GPN-Star compares human DNA with DNA from other species to rank variants for testing
GPN-Star compares human DNA with DNA from other species to rank variants for testing
On 9 September 2026, Nature published a paper on GPN-Star. The model compares a region of human DNA with the corresponding region in other species to assess individual DNA letter substitutions. A multiple genome alignment identifies these corresponding regions, while a phylogenetic tree describes how closely the species are related.
The human genome contains many single-letter DNA substitutions. Laboratory experiments and genetic studies require researchers to select candidates from this large set. Evolution provides a clue: substitutions in regions important for cell function are less likely to become established across species.
GPN-Star takes a window from the alignment and a phylogenetic tree as input. During training, it masks one DNA letter and predicts it from neighboring letters in the human sequence and the corresponding positions in other species. The model groups closely related species and uses distances between branches of the tree to weight their contributions to the prediction.
The model thus assigns a probability to each DNA letter at a given position. The authors use these probabilities to calculate an evolutionary constraint score, which measures how much a substitution departs from the pattern of conservation at that position over the chosen evolutionary timescale. This score can be used to rank variants for experimental or statistical testing.
The authors trained separate versions of the model on vertebrate, mammalian, and primate alignments. In their evaluations, the model using the deeper vertebrate comparison performed better at assessing substitutions in protein-coding regions and rare variants with large effects. Models using more closely related species performed better in some regulatory regions that control gene activity. Different alignments provide the model with different patterns of evolutionary conservation.
The authors tested the approach on 34 quantitative traits in UK Biobank, a British project that collects genetic and medical data. In an analysis of 161 822 participants of European ancestry, they added three GPN-Star scores to DeepRVAT, a statistical method that uses rare variants to identify genes associated with traits. Averaged across three runs, the number of genes identified increased from 383 to 402. The number of gene-trait associations subsequently replicated in two larger UK Biobank studies increased from 338 to 353.
GPN-Star also requires full alignments across species when calculating scores during model application. Their fixed format is less suitable for DNA insertions, deletions, and large rearrangements. Where such alignments are available, GPN-Star connects the human sequence context with the evolutionary history of the corresponding region in related species and prioritizes variants for further testing.