SNPBag

SNPBag is a foundation model for genome-scale single-nucleotide polymorphism (SNP) analysis, developed at the Research Center for Life Sciences Computing at Zhejiang Lab and first posted to bioRxiv in early 2025. National biobanks have now genotyped millions of human genomes, and training a large model across this variation data promises a universal representation of human genetic diversity. The scale, however, is daunting: genome length multiplied by population size reaches quadrillions of nucleotides, overwhelming conventional modeling approaches. SNPBag introduces an AI framework designed to scale with biobank-sized data while supporting many downstream tasks from one pretrained model.

Rather than training a separate model for each genomics task, SNPBag learns a shared representation of whole-genome SNP profiles that transfers across imputation, phasing, ancestry inference, and relatedness estimation. This multitask, single-checkpoint design places it alongside other genomic foundation models but with a distinctive focus on population-scale human variation and the practical demands of biobank analysis.

#Key Features

• Multitask from one checkpoint: A single pretrained model handles genotype imputation, haplotype phasing, genome embedding, ancestry classification, and relatedness estimation without task-specific retraining.
• State-of-the-art imputation: On genotype imputation benchmarks, SNPBag achieves state-of-the-art accuracy across multiple evaluation tasks.
• Fast, accurate phasing: In haplotype phasing it rivals the best existing method while delivering a 72-fold speedup, validated on HGDP samples.
• Compact genome embeddings: SNPBag encodes six million SNPs per genome into a 0.75 MB embedding, enabling efficient storage, transfer, and downstream reuse.
• Population-scale inference: The embeddings support rapid ancestry inference across global populations and detection of genetic relationships out to 12th-degree relatives.

#Technical Details

SNPBag is a 0.8-billion-parameter language model trained on one million synthesized human genomes, corresponding to roughly six trillion SNP tokens. It is pretrained in a self-supervised fashion over whole-genome SNP data and then applied to diverse analytical tasks through its learned embeddings and task heads. The authors benchmark it against established tools for imputation and phasing, reporting state-of-the-art imputation accuracy and phasing quality on par with leading methods at a fraction of the runtime. Ancestry and relatedness analyses use the compact 0.75 MB per-genome embedding, with population labels drawn from the 1000 Genomes Project and phasing validated on Human Genome Diversity Project samples.

#Applications

SNPBag targets human genetics and biobank-scale genomics, where imputation, phasing, ancestry inference, and kinship analysis are routine but computationally intensive. Its compact embeddings let large cohorts be stored and shared efficiently while preserving information needed for downstream analysis, and its multitask design means a single model can serve population genetics, genotype quality control, and relatedness screening pipelines. Researchers working with array-genotyped or sequenced cohorts stand to benefit from both the accuracy gains and the substantial speedups.

#Impact

SNPBag proposes a unified paradigm for scalable, multitask analysis of ever-growing human variation data, demonstrating that a single genome-scale foundation model can match or exceed specialized tools across several core genetics tasks. The 72-fold phasing speedup and 0.75 MB genome embeddings are particularly consequential for biobank-scale work, where compute and storage are limiting factors. As a preprint that has been iteratively expanded across multiple versions and is not yet peer reviewed, its benchmark claims await independent validation, and the training data are synthesized genomes rather than raw biobank samples.