Deep learning the cis-regulatory code for gene expression in selected model plants

Area of Science:

• Genomics
• Bioinformatics
• Plant Science

Background:

• Understanding gene regulation is key to interpreting genetic variation.
• Non-coding regulatory elements significantly influence gene expression.

Purpose of the Study:

• To develop interpretable deep learning models for predicting plant gene expression from flanking DNA sequences.
• To identify conserved and species-specific regulatory elements across multiple plant species.
• To link genetic variation to gene expression changes and phenotypic differences.

Main Methods:

• Training interpretable deep learning models on gene flanking regions from Arabidopsis thaliana, Solanum lycopersicum, Sorghum bicolor, and Zea mays.
• Utilizing predictive feature selection to identify important regulatory sequences.
• Applying models to analyze genetic variation across fourteen tomato genomes.
• Predicting genotype-specific expression for functional gene groups.

Main Results:

• Models achieved over 80% accuracy in predicting gene expression.
• Identified Untranslated Regions (UTRs) as significant determinants of gene expression levels.
• Demonstrated strong cross-species performance, distinguishing conserved and unique regulatory features.
• Revealed causal links between genetic variation and gene expression changes in tomato.
• Successfully predicted genotype-specific expression, highlighting phenotypic differences.

Conclusions:

• Interpretable deep learning models are effective tools for dissecting gene regulation in plants.
• UTR regions play a critical role in controlling gene expression.
• The approach can identify both conserved and species-specific regulatory mechanisms.
• This methodology facilitates the understanding of genotype-phenotype relationships driven by gene expression variation.