Integrating long-range regulatory interactions to predict gene expression using graph convolutional networks

Abstract

Long-range spatial interactions among genomic regions are critical for regulating gene expression and their disruption has been associated with a host of diseases. However, when modeling the effects of regulatory factors on gene expression, most deep learning models either neglect long-range interactions or fail to capture the inherent 3D structure of the underlying biological system. This prevents the field from obtaining a more comprehensive understanding of gene regulation and from fully leveraging the structural information present in the data sets. Here, we propose a graph convolutional neural network (GCNN) framework to integrate measurements probing spatial genomic organization and measurements of local regulatory factors, specifically histone modifications, to predict gene expression. This formulation enables the model to incorporate crucial information about long-range interactions via a natural encoding of spatial interaction relationships into a graph representation. Furthermore, we show that our model is interpretable in terms of the observed biological regulatory factors, highlighting both the histone modifications and the interacting genomic regions that contribute to a gene’s predicted expression. We apply our GCNN model to datasets for GM12878 (lymphoblastoid) and K562 (myelogenous leukemia) cell lines and demonstrate its state-of-the-art prediction performance. We also obtain importance scores corresponding to the histone mark features and interacting regions for some exemplar genes and validate them with evidence from the literature. Our model presents a novel setup for predicting gene expression by integrating multimodal datasets.

A generalizable framework to comprehensively predict epigenome, chromatin organization, and transcriptome

“…In RNA-seq GEP, there exists two tasks, namely binary gene expression classification and gene expression value regression for each protein-coding gene. Since gene expression can be predicted from histone marks, several models [6, 5, 31] are proposed to predict RNA-seq gene expression from several histone mark profiles within-cell type (i,e., using the cell-type specific histone marks to predict gene expression in the same cell type). Compared to these models, EPCOT utilized less epigenomic feature data (EPCOT does not require histone modification profiles as inputs) and achieved comparable performances in within-cell type prediction.…”

Chromatin interaction aware gene regulatory modeling with graph attention networks

“…In this way, GATs weigh enhancer-enhancer (E-E) and enhancerpromoter (E-P) interactions, based on the features learned in the promoter and enhancer bins, in order to predict CAGE values more accurately. However, non-attention-based GNNs such as graph convolutional networks (GCN) (Kipf and Welling, 2017;Bigness et al, 2021) fail to learn the importance of individual interactions. It has been shown that GATs outperform GCNs in other machine learning contexts as well (Veličković et al, 2018).…”

“…However, these previous linear models use a fixed assignment of regulatory elements to genes without incorporating 3D interaction data, while current deep learning models consider relatively local features, such as promoters and at most nearby enhancers, and therefore cannot capture the impact of distal regulatory elements, which can be 1Mb or farther away from gene promoters. There are two studies that tried to use 3D data to predict gene expression (Bigness et al, 2021;Zeng et al, 2019), but they failed to address important aspects of modeling gene regulation. The most directly relevant one, GC-MERGE (Bigness et al, 2021), uses histone modification and Hi-C data to predict gene expression (RNA-seq) using graph convolutional networks (GCN); however, they do not provide any insight about gene regulation rules, such as finding functional enhancers or revealing the role of TF binding motifs on gene regulation.…”

“…There are two studies that tried to use 3D data to predict gene expression (Bigness et al, 2021;Zeng et al, 2019), but they failed to address important aspects of modeling gene regulation. The most directly relevant one, GC-MERGE (Bigness et al, 2021), uses histone modification and Hi-C data to predict gene expression (RNA-seq) using graph convolutional networks (GCN); however, they do not provide any insight about gene regulation rules, such as finding functional enhancers or revealing the role of TF binding motifs on gene regulation. Our proposed method, GraphReg, is the most comprehensive and versatile model that is capable of answering interesting biological questions regarding gene regulation.…”

Learning the histone codes of gene regulation with large genomic windows and three-dimensional chromatin interactions using transformer

“…There are few emerging studies that explicitly take the 3D chromatin interactions into consideration to predict the gene expression. One such example is GC-MERGE 12 , a graph neural network (GNN) to propagate information between interacting genomic regions to predict the expression levels of genes. Although it is a proof-of-concept model that cannot be applied to genes without any chromatin interactions and only performs 10kbp genomic bin-level predictions but not at gene-level, it still underscores the promise of modeling epigenomic contexts of distal genomic regions along with that of promoters.…”