Cell-free nucleic acid fragmentomics: A non-invasive window into cellular epigenomes

Highlights

• •Cell-free nucleic acids (cfNA) have distinct characteristics in health and diseases.
• •Machine learning can identify cfNA fragment properties with clinical significance.
• •Fragmentomics can complement mutation-based disease detection and monitoring.

Abstract

Clinical genomic profiling of cell-free nucleic acids (e.g. cell-free DNA or cfDNA) from blood and other body fluids has ushered in a new era in non-invasive diagnostics and treatment monitoring strategies for health conditions and diseases such as cancer. Genomic analysis of cfDNAs not only identifies disease-associated mutations, but emerging findings suggest that structural, topological, and fragmentation characteristics of cfDNAs reveal crucial information about the location of source tissues, their epigenomes, and other clinically relevant characteristics, leading to the burgeoning field of fragmentomics. The field has seen rapid developments in computational and genomics methodologies for conducting large-scale studies on health conditions and diseases – that have led to fundamental, mechanistic discoveries as well as translational applications. Several recent studies have shown the clinical utilities of the cfDNA fragmentomics technique which has the potential to be effective for early disease diagnosis, determining treatment outcomes, and risk-free continuous patient monitoring in a non-invasive manner. In this article, we outline recent developments in computational genomic methodologies and analysis strategies, as well as the emerging insights from cfNA fragmentomics. We conclude by highlighting the current challenges and opportunities.

Graphical abstract

Introduction

The year was 1948, and the wounds of WWII could be felt around the corners of devastated mainland Europe. It was around that time that two French scientists, Mandel and Metais, made the initial discovery of DNA fragments, now known as cell-free DNA or cfDNA, in human blood plasma [1]. Typically, circulating cell-free nucleic acids (cfNAs) are fragments of DNA or RNA that are released in the body fluids by normal and diseased cells. In the 1960s a couple of studies on cfNAs postulated and demonstrated the significance of circulating DNA in illnesses such as cancer and systemic lupus erythematosus [2,3]. Later in 1977, Leon et al. revealed that many cancer patients had high cfDNA levels. The correlation between cfDNA levels and disease burden showed that at least some of this DNA is tumor-derived [4]. Subsequently, the landmark finding of fetal DNA in pregnant women by Denis Lo et al. showcased the translational potential of cfNA for non-invasive clinical applications [5]. More recent studies have found cell-free nucleic acids in various body fluids (blood, urine, cerebral spinal fluid(CSF), pleural fluid, and saliva) of healthy and diseased individuals [[6], [7], [8], [9]], indicating possibilities for early diagnosis and therapies [[9], [10], [11]].

With the advances in high throughput sequencing technologies over the last decade, cfDNA-based genomic applications have emerged as vital tools for precision medicine in cancer [[12], [13], [14], [15]]. The sequence content and fragmentation characteristics of ctDNA carry signatures of genetic and epigenetic changes from its cell of origin, enabling non-invasive detection of disease-related biomarkers for clinical management [16,17]. While efforts to identify disease-associated genetic changes such as single-nucleotide variants (SNVs) and somatic copy-number aberrations (SCNAs), and also epigenetic alterations such as methylation patterns have been established to stratify cancer patients, assess therapy efficacy, and identify treatment resistance non-invasively in clinical settings [[18], [19], [20], [21], [22], [23], [24], [25]], analyzing fragmentation characteristics of cfDNA to inform about disease associated characteristics have led to the emergence of a new field called ‘Fragmentomics’. Ivanov et al. first introduced the term to describe the investigation of cfDNA fragment size patterns [26], and since then a number of studies have identified multimodal insights from cfDNA fragmentomic signatures. In this review, we discuss the recent developments in fragmentomics research, as well as the present success and boundaries of cfNA for early cancer detection. We also emphasize the potential challenges and opportunities to translate the discoveries for early detection, diagnosis, and monitoring the patients.

