Applications of ENCODE data to Systematic Analyses via Data Integration

Abstract

Large-scale genomic data have been utilized to generate unprecedented biological findings and new hypotheses. To delineate functional elements in the human genome, the Encyclopedia of DNA Elements (ENCODE) project has generated an enormous amount of genomic data, yielding around 7,000 data profiles in different cell and tissue types. In this article, we reviewed the systematic analyses that have integrated ENCODE data with other data sources to reveal new biological insights, ranging from human genome annotation to the identification of new candidate drugs. These analyses demonstrate the critical impact of ENCODE data on basic biology and translational research.

Graphical Abstract

Introduction

The development of high-throughput technologies has generated enormous amounts of data which allow biologists to examine numerous biological hypotheses. In this review, we discuss how the Encyclopedia of DNA Elements (ENCODE) project has contributed to our understanding of many aspects of biology. The ENCODE project is an international collaborative project funded by the National Human Genome Institute (NHGRI). It aims to identify and characterize functional regions in the human genome by utilizing a variety of high-throughput approaches [1].

The pilot phase (2003–2007) of the ENCODE project was launched to explore 1% of the human genome to establish protocols for scaling up analyses to the entire genome [2–4]. During the pilot phase, a variety of experimental and computational methods were compared and refined for large-scale analyses. Then, the ENCODE project was expanded to the entire human genome during the production phase (2007–2012). In alignment with ENCODE, the modENCODE project [5] and the mouseENCODE project [6] were launched to systematically identify the DNA elements in the genomes of three model organisms including fly, worm and mouse. To date, ENCODE has generated abundant genomic data of different types as well as various tools and software. In total, 7,694 profiles of different assay categories have been released to the public so far as summarized in Table 1 [7].

Table 1.

Summary of ENCODE data [8].

Assay category		DNA binding	Transcription	DNA-accessibility RNA-binding		DNA-methylation Replication timing		Genotyping	3D-structure	Proteomics
Assay name		ChIP-seq	shRNA RNA-seq, toal RNA-seq, RNA microarray, small RNA- seq, RAMPAGE, polyA RNA-seq, CAGE, CRISPRi RNA-seq, si RNA RNA-seq, CRISPR RNA-seq, single cell RNA-seq, microRNA counts, microRNA-seq, polyA depleted RNA-seq, RNA-PET	DNase-seq, ATAC-seq, genetic modification DNase-seq, FAIRE-seq, MNase-seq	eCLIP, RIP- seq, RIP- chip, ¡CLIP, Switchgear	DNAme array, RRBS, WGBS, MRE-seq, MeDIP-seq	Repli-seq, Repli- chip	genotyping array, DNA- PET, genotyping HTS	ChlA-PET, Hi-C, 5C	MS-MS
Tier1	K562	697	506	52	237	8	6	5	11	6
	GM1287B	249	60	5	25	10	6	5	4	5
	H1-hESC	106	21	3	3	6	1	1	1	2
Tier2		978	481	38	172	34	40	13	16	1
Tier3		1065	400	255	6	252	61	86	40	0
Other		926	382	190	0	116	35	32	34	0

The number of datasets for each assay category across different Tiers, cell types and tissues are shown until May, 2018 (ENCODE data portal: https://www.encodeproject.org). The Tiers refer to the ENCODE classification system of cell lines according to priority of performing the assays. Tier1 cell lines were considered of the highest priority in ENCODE project and therefore the included cell lines (K562, GM12878 and H1-hESC) are presented individually. The assay category refers to the type of genomic features described by the assay.

Based on these data, many important biological analyses have been performed, including genome annotation, chromatin state classification, and the identification of regulatory regions in the genome [2,3]. Moreover, ENCODE data has been integrated with other data sources such as cancer data to gain new insights into cancer development and drug discovery. Up to May, 2018, 2563 articles have been published by the ENCODE project and its community, and 6,683 articles that cited the ENCODE landmark paper have been published (Figure 1) [2]. In this review, we focus on systematic analyses that have integrated ENCODE with other data sources to better understand complex biological processes and human diseases.

Figure 1.

Publications and citations based on the ENCODE project [2]. (A) The number of publications from ENCODE and its community until May, 2018 (ENCODE data portal: https://www.encodeproject.org). ~2007: all publications before 2007; 2015 ~: all publications after 2015. (B) The number of publications that have cited the ENCODE landmark paper until May, 2018 [2]. Citations are based on Google Scholar (https://scholar.google.com). ~2013: all publications before 2013.

