methylGrapher: genome-graph-based processing of DNA methylation data from whole genome bisulfite sequencing

Abstract

Genome graphs, including the recently released draft human pangenome graph, can represent the breadth of genetic diversity and thus transcend the limits of traditional linear reference genomes. However, there are no genome-graph-compatible tools for analyzing whole genome bisulfite sequencing (WGBS) data. To close this gap, we introduce methylGrapher, a tool tailored for accurate DNA methylation analysis by mapping WGBS data to a genome graph. Notably, methylGrapher can reconstruct methylation patterns along haplotype paths precisely and efficiently. To demonstrate the utility of methylGrapher, we analyzed the WGBS data derived from five individuals whose genomes were included in the first Human Pangenome draft as well as WGBS data from ENCODE (EN-TEx). Along with standard performance benchmarking, we show that methylGrapher fully recapitulates DNA methylation patterns defined by classic linear genome analysis approaches. Importantly, methylGrapher captures a substantial number of CpG sites that are missed by linear methods, and improves overall genome coverage while reducing alignment reference bias. Thus, methylGrapher is a first step toward unlocking the full potential of Human Pangenome graphs in genomic DNA methylation analysis.

Graphical Abstract

Graphical Abstract.

Introduction

Progressing from the initial draft to a complete Telomere-to-Telomere assembly of a human genome reference took more than two decades [1]. In this time, tools developed based on linear genome references have played a pivotal role in advancing our understanding of human biology, health, and diseases. The most widely used human linear reference genome (hg38, aka GRCh38), is a mosaic assembly derived from multiple individuals with hundreds of unsolved gaps that account for more than 159Mb [2]. This lack of completeness was improved upon the completion of the first Telomere-to-Telomere assembly CHM13 [3]. Despite this advance, no single linear representation can capture the full genetic diversity within a population. Particularly in sequencing-based analyses, where read alignment to the reference genome is a common practice, reference bias can significantly impact the results. To overcome these limitations, the genome graph was introduced to represent diverse sequences. The Human Pangenome Reference Consortium (HPRC) released its initial draft [4] of the human pangenome reference, marking a significant step toward transcending the constraints of linear references.

Along with this paradigm shift toward diversity forward genome-representation, there must also be a shift in tooling for functional genomics to unlock the potential of this new paradigm. Several tools have been developed to enable functional genomics analysis on genome graphs, including short-read and long-read mapping (GraphAlign [5]), transcript analysis (RPVG [6]), genotyping (PanGenie [7]), peak callings (Graph Peak Caller [8]), motif analysis (GRAFIMO [9]), and graph visualization (Sequence tube maps [10]). However, a significant gap exists in analyzing the most common DNA methylation data, whole genome bisulfite sequencing (WGBS), on a genome graph. Epigenetic modifications, particularly DNA methylation, are crucial to understanding gene regulation. DNA methylation regulates gene expression by suppressing or activating transcription ([11,12]) and influencing alternative splicing [13]. WGBS ([14,15]) is the most widely used protocol for detecting cytosine methylation patterns at genome-wide scale. In this approach, genomic DNA (gDNA) undergoes sodium bisulfite treatment, followed by sequencing to achieve precise identification of methylated cytosines within the genome at single-base resolution. Through bisulfite treatment, unmethylated cytosines are transformed into uracils, which are subsequently read as thymidines during sequencing. In contrast, methylated cytosines remain unaltered and are directly identified as cytosines during the sequencing process. The locations of methylated cytosines are determined by comparing these treated DNA sequences with the reference and inferring cytosines chemically protected by methylation. Advanced methods, such as Bis-SNP [16] and gemBS [17], also infer genotype information from the data and subsequently condition their methylation calls on the results of the genotype determination. Recently, a similar technology called EM-seq is gaining some traction [18]. Instead of using bisulfite, EM-seq uses enzymes to convert cytosines, and it generates essentially the same type of data as WGBS does.

The conversion of unmethylated cytosines into thymidines reduces sequence complexity. Modern sequencing read aligners depend on the creation of short k-mer seeds, and the alignment process involves searching for these seeds in gDNA and then extending matches to alignments. This approach makes alignments susceptible to errors caused by mismatches and by the reduction in sequence complexity. Hence, direct mapping of WGBS reads to the reference results in considerably reduced mapping rates and elevated rates of false alignments, emphasizing the need for a tool capable of performing this task on the genome graph. Here, we present methylGrapher (mG), a WGBS analysis tool designed for accurate DNA methylation calling on genome graphs, for both CpG and non-CpG contexts. MethylGrapher provides a simple and customizable command-line interface that seamlessly integrates methylation calling and conversion rate estimation in a single step. Here, we benchmark methylGrapher's performance on pangenome graph reference against mapping using linear reference, and we use several common and state-of-the-art linear bisulfite mapping tools including Bismark-bowtie2 (Bis) [19] (referred as Bismark herein), BISCUIT [20], bwa-meth [21], and gemBS [17]. Our results indicate that methylGrapher accurately reconstructs DNA methylation patterns and demonstrate the advantages offered by the pangenome graph.

Materials and methods

Materials and methods used to generate data are described below. More details about the methods can be found in the Supplementary material. For the specific command used, see https://zenodo.org/records/14108397 or https://github.com/Hendricks27/methylGrapher_paper.