The biology of cell-free nucleic acids

cfDNA in blood and other body fluids can exist in multiple forms: freely circulating, bound to proteins and protein-complexes such as nucleosomes, Argonaute, and lipoproteins like HDL and LDL, or encapsulated within or attached to extracellular vesicles, including exosomes, microvesicles, and apoptotic bodies [27]. The mechanisms behind cfDNA release include cell death—through apoptosis, necrosis, and NETosis—and active secretion [28]. The mechanisms of cfNA genesis and various forms of cfNA in circulation are illustrated in Fig. 1. The contribution of these mechanisms to the circulating cfDNA pool likely varies across different physiological and pathological states [29]. cfDNA in plasma predominately originates from hematopoietic cells, with minor contributions from hepatocytes, at least in the apparently healthy individuals [30], while in specific disease contexts, other tissue types may have additional contributions [31]. Normal plasma concentrations of cfDNA can vary widely, typically ranging from 65 to 877 ng/ml, and are much lower in urine, ranging from 0 to 215 ng/ml [32,33]. In contrast, cfDNA levels can soar to over 1000 ng/ml in the serum of cancer patients, highlighting its potential as a biomarker [34]. Notably, not all fragments of cell-free DNA in body fluids are necessarily of the host origin; cell-free DNA (or cfNAs in general) in plasma and other body fluids may also originate from the human microbiome, with predominant contributions from bacterial sources [35]. cfNAs in the body fluids are highly fragmented. cfDNA in plasma originating from the nuclear genome typically have fragment sizes ranging between 80 and 200 bp [36], while mitochondrial DNA fragments typically have sizes from 30 to 80 bps while peaking between 42 and 60 bp [27]. However, DNA fragments <50 bp and greater than 10,000 bp can also be found in blood [27]. The structural integrity of cfDNA depends on the different proteins that bind with and protect the DNA from rapid degradation by nuclease enzymes in body fluids, which ultimately define the size profile of these circulating DNA fragments. Circulating nucleosomes, which consist of double-stranded DNA fragments approximately 180–200 base pairs long wrapped around an octameric histone protein complex, exemplify this phenomenon. These nucleosomes are interconnected by DNA links ranging from 20 to 90 base pairs, while the DNA segment directly associated with the histone octamer averages 147 base pairs [37]. cfDNA in the plasma of a healthy individual typically presents a modal size of about 166 base pairs, indicative of its relation to the nucleosomal structure [38]. Nucleosome-bound DNA is naturally resistant to nuclease digestion, giving rise to the cfDNA population with size profiles predominantly corresponding to mono- and oligo-nucleosome-bound DNA fragments [39]. Another characteristic of cfDNA is a series of peaks approximately 10 base pairs apart starting from around 143 bp in the size distribution profile, suggesting protection provided by the grooves of a DNA double helix [26]. The cfDNA fragments in urine also show successive peaks at 10 base pair intervals [10]. However, the mitochondrial cfDNA fragments do not exhibit the 10-bp periodicity [40]. Additionally, extrachromosomal circular DNA (eccDNA) found in human plasma is notably resistant to exonuclease activities, further diversifying the forms of circulating nucleic acids [41]. Furthermore, the size profile of cfDNA in urine exhibits a mode around 82 bp, potentially due to heightened nuclease activity [10].

Fig. 1.

The origin and forms of cell-free nucleic acids (cfNAs). Billions of cells in our body die every day due to apoptosis, necrosis, or NETosis, during which the cells release the nucleic acid contents into the body fluids. Cells can also actively secrete nucleic acids, and microbial cfNA can also be mixed in the human cfNA pool. cfDNA can be in different forms, for cfDNA, it can be bound with nucleosome or other protein/lipid moieties, whereas for cell-free RNA (cfRNA), they mostly remain inside vesicular structures.

Similar to the fragmentation mechanism, degradation of cfDNA appears to be a non-random process, with certain genomic regions more likely to be cleaved and present at plasma DNA fragment ends, reflecting tissue-specific “preferred end sites” [42]. The activity of deoxyribonucleases (DNases) plays a critical role in shaping the characteristic length distribution, typical end motif, and levels of cfDNA in both healthy and disease states. The DNA fragmentation factor B (DFFB), also known as caspase-activated DNase (CAD) cleaves DNA at internucleosomal linker regions, initially into high molecular weight fragments (50–300 kb) and subsequently into oligonucleosomal fragments, typically multiples of 180 bp. This cleavage kinetics is like those observed with naked DNA [43]. The A-end preference of cfDNA degradation is characteristic of this enzyme, illustrating its specific cleavage pattern [44]. DNASE1 and DNASE1L3, two other significant DNases can degrade double-stranded DNA (dsDNA), single-stranded DNA (ssDNA), and chromatin. DNASE1 is particularly efficient at cleaving dsDNA, much more so than ssDNA, and it preferentially targets naked DNA [45], producing subnucleosomally sized cfDNA fragments mostly having T-ends [44]. DNASE1L3 and DFFB play a crucial role in generating the typical modal size of 166 bp observed in plasma cfDNA [46]. DNASE1L3 can cleave both nucleosome-bound and naked DNA, predominantly producing C-end fragments [44]. The presence of DNASE1L3 also reduces the jaggedness of cfDNA fragments, whereas DNASE1, found in higher concentrations in urine, increases the jaggedness [44]. According to the fragmentation model of cfDNA, DFFB and DNASE1L3 initiate cfDNA fragmentation during the cell death process within cells, and once released into the plasma, DNASE1L3 and DNASE1 further process the cfDNA, contributing to the characteristic cfDNA profiles observed in different physiological fluids and pathological conditions [44].

