Deciphering the genetic code of DNA methylation

Abstract

DNA methylation plays crucial roles in many biological processes and abnormal DNA methylation patterns are often observed in diseases. Recent studies have shed light on cis-acting DNA elements that regulate locus-specific DNA methylation, which involves transcription factors, histone modification and DNA secondary structures. In addition, several recent studies have surveyed DNA motifs that regulate DNA methylation and suggest potential applications in diagnosis and prognosis. Here, we discuss the current biological foundation for the cis-acting genetic code that regulates DNA methylation. We review the computational models that predict DNA methylation with genetic features and discuss the biological insights revealed from these models. We also provide an in-depth discussion on how to leverage such knowledge in clinical applications, particularly in the context of liquid biopsy for early cancer diagnosis and treatment.

Introduction

DNA methylation in the mammal genomes is the addition of a methyl group to cytosines to form 5-methylcytosine (5mC), primarily at CG and also at CH (CH=CA, CT, CC) sites. DNA methylation in specific loci plays important roles in many biological functions. For example, DNA methylation in promoters represses gene transcription, and DNA methylation in gene bodies is associated with transcription elongation and splicing [1–3]^,. The synergy between DNA methylation and local histone modification is also locus-specific in development, somatic cell reprogramming and tumorigenesis [4,5]. Understanding how DNA methylation is established, maintained and removed in a particular locus is thus critical.

Locus-specific DNA methylation or demethylation depends on the recruitment of specific enzymes such as TET and DNMTs to the target genomic regions [6–8] (Figure 1A). DNA methylation is catalyzed by DNA methyltransferases (DNMTs) [9]. De novo methylation on both DNA strands involves DNMT3A/3B/3 L. Existing DNA methylation is maintained by a complex of DNMT1 and UHRF1, which recognizes half-methylated DNA strand (hemimethylation) after replication. Removal of the methyl group from cytosines is catalyzed by the ten-eleven-translocation enzymes (TET1/2/3), which can oxidize 5mC to 5-hydroxymethylcytosine (5hmC) and other oxidized cytosines (5-formylcytosine, 5fC, and 5-carboxylcytosine, 5caC), and then demethylate to cytosine through various pathways [10].

Figure 1 .

Mechanisms of locus-specific DNA methylation and demethylation.

Emerging evidence has suggested that enzymes like DNMTs and TETs are recruited to specific genomic regions by factors recognizing certain DNA sequences [6,11]. Recently, we have systematically identified 313 DNA motifs that regulate DNA methylation from 34 whole-genome methylomes. We show that these motifs are functional and can be applied to improve cancer prognosis and diagnosis [12]. In this review, we first survey the mechanisms proposed in the literature that orchestrate DNA methylation. We also review machine learning models that derive the genetic features of DNA methylation and discuss the biological insights revealed from these models. Finally, we propose to combine DNA methylation associated motifs and genetic mutations in clinical applications for liquid biopsy and early cancer diagnosis. We show how this approach improves the current paradigm where the discovery of biomarkers is focused on a small number of genes.

The Emerging DNA Features of Locus-specific Methylation

Accumulating evidence has shown that certain DNA sequence patterns are associated with local DNA methylation levels, such as lower GC content, enrichment of short nucleotide combinations (2–6 bp) and longer DNA motifs [13–23]. However, a puzzling observation is that the modifying enzymes including TETs and DNMTs do not have high recognition specificity of DNA motifs [6,11]. While TET1, TET3 and DNMT1 all possess a CXXC domain [24] interacting with DNA sequences, the CXXC domain mainly recognizes unmodified CpG dinucleotide. Recently, Xu et al. [25] identified four groups of DNA sequences bound by CXXC domains, all of which are at the CpG candidates for DNA methylation. Importantly, the DNA preference of the CXXC domains cannot explain how ~80% of the 28 million CpGs in the human genome are methylated (or how ~20% of the CpGs are unmethylated).

Furthermore, numerous reports have confirmed the existence of DNA sequences that dictate where DNA methylation/demethylation occurs. For example, Lienert et al. [26] have identified methylation-determining regions, which mediate de novo methylation and demethylation. Interestingly, these regions contain cis-regulatory motifs that can be recognized by DNA-binding factors (SP1, CTCF, Rfx), and mutating these motifs alters the methylation pattern. Stadler et al. [27] have shown that introducing CTCF motifs is necessary and sufficient to lower methylation of nearby CpGs.

