Decoding coupled epigenetic-fragmentomic signals in circulating cell-free DNA for interpretable cancer diagnosis

Abstract

Background

Plasma cell-free DNA (cfDNA) whole-genome methylation sequencing enables simultaneous characterization of DNA methylation and fragmentomic features for cancer detection. However, it remains unclear whether methylation library preparation faithfully preserves fragmentomic signals, how methylation and fragmentation are biologically coupled, and how these two modalities can be systematically integrated.

Methods

We systematically evaluated the preservation of cfDNA fragmentomic features in whole-genome bisulfite sequencing (WGBS) and enzymatic methyl sequencing (EM-seq) relative to conventional whole-genome sequencing (WGS). We next investigated the biological relationship between methylation and fragmentation and delineated methylation haplotype blocks (MHBs) directly from plasma cfDNA as biologically interpretable units for organizing multimodal cfDNA signals. Guided by these findings, we developed multimodal diagnostic models integrating complementary methylation and fragmentomic features.

Results

Canonical cfDNA fragmentomic patterns were partially preserved across methylation sequencing protocols, although the degree of preservation varied among individual features. DNA methylation levels were associated with characteristic fragment length and end-motif patterns. Plasma-derived MHBs exhibited tissue-associated regulatory features and distinct patterns of chromatin accessibility, repeat enrichment, and methylation heterogeneity, providing biologically interpretable units for integrated cfDNA analysis. Within these regions, methylation and fragmentomic signals exhibited coordinated yet complementary patterns. Leveraging this framework, we developed Fragmentomics And Methylation Ensemble (FAME), an interpretable multimodal model integrating MHB-centered methylation and fragmentomic features for cancer detection and tissue-of-origin prediction, and further extended it with complementary genome-wide methylation and fragmentomic features to generate a genome-wide-augmented FAME model (FAME-GW). Across cross-validation, independent validation, and comparisons with THEMIS (THorough Epigenetic Marker Integration Solution), SPOT-MAS (screening for the presence of tumor by methylation and size), and DELFI (DNA evaluation of fragments for early interception), FAME-based models achieved robust diagnostic performance, with FAME providing a compact and interpretable representation and FAME-GW delivering the strongest overall predictive performance.

Conclusions

Our study demonstrates that cfDNA whole-genome methylation sequencing partially preserves informative fragmentomic signals and establishes plasma-derived MHBs as biologically interpretable units for organizing and integrating complementary methylation and fragmentomic information. This framework provides insight into the coupling of epigenetic and fragmentomic information and an effective strategy for interpretable cfDNA-based cancer diagnosis.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1186/s13073-026-01793-w.

Background

Liquid biopsy interrogates circulating nucleic acids to enable minimally invasive, repeatable cancer screening and monitoring, with clear potential to transform early detection and longitudinal care [1–3]. Among circulating signals, cell-free DNA (cfDNA) methylation has emerged as a particularly powerful feature space: methylation changes occur early in tumorigenesis and encode rich tissue-of-origin information, thereby offering both sensitivity and localization cues [4, 5]. Beyond single-CpG averages, read-level metrics (e.g., proportion of discordant reads (PDR); methylation entropy) [6, 7] and block-level summaries (e.g., methylation haplotype blocks, MHBs) [8] capture local co-methylation structure and improve signal-to-noise. Previous MHB-based diagnostic frameworks have often relied on MHBs defined from bulk tumor or normal tissue methylomes and then quantified these tissue-derived regions in cfDNA for classification [9]. While effective, this tissue to cfDNA transfer implicitly assumes that tissue-identified blocks faithfully represent the regulatory features circulating in plasma. Given that cancer is often accompanied by systemic physiological changes beyond focal tumor alterations, directly discovering and characterizing MHBs in cfDNA provides a complementary approach to defining plasma-derived methylation structures and evaluating their cancer-detection potential.

Complementary to methylation, cfDNA fragmentomics provides an additional layer of information shaped by nucleosome positioning, chromatin accessibility, and nuclease activity [10]. Cancer-derived cfDNA exhibits hallmark fragmentation signatures, including global enrichment of short fragments, altered nucleosomal periodicity, and genome-wide tissue-specific fragmentation landscapes that have been exploited for detection and tissue-of-origin inference [11]. Metrics such as windowed protection score (WPS) [10], DELFI profiles [12], and end motif diversity (EDM) [13] capture distinct fragmentation properties that are biologically related to, but not redundant with, methylation patterns. These fragmentomic features provide diagnostic information that is distinct from methylation-based signals and may serve as a complementary source of data. Integrating fragmentomics with methylation analysis could thus enhance both the sensitivity and biological interpretability of cfDNA-based assays. Recent multimodal cfDNA frameworks have begun to combine methylation, fragmentation, copy-number, and other genome-wide features for cancer detection, highlighting the value of integrated liquid-biopsy signals [14, 15]. Wang et al. further applied cell-free methylated DNA immunoprecipitation sequencing (cfMeDIP-seq) to jointly analyze cfDNA methylation, fragment length, end motifs, and breakpoint motifs for esophageal cancer detection [16]. The prospective GUIDE study developed GutSeer, a targeted bisulfite-sequencing assay integrating methylation and fragmentomic features for multi-gastrointestinal cancer detection [17].

Yet substantial uncertainty remains about whether fragmentomic signals can be robustly recovered from methylation-sequencing libraries. Although fragmentomics is most commonly derived from whole-genome sequencing (WGS), methylation assays (e.g., whole-genome bisulfite sequencing (WGBS) or enzymatic methyl sequencing (EM-seq)) subject cfDNA to different chemical and enzymatic conditions that may damage molecules or bias ends and lengths. Previous studies have demonstrated links between cfDNA methylation and fragmentation and explored the inference of methylation states from fragmentomic patterns [18–20]. However, the extent to which different methylation-sequencing protocols preserve diverse fragmentomic features relative to WGS remains incompletely characterized. Without such an assessment, it remains difficult to determine which fragmentomic readouts from methylation libraries are robustly retained, which are more protocol- or cohort-sensitive, and how they complement methylation itself.

To address these gaps, we first systematically investigated whether whole-genome methylation sequencing preserves cfDNA fragmentomic information through cross-platform comparisons of WGS, WGBS, and EM-seq datasets across multiple independent plasma cohorts. We then examined the biological relationship between methylation and fragmentation to determine whether these two modalities capture coordinated yet complementary information. To provide an interpretable framework for integrating these signals, we delineated methylation haplotype blocks (MHBs) directly from plasma cfDNA and comprehensively characterized their tissue specificity, regulatory activity, chromatin accessibility, repeat enrichment, and methylation heterogeneity. Within these plasma-derived MHBs, we further investigated the interplay between methylation and fragmentomic features and identified two distinct classes of MHBs with divergent epigenetic and fragmentomic characteristics. Guided by these findings, we developed FAME (Fragmentomics And Methylation Ensemble), an interpretable stacking ensemble that integrates MHB-centered methylation and fragmentomic features for accurate cancer detection and tissue-of-origin prediction using methylation sequencing data alone. Finally, to evaluate whether complementary genome-wide methylation and fragmentomic information could further improve diagnostic performance, we extended FAME by incorporating genome-wide methylation and fragmentomic features to generate FAME-GW and benchmarked both frameworks against representative cfDNA diagnostic methods.

Methods

Data collection

We used ten publicly available plasma cfDNA datasets and in-house plasma cfDNA datasets (Additional file 1: Tables S1a, S1b), totaling 2,018 samples from healthy controls and seven cancer types: breast (BRCA), colorectal (COREAD), esophageal (ESCA), gastric (STAD), hepatocellular (LIHC), lung cancer (LC), and pancreatic (PACA) in this work. These samples were generated using multiple sequencing platforms, including whole-genome sequencing (WGS), enzymatic methyl-seq (EM-seq), whole-genome bisulfite sequencing (WGBS), and cfMeDIP-seq. An et al. WGS dataset (HRA002250) [20] and Mathios et al. LUCAS WGS dataset (EGAS00001005340) [21] were used to benchmark the consistency of fragmentomic profiles and to examine the coupling between methylation and fragmentomic signals. An et al. EM-seq dataset (HRA002298) [20], Gao et al. [22] WGBS dataset (CRA001142; also deposited as PRJNA494975) and our in-house WGBS dataset (HRA021447) were used to assess fragmentomic consistency across different sequencing platforms. Bie et al. [15] EM-seq dataset (HRA003209) was used to delineate MHBs and train the cancer diagnosis model. The Tao et al. cfMeDIP-seq dataset (SRP343170) [23] was analyzed to quantify the association between methylation signals and fragmentomic features.

For the diagnostic task, three cohorts were used for model development and evaluation: the Bie et al. [15] EM-seq dataset (HRA003209), the Nguyen et al. [14] WGBS dataset (PRJNA929650), and the Zhang et al. [24] WGBS dataset (CRA001537), with sample characteristics summarized in Additional file 1: Tables S2–S3 and dataset-level information for CRA001537 summarized in Additional file 1: Tables S1a, S1b. In the original studies, both HRA003209 and PRJNA929650 were pre-divided into training and validation cohorts with balanced case-control ratios and matched age and sex distributions, supporting reliable model evaluation. In the Bie et al. dataset, which includes samples from seven cancer types, we constructed cancer-specific diagnostic models for each cancer type and evaluated their performance using 10-fold cross-validation within the training dataset. It should be noted that the cancer-specific binary models served as a controlled framework for comparing feature sets and methods, rather than as a single deployable pan-cancer detection model. The corresponding cancer and healthy samples from the Bie et al. independent validation cohort were used to assess model performance. For the hepatocellular carcinoma task, the model trained on the Bie et al. training cohort (HRA003209) was applied to the Zhang et al. dataset (CRA001537) for external independent validation without retraining. For the Nguyen et al. dataset (PRJNA929650), a breast cancer model was trained on its own training cohort and evaluated once on its held-out independent validation cohort.

In-house sample collection and processing

Peripheral blood samples were collected from 9 healthy volunteers at The Eighth Affiliated Hospital of Sun Yat-sen University and 19 STAD patients at Fujian Medical University Union Hospital. For each participant, approximately 12 mL of whole blood was drawn into EDTA tubes. To isolate plasma, the blood samples underwent a two-step centrifugation protocol: an initial centrifugation at 1,600 g for 10 minutes at 4 °C, followed by a second centrifugation of the supernatant at 16,000 g for 10 minutes at 4 °C to ensure the complete removal of residual intact cells and debris. The isolated plasma was stored at −80 °C until further processing.

cfDNA extraction and library construction

Circulating cell-free DNA (cfDNA) was extracted from plasma using a magnetic-bead-based nucleic acid extraction kit (Jifan Biotechnology, Changzhou, China) according to the manufacturer’s instructions. The concentration of extracted cfDNA was measured using the Qubit dsDNA HS Assay Kit (Invitrogen), and fragment size and integrity were assessed using the Agilent 2100 Bioanalyzer (Agilent) and/or LabChip GX system (PerkinElmer).