Whole genome bisulfite library preparation and sequencing

Library preparation and sequencing was performed by the McDonnell Genome Institute's (MGI) Genome Technology Access Center as follows. The gDNA was quantified using the Qubit 1x HS dsDNA assay kit by Invitrogen (Cat# Q33231). For the preparation, 200 ng of gDNA, including 0.2% Unmethylated Lambda DNA to monitor bisulfite conversion efficiency, was fragmented to approximately 350-bp fragments using the Covaris LE220. After fragmentation, a 1.5x AMPure clean-up was performed on the EpMotion 5075 (Eppendorf), resulting in a final volume of 20 uL. The fragmented gDNA then underwent bisulfite conversion using the EZ-96 DNA Methylation-Gold Mag Prep Kit (Zymo Research, Cat# D5043), following the manufacturer's recommendations on the EpMotion 5075. The bisulfite-converted DNA (bsDNA) was quantified using the Qubit ssDNA Assay Kit (Thermo Fisher Scientific, Cat# Q10212). Automated WGBS libraries were constructed with approximately 100ng of the quantified bsDNA using the Accel-NGS Methyl-Seq DNA Library Kit (Swift BioSciences, Cat# 30 096) and unique dual indexes (UDI) (Swift, X9096) on the EpMotion 5075. Eight polymerase chain reaction (PCR) cycles were performed during the indexing PCR step, followed by a final 0.85x AMPure cleanup. The final libraries were assessed on a GX instrument to determine the final average library size and concentration. They were then quantitated using qPCR with KAPA Library Quantification Kits (Roche) to achieve optimized cluster counts. Finally, 2 × 150 paired-end reads were generated using the S4 300 Cycle kit with XP workflow on Illumina NovaSeq 6000 instruments. This sequencing process was designed to provide comprehensive coverage and deep insights into the methylation patterns across the genome, utilizing the high throughput and accuracy of the NovaSeq 6000 system.

Surjection with Ixhchel

Ixchel (pronounced Eesh-Chel) is a graph surjection tool for converting cytosine methylation calls in genome-graph positions to linear coordinates. This conversion is necessary for comparing and visualizing graph-based results against linear-based results. Minigraph-cactus-derived ([22,23]) genome-graphs were built on top of a reference assembly. In this case, the graph was built on top of hg38. These genome graphs followed the Graphical Fragment Assembly (GFA) format described at https://github.com/lh3/gfatools/blob/master/doc/rGFA.md. In this format, segments in the graph derived from hg38 are reference segments. Each reference segment retains its reference-coordinate position. As such, reference segments are used as landmarks to facilitate conversion of graph-coordinates to linear coordinates. Ixchel operates through a two-stage process to convert cytosine coordinates: pre-computation and surjection. In the pre-computation step, it constructs a lookup object containing all cytosines in the graph (GAF format). Ixchel first identifies each cytosine-containing segment. For those segments, it determines how confidently it can convert cytosine positions within said segments by three criteria. (1) Is the segment a reference segment? (2) Is the segment within one degree of separation from a reference segment? (3) Is the segment within one degree of separation of more than one downstream reference segment? Degree of separation here refers to the number of edges (connections) in the graph required to reach a reference segment from a given segment; one degree of separation means the segment is directly connected to a reference segment. Segments closely associated with a reference segment—either by being within one degree of separation downstream from a single reference segment or by being reference segments themselves—are deemed to have ‘highest confidence’ in their positional conversion accuracy. Reference segments contain in their annotations the linear-coordinate position of the first base of the segment. As such, positions in reference segments are calculated by adding the distance of the relative position within the segment (0-based) to the linear-coordinate position of the first base. Alternatively, if the position falls within segments that are not reference segments and fulfill the second and third criteria, they are converted by calculating the distance of the position from the start of the nearest reference segment and adding this distance to the position of the first base of the proximal reference segment in linear coordinates. Positions within segments that fail either the second or third criteria cannot be converted with confidence and are flagged. The convertibility is reported as a bitwise flag to allow a user to filter based on degree of conversion confidence. This process is performed for all cytosines in the graph and saved as a SQL database to enable rapid lookup of positions to be converted. The second stage uses this database to convert methylation estimate entries in graph.methyl format to methylC format. This method was applied to a draft pangenome constructed using the hg38 as the foundational reference framework.

Comparison between linear methods and methylGrapher

Adapter trimming was performed using trim_galore v 0.6.10 with the specified parameters (-q 20 –phred33 –illumina –stringency 1 -e 0.1 –length 20 –clip_R1 10 –clip_R2 15 –paired). Subsequently, deduplication was conducted using fastuniq [24] v 1.1. Bismark alignment was carried out using v 0.23.1 with default settings. Methylation extraction was performed using Bismark with the options –paired-end –no_overlap –comprehensive –merge_non_CpG, followed by bismark2bedGraph with parameters –cutoff 1 –buffer_size = 60G. To ensure comparability, we used parameters similar to those used by Bismark. For BISCUIT v0.3.14, alignments were performed in directional mode with the -b 1 option. The pileup step was conducted using the options -b 1 -m 10 -a 10 -l 10 -5 0 -3 0, while the vcf2bed step was executed with the -k 1 option. For bwa-meth, we applied the default alignment parameters, and extraction was carried out using MethylDackel with the -q 10 -p 1 options. For gemBS v4.0.4, we started with the IHEC [25] recommended configuration file and modified the mapq_threshold to 10 and qual_threshold to 0 in the calling step, as well as the phred_threshold to 0 in the extraction step. For methylation calling with methylGrapher V0.1.2, the -minimum_mapq 10 option was used to filter out multimapped reads.

For all analysis using the four linear methods, we utilized hg38 excluding the ALT contigs, while for methylGrapher analysis, we employed the initial version of the human pangenome derived from hg38.

For comparative analysis, low-confidence CpGs with a coverage of 5 read pairs or less were excluded. Common CpGs were utilized for linear regression and subsequent plotting. A 2D density plot was generated by binning the methylation levels (ranging from 0 to 100%) into 61 bins, followed by the application of Gaussian blur to reduce noise. Methylation differences were visualized using a bar graph, with CpGs categorized into nine groups based on difference ranges (−1 to −0.35, −0.35 to −0.25, −0.25 to −0.15, −0.15 to −0.05, −0.05 to 0.05, and so on up to 0.35–1). The coverage was compared using commonly identified CpG sites (where coverage > 0 in both cases), plotted in a 2D histogram. For sites with coverage greater than 30, a value of 30 is used for plotting purposes. For PCA plotting, common high-confidence CpGs across all samples, replicates, and methods were selected. CpGs with identical methylation values (zero variance) were removed prior to analysis.

