LiLabUNC/Variant-Foundation-Embeddings

Variant Foundation Embeddings

Here we present the variant level embeddings for large-scale genetic analyis as described in 'Incorporating LLM Embeddings for Variation Across the Human Genome,' based on curated annotations using high quality functional data from FAVOR, ClinVar, and GWAS Catalog. We currently present embeddings using either OpenAI's text-embedding-3-large (3072-dimensional) or Qwen's Qwen3-Embedding-0.6B (1024-dimensional) models.

Genetic variants are identified with their chromosome, position (hg38 build), reference allele based on UKB coding, alternate allele based on UKB coding, and their respective LLM embeddings.

Currently we release datasets at the following scales:

1. HapMap3 & MEGA (~1.5 million variants, OpenAI GPT-3.5)
2. UKB Imputed (~90 million variants, OpenAI GPT-3.5)
3. All FAVOR Variants (~9 billion variants, Qwen3-0.6B)

Reference: https://arxiv.org/abs/2509.20702

Updates

03-23-26

We have uploaded a lightly compressed version of the All Favor (9 billion variants) dataset which may be easier to work with.

02-07-26

We have fixed an issue with the chromosome 3 embeddings in the UKB imputed set at /UKB-Imputed-90m/chr03/.

Dataset Schema (OpenAI text-embedding-3-large)

Field	Type	Description
chrom	string	Chromosome (e.g., "1", "X")
pos	string	Base-pair position (hg38 coordinate system)
ref_UKB	string	Reference allele (A, C, G, T)
alt_UKB	string	Alternate allele (A, C, G, T)
embedding	list[float]	Embedding vector (dimension 3072, float)

Dataset Schema (Qwen qwen3-embedding-0.6B)

Field	Type	Description
chrom	string	Chromosome (e.g., "1", "X")
pos	string	Base-pair position (hg38 coordinate system)
ref_UKB	string	Reference allele (A, C, G, T)
alt_UKB	string	Alternate allele (A, C, G, T)
embedding	list[float]	Embedding vector (dimension 1024, float)