Taken together, these reports suggest the locus-specificity of DNA methylation is encoded in the genomic sequence, recognized, and mediated by locus-specific factors. Here, we review the emerging mechanisms of locus-specific DNA methylation guided by cis-acting DNA sequences, through cross-talks between transcription factors (TFs), DNMTs, TETs, DNA secondary structures and histone modifications (Figure 1B).

TFs recruit TETs for active demethylation

TET1 and TET3 can contain DNA-binding CXXC-zinc finger domain [25]. TETs prefer CpG-rich sequences such as CpG island (CGI) which spans several kilobases [28] and can bind CpG-rich DNA sequences [6] in mammals to maintain stable demethylation [29]. In addition, TET recruitment through locus-specific TF binding has been widely reported. For example, introducing a CTCF binding site at a particular locus leads to TET recruitment and local DNA demethylation [27]. PPARG binds to promoters and recruits TET for demethylation [30]. In a recent study, Suzuki et al. [31] have designed a method to screen for TFs that can facilitate DNA demethylation in a site-directed manner. In particular, they transduced selected TFs in sub-cloned vectors to cells and evaluated the methylation levels using the HumanMethylation450 methylation array near the TF binding sites (estimated by the motif locations) with and without ectopic expression of the TFs. Using this strategy, Suzuki et al. [31] have shown that RUNX1 site-specific binding correlates with demethylation in hematopoietic cells, and they have further confirmed recruitments of critical proteins involved in DNA demethylation, including TET2, TET3, TDG and GADD45, using co-immunoprecipitation. Suzuki et al. [32] further scaled-up this strategy and found that eight (RUNX3, GATA2, CEBPB, MAFB, NR4A2, MYOD1, CEBPA and TBX5) out of 15 (plus NANOG, HNF1A, PAX4, Nkx2–5, SOX2, POU5F1, HNF4A) tested TFs can facilitate demethylation of DNA in a site-directed manner.

TFs block DNMT3s and prevent de novo methylation

Many TFs can maintain low methylation by blocking the access of DNMTs to specific regions. For example, SP1 preferentially binds to CpG-rich promoters, preventing de novo methylation in mice [33,34]. Proteins containing a CXXC domain (CFP1, MLL, KDM2A/2B, IDAX) can bind to unmethylated CpGs to keep the region from being methylated [24,35,36]. Interestingly, DNMT1 has a CXXC domain, which may facilitate its binding to hemimethylated CpGs [37]; TET1 and TET3 also have a CXXC domain, which has been shown to contribute to their locus-specificity [38,39]. However, other studies have shown that the CXXC domain failed to restrain the activity of Dnmt1 on unmethylated CpG sites [40].

TFs recruit DNMTs for de novo methylation

Similarly, many TFs have been reported to facilitate DNA methylation in particular loci. For example, NR6A1 (or GCNF) can silence Oct-3/4 by binding to its promoter and recruit Dnmt3a and Dnmt3b in the mouse, facilitating methylation [41]. Dnmt3a has been reported to interact with Myc and specifically target the promoter of p21Cip1, leading to transcription repression [42]. Dnmt3b is recruited by the TF E2F6 to silence germ-line genes in murine somatic tissues [43].

DNA secondary structure shape DNA methylation

Besides TF-directed locus-specific methylation, DNA secondary structure has also been reported to shape local DNA-methylation. For example, Clark and Smith [44] showed that variable number tandem repeats (VNTR) at a non-B DNA structure contributes to abnormal DNA methylation in human breast cancers. Mao et al. [45] reported G-quadruplex (G4) DNA secondary structures are associated with hypomethylation at the CGI in the human genome. This is because G4 sites are enriched with DNMT1 binding but inhibit DNMT1 enzymatic activity, leading to the inhibition of local CpG methylation. Other studies have shown a certain group of G4 structures play roles in both DNA methylation and histone modification [46]. Meanwhile, G4 secondary structures are characterized by strong telomeric repeats, with cis-acting DNA motifs such as (GGGGCC)(n), TG (4)T(2)G(4) T and GGGCT(4) GGGC [47–49], which are GC-rich motifs that associate recruitment of TETs and hypomethylation [12,13,50]. Taken together, the DNA secondary structure provides another mechanism of how DNA sequence maintains and alters local methylation.

