Deciphering the contribution of 3D interactions between cis-regulatory elements and promoters to regulate gene expression using graph neural networks

Gene expression is regulated by various factors including nearby histone modifications (HMs), binding of transcription factors (TFs), and interplay of diverse cis-regulatory elements (CREs) in a cooperative manner. Two previous methods (GC-MERGE and Chromoformer) have been developed to predict gene expression by constructing a promoter-centered network based on 3D physical contacts between the promoter and CREs. However, these published models lack transferability to various size of promoter-centered graphs and quantify the influences of 3D physical contacts in gene regulation. In our approach, by jointly modeling multi-omics data, we can determine the main factors that regulate gene expression and can identify important CREs for each gene.

We present a novel deep learning architecture CreGNN based on a graph neural network (GNN) to decipher the contribution of 3D interactions among cis-regulatory elements (CREs) in gene expression. The transformer attention mechanism was applied to each pair of nodes in a graph. We used ~800 ChIP-seq datasets of HMs and TFs, Hi-C/micro-C, ATAC-seq and RNA-seq data from human, mouse and fly genomes. Promoter-centered graphs were built for each gene based on the open accessible or enhancer regions across the genome that contact the promoter in 3D space.

First, we trained a model based on the HMs and calculated the Pearson correlation, AUC, and AUPRC between raw gene expression and predicted gene expression in four cell lines. CreGNN showed better prediction performance relative to other existing methods. Second, we trained another model based on TFs, which showed slightly lower performance relative to HMs. By ranking node importance, we found that TFs and HMs coregulate gene expression by binding to different CREs. Third, we test the effect of 3D contacts on gene expression with different input. The prediction performance of other CREs were significantly lower than the promoter and the model with promoter only can achieve similar performance with promoter-centered graphs. 3D contact frequency showed low correlation with attention scores. Our results suggest that gene expression is dictated by promoter region activation or repression by CREs and that 3D contacts are necessary in specific cases but provide a lower relative contribution to overall gene expression.

Compared to previously developed methods, CreGNN exhibited higher accuracy and good transferability. Our method provides new insights into the function of specific CREs on respective gene promoters based on multi-omics data, and we systematically estimated the feature importance and node importance of 3D genome contacts, CREs and epigenomic features. We aspire to use our method to identify novel specific 3d contacts among promoters and CREs that contribute to the regulation of specific genes. This predictive power should not only contribute to our knowledge of specific gene regulatory mechanisms on the 3D level but also allow us to design targeted strategies to modify gene expression within individual cells.