Improving human genome annotation

The Human Genome Project has completed the sequence of the human genome in 2003, however, at that time the annotation of the genome was far from accurate [8]. Based on ENCODE data, protein-coding and noncoding transcripts, long non-coding RNAs and pseudogenes have been carefully annotated through a combination of computational analyses, manual annotation, and experiment validation [9,10]. According to the refined annotation, it is now estimated that a total of 74.7% of the human genome is covered by primary transcripts [3,9]. High-quality annotation would facilitate functional investigation of any genomic sequence in the human genome. For example, highthroughput annotations of long non-coding RNAs make it possible to explore their properties and functions with high efficiency [*11,12].

Using ChIP-seq data for eight histone modifications and one sequence-specific insulator protein, CTCF, Ernst et al. developed an algorithm named ChromHMM that segmentized the genomes of nine human cell lines into 15 different chromosome states [13,14]. Additionally, Hoffman et al. used 31 ChIP-seq, DNase-seq and FAIRE-seq profiles to develop a method called Segway that defined 25 chromosome states [15]. Using the integration of these two segmentation methods, the ENCODE project established a consensus set of seven principle genomic states: three related to gene transcription, three related to distal regulatory elements, and one related to actively repressed or inactive genomic regions [2]. Subsequent experimental studies confirmed the defined histone modification states of distal regulatory elements [16]. Additional segmentation methods have utilized this initial segmentation effort and expanded it in various ways, including methods that integrate cell type specificity [17], evaluate species-specific confounding factors [18], or take into account higher-order chromatin structure [19,**20]. Undoubtedly, the ENCODE project has significantly improved human genome annotation and segmentation.

Predicting transcription factor targets, enhancers, and enhancer-promoter interactions

Transcription factors (TFs) regulate gene transcription by binding to specific DNA elements in open chromatin regions. Human genome annotation based on ENCODE data has identified chromatin regions highly enriched for TF binding sites (TFBSs) [2]. Many promoters contain TF binding motifs, but only a small fraction of these motifs is actually bound by TFs and thus contain TFBSs. Additionally, enhancers that play critical roles in gene transcriptional regulation are also enriched for TFBSs. ENCODE data, including chromatin accessibility, histone modification, and TF binding data, provide an excellent opportunity to systematically identify TF targets, enhancers, and enhancerpromoter interactions.

ChIP-seq assays provide a powerful method to identify TFBSs throughout the entire genome. However, it may not be feasible to apply this assay to all human TFs across all cell types under various conditions (e.g., medium, drug treatment, etc.), due to its timeconsuming properties. To overcome this challenge, a number of computational methods have been developed to predict context-specific TF target genes by integrating diverse genomic data, such as histone modifications, Position Weight Matrix (PMW) motif information and DNase I hypersensitive sites (DHSs) [21–24]. The CENTIPEDE method developed by Pique-Regi et al. constructed prediction models that integrate DHS data and PMW motif information [22]. In this initial study, CENTIPEDE identified a total of 827,000 TF binding motifs in human lymphoblastoid cell lines [22]. Follow-up studies further expanded on this knowledge or adapted CENTIPEDE to increase prediction accuracy [25–27]. Another method called protein interaction quantitation (PIQ), developed by Sherwood et al., is also based on the integration of DHS data and PMW motif information, and reported improved prediction accuracy [28].

Enhancers constitute another class of regulatory elements in the genome, which physically bind to specific TFs and interact with promoters of target genes via distant chromatin interactions. A number of computational methods have been developed to predict enhancers in the genomes [29–31]. In one study, Yip et al. identified 13,539 potential enhancers based on gene-distal regulatory modules using TF ChIP-seq data and then validated some of them experimentally [30]. In another study, Rajagopal et al. constructed a random forest-based algorithm that identifies enhancers in multiple cell lines based on histone modification patterns from ENCODE data [29]. Many other enhancer prediction methods have been developed into programs or packages that are publicly available from bioinformatics websites such as OMICtools (https://omictools.com/enhancer-prediction-category).