Whole-genome bisulfite sequencing (WGBS) libraries were constructed using the Acegen Single Strand Bisulfite-Seq Library Prep Kit (Acegen, Cat. No. AG0312) according to the manufacturer’s protocol. Briefly, cfDNA was spiked with unmethylated lambda DNA as a bisulfite-conversion control and subjected to bisulfite conversion using the EZ DNA Methylation-Gold Kit (Zymo Research). The converted DNA was heat-denatured into single strands and ligated to a 3′-dA-tailed adapter. The single-stranded DNA carrying the 3′ adapter was then extended to double-stranded form, followed by ligation of a 5′ adapter containing methylated cytosines. Libraries were purified using Agencourt AMPure XP beads (Beckman Coulter) and amplified with approximately 10 cycles of PCR using Illumina dual-index primers. Library size distribution and concentration were assessed using Agilent 2100 Bioanalyzer, LabChip GX, Qubit dsDNA HS Assay Kit, and/or qPCR before sequencing.

Illumina sequencing

Qualified WGBS libraries were pooled and sequenced on the Illumina NovaSeq platform using a 150-bp paired-end sequencing strategy with NovaSeq reagent kits. Raw sequencing quality metrics, including total reads, total bases, Q20 rate, Q30 rate, and GC content, were obtained from sequencing summary reports. Bisulfite-conversion efficiency was estimated from lambda DNA spike-in controls. Bismark deduplication statistics were extracted from deduplication reports. Additional fragment-level metrics, including usable fragment counts and insert-size statistics, were computed from processed deduplicated BAM files and therefore reflect retained aligned fragments after preprocessing. These sequencing-, conversion-, duplication-, and BAM-level QC metrics for the in-house WGBS cohort are summarized in Additional file 1: Table S1c.

Data preprocessing

For WGS data, raw FASTQ files were quality-controlled with Trim Galore [25] to remove low-quality bases and adapter sequences. Cleaned reads were aligned to the human reference genome hg19 using Bowtie2, and SAM files were converted to sorted BAM format with SAMtools. PCR duplicates were marked and removed with Picard.

For EM-seq, WGBS, and cfMeDIP-seq data, similar quality-control was applied using Trim Galore or fastp. Reads were aligned to hg19 with Bismark (for EM-seq and WGBS) or Bowtie2 (for cfMeDIP-seq). Methylation calling was performed using Bismark for EM-seq and WGBS data, while PCR duplicates were removed with deduplicate_bismark or GATK depending on the platform. Sorted BAM files were generated with SAMtools.

For all datasets, genomic coordinates of aligned reads were extracted with custom scripts using the pysam library for downstream analysis.

Consistency analysis in TSS regions

In addition to assessing cross-platform similarity of fragmentomic features using Spearman correlation, we further examined the distribution of five locus-resolved fragmentomic features [promoter fragmentation entropy (PFE), fragment coverage (COV), orientation-aware cell-free DNA fragmentation (OCF), integrated fragmentation score (IFS), and WPS] within transcription start site (TSS) regions. Specifically, each TSS was expanded to a window of [−2000,+2000] bp. For each base position in this window, a sub-window of [−100, +100] bp was defined, and the fragmentomic feature value was calculated within the sub-window. The mean values across all TSSs were then aggregated at each genomic position, yielding a profile of 4,001 positions per feature.

Fragmentomic feature computation

All fragmentomic features were computed as follows:

PFE: For each region, we computed the proportion of cfDNA fragments with lengths between 100 and 300 bp at each base, and calculated the entropy according to Eq. (1).

1

Coverage (COV): Number of cfDNA fragment midpoints within the region. For cross-modal regression, coverage was corrected for GC bias in a fragment-length-specific manner [26].

End: Number of cfDNA fragment ends within the region.

OCF: Counts of cfDNA fragment start (U) and end (D) positions at each genomic coordinate. For each region, we selected the range of [−10,+10] bp around ±60 bp upstream and downstream of the region midpoint, and computed OCF according to Eq. (2). For cross-modal regression, this calculation was repeated at multiple positions within each target region using the 120-bp scale defined by the ±60-bp offsets, and the resulting OCF values were averaged.

2

IFS: Based on the number of cfDNA fragments n, mean fragment length l, and chromosome-wide mean fragment length L, IFS was calculated as shown in Eq. (3).

3

FDI: For a specific genomic region of a given length (e.g., 200 bp), std(coverage) is the standard deviation (MATLAB function: std) of fragment coverage within the region. Here, n is the total number of endpoints in the region, x is a constant within the range (0, 1) that can be adjusted based on sequencing depth to obtain a more accurate description of endpoint dispersion; in this study, x is set to 0.5; m represents the number of endpoints within a ± 10 bp window around the jth endpoint (e.g., [−10 bp, 10 bp]). The FDI is given by the following Eq. (4):

4

WPS: Defined as the difference between the number of cfDNA fragments that fully span the region and the number of fragments that only partially cover it. In addition to the standard WPS, we also computed two size-restricted versions: L-WPS, calculated using fragments of 120–180 bp, and S-WPS, calculated using fragments of 35–80 bp. For cross-modal regression, WPS was calculated in 120-bp windows at multiple positions within each target region and averaged to obtain a continuous regional score.

EDM: We considered the 4-bp 5′ end motif of each fragment whose centroid fell within the region. The proportion of each 4-mer was counted per chromosome, yielding a 256 × 22 vector per sample. For FAME, end-motif profiles were computed using only fragments whose centroids fell within the MHB-centered feature regions.

Length: We binned cfDNA fragment lengths into 10-bp intervals and, for each chromosome, computed the proportion of fragments falling into each length bin.

FSR: We partitioned the genome into non-overlapping 5 Mb windows and, for each window, calculated the proportions of short (65–150 bp), medium (151–220 bp), and long (221–400 bp) fragments.

FSD: For each chromosome arm, we grouped cfDNA fragments within 65–400 bp into 5-bp bins and computed the proportion of fragments in each bin.

Methylation feature computation

In addition to fragmentomic features, we analyzed methylation features within specified regions (e.g., TSSs, MHBs) to capture additional epigenetic signals relevant to cancer. Using the stat module of mHapSuite (v2.0), seven methylation metrics were computed from mhap files, including mean methylation level (MM), proportion of discordant reads (PDR), cellular heterogeneity-adjusted clonal methylation (CHALM) [27], methylation concurrence ratio (MCR) [28], methylated haplotype load (MHL) [8], methylation block score (MBS) [29], and methylation entropy (ENTROPY) [30]. Default parameters were used except that we set the minimum number of CpGs per block to 3, the minimum CpG coverage to 1, required at least one k4Plus haplotype, and required at least one pairwise CpG coverage for r² calculation. These metrics collectively evaluate the stability, heterogeneity, and potential regulatory impact of DNA methylation patterns, thereby providing a rich set of epigenetic features for cancer detection.

Mean methylation level (MM) was calculated as the fraction of methylated CpG observations among all covered CpG observations in a given region:

5

where and denote the numbers of methylated and unmethylated CpG observations in the region, respectively.

Methylation concurrence ratio (MCR) was used to quantify the co-occurrence of methylated and unmethylated CpGs within the same DNA molecules. For a given region, discordant reads were defined as reads containing both methylated and unmethylated CpGs. MCR was calculated as:

6

where denotes the number of unmethylated CpG observations located within discordant reads, and denotes the total number of CpG observations covered by all reads in the region.

Proportion of discordant reads (PDR) was used to quantify read-level methylation disorder within a genomic region. Discordant reads were defined as reads containing both methylated and unmethylated CpGs. PDR was calculated as:

7

where denotes the number of discordant reads in the region, and denotes the total number of eligible reads covering the region.

Cellular heterogeneity-adjusted clonal methylation (CHALM) was used to quantify molecule-level methylation while accounting for cellular heterogeneity in bulk sequencing data. Unlike mean methylation level, which averages methylated CpG observations across all covered CpG sites, CHALM first converts each read into a binary methylation state. For a given region, let x_ij denote the methylation state of the j_th CpG covered by read i, where x_ij = 1 indicates methylated and x_ij = 0 indicates unmethylated. A read was defined as methylated if it contained at least one methylated CpG. CHALM was calculated as:

8

where N denotes the total number of eligible reads covering the region. Thus, a read contributes 1 to CHALM if it contains at least one methylated CpG, and contributes 0 otherwise.

Methylated haplotype load (MHL) was used to quantify the degree of coordinated methylation across multiple CpGs on the same DNA molecule. Unlike mean methylation level, which averages CpG methylation independently, MHL considers all consecutive CpG substrings within methylation haplotypes and gives higher weight to longer fully methylated substrings. For a given region, MHL was calculated as:

9

where L denotes the maximum haplotype length considered, denotes the fraction of fully methylated consecutive k-CpG haplotypes among all observed k-CpG haplotypes in the region, and is the weight assigned to k-CpG haplotypes. Following the original definition, we used , thereby giving greater contribution to longer fully methylated haplotypes.

Methylation block score (MBS) was used to quantify continuous methylation patterns within individual DNA molecules. For a given region, each read was examined for consecutive methylated CpG segments. MBS was calculated as:

10

where n denotes the total number of eligible reads covering the region, L_i denotes the number of CpGs covered by the i-th read, l_ij denotes the length of the j-th consecutive methylated CpG segment on the i-th read, and m_i denotes the number of consecutive methylated segments on that read. The squared segment length gives greater weight to longer continuous methylated stretches, allowing MBS to capture block-like methylation patterns rather than only average methylation levels.

Methylation entropy (ENTROPY) was used to quantify the diversity and randomness of methylation haplotype patterns within a genomic region. For a region with K CpG sites considered in each methylation haplotype, let denote the frequency of the h-th observed methylation pattern among all eligible reads. ENTROPY was calculated as:

11

where H denotes the number of observed methylation patterns, and denotes the relative frequency of pattern h. A lower ENTROPY indicates more homogeneous methylation patterns, whereas a higher ENTROPY indicates greater methylation pattern diversity.

