Evolutionary trees power a genome-wide variant predictor
GPN-Star is a phylogeny-aware genomic language model trained on whole-genome alignments across vertebrates to learn functional constraints and predict genome-wide variant effects. Using a transformer with phylogeny-informed cross-attention, it achieves state-of-the-art performance in coding and non-coding variant interpretation, improves pathogenic variant prioritization and trait heritability enrichment, and extends to five model organisms. The results show that different evolutionary timescales inform different regulatory versus coding regions; while requiring WGAs for inference, GPN-Star delivers higher accuracy with modest compute compared to single-sequence models and provides public predictions and code to advance human and comparative genomics.
