Prioritizing non-coding regions based on human genomic constraint and sequence context with deep learning.

Area of Science:

• Genomics
• Computational Biology
• Human Genetics

Background:

• Elucidating the function of non-coding genomic regions remains a significant challenge.
• Sequence variation intolerance in coding and proximal non-coding regions predicts human disease relevance.
• Existing methods often rely on evolutionary conservation, which may not fully capture functional constraint.

Purpose of the Study:

• To develop a deep learning model (JARVIS) for prioritizing functional non-coding genomic regions.
• To introduce a novel metric, the genome-wide residual variation intolerance score (gwRVIS), to quantify constraint.
• To assess the performance of JARVIS and gwRVIS against existing scores in identifying pathogenic variants.

Main Methods:

• Integration of variation intolerance, functional genomic annotations, and primary genomic sequence.
• Development of JARVIS, a deep learning model for non-coding region prioritization.
• Calculation of gwRVIS using whole-genome sequencing data from 62,784 individuals via a sliding-window approach.

Main Results:

• JARVIS outperforms existing human lineage-specific scores in prioritizing non-coding regions.
• JARVIS performs comparably to or better than conservation-based scores in classifying pathogenic variants, despite being evolutionarily agnostic.
• gwRVIS effectively distinguishes Mendelian disease genes from tolerant regions and identifies ultra-conserved non-coding elements as highly intolerant.

Conclusions:

• JARVIS and gwRVIS provide novel insights into human lineage-specific constraint in the non-coding genome.
• These tools enhance the ability to identify functionally important non-coding regions.
• The findings will advance our understanding of the non-coding genome and its role in human health and disease.