Association analysis between fragmentomic and methylation features

We further investigated the relationship between cfDNA methylation and fragmentomic features. By comparing fragmentomic characteristics across regions with different methylation states, we aimed to assess the association between methylation and cfDNA fragmentation, particularly fragment length and coverage. Healthy cfDNA samples from HRA002298 (EM-seq, An et al.) and our in-house WGBS datasets were analyzed. For each sample, TSSs were defined as [−100,+100] bp windows, and average methylation levels across TSSs were computed using mHapSuite (v2.0) [31]. Regions within the top 20% of average methylation were defined as high-methylation regions, while those in the bottom 20% were defined as low-methylation regions. In each set, we quantified the proportion of short fragments (35–80 bp), long fragments (120–180 bp), and overall fragment coverage. Group differences were evaluated using the Mann-Whitney U test, a nonparametric method implemented with the mannwhitneyu function in SciPy (v1.11.4) under Python 3.9.7.

In addition, we compared unenriched and methylation-enriched cfDNA using paired datasets SRP343169 (WGS, Tao et al. [23]) and SRP343170 (cfMeDIP-seq, Tao et al.), which were generated from the same plasma samples. WGS profiles both methylated and unmethylated fragments, whereas cfMeDIP-seq enriches for methylated DNA [32]. We first compared fragment length distributions between the two datasets to evaluate methylation-dependent length differences. We then analyzed the composition of 4-mer end motifs (EDM) to assess whether methylation enrichment is associated with differences in fragment end sequence patterns. Since the paired samples did not pass normality tests, group differences in end motif proportions were assessed using the Wilcoxon signed-rank test, implemented with the wilcoxon function in SciPy (v1.11.4).

Cross-regression between fragmentomic and methylation features

Cross-modal regression analyses were performed separately in the An et al. EM-seq dataset (HRA002298) and the in-house WGBS dataset using TSS-centered regions. Five fragmentomic features (WPS, OCF, COV, FDI, and IFS) and seven methylation features (MM, PDR, CHALM, MHL, MCR, MBS, and ENTROPY) were evaluated. Each fragmentomic feature was individually used to predict each methylation feature, and reciprocal regressions were performed with methylation features as predictors and fragmentomic features as targets. The processed WPS, OCF, and COV values described above were used in these analyses. Beyond this fragmentomics to methylation setup, we further extended the analysis into a fully reciprocal framework: each fragmentomic feature was used to predict all methylation features as well as the other fragmentomic metrics, but not itself; likewise, each methylation feature was used to predict all fragmentomic features and the remaining methylation metrics, excluding self-prediction. This design ensured that every feature was evaluated both as a predictor and as a target, except for trivial self-regression, thereby providing a balanced reference for quantifying shared versus unique information across the two feature classes.

We employed a gradient boosting regression model based on decision trees to characterize nonlinear relationships between fragmentomic and methylation features [33]. The model was implemented in MATLAB using the fitrensemble function with the method set to LSBoost. The analyses used 50, 100, 500, 1,000, 2,000, 5,000, and 10,000 boosting trees. Model fit was assessed using R² calculated on the regions used for fitting. Residuals were defined as observed minus fitted target values, and their Spearman correlations with the observed target values were calculated. Results were summarized separately for each cohort, with predictor-specific summaries averaged across targets in the opposite modality. Directional summaries at 10,000 trees were additionally averaged across predictors, and the corresponding unexplained fractions were calculated as 1 − mean fitted R².

Identification and analysis of methylation haplotype blocks (MHBs)

Methylation haplotype blocks (MHBs) are defined as genomic regions where adjacent CpG sites exhibit highly correlated methylation states, forming patterns of co-methylation. Because MHBs have been implicated as potential biomarkers in cancer and other diseases, accurate identification and characterization of these regions are critical.

In this study, MHBs were identified using mHapSuite (v2.0), a Java implementation of mHapTk specifically designed for methylation haplotype analysis [31]. Compared to the original mHapTk, mHapSuite provides improved efficiency and functionality, and includes the MHBDiscovery module for de novo detection of locally co-methylated regions across the genome.

For MHB identification, we selected the Bie et al. dataset, which comprises seven cancer types (breast, colorectal, esophageal, gastric, lung, liver, and pancreatic cancer) as well as healthy controls. This dataset was chosen due to its large sample size and high sequencing quality. After preprocessing, aligned BAM files were first converted to mhap format using the convert module of mHapSuite. The mhap format encodes the methylation states of CpG sites on individual cfDNA fragments, substantially reducing file size while preserving fragment-level methylation information. Next, mhap files from training samples of the same type were merged using the merge module (e.g., all breast cancer samples were combined into a single mhap file). Finally, MHBs were detected for each cancer type and the healthy group using the MHBDiscovery module with default parameters: –window 5 –r_square 0.5 –p_value 0.05. This process yielded sets of MHBs specific to each cancer type as well as to healthy controls.

Genomic distribution of methylation haplotype blocks (MHBs)

After identifying MHBs across different sample types, we next examined their genomic distribution. The localization of MHBs within functional genomic regions, such as promoters, introns, and exons, may reveal potential associations with regulatory activities and disease mechanisms.

To systematically evaluate the enrichment of MHBs in different genomic regions, we used ChIPseeker (v1.40.0) [34]. Originally developed for chromatin immunoprecipitation sequencing (ChIP-seq) data, ChIPseeker provides precise genomic annotation and functional categorization of genomic intervals. In this study, it was applied to annotate the positions of MHBs and quantify their enrichment across functional categories including promoters, introns, and exons, thereby enabling further investigation into their potential roles in gene regulation.

Epigenomic features and potential regulatory roles of MHBs

We further investigated the epigenomic features of MHBs and their potential regulatory functions. In addition to reflecting co-methylation patterns among adjacent CpG sites, MHBs may also be associated with gene expression regulation and chromatin accessibility. Previous studies have shown that the genome can be partitioned into distinct DNA methylation domains, such as unmethylated regions (UMRs), low-methylated regions (LMRs), partially methylated domains (PMDs), and high-methylated regions (HMRs), each enriched for different regulatory activities [35]. To assess the distribution of MHBs across these domains, we used tissue-specific DNA methylation annotations provided by Feng et al. [36] and calculated overlaps between MHBs and annotated regions with the intersect function in the GenomicRanges package (v1.56.2) [37]. The proportions of MHBs falling into each category (UMR, LMR, PMD, and HMR) were quantified.

To further explore the potential regulatory functions of MHBs, we analyzed their distribution within open chromatin regions, which mark chromatin accessibility and are closely associated with transcription factor binding and gene regulation. We obtained assay for transposase-accessible chromatin using sequencing (ATAC-seq) profiles of relevant tissues from Feng et al. [36] and overlapped MHBs from each sample type with the corresponding open chromatin regions. The degree of overlap was quantified to characterize the overlap between MHBs and accessible chromatin, thereby providing indirect evidence for their potential involvement in gene regulatory processes.

Identification of cancer-specific MHBs and analysis of their tissue-of-origin features

To identify cancer-specific MHBs, we compared MHBs detected in each cancer type with those from other sample groups. For a given cancer type, MHBs that overlapped with MHBs from any other group were removed, and only non-overlapping regions unique to that cancer type were retained as cancer-specific MHBs. This filtering ensured that the selected MHBs were exclusive to the cancer of interest.

To further evaluate whether MHBs identified from plasma cfDNA reflect signals originating from the corresponding tissues, we compared cancer-specific MHBs with tissue-derived MHBs reported by Feng et al., which were identified from WGBS data of multiple normal tissues. The overlap analysis was performed using the LOLA package (v1.34.0) [38] in R. LOLA is a widely used tool for genomic region enrichment analysis that assesses the overlap between input loci and reference genomic datasets. In this study, LOLA was applied to quantify the overlap between cancer-specific MHBs and normal tissue MHBs, thereby assessing whether the cfDNA-derived MHBs exhibit tissue-of-origin characteristics and exploring their potential roles in cancer development.

Motif identification and GO/KEGG pathway enrichment analysis

To investigate the potential regulatory functions of cancer-specific MHBs, we performed motif discovery as well as Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analyses. Motif enrichment was used to identify transcription factor binding sites within MHBs, providing insights into their potential role in gene expression regulation. GO enrichment analysis evaluated the biological processes, cellular components, and molecular functions associated with MHB-linked genes, while KEGG pathway analysis further characterized their involvement in signaling and metabolic pathways, thereby offering a more comprehensive view of the biological functions of MHBs in cancer.

Motif enrichment analysis was conducted using HOMER (v4.11.1) [39]. The findMotifsGenome.pl module was applied separately to cancer-specific MHBs from each cancer type to identify overrepresented transcription factor binding motifs. Known motif enrichment results were evaluated using the Benjamini q values reported by HOMER, which were used to account for multiple testing. Motifs were considered enriched only when they passed the HOMER Benjamini q-value threshold, and the enriched motifs were selected from these corrected results.

GO and KEGG enrichment analyses were performed using Cistrome-GO [40], which integrates large-scale epigenomic data to systematically evaluate functional characteristics of regulatory regions and their associated genes. Genes linked to cancer-specific MHBs were analyzed with Cistrome-GO to assess their enrichment in biological processes, chromatin regulatory pathways, and cancer-related signaling cascades. For GO and KEGG analyses, enriched terms were filtered using the false discovery rate (FDR) values reported by Cistrome-GO, with FDR < 0.2 used as the threshold for retaining enriched terms. The raw enrichment p values were used for ranking and visualization, whereas FDR values were used to account for multiple testing.

By integrating motif enrichment, GO analysis, and KEGG pathway analysis, we comprehensively characterized the potential regulatory functions of cancer-specific MHBs, thereby providing insights into their potential roles in cancer development and progression.

Tumor DNA fraction estimation

Tumor DNA fraction represents the proportion of cfDNA derived from tumor tissue within a sample and serves as an important indicator for assessing tumor burden, guiding early cancer detection, and monitoring disease progression. We estimated tumor DNA fractions using ichorCNA (v0.2.04) with parameter settings consistent with previous studies [41]. ichorCNA is a widely used copy number variation (CNV) analysis tool for low-coverage whole-genome sequencing data, which can infer the tumor-derived cfDNA fraction without requiring prior knowledge of somatic mutations [42]. By modeling genome-wide CNV patterns together with background noise, ichorCNA estimates tumor fraction by modeling genome-wide copy-number alterations.

In this study, tumor DNA fractions were estimated for all samples from the Bie et al. dataset and further analyzed in relation to MHB regions and cancer diagnostic models. The Nguyen et al. PRJNA929650 dataset was excluded from tumor DNA fraction estimation due to its shallow sequencing depth.

