Skip to content

AlphaMissense

Resource Location

/labs/SysMedBio/Q/Resources/Genomics/AlphaMissense

/Volumes/lab/SysMedBio/Resources/Genomics/AlphaMissense

Q:\SysMedBio\Resources\Genomics\AlphaMissense

AlphaMissense is a deep learning model developed by Google DeepMind designed to predict the pathogenicity of missense variants in human proteins.

Scientific Context

Missense variants alter a single amino acid in a protein and represent a major category of genetic variation associated with human disease. However, the vast majority of identified missense variants remain classified as Variants of Uncertain Significance (VUS) due to a lack of clinical evidence.

AlphaMissense addresses this challenge by combining: 1. Structural Context and Language Modeling: It integrates unsupervised protein language modeling to capture evolutionary sequence context with structural features derived from AlphaFold's 3D representations to assess how mutations disrupt protein stability and interactions. 2. Weak Labeling for Training: Instead of training on clinical databases (which are heavily biased), the model is trained on population frequency data from humans and primates, operating under the principle that common variants are likely benign, while rare or highly conserved variants are more likely pathogenic. 3. Proteome-Wide Catalog: The authors scored all 71 million possible single-amino-acid substitutions across the human proteome, classifying 57% as likely benign, 32% as likely pathogenic, and 11% as ambiguous. This significantly expands classification coverage compared to curated clinical databases.

Main Publication

  • Title: Accurate proteome-wide missense variant effect prediction with AlphaMissense
  • Authors: Jun Cheng, Guido Novati, Joshua Pan, Clare Bycroft, Akvilė Žemgulytė, Taylor Applebaum, Alexander Pritzel, Lai Hong Wong, Michal Zielinski, Tobias Sargeant, Rosalia G. Schneider, Andrew W. Senior, John Jumper, Demis Hassabis, Pushmeet Kohli, Žiga Avsec
  • Journal: Science (2023, Vol 381, Issue 6664, pp. 1302-1310)
  • DOI: 10.1126/science.adg7492

Available Files

The following files are available in the AlphaMissense resource directory:

  • AlphaMissense_aa_substitutions.tsv.gz: Proteome-wide missense variant effect predictions indexed by amino acid substitutions.
  • AlphaMissense_gene_hg19.tsv.gz: Gene-level prediction mappings for the hg19 reference genome.
  • AlphaMissense_gene_hg38.tsv.gz: Gene-level prediction mappings for the hg38 reference genome.
  • AlphaMissense_hg19.tsv.gz: Genomic coordinate-based variant pathogenicity predictions mapped to the hg19 assembly.
  • AlphaMissense_hg19.tsv.gz.tbi: Tabix index for AlphaMissense_hg19.tsv.gz, enabling fast genomic coordinate queries.
  • AlphaMissense_hg38.tsv.gz: Genomic coordinate-based variant pathogenicity predictions mapped to the hg38 assembly.
  • AlphaMissense_hg38.tsv.gz.tbi: Tabix index for AlphaMissense_hg38.tsv.gz, enabling fast genomic coordinate queries.
  • AlphaMissense_isoforms_aa_substitutions.tsv.gz: Pathogenicity predictions for substitutions across alternative protein isoforms.
  • AlphaMissense_isoforms_hg38.tsv.gz: Isoform variant predictions mapped to the hg38 genomic assembly.
  • README.pdf: Release documentation accompanying the dataset.