New phylogeny-aware genomic language model outperforms existing tools
Researchers introduced GPN-Star, a genome-wide language model that incorporates species trees and whole-genome alignments, achieving state-of-the-art performance on variant effect prediction across humans and several model organisms.
The study presents GPN-Star, a genomic pretrained network that integrates species-tree data and whole-genome alignments within a phylogeny-aware transformer, enabling explicit modeling of evolutionary relationships. Trained on three evolutionary timescales—vertebrate, mammal and primate—the model outperforms traditional conservation tools (PhyloP, PhastCons) and large single-sequence language models on a broad suite of benchmarks, including pathogenic coding and non-coding variant classification, GWAS fine-mapping, and rare-variant association testing.
GPN-Star also delivers stronger heritability enrichment for over a hundred complex traits, especially when using the primate-scale model. The framework was extended to five additional model organisms, where it consistently yielded higher rare-variant enrichment than comparable methods. Analyses of nucleotide dependencies reveal biologically meaningful co-evolutionary signals, demonstrating that the model captures functional genomic syntax without supervision. Pretrained models and genome-wide predictions have been released publicly for community use.
Why it matters
Accurate prediction of functional genomic constraints can accelerate disease gene discovery and improve genetic risk assessment.
In this story