Definition of MHB-defined red and blue domains

To define MHB-centered red and blue domains, we first grouped samples by category and computed the mean value of each methylation feature for every MHB region within each category. Regions with more than two missing values among the seven methylation features were excluded. The remaining MHB regions were clustered using k-means based on their methylation feature profiles. The two clusters were then assigned as red or blue domains according to their average methylation levels: the lower-methylation cluster was defined as the red domain class, and the higher-methylation cluster was defined as the blue domain class. These labels were used as operational annotations for cfDNA-derived MHB contexts rather than as predefined chromatin-state labels. To evaluate whether this methylation-defined organization was consistent across genomic contexts and sequencing protocols, we also applied the same procedure to TSS-centered regions in the HRA002298 and in-house WGBS datasets.

Enrichment analyses of repetitive elements and replication timing

We next evaluated the enrichment of repetitive elements within the red (low-methylation, regulatory-like) and blue (high-methylation, structural-like) regions defined from MHBs.

For repetitive elements, each MHB was assigned a midpoint m=(start+end)/2, and a symmetric window [m−w,m+w] was constructed with w = 500 bp. Regions lacking red/blue labels were excluded. Overlaps between these windows and repetitive elements annotated by UCSC RepeatMasker (e.g., SINE, LINE, LTR) were then assessed on a per-chromosome basis. A binary indicator (1 = overlap, 0 = no overlap) was assigned to each window, and hit rates were calculated separately for red and blue groups. The primary effect size was the difference in hit rates (), with odds ratios (ORs) derived from 2 × 2 contingency tables as a complementary effect measure. For PDR comparisons across SINE-overlapping blue MHBs, each cancer type was compared with healthy controls using two-sided Wilcoxon rank-sum tests, followed by Benjamini–Hochberg FDR correction across cancer types.

Uncertainty was quantified by constructing Wilson confidence intervals for within-group proportions and bootstrap percentile 95% CIs (8,000 resamplings) for the effect size Δ. Statistical significance was assessed using two-sided Fisher’s exact tests on contingency tables, as well as label permutation tests (5,000 iterations) in which group labels were randomly reassigned while preserving the red-group sample size. A continuity-corrected p value was derived from the permutation distribution of . Results were visualized using Gardner–Altman estimation plots: the left panel displayed group hit rates with Wilson CIs (with permutation p values annotated), and the right panel overlaid the bootstrap density of Δ with the null distribution from permutations, indicating the observed Δ, its 95% CI, and reporting the OR, Fisher p value. Analyses were implemented in MATLAB (R2020a) with the random seed fixed at 42. For multiple comparisons (e.g., across repeat families or window sizes), p values were adjusted using the Benjamini–Hochberg FDR procedure (effect sizes and intervals were reported without correction).

For model interpretation, replication timing enrichment was assessed by comparing positive and negative important MHB regions identified from the first-layer linear SVMs of FAME. Replication timing data were downloaded from the ENCODE Project (accession ENCFF291UPC, bigWig format). The top 150,000 genomic regions with the highest signal intensities were defined as early-replicating regions, and the bottom 150,000 regions were defined as late-replicating regions.

Methylation and fragmentomic analysis in red and blue domains

We evaluated the behavior of methylation and fragmentomic features in red versus blue regions. In the An et al. EM-seq (HRA002298) datasets, four fragmentomic metrics related to chromatin accessibility (WPS, OCF, IFS, and FDI) were compared with DNase-seq signals from monocytes and B cells. Spearman correlations were computed as in a previous study [43]. To ensure balanced comparison, red and blue regions were separately ranked by DNase-seq signal intensity and partitioned into bins of equal size, and median feature values were calculated within each bin.

For diagnostic analyses across cancer types, the Bie et al. training and independent validation datasets were used. MHBs were divided into red and blue regions without subsampling, as the region counts were similar. Classifiers were trained with MATLAB’s default fitcsvm function, using the same procedures as for single-modality models. Classification performance was then evaluated separately in red and blue regions.

Selection of cfDNA methylation and fragmentomic features

Based on the preceding analyses, we selected multi-cancer and control MHBs identified from the Bie et al. training cohort (HRA003209) as the feature regions. Accordingly, only features defined within these specific regions were considered, while global descriptors such as FSR and FSD were excluded. These MHB-centered regions allowed methylation and fragmentomic signals to be quantified within matched genomic intervals.

For cfDNA fragmentomic features, we focused on those originally proposed to capture chromatin accessibility. Among WPS, FDI, IFS, OCF, and COV, we retained only WPS, as these features describe related aspects of local chromatin accessibility, nucleosome protection, coverage, or endpoint dispersion, and are therefore biologically related and partially redundant. WPS was selected because it is a widely used and directly interpretable measure of local nucleosome protection, defined by the balance between fragments spanning a region and fragments ending within it. In contrast, IFS integrates regional fragment abundance with fragment-length information normalized by chromosome-wide average length, which can provide strong predictive signal but is less directly tied to a single local protection mechanism. To complement this, we further included EDM, which was treated as a distinct fragmentomic feature class because it captures 4-bp fragment-end motif composition and reflects nuclease-associated cleavage preferences rather than local protection or coverage patterns. PFE was originally designed for ultra-deep whole-exome sequencing data. When applied to shallow WGS data in this context, its performance exhibited greater variability, and therefore it was not included in the diagnostic model.

For methylation feature selection, we organized the seven metrics into two broad, descriptive groups based on their definitions and observed profiles across MHBs: methylation-level-related features (MM, CHALM, MBS, and MHL) and methylation heterogeneity/discordance-related features (PDR, ENTROPY, and MCR). MHL and MBS emphasize continuous methylation patterns, whereas MCR quantifies unmethylated CpGs within partially methylated reads. From the first group, we retained MBS owing to its higher signal-to-noise ratio, haplotype-aware definition, and ability to capture continuous methylation patterns. From the second group, we selected PDR to capture read-level methylation discordance and complement the methylation-level information represented by MBS. MHL was not retained because it showed lower training-set performance than PDR and ENTROPY.

Together, the final feature set consists of MBS, PDR, WPS, and EDM, capturing four complementary dimensions of cfDNA variation: methylation-level-related information (MBS), methylation heterogeneity/discordance (PDR), chromatin-accessibility-related fragmentation (WPS), and fragment-end motif composition (EDM). This combination balances interpretability, robustness, and discriminative power while minimizing redundancy. We further evaluated this selection using single-feature, methylation-only, fragmentomics-only, feature-replacement, leave-one-feature-out, and all-feature models, as summarized in Additional file 1: Table S4.

Construction of FAME, FAME-GW, and FAME-multi models

For single-feature models, we independently trained support vector machine (SVM) classifiers using each feature type, including methylation block score (MBS), proportion of discordant reads (PDR), windowed protection score (WPS), and end motif diversity (EDM). All single-feature models were implemented using MATLAB’s fitcsvm function with a linear kernel and otherwise default hyperparameters. Model performance was evaluated by 10-fold cross-validation within the training set and validated on the independent test set for cancer-specific binary classification tasks.

To integrate complementary signals across methylation and fragmentomic dimensions, we developed the FAME model. In each training fold, predicted scores from the four single-feature SVMs were retained and concatenated as input to a second-layer SVM, which was also implemented using MATLAB’s fitcsvm function with a linear kernel and otherwise default hyperparameters. This stacking strategy leverages complementary feature representations to enhance classification robustness. Predicted outputs from FAME were used for downstream analyses, including comparisons across cancer types, disease stages, and tumor DNA fraction strata.

We also constructed FAME-GW as a genome-wide extension of the FAME framework. In addition to the four MHB-centered FAME features, including PDR, MBS, WPS, and EDM, FAME-GW incorporated four genome-wide cfDNA features derived from existing multimodal cfDNA frameworks: genome-wide methylation (GWM), methylated fragment ratio (MFR), chromosomal aneuploidy of featured fragments (CAFF), and genome-wide 4-mer end motif (EM). For each of the eight feature groups, a first-layer linear-kernel SVM classifier was trained independently, and each sample was represented by the corresponding prediction score from each feature-specific model. The eight first-layer prediction scores were then concatenated and integrated using a second-layer random forest classifier to generate the final FAME-GW prediction score. This design was used to evaluate whether MHB-centered methylation-fragmentomic signals could be further strengthened by complementary genome-wide methylation, copy-number, and end-motif information.

To extend the framework to multi-class classification, we implemented a variant termed the FAME-multi model. The first layer remained a multi-class SVM implemented with MATLAB’s fitcecoc function, while the second-layer classifier was replaced by k-nearest neighbors (KNN) using MATLAB’s fitcknn function. This design outputs class probability distributions across cancer types, enabling direct multi-class predictions while retaining the same feature representations as the original FAME model. KNN was chosen for the second layer because, unlike SVM, which requires explicit multi-class decomposition, KNN natively handles multi-class settings and provides probabilistic outputs based on neighborhood voting. This makes it suitable for capturing local similarity among samples across cancer types, thereby improving interpretability and stability in multi-class tasks.

Benchmark implementation of THEMIS

THEMIS was implemented as a genome-wide multimodal cfDNA benchmark based on the feature definitions reported by Bie et al. Briefly, four published THEMIS feature types were extracted: methylated fragment ratio (MFR), defined as the fraction of fully methylated fragments in 1-Mb genomic windows; fragment size index (FSI), defined as the ratio of short fragments (100–166 bp) to long fragments (169–240 bp) in 5-Mb windows; chromosomal aneuploidy of featured fragments (CAFF), summarized by chromosome-arm-level copy-number changes using size-selected short and long fragments; and fragment end motif (FEM), calculated from the frequencies of 256 possible 4-mer motifs at the 5′ fragment ends. For classification, MFR and FSI were modeled using support vector machines, FEM was modeled using logistic regression, and CAFF was represented by the log-transformed plasma aneuploidy score. These modality-level outputs were integrated using a regularized logistic-regression classifier to generate the final THEMIS prediction score.

For a genomic window w, MFR was calculated as

12

where eligible fragments were required to cover at least three CpGs. FSI was calculated as the ratio of short to long fragments in each 5-Mb window:

13

FEM was represented by the relative frequency of each 4-mer 5′ end motif:

14

where denotes the number of fragments carrying the k-th 4-mer end motif. CAFF was calculated following the published THEMIS definition, using chromosome-arm-level copy-number changes from size-selected fragments and summarized as a plasma aneuploidy score.

