB models: CC BY 4.0, from their providers.

Citation

Abo-Dahab Y, Chun J, Giacomini KM. Calibrated pathogenicity predictions for 783 human transporter and ion-channel genes. Version 1.0.0, 2026. See docs/downloads/CITATION.cff. If you use the routed track, also cite AlphaMissense (Cheng et al., Science 2023).

Predictions are for research use and are not a clinical classification.