Machine-learning algorithm uses structure prediction to spot disease-causing mutations

DeepMind is releasing AlphaMissense, a tool researchers can use to learn more about the effects that missense mutations – which make up the majority of “variants of uncertain significance” – have on disease. AlphaMissense could help identify pathogenic missense mutations and previously unknown disease-causing genes. Uncovering the root causes of disease is one of the greatest challenges in human genetics. Many of the genetic mutations that cause disease in humans occur in protein-coding regions. Although the capacity to sequence DNA and identify disease-causing variants has substantially increased, the ability to interpret their effects remains limited. This problem is particularly acute for missense variants, genetic mutations that alter the amino acid sequence of proteins. The average person carries thousands of missense variants, most of which are benign, but others of which are pathogenic. Very few missense variants have been confirmed by experts, so their pathogenic versus benign status remains an open question.

To develop a better method to classify missense variants, Jun Cheng and colleagues created AlphaMissense based on the DeepMind AlphaFold methodology for predicting protein structures from gene sequences. AlphaMissense works not by predicting the change in protein structure upon mutation, but by leveraging databases of related protein sequences and structural context of variants to produce a score. The score rates the likelihood of a variant being pathogenic, or disease-causing. Cheng and colleagues used AlphaMissense to predict the pathogenicity of all 216 million possible single amino acid changes across the 19,233 canonical human proteins, resulting in 71 million missense variant predictions. This catalogue of the predictions of all possible missense variants could assist clinicians in prioritizing variants for rare disease diagnostics. Next, the authors applied AlphaMissense to classify 89% of these missense variants, predicting 57% were likely benign and 32% were likely pathogenic. “Although this will undoubtedly be helpful for variant interpretation and prioritization,” say Joseph A. Marsh and Sarah A. Teichmann in a related Perspective, “it is impor­tant not to confuse these labels with the very specific clinical definitions of these terms, which rely on multiple lines of evidence.” When com­paring AlphaMissense to many existing and similar tools (called variant effect predictors, or VEPs), the new DeepMind tool had superior performance, say the authors.

***A related embargoed news briefing was held on Monday, 18 September, 2023, as a Zoom Webinar. Recordings of the briefing can be found here.***