Benchmark implementation of SPOT-MAS

SPOT-MAS was implemented as a representative multimodal cfDNA benchmark based on the feature definitions reported by Nguyen et al. Nine published SPOT-MAS feature groups were extracted, including target methylation (TM), genome-wide methylation (GWM), copy number aberration (CNA), 4-mer end motif (EM), fragment length distribution (FLEN), short fragment count (SHORT), long fragment count (LONG), total fragment count (TOTAL), and short-to-long fragment ratio (RATIO). For methylation-based features, the methylation ratio was calculated as:

15

TM was computed within the published cancer-associated target regions, whereas GWM was computed in genome-wide 1-Mb bins. CNA was derived from normalized genome-wide bin coverage. FLEN represented the fragment length frequency distribution from 100 to 250 bp:

16

For 5-Mb genomic bins, SHORT, LONG, and TOTAL were calculated from the counts of short fragments, long fragments, and all fragments, respectively, and RATIO was calculated as:

17

where short and long fragments were defined as 100–150 bp and 151–250 bp, respectively. EM was represented by the relative frequency of each 4-mer end motif:

18

where denotes the number of fragments carrying the k-th 4-mer motif. For classification, we followed the main SPOT-MAS strategy by training base classifiers for individual feature groups and integrating their prediction scores using a stacking model.

Benchmark implementation of DELFI

DELFI was implemented as a representative genome-wide fragmentomics benchmark. The genome was partitioned into non-overlapping 5-Mb windows. Within each window, cfDNA fragments of 100–150 bp were defined as short fragments, and fragments of 151–220 bp were defined as long fragments. For each genomic window, the short-fragment ratio was calculated and then corrected for GC-content bias to generate the DELFI feature profile. For each cancer-specific binary task, DELFI feature matrices were constructed using the same training and validation splits as the other benchmark methods. An SVM classifier was trained on the training set or within each cross-validation fold, and the resulting prediction scores were used to calculate the area under the receiver operating characteristic (ROC) curve (AUC) and sensitivity at 95% specificity.

Leakage-control analysis using test-derived MHB regions

To evaluate the possible influence of training-cohort MHB discovery on cross-validation performance, we performed an additional leakage-control analysis using MHBs independently discovered from the Bie et al. independent validation cohort (HRA003209-test). In the original FAME analysis, 115,759 MHB-centered 500-bp regions derived from the HRA003209 training cohort were used as the fixed feature atlas. These regions and the four feature types (PDR, MBS, WPS, and EDM) were fixed across all cancer-specific binary tasks and were not reselected for individual cancer types.

For the leakage-control analysis, MHBs were discovered from the held-out HRA003209 test samples, yielding 116,335 test-derived MHB intervals. To construct an alternative feature atlas with a region number comparable to the original FAME atlas, these intervals were converted into MHB-centered 500-bp regions without additional union merging, resulting in 115,501 MHB-TEST feature regions. The four FAME features, including PDR, MBS, WPS, and EDM, were then recalculated on the HRA003209 training samples using the MHB-TEST atlas.

Using these MHB-TEST features, we repeated the same seven cancer-versus-healthy 10-fold cross-validation tasks with the unchanged FAME architecture, consisting of four first-layer linear SVM classifiers and a second-layer linear SVM. Model performance was evaluated by AUC and sensitivity at 95% specificity. This analysis was designed to test whether the main diagnostic conclusions were robust to replacing the original training-derived MHB atlas with an independently discovered test-derived MHB atlas.

Statistical analysis

Group comparisons were performed using nonparametric tests. Differences between two groups were assessed using the Mann–Whitney U test, whereas differences among three or more groups were assessed using the Kruskal–Wallis test. These tests were implemented in SciPy (v1.11.4). For diagnostic performance, model performance was evaluated using the area under the receiver operating characteristic curve (AUC) and sensitivity at 95% specificity.

To assess whether performance differences between FAME-based models and benchmark methods were statistically supported, paired comparisons were performed on the same evaluated samples. For each cancer type and evaluation setting, performance differences were calculated as the metric of the FAME-based model minus that of the benchmark method. When paired sample-level prediction scores were available, uncertainty was estimated using stratified paired bootstrap resampling, in which cancer and control samples were resampled with replacement separately to preserve the original case–control ratio. AUC and sensitivity at 95% specificity were recalculated for both methods in each bootstrap iteration. Two-sided p values were derived from the bootstrap distribution of paired differences, and 95% confidence intervals were calculated from the 2.5th and 97.5th percentiles. For these benchmark comparisons, 3,000 bootstrap resamples were used, and p values were adjusted using the Benjamini–Hochberg procedure across the 42 comparisons for each target model and metric. For cross-validation analyses where only fold-level metrics were available, paired Wilcoxon signed-rank tests across folds were used as a complementary analysis.

For enrichment analyses involving multiple tested categories or comparisons, multiple-testing correction was applied using false discovery rate control. For motif enrichment analysis, the Benjamini q values reported by HOMER were used, and enriched motifs were retained using a q-value threshold of 0.2. For GO and KEGG pathway enrichment analyses, the FDR values reported by Cistrome-GO were used, with FDR < 0.2 used as the threshold for retaining enriched terms. For repetitive-element enrichment analyses in red and blue MHB regions, overlap proportions were compared using two-sided Fisher’s exact tests and label permutation tests. Multiple-testing correction across repetitive-element families was performed using the Benjamini–Hochberg FDR procedure. For comparisons of PDR levels across SINE-overlapping blue MHBs, each cancer type was compared with healthy controls using two-sided Wilcoxon rank-sum tests, followed by Benjamini–Hochberg FDR correction across cancer types. For replication-timing enrichment analyses, overlap proportions between top positive-contributing and bottom negative-contributing MHBs were compared using two-sided Fisher’s exact tests, and p values across the early- and late-replication comparisons were adjusted using the Benjamini–Hochberg FDR procedure.

Feature interpretation

For region-level interpretation of the FAME model, we focused on the linear SVM classifiers trained for individual feature modalities. Feature contribution was estimated using the signed weights assigned by the linear SVM. Features were ranked according to the absolute values of their weights. Regions with positive weights in the top 5% were defined as positive important regions, whereas regions with negative weights in the top 5% by absolute value were defined as negative important regions. These two sets of regions were used to examine genomic contexts that positively or negatively contributed to model prediction.

Inter-modality interaction analyses

Modal rescue contributions were evaluated separately in cross-validation and independent validation of the Bie et al. dataset. Within each cancer-specific task and evaluation setting, a separate classification threshold was determined for each single-modality model and FAME from the evaluated samples, using the last ROC operating point with specificity of at least 95%. A sample was defined as rescued when FAME classified it correctly but at least one single-modality model did not. For each rescued sample, the positive margin for modality m was calculated as max[0, (2y − 1)(sₘ − tₘ)], where y is the binary class label (1, cancer; 0, healthy), sₘ is the prediction score, and tₘ is the corresponding threshold. These margins were normalized by their sum across the four modalities to obtain fractional contributions. Contributions were summed within each cancer task and across all seven tasks for the “All” group. Records with no positive margin were not assigned to any single modality. The pooled summary counts sample–task records, so healthy participants shared across tasks can contribute more than once.

To assess whether relative single-modality performance varied across cancer types, we performed an AUC-based modality × cancer interaction analysis separately for cross-validation and independent validation. AUCs were arranged in a seven-cancer-by-four-modality matrix. Interaction terms were obtained by subtracting the corresponding row and column means and adding the grand mean; their Frobenius norm was used as the global interaction statistic. Uncertainty was evaluated using 8,000 paired subject-level bootstrap resamples, with cancer cases resampled within each cancer type and shared healthy participants resampled jointly across all tasks. The same resampled subjects were used for all four modalities. Null statistics were calculated from the Frobenius norms of the bootstrap interaction matrices after subtracting the observed interaction matrix. Empirical upper-tail p values were calculated as (k + 1)/(8,000 + 1), where k is the number of null statistics at least as large as the observed statistic. Single-modality AUC rankings were reported descriptively.

Pairwise modality rescue influence was evaluated separately in cross-validation and independent validation after pooling sample–task records across the seven cancer tasks. For each model, a classification threshold was determined from the pooled records using the last ROC operating point with specificity of at least 95%. For each modality pair, we selected records correctly classified by the row modality but misclassified by the column modality. We then subtracted the number of incorrect FAME predictions from the number of correct FAME predictions among these records, divided this difference by the total number of pooled sample–task records, and multiplied by 200 to obtain the signed rescue influence. Positive values indicate an excess of correct FAME predictions in the selected records. Self-comparisons were excluded.

Results

Whole-genome methylation sequencing partially preserves canonical cfDNA fragmentomic features across independent plasma cohorts

To systematically evaluate whether whole-genome methylation sequencing preserves cfDNA fragmentomic information, we compared WGS (as the reference for native cfDNA fragmentation patterns) with WGBS and EM-seq datasets across three independent plasma cohort sets (Fig. 1a). Because matched samples profiled by all three sequencing platforms were not available, we adopted a conservative cross-cohort strategy in which biologically comparable plasma samples from independent datasets were used to assess whether fragmentomic signals observed in WGS can be recapitulated in methylation sequencing data under realistic non-matched conditions.

Fig. 1.

Study design and feature-dependent preservation of cfDNA fragmentomic signatures in whole-genome methylation sequencing. (a) Overview of the study design. Public and inhouse cfDNA datasets were used to evaluate the consistency of fragmentomic profiles across WGS, EM-seq, and WGBS platforms, investigate the biological association and non-redundancy between methylation and fragmentation signals, identify and characterize cfDNA-derived methylation haplotype blocks (MHBs), and develop the FAME framework for improved cancer diagnosis and benchmark validation. (b) Fragment length distributions of plasma cfDNA across WGS, EM-seq, and WGBS datasets. The left panel shows the global fragment length distribution, and the right panel zooms in on the mononucleosomal fragment-size range. All three platforms showed similar cfDNA length profiles, including the major mononucleosomal peak and local periodic patterns. (c) Cross-platform concordance of representative fragmentomic features. Three-dimensional scatter plots compare length, EDM, FSD, and FSR profiles among WGS, EM-seq, and WGBS. Each point represents a matched feature dimension, such as a fragment-length bin, end-motif feature, fragment-size bin, or genomic feature unit. Length, EDM, and FSD showed strong cross-platform concordance, whereas FSR showed weaker agreement, indicating feature-dependent preservation of fragmentomic information. Spearman correlation coefficients are shown in each panel. (d) Distribution of locus-resolved fragmentomic features around transcription start sites (TSSs). Coverage, WPS, OCF, and IFS showed generally consistent spatial patterns across WGS, EM-seq, WGBS, and the BH01 benchmark sample from Snyder et al. (SRR2129993), supporting the retention of functional fragmentomic signals in regulatory regions by whole-genome methylation sequencing

The primary cohort set consisted of healthy plasma samples combining the An et al. WGS datasets [20], our in-house WGBS cohort (HRA021447), and the An et al. EM-seq datasets [20] (Fig. 1b–d). To assess robustness across independent populations, we further analyzed a second healthy cohort set comprising the LUCAS WGS datasets [21], the Gao et al. WGBS datasets [22], and the Bie et al. EM-seq datasets [15]. In addition, a third cohort set focused on gastric cancer plasma samples, including the Cristiano et al. [12] WGS datasets, gastric cancer samples from our in-house WGBS cohort, and the Bie et al. EM-seq datasets, was used to evaluate whether preservation of fragmentomic signals relative to WGS extends to disease contexts (Additional file 2: Fig. S1). Together, these cohort sets enable a robust assessment of fragmentomic preservation across independent biological settings.

