Predicting variant pathogenicity with AlphaMissense

A team of researchers from Google DeepMind report in Science the development of AlphaMissense, an artificial intelligence-based approach to predict the pathogenicity of missense variants adapted from their previously described protein structure prediction tool AlphaFold. Given that only ~2% of the 4 million missense variants observed in the human genome have been clinically classified, AlphaMissense has enormous potential for informing diagnosis and, possibly, therapy of rare genetic disease.

AlphaMissense is trained in two stages. The first stage, which is similar to that of AlphaFold, trains the network to predict the 3D structure of a reference sequence. In the second stage, the model is trained on human proteins to predict the pathogenicity of variants of the reference sequence. Sequences in the training set are labelled as benign if they occur frequently in human populations or pathogenic if they are absent. However, the training set is inherently noisy, because variants that are not observed in human populations could nevertheless be benign, so the preliminary model generated in stage 2 is then used to filter out unobserved variants that are likely to be benign, and the model is refined by training on the filtered dataset. Pathogenicity predictions from AlphaMissense were calibrated against an evaluation set of ClinVar variants to generate a pathogenicity score that is defined as ‘likely pathogenic’, ‘ambiguous’ or ‘likely benign’ on the basis of threshold values assigned with 90% precision on ClinVar variants. Thus, AlphaMissense combines the structural context of variants with their evolutionary conservation to predict functional effects, without direct training on previously assigned databases such as ClinVar that have curation biases.