Chromatin interaction-aware gene regulatory modeling with graph attention networks

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Linking distal enhancers to genes is crucial for understanding gene regulation and noncoding genetic variation.
• Current methods struggle to accurately model the impact of enhancers on target gene expression.

Purpose of the Study:

• To introduce GraphReg, a novel deep learning approach for predicting gene expression using 3D genome interactions.
• To improve the modeling of gene regulation by incorporating chromosome conformation capture data.

Main Methods:

• GraphReg utilizes graph attention networks to analyze 3D interactions from chromosome conformation capture assays.
• It predicts gene expression from 1D epigenomic data or DNA sequence, considering genomic elements up to 2 Mb apart.
• Feature attribution methods are employed to identify functional regulatory elements.

Main Results:

• GraphReg outperforms existing state-of-the-art deep learning methods in predicting gene expression levels.
• Feature attribution with GraphReg accurately identifies functional enhancers, validated by CRISPRi-FlowFISH and TAP-seq assays.
• Sequence-based GraphReg accurately predicts direct transcription factor targets, confirmed by CRISPRi TF knockout experiments.

Conclusions:

• GraphReg offers a significant advancement in modeling the regulatory impact of epigenomic and sequence elements.
• The tool enhances the interpretation of noncoding genetic variation by accurately linking enhancers to genes.
• GraphReg provides a powerful new method for regulatory genomics research.