Summary
A bioRxiv preprint introduces hEVO, a genomic model trained on 4.784 billion haplotype sequences reconstructed from whole-genome data. Its authors report strong results across variant-interpretation benchmarks, including tests that required no task-specific training.
Researchers have introduced hEVO, a genomic foundation model trained on human genetic variation. In a bioRxiv preprint posted on October 21, 2026, the authors report that it ranked first in 18 of 19 primary comparisons across functional annotation, variant interpretation and human-evolution tasks. The work uses genome sequences from tens of thousands of people to train a model intended to interpret both the effects of genetic differences and patterns of variation across human populations.
Learning from haplotypes
The researchers trained hEVO on 4.784 billion distinct haplotype sequences reconstructed from whole-genome sequencing data for 63,809 individuals selected from 550,237 samples across eight cohorts. A haplotype is a set of genetic variants inherited together on one chromosome. Learning from these sequences can expose a model to combinations of variants as they occur in genomes, rather than treating every variant as an isolated change.
The pretraining integrated shared sequence features and naturally occurring differences between alleles without clinical or functional labels. In other words, the model was not trained with those labels as its learning target. The authors then evaluated its pretrained allele scores on several kinds of genomic tasks.
Reported results across variant tasks
The preprint compares hEVO with representatives of the Evo 2, NT and GPN model families, among others. The authors report that hEVO ranked first in 18 of 19 primary comparisons. Its allele scores also generalized across six complementary variant benchmarks without task-specific training, with the highest reported performance across tests of clinical pathogenicity, fine-mapped causal variants and experimentally measured regulatory effects.
The authors also describe a retrospective analysis of ClinVar classifications. They report that hEVO helped identify variants whose earlier pathogenic classifications were later downgraded, a result they say supports reassessing some past clinical interpretations. This is a retrospective database analysis, rather than a report of prospective use in patient care.
The model’s reported performance extended to ancestry inference, human accelerated sequences and preference for reported selected alleles. The authors say hEVO achieved this breadth with substantially fewer parameters than billion-parameter comparators and low inference overhead in the workloads they evaluated. The abstract does not give the metric values or benchmark sample sizes, so it is not possible to assess the size of the reported performance differences from the available text.
What the approach could add
Human genomes contain both shared biological patterns and variation that can affect traits, disease susceptibility and evolutionary history. A model that learns from naturally occurring haplotypes could provide transferable signals for multiple genomic tasks, rather than requiring a separate task-specific training step for each one. The reported results make hEVO a candidate approach for broadening computational analyses of genetic variation.
The findings are presented in a bioRxiv preprint. They describe model evaluations and retrospective analysis; the source does not report prospective clinical deployment.