Same factors involved in both methylation and demethylation

Some factors are involved in both site-specific methylation and demethylation. For example, SPI1 can mediate both de novo methylation (by interacting with DNMT3B) and demethylation (by interacting with TET2) in a site-specific manner [51,52]. CTCF is another example that has opposite roles in regulating DNA methylation. CTCF can promote unmethylation through blocking DNMTs. For example, Schoenherr et al. [53] showed that mutating CTCF-binding sites resulted in the recruitment of DNMTs, leading to increased methylation at the imprinting control region of Igf2/H19 locus in mouse. Stadler et al. [27] reported that CTCF binding creates a low methylation region through the presence of TETs. Other studies showed that CTCF facilitates histone modification and open chromatin, although the causality in relation to DNA methylation remains unclear [54–56].

Crosstalk with histone modification

The maintenance of DNA methylation also involves crosstalk with histone modification. For example, studies have established DNA maintenance on Uhrf1, where Dnmt1 and ubiquitination of histone H3 are involved to convert hemimethylated DNA to fully methylated DNA [57]. DNA methylation is also linked to H3K9me3 and H3K27me3, where the H3K9 methyltransferase SETDB1 interacts with DNMT3A and 3B [58,59]. Interestingly, SETDB1 does not bind to DNA but forms a repression complex with TRIM28 and zinc fingers such as ZNF274 to achieve locus-specificity [59,60]. Furthermore, Viré et al. [61] showed that the H3K27 methyltransferase EZH2, a component of the polycomb repressive complex PRC2, can interact with DNMT1, DNMT3A and DNMT3B. A more recent study by Baubec et al. [62] using genome-wide ChIP-seq and methylome measurements confirmed that DNMT3A and DNMT3B are localized to methylated CpG-dense regions in mouse stem cells; notably, they found that the PWWP domain of DNMT3B recognizes the SETD2-mediated H3K36 methylation, leading to DNMT3B preferential binding and methylation of the bodies of actively transcribed genes [62].

DNA methylation can also be co-repressed by TF binding and H3K4 methylation. Cfp1 has been reported to recruit H3K4 methyltransferases to promote H3K4me3, preventing CGI from methylation in mouse embryonic stem cells. However, Cfp1 knockout is insufficient to remove local hypomethylation, suggesting other factors are involved in this process [35,63]. In another study, unmethylated H3K4 tails were shown to interact with the de novo methylation machinery, such as Dnmt3L and Dnmt3a [64]. The association between H3K4 methylation and allele-specific DNA methylation has been shown at imprinted loci as well [65], guided by factors like KDM1B [66].

Cell-type specificity and methylation dynamics

The above-mentioned evidence showed that locus-specific methylation is tied to genetic features. Although the DNA sequences remain unchanged for a given genome, the readout of the motifs is dynamic and dependent upon cellular conditions. For example, the expression of a modifying enzyme (e.g. TETs and DNMTs) or the activity of a DNA-binding regulator and its access to DNA is cell-type/condition-dependent, which leads to the dynamic and cell-type-specific modification of epigenome [67]. Such recognition is similar to the binding of TFs to their motifs: the TF motifs in the promoters and enhancers remain the same, but the transcriptional regulation is tissue-specific and dynamic. However, although we have seen increasing evidence for motif-directed recruitment of effectors that can both promote and inhibit DNA methylation [68,69], further study is required toward a systematic characterization of the relationship between the expression of these effectors and cell-type-specific methylation level of their interacting regions.

The Models: Prediction and Revelation

The molecular mechanisms described above have laid the foundation for many studies that use genetic features to predict local DNA methylation. These studies have shed light on the sequence features of locus-specific methylation and demethylation. Below, we review the development of these studies and discuss the perspectives (Supplementary Table S1).

