GenoScan-LLM: Whole-Genome Variant Calling

GenoScan-LLM is an 800M parameter foundation model for variant calling. Given aligned reads or a target region, it identifies SNVs, small indels, and structural variants, then classifies them by pathogenicity with clinical-grade sensitivity.

At a glance

Field	Value
Category	Genomics
Parameters	800M
Version	v2.0
Endpoint	`POST /v1/models/genoscan-llm/predict`
Training data	200,000+ genomes across ancestry-balanced cohorts

Install

pip install bunker-fold

See Installation for full setup, including uv and poetry.

When to use

Call variants from whole-genome or targeted panel data.
Flag pathogenic variants in a clinical reporting workflow.
Reprioritize a VCF against updated pathogenicity evidence.

Example

curl -X POST https://api.bunkerbio.com/v1/models/genoscan-llm/predict \
  -H "Authorization: Bearer $BUNKER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "region": "chr17:41244000-41277500",
    "gene": "BRCA1",
    "coverage": 42,
    "format": "CRAM",
    "input_url": "s3://my-bucket/sample.cram"
  }'

import requests, os

r = requests.post(
    "https://api.bunkerbio.com/v1/models/genoscan-llm/predict",
    headers={"Authorization": f"Bearer {os.environ['BUNKER_API_KEY']}"},
    json={
        "region": "chr17:41244000-41277500",
        "gene": "BRCA1",
        "coverage": 42,
        "format": "CRAM",
        "input_url": "s3://my-bucket/sample.cram",
    },
)
print(r.json())

Response

{
  "variants_detected": 3,
  "variants": [
    {
      "id": "rs80357906",
      "hgvs": "c.5266dupC",
      "classification": "pathogenic",
      "acmg_class": 5,
      "confidence": 0.994
    },
    { "id": "rs1799966", "classification": "benign" },
    { "id": "rs16942", "classification": "benign" }
  ]
}

Parameters

string

required

Genomic region in chrN:start-end format.

string

required

Signed URL or S3 path to a BAM, CRAM, or VCF file.

string

required

One of BAM, CRAM, or VCF.

string

Optional gene symbol used to prioritize functional annotation.

number

Average coverage depth. Used to calibrate confidence thresholds.

Response fields

integer

Total number of variants called in the region.

array

List of variants with identifier, HGVS notation, and pathogenicity classification.

Show variant fields

string

dbSNP identifier where available.

string

HGVS coding notation.

string

One of pathogenic, likely_pathogenic, uncertain, likely_benign, benign.

integer

ACMG classification from 1 (benign) to 5 (pathogenic).

number

Model confidence in the classification, between 0 and 1.

Get Started

Models

API Reference

Installation

GenoScan-LLM: Whole-Genome Variant Calling

At a glance

Install

When to use

Example

Parameters

Response fields