Predicting functional constraints across evolutionary timescales with phylogeny-informed genomic language models

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Genomic language models (gLMs) show promise for functional genomics but often require large resources and underperform evolutionary models.
• Standard gLMs struggle to explicitly incorporate evolutionary relationships, limiting their predictive power.

Purpose of the Study:

• Introduce GPN-Star, a novel gLM with a phylogeny-aware architecture.
• Leverage whole-genome alignments and species trees to explicitly model evolutionary relationships.
• Improve variant effect prediction and prioritize clinically relevant genetic variants.

Main Methods:

• Developed GPN-Star, a biologically grounded gLM incorporating species tree and alignment representations.
• Trained GPN-Star on whole-genome alignments across various evolutionary timescales (vertebrate, mammalian, primate).
• Evaluated GPN-Star on diverse variant effect prediction tasks in coding and non-coding human genomic regions.

Main Results:

• GPN-Star achieved state-of-the-art performance in variant effect prediction across multiple genomic regions.
• The model demonstrated superior performance in prioritizing pathogenic and fine-mapped GWAS variants.
• GPN-Star showed significant enrichments of complex trait heritability and improved rare variant association testing power.
• The framework demonstrated robustness and generalizability across five diverse model organisms.

Conclusions:

• GPN-Star offers a scalable, powerful, and flexible tool for genome interpretation.
• The model effectively leverages comparative genomics data for enhanced functional constraint learning.
• GPN-Star has the potential to significantly advance human genetics and related fields.