Earlier methylation studies typically employ enzymatic fractionation assays. For example, McrBC digests methylated sequences while many methylation-sensitive restriction endonucleases remove unmethylated sequences [70]. Due to the limited data coverage and resolution, these studies tend to focus on the methylated CGIs. The CGIs reside in the promoters and their demethylation facilitates the binding of TFs [71]. To distinguish unmethylated CGI (non-CGI) from methylated CGI, a variety of predictive features have been found using machine learning methods. For example, Yamada et al. [19] have determined the methylation status of CGIs (from fully methylated to fully unmethylated) using the HpaII-McrBC PCR method in human peripheral blood leukocytes, and then used Support Vector Machine (SVM) and random forest to identify the enriched nucleotide k-mers. They showed CG, CT and CA are the most predictive dinucleotide features for human CGI states. Similarly, Das et al. [15] have separated methylated and unmethylated CpGIs using methylation-sensitive restriction endonucleases and McrBC in the normal human adult brain, and showed that Alu coverage and certain hexamers are the most predictive (86% accuracy) among ~100 predefined features such as CG content, dinucleotide counts and trinucleotide counts. Performance is further improved when including non-sequence features such as trinucleotide physicochemical properties [16] (i.e. bendability, nucleosome rigidity and nucleosome positioning), histone modification [18,72] and the methylation states of flanking CpGs [22]. Note that while both studies ([72] and [18]) used multiple mammalian tissues and cell lines, the prediction accuracy and selected sequence features generalize well across them, with top predictive features being CpGI properties, DNA sequence composition, DNA structure patterns and histone modification status.

Recent studies take advantage of genome-wide methylation assays, such as 450 K array, RRBS and WGBS. The expanded coverage of methylomes has profoundly changed the locus-specific analysis of DNA methylation in several ways. For example, functional motifs have been found outside of CGIs, extending into non-coding regions [12, 23]. In addition, genomic and epigenomic data from multiple cell lines and tissues have been made available by consortium efforts such as ENCODE [73], ROADMAP [74], TCGA [75] and iHEC [76]. Methylation levels are compared across multiple tissues, cell lines and species to establish variability. For example, Zeng et al. [23] have analyzed 50 RRBS +1 WGBS datasets and established the impact of DNA variants on local methylation. Wang et al. [12] have identified genomic regions and motifs associated with common and variable methylation across 34 WGBS, validated in 32 450 K array data sets. Scala et al. [77] examined variance and aberration of CpGs from various cancer types and blood samples across 450 K data sets in TCGA [75] and the Epic cohort [78], and have identified motifs associated with methylation stability, instability and aberration in cancers. More datasets have also allowed more sophisticated machine learning models, such as neural networks [17, 21, 23], to outperform previously best-performing machine learning models like SVM and random forest [15, 19, 22, 72]. DNA sequence features have shifted from using predefined sequences and short k-mer combinations (usually 2–5 bp) [15, 18, 19, 21, 22, 72] to using longer de novo motifs (>9 bp) [12, 13, 17, 23, 79, 80]. These studies revealed novel perspectives on how certain genetic patterns can play important roles in regulating DNA methylation.

The most fundamental change in methylation motif studies is from making predictions to exploring the functional mechanisms of DNA motifs. A natural first step to illustrate the functions of the found de novo motifs is to match them to known TFs [12, 13, 17, 23, 79, 80]. As a result, while earlier studies have associated hypomethylation with high GC contents [18, 19], recent studies have revealed that the contributing DNA motifs with repeating GC tandems are matched to known TFs associated with TET recruitment, such as CTCF, SP family and WT1 [12, 23]. Furthermore, contrary to the previous belief that methylated regions have aberrant TF binding, some TFs have also been found to preferentially bind to highly methylated regions. For example, Wang et al. [12] have identified 92 motifs associated with high methylation in WGBS with enriched bindings of DNMTs. Xuan-lin et al. [80] cross-referenced ChIP-seq of TFs and WGBS to characterize over 500 known TFBS with cell-type-specific CpG methyl-level in their motifs, and have shown some TFs, such as ZBTB33, have high binding affinity to methylated DNA. Ngo et al. [79] have proposed a high-throughput pipeline that not only revealed de novo methylated motifs but also discovered known TFs like CEBPB, NRF1, CTCF and EGR1 that can bind to highly methylated motif patterns (e.g. [GT]ATT [AG]mCGCAAT for CEBPB) which are sequentially and locationally distinct from their canonical motifs. Whitaker et al. [13] have further provided a computational framework to identify DNA motifs representing cis-acting elements with the site-specific DNA-binding factors that establish and maintain epigenomic modifications, including DNA methylation and six histone modifications, and have shown that motifs like CFP1 are found to prefer the center of DNA methylation valleys, with a specific association to H3K4me3 and H3K27me3 modification. Finally, recent studies have highlighted crosstalk between DNA methylation and histone modification among these motifs, especially between H3K36me3 and methylation motifs, as well as between H3K27ac/H3K4me3 and unmethylation motifs [12, 81]. It is worth noting that many de novo motifs found relevant to DNA methylations do not match any known motif [12] and the mechanisms of these motifs await further investigation.

