Predicting genome-wide functional constraints with GPN-Star.

GPN-Star, a phylogeny-aware genomic language model trained on vertebrate, mammal, and primate alignments, achieves state-of-the-art variant effect prediction across coding and non-coding human genome regions.

• Why it matters: Standard genomic language models are computationally heavy and still lag classical evolutionary models, leaving a gap in accurately learning genome-wide functional constraints from DNA sequence.
• What they did: GPN-Star uses whole-genome alignments and species trees in a biologically grounded architecture, then is evaluated on variant effect prediction tasks and human genetics prioritization, fine-mapping, and association testing.
• The result: It substantially outperforms prior methods for prioritizing pathogenic and fine-mapped GWAS variants, yields strong enrichments of complex trait heritability, and improves power in rare variant association testing, while generalizing to five model organisms.