AI for Bio · Model

AlphaMissense

Structure-augmented missense pathogenicity model with a large public variant catalogue

Genomics & RNABig-tech labUnited KingdomMaintained only中文

Overview

NameAlphaMissense · aliases: AlphaMissense
Full nameThe paper gives no expanded name
Ambiguous names (not auto-linked)AlphaFold
FamilyNot a family entry. Record covers the missense model only; the sibling genomics and protein-design models have their own records
OrganizationGoogle DeepMind
Organization typeBig-tech lab
CountryUnited Kingdom
Model classVariant pathogenicity prediction (structure-augmented protein LM)
DomainGenomics & RNA (also: Proteins)
First release · journal
Basis: Journal paper and developer blog on the same day source
Other milestones
  • First public evidence: · Journal paper published source
  • Journal online: · Journal online source
  • Code release: † · Code and prediction catalogue released with the paper source
  • Official announcement: · Developer project page and blog source
Latest update · no later event
Basis: Last repository commit: predictions-database licence update
The latest sourced event for this model is the 2024-03 predictions-database licence update, after which the repository was archived. Note (not an update to this model): a 2026-09-08 variant-impact score is a downstream product of a different line and is sourced only second-hand
Note: The site recorded 2026-09-08 as an update to this model; corrected source
The most recent sourced event, not a claim that nothing changed since.
StatusMaintained only · No substantive event for the model itself since 2024-03; repository archived
OpennessCode: Open
Weights: Closed
Licence change: · docs update
Basis: Predictions-database licence update (last repository commit) source
The prediction catalogue is open including commercial use and the code is open; no weights release or licence text was retrieved this round
Benchmarks reported2 counted (modified protocol 1, retrospective 1)
Wet-lab validationNone recorded
Unverified fields† background.products_partnerships, release_milestones.code_release, open_status.weights, open_status.code_license
As of2026-11-16

What it does

Starts from a co-folding protein structure model and fine-tunes on weak labels from variants that are common or absent in human and primate populations, scoring every possible single amino-acid substitution and publishing most of them as a precomputed catalogue for clinical and research lookup. Like popEVE it predicts missense effects, but it emphasises a proteome-wide catalogue while the other emphasises cross-gene calibration and rare-disease cohorts; sibling lines are AlphaGenome for genomics and AlphaProteo for protein design.

Benchmarks reported

Every row is one pairing from the survey's relation file; a score belongs to the benchmark page's ranking, where it is shown with metric, eval set, source type and date. 2 of 2 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
11 ClinVar hold-out protocolDeveloper report, modified protocol—Self-reportedIts own missense test set · evidence
13 CAGIRetrospective dataset use—Third-partyPost-hoc third-party evaluation on CAGI6 (checked at abstract level) †partly verified · evidence

Other evaluations (outside the 44)

  • Deep mutational scanning correlation (Public benchmark not among the 44) Self-reported: The developer project page reports correlation on DMS data for about two dozen proteins, outside the 44
  • Third-party variant-model comparison (preprint, abstract level) (Comparison against baselines only) Third-party: Another variant model's preprint uses it as a baseline and concedes partial tuning on a clinical variant database; abstract level only

Developers report: combining a structure model with population-frequency weak labels yields directly queryable pathogenicity predictions for proteome-wide missense substitutions and outperforms earlier evolutionary predictors.

Independent evidence: A third-party post-hoc evaluation exists (see the CAGI entry): it is listed as the best single non-meta predictor but behind meta predictors, and it was not a participant.

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesEvolutionary information from protein sequence and structure plus allele frequencies from public population resources source
b. Private or licensed dataNo source
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationNoNo own wet lab; model selection used a clinical variant database and the developers' own figure note concedes such methods may be overestimated on that benchmark source

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: The developer is the AI research unit of a listed US technology group (headquartered in London), funded internally with no separate financing; the prediction catalogue is free to use including commercially.
  • Funding: Not disclosed
  • Products and partnerships: The model derives from the group's co-folding structure series; a variant-impact score released on 2026-09-08 by a separate genomics product line is reported to combine this model's outputs, but only in second-hand sources and is not treated as an event for this model.
  • Related models: AlphaFold 3, AlphaGenome, AlphaProteo, popEVE

Reading notes

  1. The self-reported clinical-variant figures come from a developer-built held-out set; same-named protocols in different papers are not directly comparable.
  2. High discrimination is not clinical usability; evidence strength depends on the likelihood ratio at the chosen threshold.
  3. Nearly every later variant model cites it as a baseline, yet it does not appear in the official ProteinGym summary tables, so "cited as a baseline" must not be read as "on the board".
  4. The leakage concern stems from model selection on a clinical variant database, not from training on it directly.

Caveats

  1. The CAGI entry is a post-hoc evaluation, not participation, and the source gives no overall rank.
  2. The held-out clinical variant set is developer-built, not an official split.
  3. No substantive event for the model itself since 2024-03; the repository is archived.
  4. The 2026 downstream score belongs to another product line and is sourced only second-hand.

Sources

  1. Journal paper 2023 (journal)
  2. Developer project page (official site)
  3. Code and predictions repository (repository)
  4. Third-party post-hoc evaluation paper (third-party paper)
  5. Downstream product line coverage (second-hand) (second-hand) second-hand

Corrections

3 corrections
FieldWasNowBasisFound
benchmarks_reported[bench-13-cagi]Ranked first among 60 predictors in CAGI6Post-hoc evaluation, not a participant; the paper calls it the best single (non-meta) predictor, tied with another non-meta predictor and behind two meta-predictors, with no overall ranking givensource2026-11-16
latest_update2026-09-08, product update2024-03-13, no later sourced event found (the 2026 event belongs to another product line and is second-hand only)source2026-11-16
benchmarks_reportedCandidate pairing ProteinGymNot a pairing: no entry for this model in either the official clinical or the DMS zero-shot summary tablessource2026-11-16

Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.