The overall dynamics of cfDNA in blood depends on the generation, protection, and clearance processes. The clearance of cfDNA from the circulation is a complex process involving various mechanisms namely, active uptake by the reticuloendothelial system primarily in the liver (70–90 %) and spleen (3 %), passive renal filtration (4 %), and direct enzymatic degradation [47,48]. These cfNA clearance pathways are illustrated in Fig. 2. Rapid clearance of cfNA is essential for maintaining homeostasis, with its half-life estimated between 16 min and 2 h [49]. Additionally, a minor portion of fetal cfDNA clearance in pregnancy occurs through transrenal pathways, accounting for 0.2–19 % of the total [49]. Moreover, the clearance and accessibility of cfDNA are influenced by plasma proteins distinct from nucleases. Proteins such as factor VII-activating protease (FSAP) and Complement Factor H facilitate the release and processing of cfDNA by increasing its availability to nucleases and aiding in its uptake by Kupffer cells in the liver [50,51]. Additionally, the dynamics of cfDNA are affected by its encapsulation in apoptotic bodies or other vesicular structures, which impacts its half-life, exposure to nucleases, and fragmentation patterns [52]. At least a proportion of cfDNA is encapsulated in extracellular vesicles. Within the extracellular vesicles (EVs), nucleic acids are safeguarded against degradation [53]. Virtosomes—complexes comprising DNA, RNA, and lipoproteins—shield nucleic acids from nuclease degradation [54]. Similarly, DNA-binding proteins, including transcription factors, provide protection, though regions of open chromatin remain susceptible to enzymatic cleavage [37]. Such complexity underscores the intricate balance and interaction of structural, protective, and degradative mechanisms governing the presence and persistence of cfDNA in the circulatory system.

Fig. 2.

The clearance mechanisms of circulating cfNA released by different host tissues, microbes, and fetus in the case of a pregnant woman. The degradation of cfNA by nucleases depends on the accessibility of the enzymes to cfNAs. The kupffer cells in the liver are responsible for the majority of the cfNA clearance, while the kidney and spleen also play major roles in removing cfNA from circulation.

The cfDNA fragment size distribution and sequence characteristics in blood and other body fluids tend to change under some health conditions and disease states. cfDNA in plasma from cancer patients shows more variable and shorter median fragment lengths compared to those from healthy individuals, and the shorter fragments have been shown to be predominantly originating from tumor tissues - suggesting alterations in chromatin organization and nucleosomal structure associated with malignancy [25]. In pregnant women, the circulating cell-free DNA of fetal origin displays a modal size profile shorter than the maternal cfDNA, pointing to differences in nucleosomal organization and chromatin accessibility between fetal and maternal tissues [38]. It may be partly due to the generation and degradation of cfDNA in the specific disease context. DNase activities vary markedly across different types of cancer, reflecting distinct patterns of DNA degradation. For instance, reduced DNase activity has been observed in colon, stomach, pancreatic, and prostate cancers, suggesting a potential accumulation or altered degradation of cfDNA in these conditions [[55], [56], [57]]. Conversely, elevated DNase activity in oral and breast cancers indicates increased DNA fragmentation [58,59]. Furthermore, DNase activity correlates with treatment responses; increased activity is associated with positive treatment outcomes in head and neck cancer, whereas lower activity suggests effective treatment in lymphoma [60,61].

Computational tools for analysis of cfNA fragments

Streamlined extraction and high throughput sequencing of liquid biopsy samples from body fluids have motivated computational genomics method development to biological signals from cfDNA fragmentation characteristics to non-invasively learn about their likely tissue of origins, as well as cellular epigenomes and cellular processes. There has been a steady development of innovative computational and genomic tools, designed to tease apart the complex signals that the sizes, (epi)genomic contexts, and end sequences of fragments carry (Table 1).

Table 1.

Examples of computational frameworks for analyzing cfNA fragmentomics in body fluids.

