Cancer3.AIAI in OncologyAI Models › AlphaMissense
Task-specific model Open weights Research use only genomics

AlphaMissense

Classifies the pathogenicity of every possible single amino-acid substitution in the human proteome — 71 million missense variants — by fine-tuning AlphaFold on population variant frequencies; a resource for interpreting variants of uncertain significance.

At a glance

DeveloperGoogle DeepMind
Released2023-09-19
LicenceApache-2.0 (code); predictions CC-BY-NC-SA 4.0
AvailabilityOpen weights
KindTask-specific model
Regulatory statusResearch use only

What it does

Provides a pathogenicity score (0–1) and a likely benign / ambiguous / likely pathogenic call per variant; the full prediction table is downloadable and integrated into variant annotation pipelines.

Notable uses in oncology

Classified 89% of all missense variants (32% likely pathogenic, 57% likely benign) versus ~0.1% annotated by human experts at the time; used as evidence in germline and tumour variant interpretation.

Tasks, data types and cancers

Clinical settingDiagnosisBasic research
CancerPan-cancer
Inputprotein sequence + substitution
Outputpathogenicity score

Architecture

FamilyAlphaFold-derived network fine-tuned for variant classification
Pre-trainingAlphaFold structure prediction, then weak labels from population allele frequencies (gnomAD, primates)

Training data

AlphaMissense is an adaptation of AlphaFold fine-tuned on population frequency databases of human and primate variants. The authors state the model reaches its results on genetic and experimental benchmarks 'all without explicitly training on such data' — that is, without clinical pathogenicity labels as the training target; the training signal comes from which variants are observed, and at what frequency, in human and primate populations. ClinVar variants held out from training are used for evaluation (AUROC 0.94), not for training. Practical consequence for anyone reusing the model: the signal is oriented towards germline variation, so somatic driver status in a tumour is a different question and is not what the model was trained to answer.

Training set sizepredictions cover 71 million possible missense variants across the human proteome
InstitutionsGoogle DeepMind

Evaluation

Benchmark / datasetMetricValueExternal validationSource
ClinVar held-out variants AUROC 0.94 (paper) yes Source

How to run

Librarypredictions table (TSV) / repo code

Most users download the precomputed table rather than run the model; note the non-commercial licence on the predictions.

Regulatory status and intended use

Regulatory statusResearch use only
Intended useResearch evidence for variant interpretation; not a diagnostic.

Source →

Regulatory status is quoted from the source linked above and can change. Research-use-only models must not be used for clinical decisions.

Limitations and bias

  • Missense only; no indels, splice or non-coding variants.
  • Germline-oriented training signal; somatic driver status is a different question.

Sources

  1. Cheng J et al. Accurate proteome-wide missense variant effect prediction with AlphaMissense. Science 2023
  2. DeepMind blog — A catalogue of genetic mutations

This page is educational — it is not medical advice and does not replace consultation with an oncologist. Diagnostic and treatment decisions are made solely by specialist physicians.