Predicting genome-wide functional constraints with GPN-Star.

Genomic language model with a phylogeny-aware architecture, trained on whole-genome alignments and species trees

GPN-Star is a genomic language model whose architecture explicitly encodes whole-genome alignments and species trees. Trained on alignments spanning vertebrate, mammal and primate timescales, it reaches state-of-the-art performance across a wide range of variant effect prediction tasks, in both coding and non-coding regions of the human genome. It substantially outperforms previous methods at prioritizing pathogenic and fine-mapped GWAS variants, yields strong complex-trait heritability enrichments and improves power in rare variant association testing. Analyses across evolutionary timescales show task-dependent advantages of modelling recent versus deeper evolution. The framework was also trained for five model organisms (mouse, chicken, Drosophila, C. elegans, Arabidopsis), documenting its generalizability.

Synthesis written by Geno'X. For the full original abstract, please refer to the source publication.

What matters diagnostically is less the headline benchmark than the fact that the gain extends to non-coding regions, still the blind spot of exome and genome interpretation. The demonstration nonetheless rests on reference variant sets and association data rather than diagnostic cohorts: routine use would require ACMG/AMP threshold calibration (PP3/BP4) and gene-by-gene evaluation. For now this is a prioritization aid, not a classification criterion.

Analysis by Dr Thibaut Benquey