Creating benchmarking samples pangenome graph using minigraph-cactus for haplotype-specific surjection

The benchmarking samples pangenome graph (BSPG) was made from the phased chromosomal parental assemblies derived from the samples: HG002, HG00621, HG00721, HG01952, HG01978, and HG03516. The reference used was GRCh38. These samples were selected by the CIG@MGI group as benchmarking samples to use in analyses of epigenetic sequencing techniques involving the pangenome graph. This smaller graph, constructed with fully phased chromosomal assemblies, enables the use of walk information to map graph coordinates back to the individual assemblies. These assemblies were used as inputs to the cactus-minigraph pipeline running cactus v2.5.0. The pipeline was written in WDL and can be found at the CIG@MGI Group's pipeline GitHub (https://github.com/twlab/cig-pipelines/blob/f7efa68bae738dbe13c8153d04ad5db4bef7094f/wdl/pipelines/pangenome/mcgb.doc.md). The pipeline follows the instructions given at the time for building pangenome graphs. The basic workflow is:

cactus-minigraph –> cactus-graphmap –> cactus-graphmap-split cactus-align –> cactus-graphmap-join.

To facilitate the conversion of graph coordinates into phased genome space, we developed haplotype-specific surjection (HSS). This method utilizes the walk or path encoded in the genome graph in GFA format to compute coordinates based on the offset. Unlike surjection to hg38, HSS does not attempt to convert CpGs that are not situated along the intended path. Subsequently, the graph result is HSS to the maternal and paternal genomes, and linear tools were executed with the phased genome for comparison.

Processing of EN-TEx WGBS samples and shared SV identification

The EN-TEx ENC001 genomes were obtained from the ENCODE database under accession ENCFF477YTR. The VCF file containing structural variations was also sourced from ENCODE, using accession ENCSR767MYF. Accession numbers for the six WGBS libraries were ENCSR757JTX, ENCSR128RMY, ENCSR806NNG, ENCSR403DTW, ENCSR996BJT, and ENCSR224CTR. All libraries were deeply sequenced but lacked both biological and technical replicates. Each library was randomly downsampled to 200 million read-pairs, deduplicated using FastUniq, and then mapped to hg38, both hap1 and hap2 using Bismark with default parameters, as well as to the pangenome using methylGrapher with default settings. Subsequently, methylGrapher result was then surjected to hg38.

Initially, to identify insertions already catalogued by the pangenome from individual ENC001, we utilized the insertions (>50 bp) identified by long-read sequencing and aligned them back to the pangenome using GraphAligner. To minimize false positives where similar sequences could be inserted at different loci, we incorporated 1 kb of both flanking regions from the hg38 reference of the insertion to guide the alignment [see Supplementary Fig. S15], ensuring that the insertions occurred at the correct locations. A sequence identity of greater than 90% was considered a common structural variant (SV) between the individual and the pangenome [See Supplementary Fig. S16]. These insertions were used as proxies to lift-over CpGs identified by methylGrapher to haplotype-specific coordinates for further comparison.

Results

Overview of the methylGrapher workflow

MethylGrapher is a Python-based software library tailored for rapid WGBS analysis on genome graphs (Fig. 1). Drawing inspiration from the well-established WGBS analysis pipeline, Bismark, methylGrapher applies a similar strategy. It addresses the issue of modern aligners’ sensitivity to mismatches by initially transforming the genome graph into two fully converted versions: one featuring cytosine-to-thymine conversion and the other with guanine-to-adenine conversion. Sequencing reads are subjected to the same conversion process, and methylGrapher aligns these converted reads to the converted graphs by VG Giraffe [26]. As a result, it maximizes the ability to map reads. Many WGBS aligners, including gemBS [17], BRAT [27], BS-Seeker2 [28], and BWA-meth [21], have adopted Bismark's methodology of converting the reference genome. A few other paradigms also exist. For example, BISCUIT [20] uses three-letter indexes during initial seeding and then employs four-letter alignment for seed extension. Wild-card aligners like segemehl [29], BSMAP [30], and Pash [31] substitute Cs in the reference with a wildcard letter that matches both Cs and Ts. Additionally, abysmal [32] uses two-letter alignment, treating As and Gs as purines and Cs and Ts as pyrimidines during alignment. Implementing and testing these alternative paradigms could be an exciting future direction for graph-based WGBS aligners.

Figure 1.

Overview of methylGrapher's workflow. (A) The flowchart of the workflows. The middle panel illustrates the library preparation workflow. This process takes either single-end or paired-end FastQ file(s) as input. Initially, the FastQ files are trimmed using Trim Glore, deduplicated, and then fully converted prior to alignment. It is advisable to use FastQC after the trimming and deduplication steps. The left panel presents the genome preparation workflow. This workflow begins with the construction of a genome graph using the minigraph-cactus pipeline in GFA format. The lambda phage genome is incorporated as a distinct segment. Subsequently, two separate fully C-T and G-A converted graphs are generated. These can optionally have duplicated SNVs trimmed. VG Giraffe alignment indexes are then built using these two graphs. The right panel presents the alignment and methylation calling workflow. Alignments of all converted reads against all converted references were performed for non-directional libraries, and methylGrapher selects the optimal alignment for each read. If the WGBS experiment is directional, only two alignments are carried out. Using these alignments, the methylation percentage for every cytosine is calculated. After methylation extraction, read statistics are reported, and, optionally, the estimated conversion rate can be output using built-in commands. (B) A simple illustration of DNA methylation level calculation, which involves aligning reads to specific cytosines. The top half of the figure depicts a genome graph where genetic variations are represented as alternative paths in the graph. The bottom half of the figure depicts three aligned, bisulfite converted reads. After alignment, the reads are piled-up to determine how many cytosines have been converted to thymine and how many remain unchanged. (C) An example of a result table. The columns are segment ID, 0-based position on that segment, strand of the segment, cytosine context, methylated count, unmethylated count, read coverage, and methylation percentage. Here, the specific CpG with a segment ID of 10 002 was extracted from the graph in (B), and the three reads aligned to this CpG contain two CGs and one TG, thus the methylation level is estimated to be 66.67%.

