Skip to content

AlphaMissense

Summary

AlphaMissense adapts the AlphaFold 2 architecture to predict the pathogenicity of human missense variants, fine-tuning on human and primate population-frequency data rather than clinically curated labels. It achieves state-of-the-art performance across clinical and experimental benchmarks. The resulting resource provides proteome-wide predictions for all possible single amino acid substitutions.[1]

Method

AlphaMissense closely follows the AlphaFold architecture (Evoformer trunk with MSA and pair representations), with two training stages:

  1. Structure pretraining, identical to AlphaFold except with higher weight on the masked-MSA reconstruction loss, teaching the model amino-acid distributions conditioned on sequence context.[1]
  2. Variant fine-tuning, in which the model is optimized to jointly predict variant pathogenicity and reference-sequence structure. Following the PrimateAI approach, variants frequently observed in human/primate populations are labeled benign and variants absent from those populations are sampled as pathogenic (weighted by trinucleotide context and gene); the pathogenicity score is the log-likelihood difference between the reference and alternate residue.[1]

Final predictions average six models (three independently trained, each run with and without MSA diversity filtering), calibrated against a held-out balanced set of 2,526 ClinVar variants via univariate logistic regression so scores approximate a probability of pathogenicity.[1]

Performance

  • ClinVar (18,924 test variants, balanced per gene): auROC 0.940, vs. 0.911 for EVE, the next-best model not trained directly on ClinVar (p=0.001, bootstrap); AlphaMissense also exceeds models trained directly on ClinVar despite those models' data leakage.[1]
  • Per-gene ClinVar performance (612 genes with ≥5 benign and ≥5 pathogenic test variants): average gene-level auROC 0.950 vs. 0.921 for EVE (p=0.001).[1]
  • De novo variants (DDD cohort patients vs. controls, 353 vs. 57 variants across 215 genes): outperforms competing predictors at distinguishing patient from control variants.[1]
  • Confident classification rate: using score cutoffs calibrated to 90% precision on ClinVar, the fraction of ClinVar test variants confidently classified (as likely pathogenic or likely benign) rises by 25.8 percentage points versus EVE, from 67.1% to 92.9%.[1]
  • Ablations: both AlphaFold structural pretraining and the variant fine-tuning stage are individually necessary for strong performance; pretraining with masked-MSA loss alone (no structure loss) is insufficient.[1]

Proteome-Wide Resource

AlphaMissense scored all 216 million possible single amino acid substitutions across the 19,233 canonical human proteins, yielding 71 million missense variant predictions. Using the 90%-precision ClinVar cutoffs, 32% (22.8 million) are classified likely pathogenic and 57% (40.9 million) likely benign; among the subset unobserved in gnomAD, 88.8% receive a confident classification (56.0% likely benign, 32.8% likely pathogenic). The authors also release gene-level average pathogenicity scores and predictions across ~60,000 alternative transcript isoforms.[1]

Downstream clinical-utility findings reported elsewhere (e.g. VUS reclassification counts in specific diagnostic cohorts) were not verified against a primary source in this pass and are not included here.

See Also

  • AlphaFold 2 — the architecture AlphaMissense is fine-tuned from.
  • DETANGO — a complementary model that, given a variant already flagged as deleterious by a protein language model, disentangles whether the effect is stability- or function-driven.
  • AlphaGenome — the non-coding counterpart from the same lab. The two are complementary in both directions: AlphaMissense covers the ~2% of variation that is coding, while AlphaGenome covers the remainder — and AlphaGenome additionally recovers splice-disrupting variants among those AlphaMissense classifies as likely benign (ClinVar missense auPRC 0.18).[2]

Citations

[1] Cheng, Novati, Pan, Bycroft, Žemgulytė et al. (2023), "Accurate proteome-wide missense variant effect prediction with AlphaMissense", Science, 381(6664), eadg7492. Supports: method, all benchmark and proteome-wide-resource figures above. Location: Abstract; Results ("AlphaMissense" architecture section; ClinVar/DDD benchmarking; proteome-wide predictions); Methods.

[2] Avsec, Ž. et al. (2026). Advancing regulatory variant effect prediction with AlphaGenome. Nature, 649, 1206–1217. Supports: the ClinVar missense benchmark restricted to variants AlphaMissense predicts likely benign, on which AlphaGenome scores splicing effects (auPRC 0.18). Location: "Improved splicing variant predictions"; Fig. 3h. Source paper: s41586-025-10014-0.pdf