Type	Framework	Notable feature
Length/GC content-based	DELFI [25]	Correlates short vs long cfDNA fragment count ratio with cancer
	The framework developed by Liu and colleagues [62]	Compares between mono, di, and tri-nucleosome derived cfDNA fragment distributions
	The framework developed by Che and colleagues [63]	Focuses on long cfDNA fragments
End seq-based	End-motif analysis [64]	Aberrant end-motif in cancer patients
	F-profile [65]	6 F-profiles, each can describe underlying nuclease activity and disease
	Jagged-end analysis [66]	Associates jaggedness with nucleosomal patterns
	OCF [67]	OCF values correspond to the tissue of origin of the cfDNA fragments
Methylomics-based	FRAGMA [68]	Predicts methylation status using CGN/NCG motif ratios
	FinaleMe [69]	Non-homogeneous Hidden Markov Model predicting methylation of cfDNA at CpG islands
Other epigenetic- based	WPS [37]	Detects nucleosome position using the distribution of fragment endpoints
	The framework developed by Stanley and colleagues [70]	Utilizes WPS score to link nucleosome spacing with gene expression patterns
	Fragmentomics-based identification of transcription factor binding site [71]	Early tumor detection and subtype prediction
	EPIC-seq [72]	Correlates promoter fragmentation entropy with gene expression level
	Griffin [73]	Profiles nucleosome accessibility from cfDNA fragments
	Enhancer profiling [74]	Uses cfDNA fragments as a proxy for transcriptional activity
Multimodal	FrEIA [75]	ML model combining the proportion of short fragment, GINI index of end-motif, tumor fraction
	THEMIS [76]	ML model integrating MFR, FSI, CAFF, FEM
	GALYFRE [77]	ML model combining iwFAF score and end-motifs of RPR
	SPOT-MAS [78]	Simultaneously profiles methylomics, fragment length, copy number aberrations, and end motifs
	Cancer score [79]	Combines 6-bp end motif, 6-bp breakpoint motif, fragment size, and CNV
cfNA Database	cfOmics [80]	cfDNA-omics, cfRNA-omics, proteomics, metabolomics of 11,345 samples (69 diseases and 13 sample-type)
	FinaleDB [81]	cfDNA database comprising 2505 samples across 23 pathological conditions

Computational resources for analysis of cfDNA fragmentation characteristics were influential in identifying epigenetic signals from the cell of origin. Cristiano et al. [25] introduced DNA evaluation of fragments for early interception (DELFI), a novel gradient tree boosting machine learning model that correlates the fraction of short to long cfDNA fragments for each sample with the median fragment length profile of cfDNA fragments within 5-megabase windows, to differentiate healthy versus cancer-derived cfDNA. For seven cancer types, the model showed a sensitivity of detection ranging from 57 % to >99 %, with 98 % specificity and an area under the ROC curve (AUC) of 0.94. The computational pipeline developed by Liu et al. [62] adopted a more nuanced analysis of the cfDNA fragment lengths, using the test statistic (D) of the Kolmogorov-Smirnov (KS) test to compare fragment distributions ranging from 100 to 650 bp that represents mono (100–250 bp), di (251–450 bp), and tri-nucleosome (451–650) derived cfDNA segments. While most of the research focused on the most prevalent cfDNA fragment length (<500 bp) found in blood, Che et al. concentrated on the long cfDNA molecules (>500 bp), and demonstrated that these fragments are preferentially derived from transcriptionally active, GC-rich, gene-rich euchromatic regions, marked by overrepresentation in light bands of Giemsa-stained chromosomes and a strong association with DNase I hypersensitive sites (DHSs) and CCCTC-binding factors (CTCFs). Their findings further reveal a positive correlation between long cfDNA molecule abundance and factors such as gene density, transcriptional activity, and lower methylation levels. The Windowed Protection Score (WPS) infers nucleosome positions in the genome by assessing the distribution of fragment endpoints relative to nucleosome core particle (NCP) boundaries [25]. Capitalizing on the concept of the WPS score, Stanley et al. introduced a novel method for disease-agnostic cell-of-origin analysis of cfDNA by integrating single-cell transcriptome data with cfDNA fragmentation patterns and then analyzed ultra-low coverage cfDNA data using the WPS score analyzed through fast Fourier transformation to associate nucleosome spacing with gene expression across various cell types [70].

Several recently developed computational tools can decode the fragmentation pattern to predict the DNA methylation status, in order to predict tumor presence and the tissue of origin of cfDNA. Using convolutional neural networks (CNNs), Zhou et al. developed a fragmentomics-based methylation analysis (FRAGMA) method leveraging cleavage patterns near CpG sites in cfDNA fragments to infer methylation status, utilizing the differential fragmentation indicated by CGN/NCG motif ratios within an 11-nucleotide window [68]. Liu et al. introduced FinaleMe, a non-homogeneous Hidden Markov Model designed for analyzing plasma whole-genome sequencing data to predict DNA methylation levels of cell-free DNA at individual CpG sites, especially in CpG-rich areas, and to deduce the tissue of origin of cfDNA fragments from these methylation patterns [69].

