Nucleotide dependency analysis of genomic language models detects functional elements.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Genomic language models (gLMs) learn from DNA sequences but lack interpretability for discovering functional elements.
• Identifying regulatory instructions and molecular machines encoded in genomes remains a significant challenge.

Purpose of the Study:

• To develop an interpretable method for deciphering functional genomic elements from gLMs.
• To introduce and validate nucleotide dependencies as a tool for genomic analysis.

Main Methods:

• Developed nucleotide dependency analysis to quantify the impact of nucleotide substitutions on genomic sequence probabilities.
• Compared nucleotide dependencies with alignment-based conservation and gLM reconstruction for variant deleteriousness prediction.
• Applied dependency analysis to identify regulatory motifs and RNA structural elements.

Main Results:

• Nucleotide dependencies outperform existing methods in predicting the deleteriousness of genetic variants.
• The method accurately detects regulatory motifs and reveals novel RNA structures, including pseudoknots and tertiary contacts.
• Dependency analysis identified limitations in current gLM architectures and training strategies.

Conclusions:

• Nucleotide dependency analysis provides a novel, interpretable approach for discovering and studying functional genomic elements and their interactions.
• This method enhances the utility of gLMs for understanding genome function and disease.
• The findings open new avenues for RNA structure prediction and analysis.