In directional WGBS sequencing, the first read in paired-end sequencing originates from either the Crick + or Watson- strands, resulting in C-depletion, while the second read in paired-end sequencing originates from the complementary strand (Crick- or Watson+), leading to G-depletion. On the other hand, non-directional sequencing yields a mix of C-depleted and G-depleted reads from any of the four strands [17]. MethylGrapher supports both directional and non-directional WGBS library preparation. In the case of directional library, only a 2-pass alignment is required, which removes the need to align G-to-A converted read 1 and C-to-T converted read 2. Post-alignment, methylGrapher tracks bisulfite-treated reads back to the original genome graph and conducts a straightforward count of converted and unconverted bases to determine the level of support for unmethylated versus methylated cytosines. Compared to the three-letter linear WGBS mapper, three-letter graph alignment introduces a new challenge: cytosine and thymine may represent a single nucleotide variation (SNV) occurring at the same locus within the graph. Genomic variations frequently display genetic linkage [33], and genome graphs capture both these variants and the associated haplotype information. To overcome the challenge of resolving ambiguous alleles, haplotype-aware aligners, such as VG giraffe [26], leverage this haplotype data to distinguish between closely related variants. For multi-mapped reads, methylGrapher provides two options for handling them. By default, methylGrapher generates multiple possible alignments, and if one alignment has a higher score than all others, it is considered uniquely aligned. This strategy is adopted based on the Bowtie2 version of Bismark, and it generally retains more reads. Alternatively, methylGrapher can treat reads as uniquely aligned if their alignment surpasses a specified MapQ [34] threshold. Reads considered multi-mapped are discarded for methylation quantification.

Bisulfite conversion rate is the efficiency of converting unmethylated cytosines into uracils, which are later read as thymines during sequencing. This rate is typically calculated by spiking in a known unmethylated control, such as lambda DNA, and determining the percentage of cytosines in CpG, CHH, or CHG contexts (where H represents any base except for guanine) that have been successfully converted to thymines. MethylGrapher integrates the genome of spike-in DNA into the genome graph (Fig. 1A) and can output both the overall conversion rate and the conversion rate in different sequence contexts.

MethylGrapher offers a user-friendly one-line command that streamlines the entire process, filling a notable gap in the field of pan-genomics. The methylGrapher extraction output (Fig. 1C) provides a detailed list of each cytosine mapped to the genome graph, specified by its coordinates (segment ID, 0-based position on the segment, and strand relative to the segment). For each cytosine, the output includes its context, the number of read-pairs supporting whether the cytosine is methylated or unmethylated, and the read coverage, which is the sum of methylated and unmethylated read-pairs. The methylation percentage is calculated as the ratio of methylated read-pairs to the total coverage.

Benchmark of five different WGBS alignment tools

To showcase methylGrapher's ability to reconstruct methylation profiles, we employed data from five WGBS experiments conducted by the HPRC (see ‘Materials and methods’). We analyzed this data using methylGrapher with version 1.0 of the human pangenome graph, which was constructed using hg38 as the backbone and incorporated 44 diploid genomes (88 haplotypes). Additionally, we performed a parallel analysis using Bismark, BISCUIT, bwa-meth and gemBS with hg38 as the reference genome. In order to ensure a direct and impartial comparison, all read pairs underwent the same trimming conditions, followed by identical pre-alignment deduplication, similar alignment, and similar methylation extraction procedures (see ‘Materials and methods’).

We first reported typical CPU and memory usage on a standard methylGrapher run. Due to the substantial size and complexity of the human pangenome graph, the computational resources required for analysis are unavoidably more demanding than using linear methods. We used sample HG00621 replicate 1 as a representative WGBS experiment which was sequenced at ∼15X. We run each pipeline with recommended parameters by their documentation or default setting three times on an identical server architecture featuring dual Intel Gold 6336Y processors and 512 GB of memory (refer to the Supplementary Materials for details). Overall, gemBS was the fastest in terms of end-to-end processing, primarily due to its rapid alignment step. Notably, methylDackel (recommended extractor of bwa-meth) required only five minutes for the methylation extraction step (Fig. 2A). In terms of peak memory usage (Fig. 2B), Bismark-bowtie2 launched multiple Bowtie2 instances to enable parallel computing, resulting in a significantly larger memory usage of 271.8 GB. In contrast, BISCUIT and bwa-meth were more memory efficient. GemBS used a larger index, which was a trade-off for its much faster alignment speed. Despite the overhead of manipulating the graph structure and tracing positions along the alignment path back to specific segments on the graph, the amount of memory and runtime of methylGrapher was generally on par with other state-of-the-art linear aligners.

Figure 2.