The integration of ENCODE data and other data sources allows computational prediction of enhancer-promoter (EP) interactions in a systematic manner. In a largescale study, 1,046 regulatory elements, including enhancers, were connected to their target promoters based on data from DHS, ChIP-seq and Chromosome Conformation Capture Carbon Copy (5C) experiments [23]. Cao et al. developed a computational method called joint effect of multiple enhancers (JEME) and applied it to predict EP interactions in 935 human primary cell types, cell lines and tissues by the integration of histone modification, DNase-seq, RNA-seq and other data types [*32]. These analyses have revealed critical insights into EP interactions, including that i) they are not simply determined by genomic proximity [33–35]; ii) multiple enhancers might control the same promoter [36]; and iii) EPs are cell-type specific [37,*38].

Quantifying the relation between gene expression, transcription factor binding and histone modifications

Inappropriate gene regulation underlies a variety of human diseases and therefore there has been a long-standing interest in the prediction of gene expression. TFs are essential in gene expression regulation and different approaches have been undertaken to utilize TFs as an indicator of gene expression.

The work of Ouyang et al. was one of the first studies to systematically investigate the contribution of TF binding signals captured by ChIP-seq assay to gene expression [39]. Using the intensities of ChIP-seq peaks near genes of interest as a measure of TF association strength, they suggested that the activity of 12 TFs explains a large portion of the gene expression variation observed in mouse embryonic stem cells [39]. A subsequent study used a similar log-linear regression-based approach to infer TF regulation and also incorporated histone modification data, which further improved gene expression prediction [40]. Following this, other studies based on different machine learning approaches have confirmed the quantitative relationship between gene expression levels, TF binding and histone modification signals [41,42]. As reported in the ENCODE landmark paper [2], more than 50% variation of gene expression can be explained by TF binding and/or histone modification signals in the promoter proximal DNA regions in the human genome. The same conclusion has also been made in several other species including yeast, fly, worm and mouse [5]. Thus, by utilizing genomic data from ENCODE and modENCODE, it has become possible to model transcriptional regulation of genomes in a quantitative, and likely, a dynamic manner.

Inferring regulatory activity of transcription factors

Inferring the regulatory activity of TFs has been another application of ENCODE data. TF genes tend to have relatively low expression levels and TF activity is intensively regulated at the post-transcriptional and post-translational level [45,46]. As such, the mRNA levels of TFs often do not accurately reflect their regulatory activity and therefore may not manifest their functions [43,44]. To overcome this issue, computational methods have been developed to infer TF activity based on the expression level of their target genes. The ENCODE project has provided 1,864 ChIP-seq profiles for over 100 human TFs in different cell lines, enabling the identification of TF target genes in an unbiased and systematic fashion.

Cheng et al. developed a rank-based statistical framework that summarizes the relative expression levels of genes regulated by a TF to infer its activity in each sample. In this framework, target genes of a TF are modeled as probabilistic events rather than being defined as a definite gene set [45]. Jiang et al. developed a multivariate linear model to integrate all ENCODE ChIP-seq TF binding profiles with 7,484 tumor gene expression profiles to systemically estimate the activities of 150 human TFs in different cancer types. Some of the inferred TF activities were then validated using knock out experiments in the K562 and HL60 cell lines [46]. Additionally, several other methods that rely on gene expression to infer TF activities have been developed based on different statistical models [47,48]. In summary, the integration of ENCODE ChIP-seq data and gene expression profiles has enabled statistical inference of TF regulatory activities under different biological settings, and therefore led to more accurate investigation of TF functions and transcriptional regulatory mechanisms.

Characterizing genomic alterations in cancer

The studies described in the previous sections have been instrumental to our understanding of genomic alterations in human diseases, including cancer. ENCODE data have been used in various pivotal studies that have characterized cancer-specific alterations in regulatory elements through the development of methods and accumulation of sequence-based studies. For example, global enhancer activation has been shown in almost all cancer types and might play a role in genomic rearrangements that are characteristic of cancers [**49]. The elucidation of EP interactions can identify enhancer-target networks for specific cancer types and point to cancer-specific alterations [*32].

Additionally, it is now recognized that alterations affecting regulatory regions are potentially as important in cancer progression as alterations in protein coding regions or those that directly alter functional RNA molecules [*50]. Epigenetic changes have been proposed to play a major role in the mutational landscape of cancers and might explain up to 86% of mutational rates in cancer genomes together with replication timing [51]. This association between chromatin features and mutation density is highly cell-type specific and mutations tend to be enriched in heterochromatin [52]. Accordingly, coding regions and regulatory elements seem to contain fewer mutations compared to intergenic, non-regulatory regions [53,54]. Due to the complexity of mutations in noncoding regions, several algorithms have been developed to assess tumor-specific somatic variants from tumor genomes and obtain a short list of candidate driver mutations or cancer-specific TFs [46,55–57].

