README.md

---
license: cc-by-4.0
task_categories:
  - tabular-classification
tags:
  - genomics
  - alphamissense
  - variant-pathogenicity
  - proteomics
  - bioinformatics
  - sqlite
size_categories:
  - 10M<n<100M
---

# Allelix AlphaMissense Pathogenicity Cache

Pre-built SQLite cache of missense variant pathogenicity predictions from [DeepMind's AlphaMissense](https://github.com/google-deepmind/alphamissense) for use with [Allelix](https://github.com/dial481/allelix).

## What's in the file

`alphamissense.sqlite.gz` is a gzipped SQLite database containing the `alphamissense_scores` table with all ~71M missense variant predictions from AlphaMissense, 7.5M of which carry rsID mappings via a gnomAD v4.1 coordinate join.

### Schema

```sql
CREATE TABLE alphamissense_scores (
    chrom TEXT NOT NULL,
    pos INTEGER NOT NULL,
    ref TEXT NOT NULL,
    alt TEXT NOT NULL,
    rsid TEXT,
    uniprot_id TEXT,
    transcript_id TEXT,
    protein_variant TEXT,
    am_pathogenicity REAL NOT NULL,
    am_class TEXT NOT NULL,
    PRIMARY KEY (chrom, pos, ref, alt)
);

CREATE INDEX idx_am_rsid ON alphamissense_scores(rsid);
```

### Columns

| Column | Description |
| --- | --- |
| `chrom` | Chromosome (1-22, X, Y, no "chr" prefix) |
| `pos` | Genomic position (GRCh38) |
| `ref` | Reference allele |
| `alt` | Alternate allele |
| `rsid` | dbSNP rsID (from gnomAD v4.1 coordinate join, NULL for ~63.5M variants without exome coverage) |
| `uniprot_id` | UniProt protein accession |
| `transcript_id` | Ensembl transcript ID |
| `protein_variant` | Protein-level change (e.g., K2N) |
| `am_pathogenicity` | Pathogenicity score (0.0-1.0, higher = more pathogenic) |
| `am_class` | Classification: `likely_pathogenic`, `likely_benign`, or `ambiguous` |

### Multi-allelic sites

The composite primary key `(chrom, pos, ref, alt)` preserves multi-allelic sites. A single rsID may have multiple rows with different alternate alleles and divergent pathogenicity scores. For example, chr1:69094 has G→T (0.294, likely_benign) and G→C (0.981, likely_pathogenic) at the same position.

## How Allelix uses this

Allelix downloads this file automatically during `allelix db update`. The enrichment annotator looks up variants by rsID and returns `MAX(am_pathogenicity)` across alleles for a given rsID. The AM Score column appears in terminal, HTML, and JSON reports. AlphaMissense scores reflect predicted protein structural impact only — not clinical or pharmacogenomic significance.

## Building from source

The cache can be rebuilt from the AlphaMissense Zenodo TSV using the build script included with Allelix:

```bash
python scripts/build_alphamissense_cache.py --tsv AlphaMissense_hg38.tsv.gz --output alphamissense.sqlite
gzip alphamissense.sqlite
```

The build script joins against the gnomAD cache (if present) to populate rsID mappings. Without gnomAD, all rsID fields will be NULL and the annotator will not match any variants by rsID.

## Source and license

- **Source:** [AlphaMissense](https://zenodo.org/records/8208688) (Cheng et al., *Science* 2023, Google DeepMind)
- **License:** [Creative Commons Attribution 4.0 (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/)
- **Citation:** Cheng J, Novati G, Pan J, et al. Accurate proteome-wide missense variant effect prediction with AlphaMissense. *Science*. 2023;381(6664):eadg7492.