Performance comparison across 5 methods. Here we use sample HG00621 replicate 1 as a representative WGBS experiment which was sequenced at ∼15X. We run both methylGrapher and Bismark using default parameters three times on an identical server architecture featuring dual Intel Gold 6336Y processors and 512 GB of memory. (A) Runtime breakdown comparison across five methods over three separate runs. Here, ‘Total’ represents the overall runtime required for processing raw FastQ reads. ‘Trim’ denotes adapter-trimming, ‘Dedup’ refers to deduplication, ‘Align’ indicates the time taken for alignment including fully converting the FastQ reads, and ‘Extract’ represents the time required to extract cytosine methylation levels from the alignment file. The Y-axis displays the runtime for each step in hours. (B) Peak memory usage comparison across five methods at each step, reported from the job scheduler. The breakdown includes memory usage for ‘Trim’ (adapter-trimming), ‘Dedup’ (deduplication), ‘Align’ (alignment and conversion of FastQ reads), and ‘Extract’ (extraction of cytosine methylation levels from the alignment file). The Y-axis indicates the peak memory usage for each step in GBs. (C) The percentage of reads aligned to the reference genome is shown, with each dot representing a WGBS library. The number displayed indicates the average mapping rate. (D) The percentage of reads aligned with MapQ greater than 10, with each dot representing a WGBS library. The number displayed indicates the average mapping rate.

To compare the alignment results, we first examined the fraction of deduplicated reads that aligned to the reference genome (Fig. 2C). With the exception of Bismark, the other four mappers aligned nearly all of the reads. Bismark, however, aligned only about 65.3% of the total reads. As mentioned earlier, MapQ scores were used to remove multi-mapped reads. We applied a filter with MapQ greater than 10 to define uniquely mapped reads. Interestingly, we found that nearly half of the reads from gemBS did not pass this filter (Supplementary Fig. S1, Fig. 2D). The other mappers showed only a small percentage of reads as multi-mapped. Overall, methylGrapher had satisfactory computational efficiency and alignment rates.

Linear methods and methylGrapher yield comparable genome-wide functional outcomes

To compare the results obtained from the five methods, we first analyzed the genome-wide distribution of CpG methylation levels (Fig. 3A and Supplementary Fig. S2). All methods produced the anticipated bimodal pattern and largely superimposed each other. Additionally, all methods demonstrated a similar coverage distribution pattern (Fig. 3B and Supplementary Fig. S3). However, methylGrapher, BISCUIT, bwa-meth consistently exhibited much higher coverage compared with Bismark and gemBS, primarily due to the substantial increase in uniquely mapped reads.

Figure 3.

Comparison of genome-wide analysis characteristics across the five methods. Here we use HG00621 replicate 1 as a representative in (A) and (B). All five methods generate a similar distribution of methylation. However, methylGrapher, BICUIT, and bwa-meth identify a greater number of CpG sites with higher coverage compared with Bismark and gemBS. (A) The distribution of CpG methylation levels across the five methods. Result of all methods exhibit a bimodal distribution and a similar trend in CpG genome-wide methylation distribution. The X-axis represents methylation levels from 0% to 100%, while the Y-axis indicates density in a kernel density plot. (B) The distribution of CpG coverage across all five methods. The X-axis represents CpG coverage, while the Y-axis indicates the fraction of CpG sites at specific coverage levels. (C) The number of CpG sites identified by all methods. There are 29.2 million CpGs in hg38, while the first draft of the human pangenome graph, constructed using Minigraph-Cactus for the HPRC, contains 35.8 million CpGs, as indicated by the black thin lines on top. Each bar displays the counts of CpGs identified by each of the five methods (Coverage >0), grouped by sample and replicates. The different colors illustrate CpGs called by different methods. Solid coloring indicates the count of CpGs with coverage exceeding 5X. The bar graph demonstrates methylGrapher identifies more CpG sites than linear methods.

We then compared the total number of CpG sites detected by each method. In Fig. 3C, we displayed the total number of CpGs detected across all five samples, two replicates, and by five methods, and highlighted those that had greater than 5X coverage (i.e. CpGs with higher confidence methylation calls). Overall, methylGrapher consistently identified more CpG sites than linear methods did.

The hg38 reference genome encompasses approximately 29 million CpG sites [15], yet our WGBS experiments can only detect 27.0–27.9 million with linear methods. Upon examination of individual genomes, we found that each haplotype genome harbors approximately 30.5–32 million CpG sites, with CHM13 T2T 2.0 featuring around 33.9 million CpG sites. The first draft of human pangenome graph made by Minigraph-cactus contains 35.8 million CpG sites (see Supplementary Table S1). This observation underscores the substantial number of CpG sites that remain undetectable in hg38 but are now accessible by the Pangenome Graph. Next, we compared individual CpG methylation levels between technical replicates, using Bismark (Supplementary Fig. S4), BISCUIT (Supplementary Fig. S5), bwa-meth (Supplementary Fig. S6), gemBS (Supplementary Fig. S7), and methylGrapher (Fig. 4A and B and Supplementary Fig. S8). The genome-wide CpG methylation level correlations between replicates were remarkably similar regardless of the aligner (Pearson correlation coefficient (r) ranging from 0.85 to 0.91 across five samples, Fig. 4E). Figure 4B plotted the distribution of differences of methylation calls between two replicates, and 74% CpGs had methylation calls within 15% difference between replicates by methylGrapher (defined as 15% concordance [35]). Similar 15% concordance from 63% to 74% was observed in other samples using all five aligners (Supplementary Fig. S13).

Figure 4.

