SNPBag

Towards a universal foundation model for biobank-scale human genome variation.

• bioRxiv

Overview

SNPBag is a large language model pre-trained with whole-genome SNP data from 1 million synthetic individuals. It is currently fine-tuned for 5 tasks:

• genotype imputation
• haplotype phasing
• genome embedding
• ancestry inference
• relatedness inference

Inference service (Demo)

• Four tasks are available: imputation, embedding, ancestry and relatedness inference.
• Service on server0 with Nvidia A800.
• Service on server1 with Nvidia RTX3090.
• Service on server2 with Nvidia RTX3090.

Folders

Folders contain code for reproducing figures on the manuscript.

• emb: generating genome embedding.
• imputation: genotype imputation for SNP array data.
• phasing: haplotype phasing, tested with physically phased HGDP samples.
• pop_classifier: classify individuals among human populations defined by 1000 Genome Project.
• relatedness: estimate degree of genetic relatedness between individuals.

Schema