AlphaMissense pathogenicity scores are a missense prior
AlphaMissense pathogenicity scores come from a model that ranked amino-acid substitutions across the proteome. High scores mean the model treats the change as likely deleterious. Low scores mean likely benign. A middle band is ambiguous.
The score does not know your family history, zygosity, or the rest of the gene. It does not know whether the variant is even present in a person. It is a prior over missense edits, not a classification of a case.
Cheng et al. published the resource as a genome-wide table. genome.sh can ship those scores in a local database tier. That still does not make the number a ClinVar clinical significance value.
- Input: a missense substitution at a protein position
- Output: a score plus a model label (likely deleterious, ambiguous, likely benign)
- Not input: splice, regulatory, structural, or synonymous variants
What the model scored and what it ignored
Missense means an amino-acid change. Frameshifts, stop-gained, canonical splice, copy-number, and promoter variants are outside the table. A high score on a nearby missense does not annotate those events.
The model does not apply ACMG/AMP criteria. It does not read a VCV submission. It does not see gnomAD frequency unless you join that column yourself.
Use it where ClinVar evidence is thin and the variant is actually missense. Do not use it to override a well-reviewed ClinVar aggregate without reading the record.
A stop-gained or frameshift in BRCA1 will not get an AlphaMissense number. If your pipeline prints a score of zero for those rows, you invented a field. Leave the score absent and keep ClinVar.
Read the score next to ClinVar, not instead of it
ClinVar clinical significance is a submitted label with a review status. AlphaMissense is a predictor. If they disagree, you have a fact to investigate, not a reason to delete one column.
A high score plus Pathogenic, multiple submitters, no conflicts is consistent and still not a diagnosis. A high score plus VUS is a prior, not a promotion to likely pathogenic in your PDF.
Frequency still matters. A predicted-deleterious missense that is common in a continental gnomAD group deserves extra reading, the same way a ClinVar pathogenic founder allele does.
When you annotate a VCF, join the score only onto missense rows. genome annotate file.vcf.gz --format json still only sees called sites. A missing chip probe is not a low score.
Query AlphaMissense with genome.sh
Install a database tier that ships AlphaMissense. Then query the rsID as usual. JSON is the form that exposes the score field. Human output may mention it when present.
lite is ClinVar-oriented and may omit scores. standard or full are the usual tiers when you want predictors and frequencies together. genome db stats and GET /v1/sources tell you what the snapshot includes.
The HTTP API returns the score when the server database includes it. There is no API key. The API still rejects genome files.
cargo install genome-sh genome db install standard genome query rs1800562 --format json | jq . curl -s https://api.genome.sh/v1/query/rs1800562 | jq . curl -s https://api.genome.sh/v1/sources | jq .
Missense only: splice, stop, and regulatory sites
If the variant is not a single amino-acid substitution, AlphaMissense has nothing to say. Do not impute a score from a nearby codon.
BRCA1 pathogenic catalogs are full of frameshifts and splice variants that a missense table will not rank. A consumer chip that misses those sites is still not a BRCA test, score or no score.
For noncoding effect prediction, genome.sh has a separate opt-in AlphaGenome command. That command needs a key and explicit consent. It is not the default query path and it is not AlphaMissense.
- AlphaMissense: static missense score table
- ClinVar: submitted clinical significance
- AlphaGenome: opt-in, key-gated effect prediction, not default query
AlphaMissense is not AlphaGenome
The names share a prefix and confuse search logs. AlphaMissense is a published score table you can index locally. AlphaGenome prediction in genome.sh is a network call to an external model.
Do not set an AlphaGenome key because you wanted missense scores. genome query already returns AlphaMissense when the database ships it. genome predict is the other product surface.
Keep genotypes local either way. Identifier lookup can go over HTTPS. File annotation stays on disk with genome annotate file.vcf.gz --format json.
genome query rs1800562 --format json | jq '{rsid, alphamissense, clinvar}'
genome annotate sample.vcf.gz --format json
# genome predict is opt-in, key-gated, and not this tableChip files, missing sites, and scores that do not apply
A score on an rsID is still a public fact about a substitution. It becomes personal only after you match the rsID to a called genotype. If the chip did not call the site, write not called.
Consumer arrays miss most rare pathogenic variants. A file with no high AlphaMissense hits is not a negative exome. It is a sparse genotype table.
genome.sh prints public annotations. It does not diagnose. Use the score as a triage prior next to ClinVar and gnomAD, then stop.
Tiers, missing scores, and how to cite the table
If JSON has no AlphaMissense field, check the tier and GET /v1/sources. lite may omit the table. A missing field is not a benign prediction.
Cite the snapshot. AlphaMissense versions differ. genome.sh reports the build it shipped, not a live DeepMind API. That is the point of a local table.
Do not mix AlphaMissense labels with ClinVar words in one column. Keep predicted_label and clinical_significance as separate keys in anything you store.
genome db stats
curl -s https://api.genome.sh/v1/sources | jq .
genome query rs1800562 --format json | jq '{rsid, alphamissense, significance: .clinvar}'Questions
What are AlphaMissense pathogenicity scores?
Model scores for amino-acid substitutions. High means the model treats the change as likely deleterious. They are not ClinVar clinical significance and not a diagnosis.
Does a high AlphaMissense score mean pathogenic in ClinVar?
Not necessarily. Check ClinVar review status and gnomAD frequency. The score is a prediction. ClinVar is a submitted interpretation. A disagreement is a reason to read the VCV page, not to delete a column.
Does AlphaMissense work for noncoding variants?
No. It is for missense substitutions. Splice, regulatory, and structural variants need other methods.
Is the score on the genome.sh HTTP API?
When the API database includes AlphaMissense, it appears in query JSON. Check GET /v1/sources. There is no API key for ordinary identifier lookup.
How do I query AlphaMissense with genome.sh?
Install a tier that ships the table, then genome query rs1800562 --format json. lite may omit it. standard or full are the usual choice. Confirm with genome db stats or GET /v1/sources.
Is AlphaMissense the same as AlphaGenome?
No. AlphaMissense is a static score table. AlphaGenome prediction is a separate opt-in CLI command that needs a key and explicit consent.
Can I use AlphaMissense instead of ClinVar for BRCA1?
No. Many pathogenic BRCA1 variants are not missense. A consumer chip is not a BRCA test. Read ClinVar on sites that were actually called.
Is an AlphaMissense score medical advice?
No. genome.sh prints public annotations. Clinical care needs a clinician and the right assay.