Others have attempted to integrate several properties of cfDNA fragments to elucidate the underlying biological signals. Moldovan et al. developed the fragment-end integrated analysis (FrEIA) metric to analyze the cfDNA fragment-end composition of healthy and cancer patients. Their research also included the analysis of tumor fraction (ichorCNA TF) from low-coverage whole-genome sequencing data, calculating the proportion of short fragments (P20–150) and trinucleotide sequence diversity using the Gini index - all combined in a machine learning model designed to detect cancer, with an accuracy of 82 % and AUC of 0.96 [75]. Bie et al. introduced THEMIS (THorough Epigenetic Marker Integration Solution), a novel blood-based multicancer detection approach utilizing whole-methylome sequencing (WMS) of cfDNA from just 4 ml of plasma [76]. The computational tool is developed through ensemble machine learning by combining 4 different scoring systems, such as methylated fragment ratio (MFR, the ratio of fully methylated fragments within each 1-Mb genomic segment), fragment size index (FSI, the ratios of short [100–166 bp] to long [169–240 bp] fragments for each 5-Mb windows), chromosomal aneuploidy of featured fragments (CAFF, calculation of plasma aneuploidy score to measure the copy number alterations of chromosome arms), and fragment end motifs (FEM, the frequencies of 256 4-mer motifs at the 5′ end of fragments). The machine learning model achieved a detection sensitivity of 73 % at 99 % specificity for early-stage patients [76].

Abnormal cfDNA fragmentation patterns are common in some health conditions and diseases such as cancer. Budhraja et al. aimed to assess the extent of fragmentation anomalies in a given sample utilizing both the length and GC content information of cfDNA fragments using the information-weighted fraction of aberrant fragments (iwFAF) scoring metric of the recurrently protected regions (RPRs) across the genome [77]. Their fragmentomics-based machine learning model named GALYFRE is trained based on the iwFAF score and the frequencies of fragment end nucleotides that correlated the most with the iwFAF score of the cfDNA molecules to infer tumor fraction and differentiate healthy individuals from cancer patients. GALYFRE demonstrated an AUC of 0.91 for the detection of cancer at any stage, and 0.87 for the detection of stage I cancer patient samples [77].

Several databases have been developed to not only compile the cfDNA fragmentomics results from large-scale studies but also to offer intuitive browsing and visualization functionalities. Zheng et al. introduced FinaleDB, a database (http://finaledb.research.cchmc.org/) designed to consolidate thousands of uniformly processed and curated cfDNA whole-genome sequencing (WGS) datasets from various pathological conditions [81]. Equipped with a fragmentation genome browser, FinaleDB enables users to overlay these cfDNA datasets with a multitude of omics data across different cell types, offering a holistic view of the gene-regulatory landscape and cfDNA fragmentation dynamics for comprehensive analysis [81]. Li et al. introduced cfOmics, a web-accessible database (http://111.198.139.65/cfomics/) combining the four omics categories (cfDNA omics that include methylation, end motif, fragment size and nucleosome occupancy of cell-free DNA, as well as microbial DNA abundance; cfRNA omics which include but not limited to gene expression, alternative promoter, chimeric RNA, and RNA SNP, as well as microbial RNA abundance; proteomics; and metabolomics based on mass-spectrometry data) [80]. They integrated publicly available data from 69 disease conditions (28 cancer type and 41 non-cancerous diseases) and 13 specimen variations (plasma, platelet, serum, urine, etc.) totaling 11,345 samples [80]. Taken together, the innovative computational genomics methods and database utilities offer rich resources to capture complex signals from cell-free nucleic acid fragments to learn about the underlying biological interpretations. In Fig. 3, we have shown a chronological schematic of the computational tools developed in recent years, categorized by their salient features.

Fig. 3.

Schematic of some of the cfNA-focused computational tools developed in recent years (2016–2024), ordered in chronology and categorized by their salient features (length/GC-based, end seq-based, methylomics-based, other epigenetic-based, multimodal, and cfNA database). The citation numbers of these tools are included, for multiple tools of similar categories developed in the same year, the citation numbers are stacked on top of each other.

Multimodal biological insights from fragmentomics

Although a majority of the initial liquid biopsy-based cfDNA studies in the past decade focused on detecting disease-associated genetic alterations e.g. CNVs, aneuploidies, and mutations [82,83], those approaches do not harness the full spectrum of non-genetic signals that can be derived from cfDNA fragments [84]. Recent investigations of fragmentation characteristics of cell-free nucleic acid sequences (length, end motif, etc.) have offered insights into the epigenetic makeup, potential regulators of transcriptional states, and cellular processes in cells of origin, complementing the genomic information. cfDNA fragments in blood typically follow a size distribution with a peak of ∼167 bp and multiples thereof, indicating the positioning of nucleosomes in the cell of origin. Using the WPS score, Snyder et al. mapped genome-wide in vivo nucleosome occupancy, which provided insights into nuclear architecture, gene structure, and expression [37]. Cristiano et al. showed that the machine learning algorithm of DELFI can identify the tissue source of cfDNAs using their fragment length profile [25]. Sun et al. performed orientation-aware cfDNA fragmentation (OCF) analysis to study nucleosome positioning and the tissue-specific origins of plasma DNA [67]. This method capitalizes on characteristic fragmentation patterns within open chromatin regions, detecting sequencing coverage imbalances and orientation-specific fragment end signals that reflect the relative contributions of different tissues to the plasma DNA pool [67].

