Predicting gene expression state and prioritizing putative enhancers using 5hmC signal.

Area of Science:

• Epigenetics
• Genomics
• Computational Biology

Background:

• 5-hydroxymethylcytosine (5hmC) is a key epigenetic modification of DNA, derived from 5-methylcytosine (5mC).
• TET dioxygenases generate 5hmC at gene bodies and enhancers, correlating with active transcription and specific cell lineages.
• Predictive models for gene expression or regulatory regions based on 5hmC have been lacking.

Purpose of the Study:

• To develop predictive models of gene expression state using 5hmC enrichment data.
• To identify putative regulatory regions, including distal enhancers, by leveraging 5hmC signals.
• To establish a framework linking 5hmC to genome function and gene regulation.

Main Methods:

• Development of deep neural network models utilizing 5hmC enrichment in genic regions to predict gene expression.
• Application of an Activity-by-Contact model and a graph convolutional neural network incorporating Hi-C data and 5hmC for enhancer-promoter link prioritization.
• Validation of models across 49 diverse cell types.

Main Results:

• Neural network models accurately predict high vs. low gene expression states based solely on 5hmC levels.
• Models demonstrate generalization capabilities to unseen cell types.
• Prioritization of enhancer-promoter links identified known and novel putative enhancers for key genes in immune cells.

Conclusions:

• 5hmC plays a crucial role in gene regulation via both proximal and distal mechanisms.
• The developed methods provide a framework for linking 5hmC to genome function.
• Advances in DNA sequencing will facilitate routine 5hmC profiling, enabling broad applications of these predictive models.