Along with the mechanistic insights on the shaping of the methylome, recent studies also highlight the functional validation of the identified motifs through DNA variants. For example, Wang et al. [12] have shown motifs with enriched methylation quantitative trait loci (mQTL) and expression quantitative trait loci (eQTL), and somatic mutation on the motifs correlates with altered local CpG methylation. Similarly, Banovich et al. [82] have characterized the mQTL in relation to TF binding and expression, and shown that STAT5 and ZNF274 have positive associations between TF expression and DNA methylation nearby binding sites. Further, Zeng et al. [23] have proposed a deep learning framework, CpGenie, to systematically predict methylation change from sequence variant, given the neighboring methylation and DNA sequences.

Taken together, we have observed explosive growth of computational models that explain DNA methylation based on sequence features, in combination with the traditional usage of physicochemical properties, nearby CpG states, TF occupancies and histone states. The improved model performance and the revealed genetic-epigenetic association have made the clinical application possible.

Clinical Application

DNA methylation is closely linked to development, aging and cancer [83, 84]. A common observation in cancer is that methylation on the promoter of a tumor suppressor gene often results in transcriptional repression and phenotypic alteration [1, 85]. Such DNA methylation patterns can thus be used for diagnosis and prognosis purposes. Notably, the recent development of early cancer diagnosis and treatment guidance has been enabled by liquid biopsy [86, 87], whose successful application depends on differentiating the tumorous circulating tumor DNA (ctDNA) from the ‘normal’ call-free DNA (cfDNA) fragments [86, 87]. However, the major challenge is that ctDNA is a small fraction (0.01%–10%) of the total cfDNA [86–88]. Therefore, to achieve sensitive and selective tumor variant detection, current strategies rely heavily on carefully selecting a collection of features (or biomarkers) combinatorically most predictive of the target phenotypes.

Over the years, the choice of biomarkers has shifted from focusing on genetic mutations on tumor suppressors (such as TP53 [89] and PTEN [90]) to leveraging epigenetics signatures [91]. For example, BRCA1, PTEN, HRK, APC and RASSF1A have been found methylated in cancer, and some related to prognosis and reflect on the efficacy of therapy [92–94]. DNA methylation patterns derived from RRBS have also been used as a predictor for breast cancer dissemination [95]. Other studies have reported success with DNA methylation cfDNA assay outside plasma for specific cancer types, such as urine-based assays for prostate cancer [96, 97] and stool-based assays for colorectal cancers [98]. Guo et al. [99] reported segments of DNA methylation (termed haplotype blocks) from plasma DNA can aid the deconvolution of heterogeneous tissue samples. A more recent study by Grail has successfully mapped and identified tumor origin by cfDNA methylation in 25 human tissues and cells [100]. Notably, Shen et al. [101] have developed an immunoprecipitation-based genome-wide cfDNA methylome screening protocol (cfMeDIP–seq). They showed sensitive tumor detection and classification among several tumor types, using differentially methylated regions and CpGs. Overall, the adoption of cfDNA methylation analysis has greatly improved the diagnosis power in previously low-performing cancer types.