Analysis of five WGBS libraries from HPRC benchmark samples. Differences in DNA methylation calls between any methods are much smaller than differences between biological replicates. (A) The 2D histogram illustrates CpG methylation comparison between HG00621 replicate 1 and replicate 2, based on the methylGrapher results. Linear regressions are displayed in the figure. (B) The distribution of DNA methylation differences called on the same CpG between replicates. The differences are defined as ‘delta methylation’. For each CpG site between replicate 1 and replicate 2 of HG00621, delta methylation is the value of subtracting replicate 2 from replicate 1. The delta methylation is visualized using a bar graph, with CpGs categorized into 9 groups based on difference ranges (−1 to −0.35, −0.35 to −0.25, −0.25 to −0.15, −0.15 to −0.05, −0.05 to 0.05, and so on up to 0.35–1). Approximately 73.6% of CpGs exhibit a difference of less than 0.15. (C, D) Similar to Figs. 4A and B, but the comparison examines methylation differences of common CpGs’ methylation levels between the Bismark and methylGrapher, using HG00621 replicate 1 as a representative. Approximately 96.3% of CpGs exhibit a difference of <0.15. The linear regression indicates better agreement between the methods than between replicates. (E) Pearson correlation of methylation levels between replicates on common CpG sites, with sample names on the X-axis and method names on the Y-axis. (F) Pearson correlation of methylation levels between different methods on common CpG sites, with sample names and replicate numbers on the X-axis and the method names on the Y-axis. (G) PCA on common CpGs. PCA is based on the shared CpGs identified across all samples, replicates, and methods (only including Bismark and methylGrapher here) with greater than 5X coverage. The individual methylomes were separated by sample identity, then to a much lesser degree by replicates, and had very little differences between the two methods (Bismark and methylGrapher).

To directly compare methylation calls at individual CpG level across methods, we used a procedure called surjection to transform graph-coordinates into a shared, linear coordinate system (See ‘Materials and methods’). This graph-to-linear surjection process is conceptually analogous to lift-over, which calculates the corresponding nucleotide position in one linear haplotype (path indicated by a walk or path) to a second linear haplotype. Surjection is different in that it calculates this conversion from all linear haplotypes down to a single reference haplotype.

A substantial portion of methylGrapher's output, ∼98.8%, can be surjected with the highest level of confidence to the hg38 coordinate system. This allowed us to directly compare individual CpG’s methylation calls between methylGrapher and linear methods. Focusing on CpGs with coverage greater than 5X, we found that the per-CpG methylation levels were strongly correlated between methylGrapher and linear methods, with Pearson correlation coefficient (r) ranging from 0.97 to 0.99 (Fig. 4C and F). Figure 4D plotted the distribution of differences of methylation calls between Bismark and methylGrapher, and 96% CpGs had methylation calls within 15% concordance between the methods (Fig. 4D). Overall, a 15% concordance of 90–99% was observed between the linear method compared with methylGrapher (Supplementary Fig. S13). A similar pattern was observed when examining coverage for the same CpGs. For the coverage of CpGs identified in both replicates (with at least one supporting read), we observed Pearson correlation coefficients ranging from 0.32 to 0.55 across five samples using five different methods (Supplementary Figs S4–S8). When comparing coverage between each linear method and methylGrapher, the Pearson correlation coefficients increased to a range of 0.7–0.89 (Supplementary Figs S9–S12). Notably, BISCUIT and bwa-meth showed stronger agreement with methylGrapher (Supplementary Figs S10–S11) than Bismark (Supplementary Fig. S9) and gemBS (Supplementary Fig. S12) did in both methylation levels and coverage comparisons. This is likely because the core alignment algorithms of BISCUIT and bwa-meth were based on bwa-mem, and methylGrapher's core aligner, VG Giraffe, employed a similar alignment strategy [36]. Thus, the variation between methods was much smaller than that observed between technical replicates.

To compare the methylomes derived from the analysis of five benchmarking samples, we identified shared CpG sites across all samples, each with two replicates. We focused on CpG sites with coverage greater than 5X in all samples, replicates, or methods to conduct a principal component analysis (PCA) (Fig. 4G). Given that BISCUIT and bwa-meth showed stronger agreement with methylGrapher in both methylation levels and coverage, and considering that gemBS yielded a significantly lower percentage of CpG sites with coverage greater than 5X in some library, we focused our comparative analysis primarily on Bismark and methylGrapher to further evaluate their consistency. The individual methylomes separated by sample identity, then to a much lesser degree by replicates, and had very few differences between the two methods. These results demonstrated that methylGrapher produces very consistent results with Bismark on shared CpG sites.

MethylGrapher enhances methylation calling in regions with genomic variation

We next examined CpG sites that were uniquely assessed by Bismark or methylGrapher. Across all of our samples (Fig. 5A illustrates HG00621 Rep 1, and Supplementary Fig. S14 illustrates the remaining datasets), we found that Bismark and methylGrapher commonly mapped 26.2–27.1 million CpG sites. Bismark uniquely mapped 0.8 million CpG sites (with at least 1 read coverage), whereas methylGrapher uniquely mapped 2.2–2.9 million CpG sites (with at least 1 read coverage). Of these methylGrapher-unique CpG sites, 0.4–1.1 million were in hg38 but not identified by Bismark; 1.2–1.4 million represented SNV sites compared to hg38; and 0.4–0.6 million CpG sites were within INDELs or SVs. To investigate the approximately 0.8 million CpG sites uniquely identified by Bismark, we aligned the hg38 autosomes and sex chromosomes to the individual's maternal and paternal genomes using minimap2 [37]. For each CpG site within hg38 on the autosomes and the X and Y chromosomes, we determined whether it exists in either the maternal or paternal genome. We defined CpG sites absent from both the maternal and paternal genomes as ‘lost CpG sites’ (Supplementary Table S2 and S3). Approximately 0.9 to 1.0 million CpG sites were ‘lost’ in the individuals, which could explain why Bismark detected about 0.8 million unique CpG sites using hg38.

Figure 5.

