A Phylogenetic Approach to Genomic Language Modeling.

Area of Science:

• Genomics
• Computational Biology
• Evolutionary Biology

Background:

• Genomic language models (gLMs) have limitations in identifying evolutionarily constrained elements in mammalian genomes.
• Accurate identification of constrained elements is crucial for understanding genome function and evolution.

Purpose of the Study:

• To develop a novel framework for training gLMs that explicitly models nucleotide evolution on phylogenetic trees.
• To improve the identification of evolutionarily constrained elements and functionally disruptive variants.

Main Methods:

• Developed a new gLM training framework incorporating phylogenetic tree-based nucleotide evolution modeling.
• Integrated multispecies whole-genome alignments into the model's loss function during training.
• The model, PhyloGPN, predicts disruptive variants from single sequences without requiring alignments during inference.

Main Results:

• The novel framework significantly enhances gLM performance in identifying constrained genomic elements.
• PhyloGPN demonstrates superior accuracy in predicting functionally disruptive variants compared to existing methods.
• The model exhibits strong transfer learning capabilities, applicable across different genomic contexts.

Conclusions:

• The proposed framework offers a powerful approach for training gLMs that account for evolutionary processes.
• PhyloGPN represents a significant advancement in predicting functional impacts of genetic variants.
• This method enhances the utility of gLMs for evolutionary genomics and variant effect prediction.