Researchers at Google DeepMind adapted their protein-folding system AlphaFold into a new model, AlphaMissense, trained to judge whether a given mutation was likely to disrupt a protein enough to cause disease, using patterns from human and primate genetic variation.
AlphaMissense scored all 71 million possible missense mutations and confidently classified 89 percent of them as likely disease-causing or likely harmless, compared with roughly 0.1 percent that had been clinically classified by direct evidence before. It agreed closely with known results on variants doctors had already confirmed.
A prediction is not a diagnosis. The model gives clinicians a likelihood, not certainty, for mutations found in a patient's genetic test, and some of its calls will be wrong, especially for mutation types underrepresented in its training data. Even so, it gave genetics labs a first-pass answer for mutations that previously had none at all.
Read the original study
Cheng et al., Science, 2023 · doi.org/10.1126/science.adg7492
Seventy-one million mutations, graded almost overnight.
Source
- Cheng, J., Novati, G., Pan, J., Bycroft, C., Žemgulytė, A., Applebaum, T., Pritzel, A., Wong, L. H., Zielinski, M., Sargeant, T., Schneider, R. G., Senior, A. W., Jumper, J., Hassabis, D., Kohli, P., & Avsec, Ž. (2023). Accurate proteome-wide missense variant effect prediction with AlphaMissense. Science, 381(6664), eadg7492. https://doi.org/10.1126/science.adg7492