Improvement of methylGrapher over linear methods. Here we demonstrate that methylGrapher correctly identifies the methylome in regions with genomic variation compared with Bismark. (A) Shared and unique CpGs between the methods. The breakdown of CpG counts is as follows, from top to bottom: count of Bismark unique CpG sites, count of common CpG sites between the two methods, count of methylGrapher unique CpG sites overlap with hg38, count of methylGrapher unique CpG sites with SNV, count of methylGrapher unique CpG sites within either INDEL or SV. (B) MethylGrapher fills the gap in the inserted region. An example global view from the WashU Epigenome Browser (chr1:23 268 371–23 373 036, 7.4-kb insertion, HG00621) demonstrates that methylGrapher successfully identifies the methylome within the inserted region. The red arrow on the ruler indicates the insertion breakpoint on the hg38 coordinates, while the yellow box highlights the region that was inserted. From top to bottom: refGene track, methylGrapher result surjects to hg38 (replicate 1), Bismark with hg38(replicate 1), methylGrapher result surjects to hg38 (replicate 2), Bismark with hg38 (replicate 2), HG00621 maternal genome in comparison to hg38, methylGrapher result HSS to its maternal genome (replicate 1), Bismark with maternal genome (replicate 1), methylGrapher result HSS to its maternal genome (replicate 2), Bismark with maternal genome (replicate 2), HG00621 paternal genome in comparison to hg38, methylGrapher result HSS to its paternal genome (replicate 1), Bismark with paternal genome(replicate 1), methylGrapher result HSS to its paternal genome (replicate 2), and Bismark with paternal genome (replicate 2). Data are displayed in methylC track [42]. The height of the blue bar represents the methylation percentage, while the black line indicates the coverage. (C) MethylGrapher accurately identifies CpG sites within the pan-genome. Two close-up views from the WashU Epigenome Browser (chr1:24 009 936–24 009 998) demonstrate that methylGrapher accurately identifies CpG sites without an individualized genome, while Bismark incorrectly identifies CpG sites using hg38. On the left, a homozygous CC site in the individualized genome was misidentified as a CpG site by Bismark. On the right, a homozygous CpG site was ignored by Bismark with hg38; however, Bismark identified them in the individualized genome. From top to bottom: methylGrapher result surjects to hg38 (replicate 1), Bismark with hg38 (replicate 1), methylGrapher result surjects to hg38 (replicate 2), Bismark with hg38 (replicate 2), HG00621 maternal genome in comparison to hg38, Bismark with maternal genome (replicate 1), Bismark with maternal genome (replicate 2), HG00621 paternal genome in comparison to hg38, Bismark with paternal genome (replicate 1), and Bismark with paternal genome(replicate 2). The height of the blue bar represents the methylation percentage, while the black line indicates the coverage.

We utilized the WashU Epigenome Browser [38] to visualize the consistency and differences between Bismark and methylGrapher. When examined on a global scale (Fig. 5B, top 4 tracks), both replicates of HG00621, analyzed by methylGrapher and Bismark, demonstrated a consistent pattern. This pattern revealed hypomethylation in the vicinity of gene promoters and hypermethylation across gene bodies, thereby corroborating earlier findings. However, HG00621 contained a structural variation (a homozygous insertion) of approximately 7.4 kb comparing to hg38. Such SVs are not only not accessible when mapping using hg38, but data derived from these SVs are often falsely mapped to other parts of the genome and contribute to false positives. Using methylGrapher, we were able to successfully map reads to this SV, call DNA methylation for CpGs that were unique to this SV, and HSS (see ‘Materials and methods’), the results to HG00621’s own phased genome assemblies. HSS utilizes the walk encoded in the genome graph to compute haplotype linear coordinates from graph coordinates. Unlike surjection to hg38, HSS does not attempt to resolve CpGs that are not situated along the intended path. While the initial pangenome draft incorporated phased genomes of the five individuals in this study, it was constructed from long contigs but not from scaffolded assemblies. This makes visualization impractical. To address this, we generated scaffolded assemblies from the contigs and used them along with hg38 to construct a genome graph we called BSPG graph using minigraph-cactus (see ‘Materials and methods’). We run methylGrapher against the BSPG graph and HSS to the individual genome coordinates for visualization (Fig. 5B, tracks mark with HSS). These methylation calls were in complete agreement with calls made by running Bismark on HG00621’s own phased genome assemblies, and the results can be visually appreciated by using the comparative genome browser function of the WashU Epigenome Browser [39].

In addition to mapping data to SVs not contained in hg38, methylGrapher also reduces false positive calls and false negative calls resulted from SNVs. Figure 5C illustrated a region with both cases of false positive CpG annotations and false negative annotations. On the left, hg38 chr1:24 009 946–24 009 947 was a CpG site, and Bismark called it a hypomethylated CpG on hg38. However, the individual (HG00621) had homozygous CC dinucleotides at this position as illustrated by the genome align tracks. Applying methylGrapher on the genome graph or applying Bismark on HG00621’s own assembly correctly identified this position as non-CpG. On the right, hg38 chr1: 24 009 980–24 009 981 was TG dinucleotides. Thus, applying Bismark on hg38 predicted nothing. However, applying methylGrapher on genome graph or Bismark on HG00621’s own assembly predicted a hypermethylated CpG, which was confirmed by the genome align tracks. This discrepancy highlights how Bismark can misclassify non-CpG sites as CpG and vice versa when using hg38, underscoring the impact of reference bias. These examples illustrate methylGrapher's capacity to accurately delineate methylomes in the context of genetic variation.

methylGrapher improves methylation detection on samples whose individual genomes are not used for genome graph construction

Our benchmarking thus far relied on WGBS data derived from individuals whose genomes were incorporated in the current draft human pangenome graph. It is important to assess how methylGrapher performs when confronted with genomes not included in the reference pangenome. Toward this end, we utilized EN-TEx [40] data to demonstrate performance when the individualized genome was absent from the construction of genome graph. The diploid genomes of the EN-TEx individual ENC001 was constructed utilizing both short-read and Oxford Nanopore Technology long-read whole genome sequencing, complemented by 10x Genomics linked reads and Hi-C. Initially, the reads were aligned to hg38, with short-reads employed to identify SNVs and INDELs, and long-reads used for structural variation detection using Sniffles [41]. Subsequently, CrossStitch integrated the phased variants into pseudo-haplotype 1 (hap1) and pseudo-haplotype 2 (hap2). Comparing hap1 and hap2 with hg38, we identified 10 454 insertions, totaling 5.29Mb of sequences and encompassing 134 456 CpG sites that would be missed by any hg38 based analysis. Of these, 9 222 insertions (87.5%) representing 4.46Mb (84.3%) and 94 507 CpG sites (70.3%) were cataloged by the current draft pangenome graph, including 6 824 insertions (3.24Mb) from hap1, and 6 401 insertions (2.97Mb) from hap2 (see ‘Materials and methods’).

