Biobank-scale self-supervised genotype pretraining yields transferable representations for human genetics.

SNPBERT, a self-supervised transformer pretrained on genotypes from 485 923 UK Biobank participants, learns embeddings that recover population structure and transfer to external datasets.

• Why it matters: Biobank-scale genotyping exists, but scalable methods for learning transferable representations from largely unlabeled individual-level genotype data are still lacking.
• What they did: SNPBERT was pretrained via masked rsID prediction and evaluated by recovering known population structure across UK Biobank and three external datasets, plus ancestry-informative marker identification and disease fine-tuning.
• The result: Fine-tuning produced disease-specific risk scores with Nagelkerke’s R2 increases of 0.7% to 3.3%, independent associations for seven of eight chronic diseases, and in celiac disease attention prioritized MHC signals while elevating CCHCR1 and NOTCH4 above GWAS ranking.

AI summary of the abstract · may contain errors · not medical advice · read the paper.