Preprint: Mendel foundation model trained on human genetic variation improves regulatory variant prioritisation

A preprint posted to bioRxiv describes Mendel, a DNA foundation model trained from scratch on human whole-genome sequencing data rather than on multi-species genomic alignments — the training corpus used by most current DNA foundation models. The authors argue that models trained on conservation-dominated corpora learn features that reflect the conserved genomic background shared across all individuals, rather than the sparse variant positions where human genetic variation is actually concentrated.

By reorienting training towards human sequence variation, Mendel learns representations that are more informative about the regulatory consequences of individual alleles. The preprint reports that the model outperforms existing DNA language models in prioritising regulatory variants identified through genome-wide association studies and in predicting allele-specific gene expression — two tasks central to the functional interpretation of GWAS findings.

The work addresses a recognised limitation in the field: hundreds of thousands of variant-trait associations from GWAS remain mechanistically unresolved, in part because linking a variant to a regulatory consequence at allele resolution is technically demanding. Foundation models that better capture human-specific variation could accelerate this interpretation. The preprint has not been peer-reviewed; independent benchmarking on held-out datasets will be important to assess the generalisability of the reported performance gains. Researchers in statistical genomics, functional genomics, and machine learning applied to biology will be the primary audience for this work.