A natural progression is combining both genetic and epigenetic signals to further improve performance and detection limit in early cancer diagnosis and personalized treatment. Indeed, Westesson et al. [102] have recently shown that with combined genomic, methylation and fragmentomic signals in 162 early-stage colorectal cancer patients; they achieved an overall sensitivity of detection at 90.3% (90% Stage I; 88% Stage II; 96% Stage III) and specificity at 96.6%. The rapid adoption of the multi-omics approach evokes an emerging strategy where the knowledge of how cis-acting DNA variants impact disease-associated epigenome leads to improved diagnostics and prognostics. For example, the presence of a single-nucleotide polymorphism (SNP) at the MGMT promoter negates the promoter’s methylation in glioblastoma, correlates with worse temozolomide treatment outcome [103]. An SNP at the CpG site located at the ARPC3 promoter is associated with hypertriglyceridemia in overweight patients [104]. Three CpG-SNP pairs have been reported significant for the prognosis of breast cancer patients [105]. Multiple studies have reported DNA variants are particularly found in the CGI at the promoter of genes related to cancer [106–109]. Zeng et al. [23] have reported a model to accurately quantify how DNA variants can impact local CpG methylation and gene expression. Recently, we have discovered and characterized 313 DNA motifs that regulate DNA methylation and unmethylation and showed that DNA mutation overlapping with these motifs impacts local CpG methylation (Figure 2A). Moreover, we have demonstrated that profiling somatic mutations in cancer patients based on which DNA motifs they overlap, providing a significant performance improvement over using these somatic mutations alone, both for diagnosis and prognosis [12] (Figure 2B). Taken together, these results suggest understanding how non-coding DNA-variants change methylation can improve the re-evaluation of the existing DNA biomarkers and provide new perspectives on biomarker discovery.

Figure 2 .

Clinical application. A. An example of altered cis-acting DNA motif changes local methylation level. B. Evidence that combining the prior knowledge of DNA motifs that regulates DNA with somatic mutation significantly improves performance for both cancer diagnosis (left) and prognosis (right). (reprinted from Wang et al. 2019).

Outlook

Current research for liquid biopsy benefits from two contributing factors: the quickly increasing sequencing power [110], and clinical studies linking molecular profile to early pan-cancer diagnosis, as well as treatment outcome of late-stage cancer patients [111–113]. Therefore, while currently available assays have relatively small numbers of features (i.e. 10–100 biomarkers) due to limited variant data [114], future studies can use many more features. Furthermore, the large data set will allow the development of more powerful deep learning models to improve the prediction power. Both trends require a deeper understanding of the interplay between the existing features (i.e. DNA methylation and DNA variant).

Ideally, cancer diagnosis and prognosis could benefit from combining a diverse set of relevant molecular signatures, including DNA variants, methylome, transcriptome, proteome, HLA signature and chromatin structure. However, given the limited resources, the major challenge to distinguish between cancer and normal cfDNA is the limited number of biomarkers, and how to detect them frugally. As a result, we would argue that the most cost-effective strategy is to adopt the prior knowledge of how DNA sequence and methylation interact with each other to further improve accuracy and sensitivity. Recent technological advances have made it possible to simultaneously detect DNA variants and methylation variants on cfDNA [101]. The synergistic interplay between DNA variants and DNA methylation makes using DNA motifs advantageous and versatile in many clinical settings.

In addition to DNA methylation, we have recently discovered DNA motifs that regulate histone modifications [13, 81] and showed that the altered DNA motif leads to abolished histone modification, which is also important in cancer [115]. These cis-acting motifs can be leveraged to reveal information on the state of histones, which is not readily available in cfDNA [87]. Furthermore, DNA patterns are also important in establishing local DNA secondary structures, which have been reported as an epigenetic determinant of cancer genome [116]. Clark et al. [44] have reported a sequence pattern in the DNA secondary structures as a hotspot for DNA methylation in human breast cancer patients.

Taken together, we believe the ever-growing research revealing genetic-epigenetic interplay has opened doors to previously underexplored strategies in biomarker selection and points to new perspectives in characterizing DNA variants in combination with epigenetic signatures.

Key Points

• Increasing evidence has shown locus-specific DNA methylation is regulated by cis-acting DNA elements.

• Recently, computational models are used to predict genetic features of DNA methylation patterns.

• Biological insights have been revealed from these models.

• Future application of methylation biomarkers considering liquid biopsy for early cancer diagnosis and treatment are discussed.

Mengchi Wang is a PhD student at the Bioinformatics and Systems Biology at University of California at San Diego, with current research focus on the genetic basis for DNA methylation.

Vu Ngo is a PhD student at the Bioinformatics and Systems Biology at University of California at San Diego, with current research focus on the genetic basis for histone modification.

Wei Wang is the principle investigator and full professor at the Bioinformatics and Systems Biology, Department of Chemistry and Biochemistry, and Department of Cellular and Molecular Medicine at University of California at San Diego.

Data Availability Statement

Data sharing is not applicable to this article as no new data were created or analyzed in this study.