AlphaMissense

Overview

AlphaMissense is a deep learning model that predicts the pathogenicity of missense variants across the human proteome. It produces a pathogenicity score between 0 and 1, where higher values indicate more pathogenic predictions.

This release provides pre-computed predictions for all possible human amino acid substitutions across major transcripts and isoforms.

For more details, refer to:

Publication

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. Accurate proteome-wide missense variant effect prediction with AlphaMissense. Science (2023). https://doi.org/10.1126/science.adg7492

Parsing

AlphaMissense provides two tab-separated files:

• a canonical TSV file (required)
  • AlphaMissense_hg38.tsv.gz for GRCh38
  • AlphaMissense_hg19.tsv.gz for GRCh37
• an isoforms TSV file (optional)
  • AlphaMissense_isoforms_hg38.tsv.gz only for GRCh38

Only a subset of columns are consumed during parsing. Column positions are 0-based below.

TSV File

AlphaMissense_hg38.tsv.gz Snippet

#CHROM  POS REF ALT genome  uniprot_id  transcript_id   protein_variant am_pathogenicity    am_class
chr1    69094   G   T   hg38    Q8NH21  ENST00000335137.4   V2L 0.2937  likely_benign
chr1    69094   G   C   hg38    Q8NH21  ENST00000335137.4   V2L 0.2937  likely_benign
chr1    69094   G   A   hg38    Q8NH21  ENST00000335137.4   V2M 0.3296  likely_benign
chr1    69103   T   C   hg38    Q8NH21  ENST00000335137.4   F5L 0.9110  likely_pathogenic
chr1    69103   T   G   hg38    Q8NH21  ENST00000335137.4   F5V 0.4055  ambiguous

From the canonical file, the following columns are parsed (0-based indices):

• 0: #CHROM (reference name)
• 1: pos (1-based position)
• 2: ref (reference allele)
• 3: alt (alternate allele)
• 6: transcriptId (Ensembl transcript)
• 7: proteinVariant (amino-acid substitution, e.g., V2L)
• 8: pathogenicity (0-1)
• 9: classification (e.g., likely_benign / likely_pathogenic / ambiguous)

Columns present but not consumed:

• genome
• uniprot_id (UniProt accession; see release notes: UniProt release notes)

AlphaMissense_isoforms_hg38.tsv.gz Snippet

#CHROM  POS REF ALT genome  transcript_id   protein_variant am_pathogenicity    am_class
chr1    65568   A   C   hg38    ENST00000641515.2   K2Q 0.0938  likely_benign
chr1    65568   A   G   hg38    ENST00000641515.2   K2E 0.0766  likely_benign
chr1    65569   A   G   hg38    ENST00000641515.2   K2R 0.0756  likely_benign
chr1    65569   A   T   hg38    ENST00000641515.2   K2M 0.1732  likely_benign
chr1    65569   A   C   hg38    ENST00000641515.2   K2T 0.1186  likely_benign
chr1    65570   G   T   hg38    ENST00000641515.2   K2N 0.1432  likely_benign

From the isoforms file, the following columns are parsed (0-based indices):

• 0: #CHROM
• 1: pos
• 2: ref
• 3: alt
• 5: transcriptId
• 6: proteinVariant
• 7: pathogenicity
• 8: classification

During ingestion, transcripts present in the canonical file take precedence. Isoform records with the same transcriptId as a canonical record are skipped.

Classification labels

AlphaMissense provides am_class labels such as likely_benign, likely_pathogenic, and ambiguous.

SA Generation

dotnet SAUtils.dll \
AlphaMissense \
--r "${References}/Homo_sapiens.GRCh38.Nirvana.dat" \
--t "${ExternalDataSources}/AlphaMissense/AlphaMissense_hg38.tsv.gz" \
--i "${ExternalDataSources}/AlphaMissense/AlphaMissense_isoforms_hg38.tsv.gz" \
--o "${SaUtilsOutput}"

Notes:

• --i is optional. If omitted, only canonical records are ingested.
• Output files are written with an automatically derived version name based on the .version sidecar file.

Known Issues

Known Issues

Some transcript IDs defined in the AlphaMissense files may be obsolete, retired, or updated. They are not modified by Illumina Connected Annotations and are passed as-is from the data source.

License and Disclaimer

Disclaimer

AlphaMissense predictions have varying confidence; they are not medical advice and are not approved for clinical use. This is not an officially supported Google product.

License

We use and redistribute AlphaMissense predictions only, which are licensed under Creative Commons Attribution 4.0 (CC BY 4.0). See CC BY 4.0 legal code.

Attribution (CC BY 4.0): credit DeepMind/AlphaMissense and the authors, link to the license and source, indicate changes, and do not imply endorsement.

AlphaMissense predictions © 2023 DeepMind Technologies Limited, used under CC BY 4.0. Adapted for Illumina Connected Annotations.

Source: AlphaMissense – License and Disclaimer

Download URL

https://console.cloud.google.com/storage/browser/dm_alphamissense

Contact

For questions about the dataset, contact alphamissense@google.com.

JSON Output

"alphaMissense": [
  {
    "transcriptId": "ENST00000335137.4",
    "proteinVariant": "V2L",
    "pathogenicity": 0.2937,
    "classification": "likely_benign",
    "isIsoform": false
  }
]

Field	Type	Notes
transcriptId	string	Transcript ID (Ensembl)
proteinVariant	string	Protein change (e.g., V2L)
pathogenicity	float	range: 0 - 1.0
classification	string	e.g., likely_benign, likely_pathogenic, ambiguous
isIsoform	bool	true if the record originated from the isoforms TSV