In the primary healthy cohort, cfDNA fragment length distributions were highly consistent across WGS, WGBS, and EM-seq, with all platforms recapitulating the canonical mononucleosomal peak observed in WGS (Fig. 1b). At the genome-wide level, key fragmentomic features showed strong concordance with WGS, particularly fragment length distributions, end-motif diversity (EDM) [13], and fragment size distribution (FSD) [44]. While fragment size ratio (FSR) [44] exhibited relatively lower concordance compared to these metrics, it still showed clear cross-platform agreement, indicating that preservation of WGS-derived fragmentomic signals is generally robust but varies in degree across feature classes (Fig. 1c). At local genomic regions surrounding transcription start sites (TSSs), methylation-based profiles also closely resembled WGS-derived patterns, including coverage, IFS (integrated fragmentation score) [41], WPS [10], and OCF [45] (orientation-aware cell-free fragmentation) signals, with consistent PFE [46] (promoter fragmentation entropy) profiles (Fig. 1d; Additional file 2: Fig. S1a).

Consistent patterns were observed in both the independent healthy cohort and the gastric cancer cohort. Although absolute feature values varied across datasets, fragment length distributions remained highly similar across sequencing platforms in both settings (Additional file 2: Fig. S1b, c). Genome-wide comparisons further confirmed robust preservation of fragment length-related features relative to WGS, whereas EDM, FSD, and FSR showed cohort-dependent variability (Additional file 2: Fig. S1d, e). Despite these differences in global metrics, local nucleosome-associated fragmentation patterns around transcription start sites remained highly consistent with WGS-derived profiles across all platforms and cohorts (Additional file 2: Fig. S1f, g).

Because the comparisons used unmatched samples from different cohorts, sequencing-protocol and cohort effects could not be fully separated. Across the three cohort sets, methylation sequencing recapitulated the characteristic cfDNA fragment length distributions and local nucleosome-associated fragmentation patterns observed in WGS, although concordance varied for some genome-wide features.

Exploring the relationship between cfDNA methylation and fragmentation

Having established that methylation-sequencing libraries retain major fragmentomic readouts, we next explored the biological relationship between methylation and fragmentation. To this end, we analyzed the healthy EM-seq and WGBS plasma samples used in the platform-comparison analysis, including 5 EM-seq samples from the An et al. [20] datasets and 9 WGBS samples from the in-house datasets. Stratification of TSSs by methylation level revealed that hypomethylated TSSs were enriched for short fragments (35–80 bp), whereas hypermethylated TSSs were predominantly associated with long fragments (120–180 bp) (Fig. 2a). Independent validation using matched WGS and cfMeDIP-seq datasets corroborated this pattern: cfMeDIP-enriched methylated cfDNA showed longer fragment profiles and distinct end-motif preferences, with enrichment of DNASE1L3-associated CCCA and depletion of DNASE1-preferred TGTG motifs (Fig. 2b; Additional file 2: Fig. S2a).

Fig. 2.

Relationships between cfDNA methylation and fragmentomic features. (a) Length distribution of cfDNA fragments in high- versus low-methylation regions from healthy samples. High-methylation regions were enriched for long fragments (120–180 bp), whereas low-methylation regions favored short fragments (35–80 bp) in both EM-seq and WGBS. (b) Comparison of cfDNA end motif proportions in WGS and cfMeDIP-seq, highlighting several methylation-associated motif shifts. (c) Summary of bidirectional cross-modal regression in EM-seq. at 10,000 boosting trees, the mean fitted R² was 50.0% for fragmentomic features predicting methylation and 42.9% for methylation features predicting fragmentomic features, averaged across five fragmentomic and seven methylation predictors, respectively. The percentages outside the schematic overlap are the corresponding 1-R² values (57.1% for fragmentomics and 50.0% for methylation); the circle areas are illustrative. (d) Model fit in EM-seq as the number of boosting trees increased from 50 to 10,000. For each fragmentomic predictor (WPS, COV, FDI, OCF, and IFS), the upper plot shows fitted-region R² averaged across seven methylation targets, and the lower plot shows the corresponding mean Spearman correlation between the residuals and observed target values. Two-sided Mann–Whitney U tests were used for (a), and two-sided Wilcoxon signed-rank tests for the paired comparison in (b). *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001

To assess how much variation in one modality could be explained by the other, we used gradient-boosting regression to predict methylation features from fragmentomic features and vice versa. Each of five fragmentomic features—IFS, fragment dispersity index (FDI) [43], WPS, OCF, and fragment coverage (COV)—was used separately to predict each of seven metrics describing methylation levels and patterns (see Methods). We then reversed the predictor and target in each pair. At 10,000 boosting trees, the models explained, on average, 50.0% of the variation in methylation features and 42.9% of the variation in fragmentomic features in EM-seq, as measured by fitted R² (Fig. 2c; Additional file 2: Fig. S2b). The corresponding values in WGBS were 50.5% and 48.7% (Additional file 2: Fig. S2c). Explanatory power varied across predictors: the best-performing fragmentomic predictor in EM-seq achieved a mean fitted R² of approximately 0.75 across the seven methylation targets (Fig. 2d). Residuals remained correlated with observed target values (Fig. 2d; Additional file 2: Fig. S2b, d–e). Thus, methylation and fragmentomic features could partially explain each other, but substantial variation in both modalities remained unexplained by these models.

De Novo discovery and biological characterization of plasma cfDNA-derived methylation haplotype blocks

To organize methylation and fragmentomic signals within a common regional framework, we focused on methylation haplotype blocks (MHBs), which capture local co-methylation patterns. We identified and characterized MHBs directly in plasma cfDNA to ground this framework in circulating DNA rather than tissue-derived annotations. We identified 151,473 MHBs directly from plasma cfDNA across seven cancer types and healthy controls (Bie et al. EM-seq datasets [15], see Methods). The number of detected MHBs was not correlated with mhap counts, which roughly scale with sequencing depth, suggesting that block detection was not simply driven by sequencing depth. These cfDNA-derived MHBs had shorter median lengths (30–50 bp) than tissue-derived MHBs, consistent with extensive cfDNA fragmentation (Fig. 3a–b).

Fig. 3.

MBHs identified from cfDNA: biological characteristics and functional insights. (a) Relationship between the number of identified MHBs and mhap counts across sample groups. (b) Length distributions of cfDNA-derived MHBs across sample groups, with median lengths of approximately 30–50 bp. (c) Genomic distribution of MHBs across sample groups. A substantial fraction of MHBs localized to promoter regions in cancer groups, while healthy controls showed a lower proportion. (d) Overlap of cancer-specific MHBs with distinct methylation domains, including unmethylated regions (UMRs), low-methylated regions (LMRs), partially methylated domains (PMDs), and high-methylated regions (HMRs). (e) Proportions of cancer-associated MHBs overlapping tissue-derived ATAC-seq peaks (blue bars) and estimated tumor DNA fractions (red line; mean with 95% confidence intervals) across cancer types. (f) Overlap ranking between cancer-specific cfDNA MHBs and tissue-derived MHBs. (g) Transcription factor motif enrichment in cancer-specific MHBs across seven cancer types. (h) KEGG pathway enrichment analysis of cancer-specific MHBs across seven cancer types, highlighting pathways related to transcriptional regulation, proliferation, and signal transduction

cfDNA MHBs frequently localized to regulatory regions, with up to 50% overlapping promoters in breast and colorectal cancers (Fig. 3c). Cancer-type-specific MHBs preferentially emerged in genomic domains that are normally hypo- or hyper-methylated in the corresponding healthy tissue, and the proportion overlapping tissue-derived open chromatin regions tended to increase with higher estimated tumor DNA fraction (Fig. 3d–e). After excluding shared blocks, cancer-specific MHB sets (1,528–9,679 per cancer) showed striking enrichment for transcription factor motifs characteristic of each malignancy (e.g., GATA3 in breast [47], Snail1 in esophageal [48], ETS factors in pancreatic [49]; Fig. 3f–g) and were functionally linked to proliferation, Wnt/β-catenin, and PI3K/Akt signaling [50] (Fig. 3h; Additional file 2: Fig. S3). Thus, plasma-derived MHBs constitute biologically coherent, tissue-informed units that capture cancer-associated regulatory and epigenomic contexts in circulation.

Reproducibility of cfDNA-derived MHB segmentation

Because MHBs were identified from pooled cancer and control groups, we further evaluated their reproducibility across independent sample groups and cohorts. MHBs re-identified from the Bie et al. independent validation cohort (MHB-TEST) showed high reciprocal overlap with the original training-derived atlas. Without boundary tolerance, 86.0% of original MHB regions matched the MHB-TEST atlas, and 82.8% of MHB-TEST regions matched the original atlas. These rates increased to 88.3% and 84.9%, respectively, at a 2-kb tolerance (Additional file 2: Fig. S4a).

