Recapitulation of patient-specific 3D chromatin conformation using machine learning and validation of identified enhancer-gene targets

Abstract

Regulatory networks containing enhancer to gene edges define cellular state and their rewiring is a hallmark of cancer. While efforts, such as ENCODE, have revealed these networks for reference tissues and cell-lines by integrating multi-omics data, the same methods cannot be applied for large patient cohorts due to the constraints on generating ChIP-seq and three-dimensional data from limited material in patient biopsies. We trained a supervised machine learning model using genomic 3D signatures of physical enhancer-gene connections that can predict accurate connections using data from ATAC-seq and RNA-seq assays only, which can be easily generated from patient biopsies. Our method overcomes the major limitations of correlation-based approaches that cannot distinguish between distinct target genes of given enhancers in different samples, which is a hallmark of network rewiring in cancer. Our model achieved an AUROC (area under receiver operating characteristic curve) of 0.91 and, importantly, can distinguish between active regulatory elements with connections to target genes and poised elements with no connections to target genes. Our predicted regulatory elements are validated by multi-omics data, including histone modification marks from ENCODE, with an average specificity of 0.92. Application of our model on chromatin accessibility and transcriptomic data from 400 cancer patients across 22 cancer types revealed novel cancer-type and subtype-specific enhancer-gene connections for known cancer genes. In one example, we identified two enhancers that regulate the expression of ESR1 in only ER+ breast cancer (BRCA) samples but not in ER-samples. These enhancers are predicted to contribute to the high expression of ESR1 in 93% of ER+ BRCA samples. Functional validation using CRISPRi confirms that inhibition of these enhancers decreases the expression of ESR1 in ER+ samples.

Discovery of novel therapeutic targets in cancer using patient-specific gene regulatory networks

“…We constructed regulatory networks for 371 patients from 22 cancer types using ATAC-seq and RNA-seq data 20 . The enhancer to gene edges were constructed using a random forest model that was trained using three-dimensional CTCF/cohesin ChIA-PET data and models the joint contribution of multiple ATAC-seq peaks to gene expression while accounting for confounders such as copy-number variations 22 . We then identified TF footprints at accessible regions and built edges between TFs and their target genes 25,26 (Figure 1A).…”

“…The predicted peak-gene links represent the targets of regulatory elements and are consistent across cell lines and cancer types used as gold standards. 22 Model was applied to T-47D, MCF7 and MDA-MB-231 cell lines to predict peak-gene connections.…”

“…Integration of ATAC-seq and RNA-seq data allows construction of directional regulatory networks 20,21 . While previous methods for regulatory network construction modeled the relationship between chromatin accessibility and gene expression in simplistic linear correlation-based methods, recently we have developed a computational method that models the joint contribution of multiple ATAC-seq peaks to gene expression using a random forest model resulting in more accurate patient-specific enhancer to target gene connections 22 . Here, we find that the increased accuracy of enhancer to target gene connections results in more accurate regulatory networks (with nodes consisting of TFs and target genes) and identification of key TFs.…”