The activities of specific TFs have been linked to cancer development and prognosis based on ENCODE data. For example, the activity but not the expression of E2F4, a TF involved in cell cycle regulation, has been reported to be associated with prognosis of breast cancer patients [58]. In this study, E2F4 activity in tumor samples was inferred based on the expression of its target genes, which were determined based on ENCODE ChIP-seq data. A single-cell sequencing study in glioblastoma has revealed stemnessrelated expression states that might be driven by a set of core TFs, including POU3F2, SOX2, SALL2 and OLIG2 [59]. Lastly, the integration of ENCODE ChIP-seq data and data from The Cancer Genome Atlas (TCGA) has proposed new TF oncogenes in several cancer types [46].

These integrative analyses of ENCODE data in the cancer field showed that large data compendia like ENCODE could identify genomic regions which are potentially more strongly linked to the biology of cancer. ENCODE data served as an essential component, and through the integration of data from other sources, such as TCGA, have moved cancer research forward.

Prioritizing disease-associated genetic variants from genome wide association studies

Disease- and trait-associated genetic variants are rapidly being identified with genomewide association studies (GWAS) and related strategies. Almost 85 million single nucleotide polymorphisms (SNPs) have been identified in the human genome so far [60]. Interestingly, most disease-associated SNPs are located in non-coding regions of the genome, and are equally distributed between intronic and intergenic regions [61–63].

The ENCODE Consortium has been one of the pioneers in integrating GWAS associations with genomic data to annotate GWAS associations [2]. Up to 71% of GWAS SNPs may have a potential causative SNP overlapping a DHS, and 31% of loci have a candidate SNP that overlaps a binding site occupied by a TF [2], which is consistent with the suggestion of positive selection in TFBSs [64]. However, it has been suggested that SNPs proposed by GWAS are often not the associated functional SNP in regions of interest [65]. Additionally, disease-associated SNPs have been shown to be significantly enriched in cell-type-specific enhancers and regulatory regions [14,66]. For example, SNPs associated with hematological disorders were most enriched in erythrocyte leukemia cell enhancers [14], whereas SNPs in TFs associated with similar functions, such as glucose homeostasis or immune regulation, might predispose individuals to diabetes or autoimmune diseases [63,67]. Thus, the integration of GWAS and ENCODE data has provided intriguing results and the expansion of both platforms will hopefully lead to a better understanding of the role of SNPs in human diseases.

Discovering new candidate drugs

Finally, ENCODE data have also been used to predict new candidate drugs, providing an example of translational research that connects genomic data with biomedical studies. In one study, ENCODE ChIP-seq data were integrated with RNA-seq data to identify drugs that might alter TF activity in specific diseases [**68]. In another study, Chen et al. showed that they could use ENCODE data to identify genome-wide signatures of TF activity and proposed that these signatures could be utilized to identify new candidate drugs [69]. They validated their framework by showing that commonly used drugs against estrogen receptor positive (ER+) breast cancer, such as tamoxifen and fulvestrant, displayed the highest inhibitory potential in ER+ breast cancer and proposed novel drugs with benefit in this disease, such as the anti-inflammatory drug oxaprozin [69]. Additionally, Gayvert et al. developed a method, called Computational drug-Repositioning Approach For Targeting Transcription factors (CRAFTT), which identifies molecules that can indirectly modulate TFs of interest. For example, they validated the prediction of dexamethasone to inhibit the TF ERG, which is associated with several oncogenic translocation events [70]. Although examples of pharmacogenomic analysis using ENCODE data remain limited, these studies demonstrated the great potential of using ENCODE data to promote translational researches.

Conclusions

In summary, high-throughput data from ENCODE have provided us with an unprecedented wealth of genomic information. These data have not only significantly enhanced our understanding of the human genome, but also provided new insights into other research areas from transcriptional regulation to disease mechanisms and drug development. Although our discussion only touched upon a few examples, we envision even more applications of ENCODE data through integrations with other data sources.

Highlights.

• The ENCODE project has produced abundant genomic data of different categories.

• Systematic analyses based on ENCODE data have revealed critical biological insights.

• The integration of ENCODE with other data sources has provided novel knowledge on human diseases.