Integration of cfDNA fragmentation patterns with gene expression or functional genomic data has offered further valuable insights into cellular epigenome and underlying cell states. A disease-agnostic cell-of-origin analysis of cfDNA by integrating single-cell transcriptome data with cfDNA fragmentation patterns established nucleosome spacing with gene expression across various cell types. Besides the tissue of origin information, nucleosome occupancy predicted from cfDNA fragmentation patterns can provide valuable insights into chromosomal structure, gene expression level, and cell state. Ulz et al. adopted the cfDNA fragmentomics approach to analyze nucleosome footprints to quantify the transcription factor (TF) binding site accessibility and the degree of nucleosome phasing around these sites, thereby inferring the strength and presence of TF binding from the cfDNA fragmentation patterns observed [71]. Notably, their approach not only enables the identification and profiling of lineage-specific TFs but also facilitates dynamic monitoring of TF activity during disease progression, such as cellular reprogramming in prostate cancer, offering valuable insights into the temporal changes in chromatin accessibility associated with malignancy [71]. Esfahani et al. introduced EPIC-seq, an innovative approach leveraging promoter fragmentation entropy as a cfDNA feature to infer RNA expression levels, highlighting the potent utility of epigenomic profiling in understanding transcriptional regulation from cell-free DNA [72]. By focusing on nucleosome accessibility and positioning at transcription start sites, EPIC-seq captures key fragmentomics signals that are most pronounced at promoters of actively expressed genes, providing a direct link between epigenetic modifications and gene expression [72]. This method also detects significant signals in exonic regions during whole-exome sequencing of expressed genes, suggesting a broader applicability for EPIC-seq to explore gene expression patterns across the genome [72].

Besides the nucleosomal positioning, nuclease activities in the cells and body fluids dictate the end sequence characteristics of cfDNAs. resulting in the presence of CC-end fragments to be the most abundant end motif in healthy individuals [85,75]. cfDNA molecules can also have single-stranded 5′ DNA overhangs, often referred to as “jagged ends”. Jiang et al. showed that variance in the pattern of jagged end fragments is closely associated with nucleosomal occupancy patterns [66]. Zhou et al. leveraged "founder" end-motif profiles (F-profiles) of cfDNA to establish correlations with specific nucleases which elucidates the molecular diversity in cfDNA fragmentation [65]. Markus et al. performed computational genomic analysis of fragments and computed a per-base-nucleotide-bias score (PBNB-score) to reveal distinct enrichment patterns that correlate closely with nucleosome boundaries [86]. Wang et al. demonstrated that by quantifying cfDNA through fragmentation size coverage (FSC) and distribution (FSD), alongside sequencing-specific motifs at cfDNA fragment ends (EDM) and breakpoints (BPM), the underlying biological processes which alter the cfDNA fragmentation during tumorigenesis can be deduced [79].

DNA methylation at CpG sites is an important regulatory mechanism that controls the expression of proximal genes, and genome-wide methylation pattern varies across tissues. While several studies have used bisulfite sequencing or other sequencing-based techniques to identify methylation patterns in cfDNA, it requires additional processing of cfDNA, which may be less preferred in some contexts. Emerging studies suggest that it may be possible to infer DNA methylation patterns utilizing differential degradation of methylated and unmethylated CpGs at the fragment ends. DNA methylation status at individual CpG sites of cell-free DNA can predict their tissue of origin [80]. Jiang et al. revealed that cfDNA end motifs can be indicative of the tissue of origin, as evidenced by motif clustering from similar tissues such as the liver, placenta, and hematopoietic cells [64].

