AlphaMissense scores every possible missense variant in the human proteome

Google DeepMind published AlphaMissense in Science on 19 September 2023, an adaptation of AlphaFold that scores how likely a single amino acid change is to cause disease. It classified 89 per cent of all 71 million possible missense variants, against roughly 0.1 per cent confirmed by laboratory or clinical work.

Background

A missense variant is a single letter change in DNA that swaps one amino acid for another in a protein. The average person carries more than nine thousand of them, and almost all are harmless.

Working out which are not is slow. Every protein is different, each experiment has to be designed for it, and the work takes months. Of the roughly four million missense variants seen in human genomes, about two per cent had been clinically classified.

What happened

Google DeepMind published AlphaMissense in Science on 19 September 2023. It adapts AlphaFold, first training to predict the structure of a reference protein, then training on human and primate population data to score how likely a given substitution is to be harmful.

It scored all 216 million possible single amino acid changes across 19,233 human proteins, giving 71 million missense variant predictions. Of those, 32 per cent were classified likely pathogenic and 57 per cent likely benign, at score cutoffs chosen to reach 90 per cent precision against ClinVar, a database of variants already assessed by people.

Eleven per cent were left unclassified. That is around eight million variants on which the model declined to commit.

The predictions and the model code were released freely.

What followed

The training label is the thing to hold on to. A variant was treated as benign if it appears commonly in human and primate populations and as pathogenic if it is absent. That is a sensible proxy, because harmful variants are selected against, and it is not the same as knowing what a variant does. A variant can be absent for many reasons.

So the resource is a very large set of well-founded guesses, most of which nobody has checked. About 0.1 per cent have been confirmed by laboratory or clinical work.

A study at Memorial Sloan Kettering in 2025 examined whether the predictions could be used directly in clinical decisions and reported that this remains unclear. That is where the matter stands, and this entry does not put it further than that.

Why it mattered

The training label is not a clinical outcome. A variant counted as benign because it is common in human and primate populations, and as pathogenic because it is absent. That is a defensible proxy and it is not a diagnosis, and the distinction matters more here than in any other entry on this timeline.

Sources

• Accurate proteome-wide missense variant effect prediction with AlphaMissense. science.org. Primary source
• A catalogue of genetic mutations to help pinpoint the cause of diseases. deepmind.google. Official
• Predicting variant pathogenicity with AlphaMissense. nature.com. Secondary

Cite this page

AI Achievements. (2023). AlphaMissense scores every possible missense variant in the human proteome. Retrieved 2026-08-29, from https://achievements.ai/milestone/alphamissense-variant-effect-prediction

@misc{achievements_alphamissense_variant_effect_prediction,
  title  = {AlphaMissense scores every possible missense variant in the human proteome},
  author = {{AI Achievements}},
  year   = {2023},
  url    = {https://achievements.ai/milestone/alphamissense-variant-effect-prediction}
}