The Pulse
GPN-Star Uses Evolution to Rank Disease-Linked DNA Variants
GPN-Star uses evolutionary information from multiple species to help rank genetic variants, with different evolutionary timescales performing better on different prediction tasks. The study also reports tissue-specific heritability enrichme
AI.info Team · Written with AI ·
Nature Genetics published a research highlight on October 8 describing GPN-Star, a genomic AI model that uses evolutionary history to help rank DNA variants that may affect disease risk. The research paper behind the highlight appeared in Nature on September 9. Its central finding is that evolutionary timescale matters: the model trained on vertebrate genomes excelled at some coding-variant tasks, while a mammal-trained version performed best on several non-coding and association-study benchmarks.
Three evolutionary windows, different strengths
GPN-Star compares aligned DNA sequences from multiple species and uses their evolutionary relationships to estimate which parts of the genome have been constrained over time. The authors built human-reference models from alignments spanning vertebrates, mammals and primates. Unlike a model that reads only one DNA sequence at a time, GPN-Star’s architecture incorporates both the aligned sequences and a species tree.
Evolution can help prioritize variants because changes that disrupt important biological functions are more likely to be removed by natural selection. But that signal depends on the trait and variant being assessed. Chengzhong Ye and colleagues, authors of the Nature study, write:
“Analyses across timescales show task-dependent advantages of modelling more recent versus deeper evolution.”
The study found that the vertebrate-trained version led on classifying pathogenic versus benign missense variants in ClinVar, a curated database of clinical variant interpretations. The mammal-trained model performed best in the paper’s tests of pathogenic non-coding variants and fine-mapped variants from genome-wide association studies. Those results point to a practical distinction: broad evolutionary comparisons can be useful for some protein-altering changes, while closer relatives may provide more relevant signals for other genomic regions and traits.
Benchmark results are not clinical diagnoses
The team compared GPN-Star with established evolutionary scores, variant-effect predictors and sequence-to-function models. On the non-coding disease-variant tests, the mammal model outperformed the competing methods included in those benchmarks, including Enformer, Borzoi and AlphaGenome. For fine-mapped UK Biobank association variants, it ranked highest among the compared models across 65 traits for missense variants and 83 traits for non-coding variants.
Those are benchmark comparisons, not evidence that the model can diagnose a patient or establish that a particular variant causes disease. The study evaluates how well scores separate labeled or putatively causal variants from comparison groups; researchers would still need independent evidence to interpret a specific person’s genetic result. The paper also reports that adding population allele-frequency information improved GPN-Star’s ClinVar performance, a reminder that the model’s evolutionary signal is only one source of evidence.
From rare variants to five other species
The researchers also tested whether GPN-Star could help with rare variant association testing, which groups variants by gene to find links between genetic changes and traits. The paper reports improved power in this analysis and says primate-trained predictions enriched heritability signals across more than 100 complex traits. Separately, in its analysis of tissue-specific heritability enrichment, the paper reports that a brain-specific model showed the highest enrichment for schizophrenia, while a blood- and immune-specific model performed best for lupus.
The work extends beyond human genetics. The authors trained versions for mouse, chicken, fruit fly, roundworm and Arabidopsis, and found that the predictions captured evolutionary constraints in each. For three species, they also tested curated pathogenic or lethal variants and reported better performance than PhyloP and PhastCons, two conservation-based measures.
GPN-Star’s results make evolutionary comparisons a useful way to prioritize genomic regions for further study, particularly where non-coding DNA is difficult to interpret. The model’s scores narrow the search; they do not replace functional experiments or clinical assessment. The authors describe the approach as needing multispecies whole-genome alignments and the corresponding phylogeny—inputs that connect each prediction to the species and evolutionary history used to train it.