CD Genomics provides epigenomics services for research use only and they are not intended for clinical diagnosis or treatment. Liquid biopsy has moved beyond mutation panels. The most informative biomarker programs now extract multiple layers of epigenomic information from a single plasma sample — DNA methylation, cfDNA fragmentation patterns, and chromatin-level signals such as nucleosome positioning. Each modality captures a different dimension of tissue biology, and when they are integrated, the composite signal outperforms any single data type alone.

This article is for teams designing cfDNA-based biomarker studies who need to decide which modalities to measure, how to integrate them, and what study architecture supports translation from discovery to a validated panel. It does not replace a protocol — it provides the design logic that makes protocols produce interpretable results.

Figure 1: Three epigenomic modalities accessible from cfDNA — methylation, fragmentation, and chromatin signals — each reflecting a different dimension of cell-of-origin biology.

The Three Epigenomic Modalities in cfDNA

Cell-free DNA in plasma is not random degradation debris. It is chromatin that has been processed by nucleases during cell death, and its fragmentation pattern, methylation status, and nucleosome positioning all reflect the epigenomic state of the cells from which it originated. Three modalities are accessible from cfDNA:

DNA methylation is the most extensively validated cfDNA epigenomic marker. Tissue-specific differentially methylated regions (DMRs) mark cell identity, and aberrant methylation at promoter CpG islands is a hallmark of many cancers. Unlike somatic mutations, which are sparse and heterogeneous, methylation changes are abundant — thousands of CpG sites can be informative in a single sample. This density makes methylation the highest-sensitivity single modality for early detection, but methylation alone can be confounded by clonal hematopoiesis, aging-related drift, and technical variability in bisulfite conversion.

cfDNA fragmentation (fragmentomics) describes the non-random size distribution, end-motif frequencies, and genomic positioning of cfDNA fragments. Healthy cfDNA is dominated by mononucleosomal fragments of approximately 167 bp with a characteristic 10-bp periodicity reflecting the nucleosome core particle. In cancer, cfDNA fragments are shorter on average, and the distribution of fragment endpoints shifts in ways that correlate with chromatin accessibility and transcription factor occupancy in the tumor of origin. Fragmentation is a passive readout — it does not require enrichment or chemical conversion — and can be extracted from any genome-wide sequencing dataset, including those generated primarily for methylation or copy-number analysis.

Chromatin signals — nucleosome positioning, inferred histone modification landscapes, and transcription factor footprints — are the third layer. Deep sequencing of cfDNA reveals nucleosome occupancy patterns that mirror the chromatin architecture of living cells. At actively transcribed genes, nucleosome-depleted regions appear at transcription start sites. At enhancers, nucleosome positioning reflects tissue-specific regulatory activity. These signals were first described by Snyder et al. in 2016 and have since been extended to cancer classification, tissue-of-origin deconvolution, and treatment response monitoring.

Why Single-Modality Analysis Falls Short

Each modality, used alone, has a performance ceiling determined by biology and technology.

Methylation alone can detect cancer signals at low tumor fractions — cfMeDIP-seq has demonstrated detection from as little as 1 ng of cfDNA — but methylation-based classifiers can flag benign conditions that share methylation features with malignancy, such as chronic inflammation or aging-related epigenetic drift. Specificity at very high sensitivity remains a challenge.

Fragmentation alone is appealing because it is essentially free — it can be extracted from any sequencing dataset — but its standalone performance is modest. In the EMMA study (Liu et al., 2024), fragmentation features alone achieved an AUC of approximately 0.54 for esophageal cancer detection. The signal is real, but it is subtle and easily drowned out by technical variation in library preparation and sequencing.

Chromatin signals require deep sequencing to resolve nucleosome-level features, and their interpretation depends on reference epigenomic datasets that may not exist for the tissue or condition under study. They are powerful for tissue-of-origin inference but less sensitive than methylation for initial detection.

The core argument for multi-modal integration is that these modalities fail in different ways. Methylation provides sensitivity; fragmentation and copy-number features provide specificity. Chromatin signals provide biological context that helps distinguish true positives from confounding signals. When combined, the composite classifier outperforms any individual input.