We next assessed cross-cohort concordance using MHBs independently identified from the Nguyen et al. dataset. Despite differences in cohort composition and sequencing depth, a substantial fraction overlapped the Bie et al. training-derived MHB atlas. At a 2-kb tolerance, 79.4% of healthy-derived and 75.2% of breast cancer (BRCA)-derived MHBs from the Nguyen et al. dataset matched this atlas (Additional file 2: Fig. S4b). These results support the reproducibility of plasma cfDNA-derived MHB segmentation across independent sample sets and reduce the concern that MHB identification was driven primarily by cancer-group pooling or cohort-specific bias.

cfDNA-derived MHBs delineate “red” and “blue” genomic domains with distinct epigenetic–fragmentomic behaviors

Because DNA methylation, chromatin accessibility, repeat content, and chromatin compaction are biologically coupled genomic properties, we next asked whether cfDNA-derived MHBs could organize these relationships into interpretable chromatin-context domains. Unsupervised clustering (see Methods) of MHB methylation profiles partitioned the genome into two major classes (Fig. 4a–b): (i) lower-methylation “red” blocks that also showed strong DNase-seq and H3K4me3 signals in immune cells, and (ii) hypermethylated, repeat-enriched “blue” blocks characterized by H3K9me3 and dense short interspersed nuclear element (SINE)/long interspersed nuclear element (LINE)/long terminal repeat (LTR) content (Fig. 4c–h; Additional file 2: Fig. S5) [51–53]. This red–blue dichotomy was recapitulated at TSSs using independent datasets and protocols (Additional file 2: Figs. S6–S7). We therefore use “red” and “blue” as operational labels for MHB-defined genomic contexts, rather than as universal chromatin-state annotations.

Fig. 4.

MHB-defined red and blue domains organize regulatory-open and repeat-enriched cfDNA contexts. (a–b) K-means clustering of methylation feature profiles separated MHB-centered regions into two operational domain classes: low-methylated red domains and high-methylated blue domains. (c) DNase-seq signal distributions in monocytes across red, blue, and heterochromatin (HC) regions. (d) H3K4me3 signal distributions in monocytes across red, blue, and heterochromatin regions. (e) H3K9me3 signal distributions in monocytes across red, blue, and heterochromatin regions. (f–h) Repetitive-element enrichment in red and blue regions defined from all MHBs identified in the Bie et al. dataset, including all eight sample groups. The proportions of blue regions overlapping SINEs (f), LINEs (g), and LTRs (h) are shown, reflecting the enrichment of blue domains in repeat-associated genomic contexts. (i) PDR heatmap across SINE-overlapping blue MHBs, showing higher methylation heterogeneity in cancer samples than in healthy controls. (j) Violin plot of PDR values in SINE-overlapping blue MHBs, showing elevated methylation heterogeneity in cancer samples compared with healthy controls. Two-sided Fisher’s exact tests were used for overlap comparisons (f–h). PDR comparisons in (j) used two-sided Wilcoxon rank-sum tests with Benjamini–Hochberg correction. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001

This organization also helps explain why different cfDNA feature classes show distinct behaviors across genomic contexts. In red domains, high accessibility is expected to expose DNA to nuclease cleavage and to generate stronger local fragmentomic signals; fragmentomic signals of openness (WPS, IFS, FDI) correlated more strongly with accessibility in red than blue regions, whereas cancer-associated methylation heterogeneity (PDR) was more strongly elevated in blue regions, particularly those overlapping repetitive elements, consistent with repeat-associated methylation disruption (Fig. 4i–j; Additional file 2: Figs. S5c–h, S8) [54]. Consequently, red blocks contributed strong methylation-level and chromatin-accessibility signals, while blue blocks provided powerful heterogeneity-based discrimination. This context-dependent complementarity suggests a mechanistic rationale for multimodal integration.

Development and benchmarking of FAME-based models for cancer detection

Guided by the MHB-defined domain framework above, we developed FAME (Fragmentomics And Methylation Ensemble), a two-layer support vector machine (SVM) integrating four complementary MHB-based features: PDR and methylation block score (MBS) as methylation features, and WPS and EDM as fragmentomic features. These features were selected to represent four major classes of MHB-based cfDNA signals: methylation heterogeneity/discordance (PDR), methylation-level-related information (MBS), chromatin-accessibility-related fragmentation (WPS), and fragment-end motif composition (EDM). The feature types were fixed across all cancer-specific binary tasks rather than reselected for each cancer type. Detailed feature-selection rationale and ablation analyses are provided in the Methods section “Selection of cfDNA methylation and fragmentomic features” and Additional file 1: Table S4.

To place FAME in the context of existing cfDNA diagnostic frameworks, we benchmarked it against representative fragmentomics, methylation-fragmentomics, and multimodal methods, including DELFI [12], THEMIS [15], and SPOT-MAS [14] (see Methods). For each method, we followed the original feature design, model architecture, and classifier choice as closely as possible, while evaluating all methods on the same cancer-detection tasks, train/test splits, and metrics. We also constructed FAME-GW, a genome-wide-augmented implementation that combines FAME-derived MHB signals with broader genome-wide methylation and fragmentomic features using a random forest second-layer classifier, to assess whether MHB-centered signals are complemented by genome-wide cfDNA information.

Across the seven cancer-type-specific classification tasks on the Bie et al. dataset, FAME retained competitive performance in both cross-validation and independent validation, with numerically higher AUCs in multiple comparisons with the benchmark methods, while FAME-GW achieved the strongest overall performance across most AUC and sensitivity-at-95%-specificity comparisons (Fig. 5a–f; Additional file 1: Table S5; Additional file 2: Fig. S9). Paired bootstrap analyses further supported statistically significant performance gains of FAME-GW over benchmark methods in multiple cancer-type comparisons, with detailed ΔAUC and ΔSensitivity@95% specificity visualized in Additional file 2: Fig. S10.

Fig. 5.

Benchmarking and diagnostic performance of FAME-based models. (a–d) performance comparison of FAME, FAME-GW, THEMIS, SPOT-MAS, and DELFI across seven cancer-specific binary detection tasks in the Bie et al. cohort. AUCs are shown for ten-fold cross-validation in the Bie et al. training cohort (a) and independent validation in the Bie et al. independent validation cohort (b). Sensitivity at 95% specificity is shown for the same cross-validation (c) and independent validation (d) settings. Error bars in (a, c) indicate variability across cross-validation folds. (e–f) rank-based summary of model performance across validation settings. Models were ranked within each cancer task according to AUC (e) or sensitivity at 95% specificity (f). The ranking analysis includes the seven Bie et al. cross-validation tasks, the seven Bie et al. independent-validation tasks, and two additional independent-validation tasks where available. Darker colors indicate more frequent top-rank performance. Mean ranks are shown above each model. (g–h) ROC curves for the two additional independent-validation cohorts included in the ranking analysis: BRCA detection in the Nguyen et al. independent validation cohort (g) and LIHC detection in the Zhang et al. external validation cohort (h). (i–j) multi-cancer classification performance of FAME-multi in Bie et al. dataset, showing top-1 and top-2 accuracies in cross-validation (i) and independent validation (j)

Evaluation in additional datasets further supported the performance of FAME-based models. In the BRCA independent validation cohort (Nguyen et al. [14]), FAME-GW achieved the highest AUC among the compared methods, while SPOT-MAS also performed strongly (Fig. 5g). In the LIHC external validation cohort (Zhang et al. [24]), FAME-GW remained among the top-performing methods, although SPOT-MAS achieved the highest AUC in this small cohort (Fig. 5h). Together, these results indicate that MHB-centered features provide a compact and interpretable core representation, whereas genome-wide methylation and fragmentomic features offer complementary information that can further improve predictive performance.

Robustness of feature design and stability of MHB-based modeling

We next evaluated the robustness of the FAME feature composition through systematic ablation analyses. Models using only methylation-related features (PDR + MBS) or only fragmentomic features (WPS + EDM) both showed reduced performance compared with the full FAME model, indicating that methylation and fragmentomic signals provide complementary information. Additional feature-replacement analyses further supported the stability of the selected feature set (Additional file 1: Table S4).

To assess potential information leakage arising from MHB construction, we reconstructed the MHB atlas using an independent held-out subset of the HRA003209 cohort (HRA003209-test) and repeated the full cross-validation pipeline. This alternative MHB definition resulted in only marginal changes in AUC and sensitivity at 95% specificity across cancer types (Additional file 1: Tables S6a–S6b; Additional file 2: Fig. S9b), suggesting that model performance is not driven by dataset-specific MHB identification.

We further examined whether the source of MHB definition influences predictive performance by comparing plasma-derived MHBs with tissue-derived MHBs reported by Guo et al. [8] (based on 61 WGBS datasets). For a fair comparison, tissue-derived MHBs were downsampled to match the number of plasma-derived regions and evaluated under the same modeling framework. Plasma-derived MHBs yielded higher AUCs in six of seven cancer-specific cross-validation tasks in the Bie et al. dataset, while their relative performance in independent validation varied across cancer types and cohorts (Additional file 1: Table S7; Additional file 2: Fig. S9c). These results support our choice of plasma-derived MHBs as the regional basis for FAME.

Multi-cancer classification and clinical relevance of FAME scores

We further extended FAME to multi-class cancer classification. Due to limited sample size and unstable performance, the ESCA cohort was excluded from this analysis. The resulting FAME-multi model achieved strong top-1 and top-2 accuracy across six cancer types in both cross-validation and independent validation settings (Fig. 5i–j; Additional file 1: Table S8), supporting the applicability of the MHB-centered framework for pan-cancer classification and tissue-of-origin inference.

Finally, we evaluated whether FAME scores reflect clinically meaningful tumor-associated signals. FAME scores were positively correlated with tumor DNA fraction and increased with tumor stage (Additional file 2: Fig. S11). In contrast, in non-malignant liver disease samples (hepatitis, n = 16; cirrhosis, n = 17), FAME scores did not show cancer-like elevation, suggesting specificity for malignant processes rather than general inflammatory or fibrotic conditions.

Mechanistic interpretability of multimodal integration

To investigate which genomic regions contribute to FAME predictions, we compared model-important regions with early- and late-replicating regions, which generally differ in chromatin accessibility. In the breast cancer analysis of the Bie et al. dataset, the selected important regions with positive SVM weights showed greater overlap with late-replicating regions for PDR and with early-replicating regions for WPS and MBS than those with negative weights (Fig. 6a–c). These findings suggest that the feature modalities draw diagnostic information from distinct genomic contexts. The four single-modality classifiers showed cancer-dependent differences in relative AUC performance, with PDR frequently ranking highest. Modality × cancer interaction tests supported this variation in both cross-validation (p = 0.0051) and independent validation (p = 0.0101; Fig. 6d). These findings suggest that the modalities have distinct diagnostic strengths across cancer types, supporting their complementary use in FAME. Analysis of “rescued” samples, correctly classified by FAME but misclassified by at least one single-modality model, showed substantial contributions from all four modalities—PDR, MBS, WPS, and EDM—in both cross-validation and independent validation (Fig. 6e). Pairwise rescue influence matrices showed positive values across modality pairs in both validation settings, consistent with complementary prediction patterns within FAME (Fig. 6f). Together, these results support the interpretation that FAME benefits from integrating complementary methylation and fragmentomic signals.

