Species-aware DNA language models capture regulatory elements and their evolution.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Large-scale multi-species genome sequencing advances understanding of gene regulation.
• New algorithms are required to identify regulatory elements using evolutionary conservation.

Purpose of the Study:

• Introduce species-aware DNA language models for regulatory element identification.
• Evaluate their ability to capture conserved regulatory elements across evolutionary distances.

Main Methods:

• Trained DNA language models on over 800 species.
• Assessed model performance in predicting masked nucleotides and distinguishing functional motifs.
• Compared model capabilities with sequence alignment for capturing distant evolutionary conservation.

Main Results:

• DNA language models distinguish transcription factor and RNA-binding protein motifs from non-coding sequences.
• Models capture conserved regulatory elements over greater evolutionary distances than sequence alignment.
• Models accurately reconstruct in vivo bound motif instances and account for motif evolution.

Conclusions:

• Species-aware DNA language models are powerful tools for integrating data from highly diverged genomes.
• These models offer a flexible and scalable approach to analyzing genomic regulatory information.
• The models demonstrate a sophisticated understanding of sequence and evolutionary context in gene regulation.