To evaluate the performance of methylGrapher, we mapped the WGBS libraries from six different tissues of this donor to the pangenome using methylGrapher, and mapped the same data to hg38 as well as both pseudo-haplotypes using Bismark. Consistent with results reported in Fig. 3C, methylGrapher captured 2.2–2.6 million more CpGs than Bismark did across these six datasets, representing 8.2–9.7% improvement (Fig. 6A). Next, we specifically evaluated the accuracy of methylation calls made by methylGrapher on the CpGs sites that were not in hg38 by using the insertions as a proxy to lift-over CpG sites identified on graph coordinates to individual haplotypes (see ‘Materials and methods’), and we compared them with results obtained by running Bismark on hap1 and hap2 assemblies. Overall, Bismark was able of identifying 36–41k of the CpGs within these regions with some read coverage, while methylGrapher identified 40–48k within the same insertions (Supplementary Table S4). Methylation levels of commonly identified CpGs with coverage greater than 5 had Pearson correlations ranging from 0.81 to 0.87 (Fig. 6B and C and Supplementary Fig. S17), and 15% concordance rate of 86–92% (Supplementary Table S4). We presented one example (Fig. 6D) of such an insertion (chr6_hap1:127 829 725–127 833 860) on haplotype 1. Consistent with results reported in Fig. 5C, we illustrated with additional examples (Fig. 6E and F) where methylGrapher was able to reduce false positive calls and false negative calls resulted from SNVs. These results highlight methylGrapher's capability to more accurately categorize CpG sites and construct CpG methylation patterns associated with genetic variations, even when the individual's genomes were not used to construct the genome graph.

Figure 6.

Analysis of EN-TEx data with methylGrapher. (A) The number of CpG sites identified by Bismark with hg38 and methylGrapher with pangenome graph. (B) The common CpG sites identified by both methods within the insertions on haplotype 1 for spleen tissue. X-axis shows methylation levels identified by Bismark; Y-axis show methylation levels identified by methylGrapher. Linear regression is shown on upper left of the figure. (C) Similar to Fig. 6B for haplotype 2 for the same library. (D) The browser screenshot of chr6_hap1:127 829 725–127 833 860 highlights a 4.1-kb insertion on hap1 compared to hg38. The track labels indicate tissue type and the processing pipeline. Bismark (Bis) was run on hap1, while methylGrapher (mG) was run on the genome graph, with the insertion region subsequently lifted over to hap1. The height of the blue bar represents the methylation percentage, while the black line indicates the coverage. (E) The browser screenshot of chr6:92 846 930–92 846 931 (hg38 coordinates) with track labels indicating tissue type and the processing pipeline. Bismark (Bis) was run on hg38, while methylGrapher (mG) was run on the genome graph and subsequently surjected onto hg38. At the bottom, two genome alignment tracks compare hap1 and hap2 to hg38. This region displays a homozygous CG site, whereas hg38 suggests a GG site. MethylGrapher identified a CpG site that was not detected by Bismark. The height of the blue bar represents the methylation percentage, while the black line indicates the coverage. (F) The browser screenshot of chr6:71 758 350–71 758 351 (hg38 coordinates), in a similar style to Fig. 6E, shows a homozygous AG site instead of a CG site. Bismark identified a false positive in this region.

Discussion

In this study, we present methylGrapher which is a much needed and a significant advancement in whole-genome bisulfite sequencing analysis in the evolving landscape of personalized and pan-genomics. Departing from the conventional approach of aligning WGBS reads to a standard linear assembly, methylGrapher adopts the genome graph concept, resulting in a notable improvement in methylation calling accuracy. Moreover, its intuitive and user-friendly command-line interface enhances its accessibility and usability.

As demonstrated by our benchmarks, methylGrapher produces outcomes that closely mirror those obtained through traditional linear methods across the entire genome. Notably, employing a genome graph for methylation confers the benefit of precise methylome identification within genomic regions harboring SNV and structural variations. The initial draft of the human genome graph includes the diploid genomes of 44 individuals, and the HPRC is actively working on expanding this coverage to a more diverse population of over 300 individuals in the near future. This upcoming human pangenome graph will capture a broader spectrum of genomic variation, thereby enhancing methylGrapher's ability to detect CpG sites within regions with polymorphisms. Additionally, it is important to highlight that the use of hg38 may lead to an underestimation of CpG sites in the human genome, as well as false positives and false negatives due to reference bias. Therefore, by leveraging the genome graph approach, methylGrapher achieves a more precise and comprehensive identification of CpG sites compared to linear methods utilizing hg38.

Supplementary data

Supplementary data is available at NAR online.

Conflict of interest

None declared.

Funding

This work was supported by National Institutes of Health (U41HG010972, U24HG012070, UM1DA058219, and U24NS132103) grants.

Data availability

The WGBS data produced in this research are accessible in the NCBI Gene Expression Omnibus via accession code GSE261315.

Code availability

The methylGrapher main code is available at https://github.com/twlab/methylGrapher.

The methylGrapher documentation is available at https://twlab.github.io/methylGrapher/build/html/.

Ixchel code is available at https://github.com/twlab/Ixchel.

For the specific command used in the analysis, see either https://zenodo.org/records/14108397 or https://github.com/Hendricks27/methylGrapher_paper.