Applications of fragmentomics in cancer and other disease contexts beyond mutation detection

Inference about the cell of origin and cellular states from cell-free nucleic acids enables the assessment of disease-associated signals from liquid biopsy for non-invasive detection and monitoring of health conditions, diseased patients, or those at risk. Most studies that developed computational or genomic applications and revealed novel biology, did so in the context of diseases, typically cancer, or health conditions such as pregnancy. Some fragmentation characteristics differ between healthy and diseased patients. Nguyen et al. showed that cfDNA is more fragmented in cancer patients, with a shorter overall size distribution [78]. Analysis of the information-weighted fraction of aberrant fragments (iwFAF) scoring metric of the recurrently protected regions across the genome suggests that abnormally fragmented cfDNA likely originates from the tumor cells [77]. Jiang et al. showed that plasma DNA end motifs exhibit significant diversity across various cancers, notably a decreased abundance of the CCCA motif, which contrasts with the profiles seen in healthy individuals, indicating that specific end motifs are associated with cancerous states [64]. However, Moldovan et al. analyzed the end motif of 9 cancer types and demonstrated that all cancer types showed a similar trend in fragment end sequence fold changes, with fragments starting with CCA, CCC, or CCT decreasing the most compared to healthy control fragments, while fragments starting with TTC, TTA, or ATG increased the most, suggesting common mechanisms of cfDNA cleavage in cancer irrespective of the cancer type [75].

There has been a considerable effort to translate the multimodal signals derived from cfDNA into clinical actions. Building on the utility of cfDNA fragmentation profiles, Mathios et al. demonstrated that DELFI scores correlate with cancer stage, histology, and tumor size in lung cancer, indicating that even small lesions can be effectively detected through this non-invasive approach [87]. Foda et al. revealed that changes in fragmentation profiles were significantly associated with liver cancer stages and were distinct from profiles in individuals with non-cancer liver conditions such as cirrhosis and viral hepatitis [88]. This differentiation was further emphasized as DELFI scores increased with the severity and size of liver cancer lesions, demonstrating the potential of cfDNA fragmentation as a sensitive marker for diagnosing and monitoring liver cancer progression [88]. Lu et al. demonstrated that differentially methylated regions (DMRs) in cfDNA, identified through cfMeDIP-seq, effectively distinguish between ovarian cancer (OC) stages, with markers identified for early-stage OC are significantly altered early in the disease process and are capable of detecting all stages of OC [89]. The “cancer score” formulated by Wang et al. based on cfDNA fragmentation properties increased progressively from stage I to stage IV NSCLC, demonstrating its potential for precise stage-based diagnosis [79]. Apart from staging, the study also demonstrated that cfDNA fragmentomics can be applied to detect tumor location and lymph node metastasis in lung cancer patients [79].

Besides cancer monitoring, it is also necessary to identify the cancer stage and tumor size for disease monitoring and effective treatment strategy. Intratumor and clonal heterogeneity shape the direction of treatment outcome and disease-free survival. Doebley et al. developed Griffin which leverages nucleosome protection patterns in cell-free DNA (cfDNA) for tumor subtype classification and monitoring, particularly in identifying estrogen receptor subtypes in breast cancer, offering insights into tumor heterogeneity and the dynamic nature of cancer progression [73]. Plasma enhancer profiling using cfDNA fragmentomics can detect epigenetic modifications associated with treatment resistance in cancer, such as activating the AR gene enhancer in prostate cancer which drives castration resistance [74]. This method also enables the identification of treatment-induced neuroendocrine differentiation (NE-diff) in prostate cancer, highlighting its potential in diagnosing clinically significant histologic transformations that are difficult to detect with genetic assays. cfDNA fragmentomics have also been used to infer the effectiveness of cancer treatment. For example, Cristiano et al. demonstrated that cfDNA fragmentation profiles in non-small cell lung cancer patients undergoing anti-EGFR or anti-ERBB2 therapy correlate significantly (Spearman correlation = 0.74) with mutant allele fractions, highlighting the potential of fragmentation analysis to monitor tumor dynamics and treatment efficacy [25]. The prediction of the recurrence of cancer after surgery is difficult as there might be tiny-sized tumors present in tricky anatomical locations that might not be detected by imaging techniques. Wang et al. showed that the fragmentomics-based approach has greater sensitivity in predicting patient recurrence compared with the traditional circulating mutation, especially after the surgery for early-stage NSCLC [14]. Moldovan et al. showed that a fragmentomics-based machine-learning model achieved high accuracy in correlating fragmentation properties with recurrence-free survival in patients with resectable esophageal adenocarcinoma [75].

Not only clinical management of cancer patients, but early diagnosis of cancer is also possible by integrating the fragmentomics characteristics. Bao et al. developed a computational model based on machine learning that excelled in early detection and accurate localization of primary liver cancer (PLC), colorectal cancer (CRC), and lung adenocarcinoma (LUAD), demonstrating high effectiveness even at very early stages (stages 0 and I) and low sequencing depths. This machine learning model exhibited an AUC of 0.983, with an overall sensitivity of 95.5 % and specificity of 95 % [90]. A similar approach was adopted to robustly detect early-stage non-small cell lung cancer (NSCLC) by Wang et al. [79]. Several studies have demonstrated that fragmentomics-based machine learning models of cfDNA can perform better than traditional mutation-based diagnostic approaches, or combined fragmentomics-mutation-based approaches for cancer detection. For example, the DELFI score could detect 82 % of patients from a multi-cancer cohort, compared to only 66 % of patients detected by the mutation approach. Combining both fragmentomics and mutation-based approaches increased the fraction of patients detected to 91 % [25].