Single-Modality vs Multi-Modal Study Design

A single-modality study measures one data type and builds a classifier from it. A multi-modal study extracts two or three feature types from the same or paired sequencing libraries and integrates them before classification.

The practical difference is not just statistical — it affects sample requirements, sequencing strategy, bioinformatic infrastructure, and validation design.

Figure 2: Single-modality analysis extracts one data type from cfDNA, while multi-modal analysis extracts methylation, fragmentation, and chromatin features from a single sequencing library.

Key Technologies for Multi-Modal cfDNA Profiling

A single genome-wide sequencing library prepared with a methylation-compatible protocol can support extraction of methylation, fragmentation, and copy-number features simultaneously. The following technologies are most commonly used in multi-modal studies:

Whole-genome bisulfite sequencing (WGBS) remains the reference standard for single-base methylome coverage. For cfDNA, the harsh chemical conditions are a concern — bisulfite treatment fragments already-degraded cfDNA further — but the simultaneous readout of methylation status, fragment endpoints, and read-depth-based copy number makes WGBS the most proven single-assay platform for multi-modal extraction. The EMMA framework demonstrated that methylation calls, fragment size ratios, and copy-number variants can all be extracted from a single WGBS library, and their integration raised esophageal cancer detection AUC from 0.90 (methylation alone) to 0.99.

Enzymatic methyl-seq (EM-seq) uses TET2 and APOBEC3A enzymes instead of bisulfite, avoiding the DNA damage that limits WGBS library complexity from cfDNA. EM-seq preserves longer fragments, produces more uniform coverage, and requires as little as 100 pg of input DNA. The same multi-modal features extractable from WGBS — methylation, fragmentation, and CNV — are accessible from EM-seq data, making it an increasingly preferred replacement for cfDNA multi-modal studies.

cfMeDIP-seq enriches methylated DNA by immunoprecipitation with an anti-5mC antibody, followed by sequencing. It requires only 1–10 ng of cfDNA and avoids chemical conversion entirely. The trade-off is resolution: cfMeDIP-seq reports methylation at the level of 100–200 bp regions rather than individual CpG sites. Fragmentation features can still be extracted, but single-base methylation resolution and precise fragment-end analysis are not possible.

Cell-free chromatin immunoprecipitation sequencing (cfChIP-seq) directly profiles histone modifications in circulating nucleosomes. This is the most direct chromatin-level modality and can reveal active promoter marks (H3K4me3), enhancer marks (H3K4me1), and repressive marks (H3K27me3) in circulating chromatin. Integration of cfChIP-seq with methylation and fragmentation data is an emerging frontier.

For teams planning multi-modal cfDNA analysis, cfDNA epigenetic subtyping solutions at CD Genomics support assay selection and study design. Cell-free methylation sequencing provides the methylation backbone, and epigenomic data analysis supports the multi-pipeline bioinformatic workflow required to extract and integrate all modalities.

Designing a Multi-Modal cfDNA Study

Multi-modal studies amplify both signal and complexity. A design that does not account for the interaction between modalities will produce an ensemble classifier that performs well in cross-validation on the discovery cohort and disappoints in an independent validation set. The recommendations below address the most common failure modes.

Sample type and collection. Plasma is the standard input for cfDNA studies. Serum contains higher background from leukocyte lysis during clotting and is generally not recommended. Collection in EDTA or specialized cfDNA-stabilizing tubes (e.g., Streck, CellSave) with processing within 2–4 hours for EDTA minimizes genomic DNA contamination. For detailed guidance on plasma versus serum and collection tube selection, see the cfDNA sample type guide.

Cohort design. Multi-modal studies need larger cohorts than single-modality studies — not because the effect size is smaller, but because the feature space is larger and overfitting risk scales with the number of features considered. As a rough guide: for a binary classifier with features drawn from three modalities, plan for at least 50–100 samples per group in discovery, with an independent validation cohort of comparable size. Case-control matching on age, sex, and collection site is essential — methylation in particular is sensitive to these variables.

Sequencing depth. Modalities differ in their depth requirements. Methylation calling at single-base resolution may require 10–30× coverage; fragmentation analysis is robust at 1–5× because aggregate fragment-level statistics stabilize with large numbers of fragments even at low coverage; nucleosome positioning resolution demands deeper sequencing (30–50×) to resolve occupancy at individual regulatory elements. If chromatin-level features are a priority, budget for deeper sequencing or restrict analysis to genome-wide aggregate statistics.