Fig. 6.

Mechanistic interpretability of multimodal integration in the FAME ensemble. (a–c) Feature-level contributions from the PDR, MBS, and WPS models using breast cancer (BRCA) as an example. The top 5% most positive and most negative regions from each of these models were identified, and their replication timing profiles were compared. (d) Cancer-dependent variation in single-modality performance in the Bie et al. dataset. for cross-validation (left) and independent validation (right), gray distributions show the bootstrap null distributions of an AUC-based modality × cancer interaction statistic; red lines indicate the observed statistics, and shaded upper tails indicate empirical p values. The tables below summarize the observed AUC rankings of the four modalities by cancer type. (e) Fractionally attributed contributions of PDR, MBS, WPS, and EDM to cases correctly classified by FAME but misclassified by at least one single-modality model. Stacked bars show contributions by cancer type and across all seven cancer tasks (“all”) in cross-validation (left) and independent validation (right). (f) Pairwise modality rescue influence in cross-validation (left) and independent validation (right). Each cell summarizes FAME performance on samples correctly classified by the row modality but misclassified by the column modality. Positive values indicate that FAME corrected more errors than it retained in this subset, supporting the benefit of integrating complementary modality signals. Two-sided Fisher’s exact tests were used for (a–c). Empirical upper-tail p values in (d) were calculated from 8,000 paired bootstrap resamples. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001

Discussion

Whole-genome methylation sequencing was originally developed to profile DNA methylation, and several recent studies have also begun to derive fragmentomic features directly from methylation sequencing libraries for cfDNA-based cancer detection. However, these applications have largely relied on the implicit assumption that methylation sequencing faithfully preserves biologically meaningful fragmentomic information, whereas systematic evaluation of this assumption and its biological basis has been lacking. In this study, we found that whole-genome methylation sequencing partially preserves biologically informative fragmentomic signals across independent plasma cohorts and that these signals can be integrated with methylation information within a unified analytical framework. Our analyses further indicate that DNA methylation and fragmentation are biologically coupled but not redundant. Rather than representing interchangeable measurements of the same underlying process, they appear to capture complementary aspects of chromatin organization, nuclease activity, and epigenetic regulation. Together, these findings provide support for the emerging practice of deriving fragmentomic features from methylation sequencing data and suggest that whole-genome methylation sequencing may be viewed as a source of complementary epigenetic and fragmentomic information.

An important aspect of this study is the systematic characterization of plasma-derived methylation haplotype blocks (MHBs) and their utility as biologically interpretable units for organizing multimodal cfDNA signals. We found that plasma-derived MHBs are enriched for tissue-specific regulatory elements, chromatin accessibility, repeat-associated regions, and cancer-associated methylation heterogeneity, consistent with their biological relevance in circulating DNA. Furthermore, the identification of two MHB contexts with distinct epigenetic and fragmentomic characteristics offers a biological perspective for interpreting why different classes of cfDNA features contribute differently across genomic regions. Chromatin accessibility can vary across cell types even at sites with constitutively low methylation [55]. Accordingly, the red and blue domains are operational labels rather than universal chromatin-state annotations; they may nevertheless provide biologically meaningful genomic contexts for organizing complementary methylation and fragmentomic information.

FAME integrates four complementary methylation and fragmentomic feature classes within plasma-derived MHBs, providing a biologically interpretable framework for cfDNA-based cancer detection. Feature ablation and rescue analyses support complementary contributions from all four modalities, while the improved performance of FAME-GW suggests additional value from genome-wide features. Both models obtain methylation and fragmentomic information from a single whole-genome methylation sequencing assay, without requiring separate WGS-based fragmentation profiling. The MHB-centered framework may also provide interpretable regional units for future representation-learning approaches.

Several limitations should be acknowledged. First, our cross-platform comparisons used biologically comparable cohorts rather than matched plasma aliquots profiled by WGS, WGBS, and EM-seq, so protocol-specific effects cannot be fully separated from inter-cohort biological variation. Given the scarcity of matched multi-platform cfDNA datasets, cross-cohort comparisons remain a practical approach. Our supplementary comparison of the in-house single-stranded WGBS cohort with an independent WGBS cohort generated using a different library-preparation strategy showed generally consistent fragment-length distributions and major length-related patterns, despite variation in individual features (Additional file 2: Fig. S12). Major fragmentomic patterns were also consistently observed across three cohort sets spanning healthy individuals and cancer patients. Matched-sample studies will nevertheless be needed to quantify protocol-specific biases and establish feature-specific technical robustness. Second, our same-cohort benchmarks of SPOT-MAS, THEMIS, and DELFI followed published descriptions as closely as possible, but differences in preprocessing, feature definitions, and available implementation details preclude strict reproduction of the original pipelines. Performance differences should therefore be interpreted in the context of these implementations. Third, although FAME performed well in independent and external validation cohorts, larger prospective studies with more early-stage cancers, broader benign disease controls, and longitudinal samples are needed to assess its clinical utility for screening and monitoring. Finally, future work could extend the current stacking framework through explicit MHB-domain priors, further genome-wide features, or more advanced learning architectures while preserving biological interpretability.

Conclusions

Overall, our results suggest that whole-genome methylation sequencing may serve as a practical source of complementary epigenetic and fragmentomic information in addition to its conventional role in methylation profiling. By systematically evaluating the preservation of fragmentomic information in methylation sequencing data and characterizing plasma-derived MHBs as biologically interpretable genomic contexts, this study provides additional insight into the relationship between methylation and fragmentation. These findings may contribute to the development of more interpretable, efficient, and clinically scalable multimodal liquid-biopsy strategies for cancer detection.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Abbreviations

ATAC-seq

Assay for transposase-accessible chromatin using sequencing

AUC

Area under the receiver operating characteristic curve

CAFF

Chromosomal aneuploidy of featured fragments

cfDNA

Cell-free DNA

cfMeDIP-seq

Cell-free methylated DNA immunoprecipitation sequencing

CHALM

Cellular heterogeneity-adjusted clonal methylation

ChIP-seq

Chromatin immunoprecipitation sequencing

CI

Confidence interval

CNV

Copy number variation

COV

Fragment coverage

DELFI

DNA evaluation of fragments for early interception

EDM

End motif diversity

EM

End motif

EM-seq

Enzymatic methyl sequencing

ENTROPY

Methylation entropy

FAME

Fragmentomics And Methylation Ensemble

FAME-GW

Genome-wide-augmented FAME

FAME-multi

Multi-class FAME

FDI

Fragment dispersity index

FDR

False discovery rate.

FEM

Fragment end motif

FSD

Fragment size distribution

FSI

Fragment size index

FSR

Fragment size ratio

GO

Gene Ontology

GWM

Genome-wide methylation

HC

Heterochromatin

HMR

High-methylated region

IFS

Integrated fragmentation score

KEGG

Kyoto Encyclopedia of Genes and Genomes

KNN

k-nearest neighbors

LC

Lung cancer

LINE

Long interspersed nuclear element

LMR

Low-methylated region

LTR

Long terminal repeat

MBS

Methylation block score

MCR

Methylation concurrence ratio

MFR

Methylated fragment ratio

MHB

Methylation haplotype block

MHL

Methylated haplotype load

MM

Mean methylation level.

NGDC

National Genomics Data Center

NSCLC

Non-small cell lung cancer

OCF

Orientation-aware cell-free DNA fragmentation

OR

odds ratio

PDR

Proportion of discordant reads

PFE

Promoter fragmentation entropy

PMD

Partially methylated domain

QC

Quality control

ROC

Receiver operating characteristic

SINE

Short interspersed nuclear element

SPOT-MAS

Screening for the presence of tumor by methylation and size

SVM

Support vector machine

THEMIS

THorough Epigenetic Marker Integration Solution

TSS

Transcription start site

UMR

Unmethylated region

WGBS

Whole-genome bisulfite sequencing

WGS

Whole-genome sequencing

WPS

Windowed protection score

Author contributions

X.Z., W.Z., and Q.L. conceived and administratively supported this study. Y.W., Y.H. and X.Z. developed the methods. Y.W. and Y.H. performed the data analysis. X.L. performed additional data collection and analysis during the revision. Y.W. wrote the manuscript. All authors have read and approved the final manuscript for publication.

Funding

This work was supported by the Fundamental Research Funds for the Central Universities (2662025PY024) and the project “Development of an Analysis Software System Based on Third-Generation Sequencing” to X.Z., Huazhong Agricultural University Scientific & Technological Self-innovation Foundation and Fundamental Research Funds for the Central Universities (2662024SZ006) to W.Z. The funders have no role in study design, data collection, data analysis, data interpretation, or writing of the manuscript.

Data availability

The inhouse data generated in this study have been deposited in the National Genomics Data Center (NGDC) under accession HRA021447 [56]. Because these data were generated from human samples, access is currently managed through the NGDC data-access application system. Anyone interested in using the data may submit an access request through NGDC. Requests for non-commercial research use will be reviewed in accordance with NGDC and institutional data-access policies, and an initial response is generally expected within 15 business days. The other cfDNA data utilized in this study have been sourced from previously published papers (Table S1a, b). The code generated by this study is available at https://github.com/alcindor819/Methylation_Fragmentomic [57]. Any additional information needed to re-analyze the data reported in this paper is available from the lead contact upon request.

Declarations

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the authors used ChatGPT (OpenAI) in order to assist with language refinement. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.

Ethics approval and consent to participate

This study was conducted in accordance with the principles of the Declaration of Helsinki. For the in-house cohort used in this study, healthy control samples were collected from The Eighth Affiliated Hospital of Sun Yat-sen University, and gastric cancer samples were collected from Fujian Medical University Union Hospital. The collection and use of healthy control samples were approved by the Medical Scientific Research Ethics Committee of The Eighth Affiliated Hospital of Sun Yat-sen University (Approval No. ZBKY-2020–013-01). The collection and use of gastric cancer samples were approved by the Medical Scientific Research Ethics Committee of Fujian Medical University Union Hospital (Approval No. 2022KY118). Written informed consent was obtained from all participants prior to sample collection. All samples and data were de-identified before analysis to protect participant confidentiality.

Consent for publication

Not applicable. This manuscript does not contain any individual person’s identifiable data.

Competing interests

The authors declare no competing interests.