Apart from cancer diagnosis, fragmentomics-based machine learning models have also shown promising results in terms of some other disease diagnoses and monitoring. For example, Yu et al. demonstrated that the size distribution of fetal cfDNA in maternal plasma can be utilized to detect fetal chromosomal aneuploidies such as trisomy 21 and 18 (100 % sensitivity and 100 % specificity), by identifying abnormal proportions of short DNA fragments [91]. Ding et al. investigated the impact of nuclease deletions on the jaggedness of plasma DNA in systemic lupus erythematosus (SLE), revealing that jaggedness aberrations, especially in DNA fragments over 200 bp, are indicative of disease activity and flares (AUC of 0.96 between subjects with active SLE and without active SLE) [92]. This study highlights the potential of plasma DNA jaggedness as a biomarker for monitoring SLE disease states [92].

Discussion

Recent advances in fragmentomics and liquid biopsy-based research have opened unprecedented advances in non-invasive diagnosis and disease monitoring. Nonetheless, there are still several challenges that need to be addressed before translating the discoveries to clinical practice to aid in caring for patients. Some of these are specific to fragmentomics, while others apply to cfDNA-based translational research in general.

It remains non-trivial to estimate tumor burden accurately from mutations, copy number alterations, or multimodal signatures from cfDNA fragments. The limited amount of sample and short fragment length of cfDNA makes it inherently challenging to isolate and profile it accurately. Furthermore, biological variations and technical variations in sample handling protocols [93] can bias the accuracy of the predictions. Delays in sample processing and sample mishandling may lead to cfDNA contamination by leukocyte genomic DNA and chemical damage (hydrolysis, deamination, and oxidative damage) to the sample [94,95], potentially altering cfDNA fragmentation characteristics. The low biomass of cfDNA has been a long-standing concern affecting the accuracy and reliability of computational models designed for clinical applications. Several recent studies have effectively addressed the issue by tailoring their computational tools appropriate for very low concentrations of cfDNA [70,75,90].

Furthermore, most studies focus on blood-based liquid biopsy assays, but it is possible to get cfDNA from other body fluids including urine [96] for clinical management of cancer. Urine fragmentomic signatures are very different from those in the blood, and also contributions from different tissue types therein may be different [97,98]. In a recent study, cfDNA fragmentation pattern in urine was utilized to detect glioma patients non-invasively [99]. In another study, salivary cfDNA fragmentomics profiles along with distinct oral microbiome populations were demonstrated to effectively distinguish gastric cancer patients from healthy individuals [100].

It is also difficult to precisely differentiate tumor-derived DNA from those sourced from normal cells, especially if the sequences do not carry cancer-associated somatic mutations. A majority of cancer-derived fragments remain unmutated, even in tumors with high mutation loads and individuals with severe disease burdens. This has implications for accurate estimation of tumor burden. In addition, ascertaining the cell-of-origin of non-tumor-derived DNA needs further work. Depending on inter-tumor heterogeneity in cellular processes, some tumors may release less cfDNA (dubbed non-shedders) than others of comparable sizes, rendering bias toward tumor burden estimates [72]. Apart from "non shedders," other factors such as tumor location, size, and vascularity are known to affect detectable ctDNA quantities and their fragmentation characteristics [101,102]. In addition, further work needs to be done to estimate contributions from non-tumor tissues in cfDNA and assess their clinical implications in the context of metastasis, comorbidity, and treatment-related complications.

Nonetheless, recent findings discussed above suggest that fragmentomic signatures can offer a window to epigenomic contexts of the cells of origin and complement mutation-focused genomic assessment. It is expected that simultaneous, multimodal inference from mutational and fragmentomic markers can help obtain clinically actionable signals. Non-invasive early detection of cancer and timely monitoring of treatment response can help strategize effective clinical management plans, reduce healthcare costs, and save lives. While recent advances are very promising, albeit early-stage developments towards that direction.

CRediT authorship contribution statement

Ahmad Salman Sirajee: Writing – review & editing, Writing – original draft, Visualization, Validation, Software, Methodology, Conceptualization. Debajyoti Kabiraj: Writing – review & editing, Writing – original draft. Subhajyoti De: Writing – review & editing, Writing – original draft, Validation, Supervision, Resources, Project administration, Funding acquisition, Conceptualization.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

The author is an Editorial Board Member/Editor-in-Chief/Associate Editor/Guest Editor for Translational Oncology and was not involved in the editorial review or the decision to publish this article.