Replication strategy. Multi-modal classifiers can exploit chance associations between modalities that do not replicate. The strongest safeguard is to hold out an independent sample set — collected at a different site or time, if possible — and to evaluate not just the final ensemble score but the contribution of each modality independently. A classifier in which methylation alone achieves AUC 0.92 and the addition of fragmentation and chromatin features adds 0.02 is still a methylation classifier; the multi-modal framing is only justified when each modality independently contributes at a prespecified threshold.

Controls. Include technical control samples (sheared genomic DNA, or cfDNA from a well-characterized healthy pool) processed alongside study samples in every batch. These controls detect batch effects in fragmentation that would otherwise be misinterpreted as biological signal.

Data Integration Strategies

Integrating three data types from the same sample is a bioinformatic challenge as much as a statistical one. The following workflow is typical:

Feature extraction. Methylation features include per-CpG beta values, DMR calls, and methylation haplotype blocks. Fragmentation features include fragment size distribution statistics (mean, median, short-to-long ratio), end-motif frequencies (4-mer or 6-mer), and windowed protection scores (WPS) that reflect nucleosome occupancy. Chromatin features include nucleosome-depleted region scores at transcription start sites and inferred tissue-of-origin proportions from reference deconvolution algorithms.

Within-modality modeling. Build a base classifier for each modality independently before integration. This serves two purposes: it establishes the performance floor, and it reveals whether each modality contains usable signal. A modality that performs at chance level alone is unlikely to contribute productively to an ensemble.

Integration architecture. Common approaches include early fusion (concatenating all features into a single matrix before modeling), late fusion (training separate models for each modality and combining their scores), and intermediate fusion (learning modality-specific representations before joint modeling). Late fusion is the most interpretable and the easiest to validate independently, but intermediate fusion with neural networks can capture non-linear interactions between modalities.

Cross-validation discipline. Multi-modal studies are vulnerable to information leakage between modalities during feature selection. If methylation features are selected using labels, fragmentation features are selected using labels, and then the two feature sets are combined and modeled, the cross-validation estimate is optimistically biased. The safest approach is to define the feature extraction pipeline for each modality on a training subset only, then apply frozen pipelines to the validation set.

Interpretability. A multi-modal classifier that produces a score without explaining which modality contributed what is difficult to validate, publish, or translate. At minimum, report per-modality performance, feature importance rankings, and examples of individual samples where different modalities drive the final call.

Figure 3: A multi-modal data integration workflow — feature extraction per modality, within-modality modeling, integration architecture selection, and independent validation.

From Discovery to a Validated Biomarker Panel

The path from a multi-modal discovery study to a validated biomarker panel involves three phases:

Discovery phase. Train and internally validate a multi-modal classifier on a well-annotated retrospective cohort. The output is a locked model with defined feature sets and decision thresholds.

Technical validation. Reproduce the classifier performance on an independent sample set processed in a different batch or at a different site. This confirms that the signal is not driven by batch effects, collection artifacts, or site-specific confounders. If performance degrades substantially, examine individual modality contributions — one modality may be driving the degradation — and consider whether the affected modality should be dropped or its pipeline hardened.

Biological validation. Demonstrate that the composite signal reflects the biology it claims to reflect. For cancer detection, this means showing that the classifier score correlates with tumor fraction, that tissue-of-origin calls match the clinical diagnosis, and that the signal disappears after curative treatment. For other applications such as cfDNA cell-of-origin analysis, orthogonal confirmation with independent tissue-specific markers provides supporting evidence.

Throughout this process, targeted EM-seq can validate specific methylation markers identified in the discovery phase, and cfDNA epigenetic subtyping supports the translation of multi-modal signatures into practical assays.

Summary

Multi-modal cfDNA epigenomics recognizes that methylation, fragmentation, and chromatin signals are not independent measurements — they are different views of the same underlying biology. A fragmented, hypomethylated, nucleosome-depleted region at a tumor suppressor promoter tells a coherent story that any single modality tells only partially.

