Pathogenicity predictions for 71 million human missense variants
An AlphaFold-derived model classified 89% of all 71 million possible human missense variants as likely benign or likely pathogenic, against the roughly 0.1% that carry a clinical classification today.
- Model
- AlphaMissense
- Field
- Medicine
- Date
- 2023-09-19
What was found
AlphaMissense fine-tunes AlphaFold's structural representation on human and primate variant frequency data, so it learns which substitutions natural selection tolerates without being trained on clinical labels. Applied across 19,233 canonical human proteins it scored every possible single amino-acid change, calling 57% likely benign and 32% likely pathogenic. The full catalogue and the model code were released publicly. Published in Science.
Novelty check
Variant-effect predictors are a crowded field (SIFT, PolyPhen-2, EVE, CADD). The contribution is proteome-wide coverage at state-of-the-art benchmark accuracy without training on clinical annotations, which is what makes the predictions usable where no clinical data exists.
Caveats and known objections
These are predictions, not diagnoses, and DeepMind says so: ACMG/AMP guidance treats computational evidence as supporting rather than standalone. Benchmarks have been criticised for circularity, since the model trains on population frequency and is evaluated partly against datasets encoding related signals. Calibration varies by gene, and a "likely pathogenic" call on a variant nobody has seen in a patient remains a hypothesis.
Independent checks
Ljungdahl et al., Oxford, UCSF and New York Genome Center (bioRxiv, 2023): Against multiplexed-assay data for 17,696 variants in five genes, AlphaMissense was consistently among the top five of 39 predictors and the best-correlated for two genes, a modest improvement, and like the others it overcalls variants as pathogenic. · link ↗
David Curtis, UCL Genetics Institute (European Journal of Human Genetics, 2024): Across 18 genes with rare-variant associations to hyperlipidaemia, hypertension or type 2 diabetes, AlphaMissense gave the strongest evidence on average but the best result for only 4 genes, with marked inconsistency between genes. · link ↗
Disagree with these grades?
Bring a citation: a grade moves on evidence, not on argument.
Or on GitHub: submit a check challenge the grade send a correction or send a pull request
Flag this for triage
Signals order the review queue and nothing else. They are never published, and they never move a grade: that takes a citation.
History of this entry
- CheckedRecorded two independent evaluations: one finds AlphaMissense among the best predictors against functional assays with a modest margin, the other finds inconsistent performance for common disease genes. No grade moved. source ↗
- AddedEntered the registry graded Peer reviewed and AI-led.
Entries are never deleted. A grade that does not hold up is downgraded on the record, with the reason beside it.
Graded peer reviewed for verification and ai-led for autonomy. What these mean.
Cite this entry
whataifound.org. (2023). Pathogenicity predictions for 71 million human missense variants. whataifound.org: A Registry of AI Scientific and Mathematical Discoveries. https://whataifound.org/finding/2023-09-19-alphamissenseBibTeX
@misc{whataifound-googledeepmind-2023-alphamissense,
title = {Pathogenicity predictions for 71 million human missense variants},
author = {{whataifound.org}},
year = {2023},
howpublished = {whataifound.org: A Registry of AI Scientific and Mathematical Discoveries},
note = {Result by Google DeepMind. Verification: Peer reviewed. Autonomy: AI-led.},
url = {https://whataifound.org/finding/2023-09-19-alphamissense}
}