The practical recommendations are:

• Use a single genome-wide methylation-compatible assay (WGBS or EM-seq) to extract methylation, fragmentation, and CNV features simultaneously — additional modalities come computationally, not from additional sample volume
• Build and evaluate per-modality classifiers before integrating — a modality with no standalone signal adds noise, not information
• Budget larger cohorts for multi-modal studies to control overfitting in the expanded feature space
• Hold out independent validation samples and test not just the ensemble score but each modality's independent contribution
• Lock the feature extraction and modeling pipeline before validation — information leakage between modalities during feature selection inflates performance estimates

For teams planning multi-modal cfDNA studies, CD Genomics provides end-to-end support from sample collection guidance through cfDNA sample type selection and epigenomics sequencing services to multi-pipeline bioinformatic analysis and biomarker panel development.

Frequently Asked Questions

1. Can fragmentation and chromatin features be extracted from any cfDNA sequencing dataset?

Fragmentation features — fragment size distribution, end motifs, and coverage-based nucleosome occupancy — can be extracted from any genome-wide cfDNA sequencing dataset, including those generated for methylation or copy-number analysis, provided paired-end sequencing was used and fragment lengths can be inferred. Chromatin-level features such as nucleosome-depleted regions and transcription factor footprints require deeper sequencing (typically 30–50×) and are most informative when paired-end reads preserve fragment coordinates.

2. How much additional sample is needed for multi-modal versus single-modality cfDNA analysis?

With current methods, no additional sample is needed. A single whole-genome methylation-compatible library (WGBS or EM-seq) generates data from which methylation, fragmentation, and copy-number features can all be extracted computationally. The additional cost is bioinformatic, not sample-related.

3. Does multi-modal integration improve early-stage detection specifically?

Yes — and this is where multi-modal approaches show their greatest value. Methylation signals can be sparse in early-stage disease due to low tumor fraction. Fragmentation features, while modest in standalone performance, add orthogonal information that is not correlated with methylation signal strength, improving sensitivity at the low-tumor-fraction end. In the EMMA study, integration raised stage I detection sensitivity from 70% to 78% while maintaining specificity above 95%.

4. What is the minimum cohort size for a multi-modal cfDNA biomarker study?

A reasonable minimum for discovery is 50–100 samples per group, with an independent validation cohort of comparable size. Smaller cohorts can be used for exploratory analysis, but the expanded feature space of multi-modal data inflates overfitting risk, and results from small cohorts should be treated as hypothesis-generating rather than conclusive.

5. Which modalities should I prioritize if I can only afford limited sequencing?

Prioritize methylation as the primary detection modality — it provides the highest sensitivity and the most extensive validation literature. Fragmentation features are extracted from the same data at no additional sequencing cost and should be included as a secondary modality. Chromatin-level features requiring deep coverage can be deferred unless tissue-of-origin inference is a primary study objective.

References

1. Snyder, Matthew W., Martin Kircher, Andrew J. Hill, Riza M. Daza, and Jay Shendure. "Cell-free DNA Comprises an In Vivo Nucleosome Footprint that Informs Its Tissues-Of-Origin." Cell, vol. 164, no. 1–2, 2016, pp. 57–68. DOI: 10.1016/j.cell.2015.11.050
2. Shen, Shu Yi, Rajat Singhania, Gordon Fehringer, Ankur Chakravarthy, Michael H. A. Roehrl, et al. "Sensitive tumour detection and classification using plasma cell-free DNA methylomes." Nature, vol. 563, 2018, pp. 579–583. DOI: 10.1038/s41586-018-0703-0
3. Cristiano, Stephen, Alessandro Leal, Jillian Phallen, Jacob Fiksel, Vilmos Adleff, et al. "Genome-wide cell-free DNA fragmentation in patients with cancer." Nature, vol. 570, 2019, pp. 385–389. DOI: 10.1038/s41586-019-1272-6
4. Liu, Jiaqi, Lijun Dai, Qiang Wang, Chenghao Li, Zhichao Liu, et al. "Multimodal analysis of cfDNA methylomes for early detecting esophageal squamous cell carcinoma and precancerous lesions." Nature Communications, vol. 15, 2024. DOI: 10.1038/s41467-024-47886-1

CD Genomics provides epigenomics services for research use only and they are not intended for clinical diagnosis or treatment.