Research Article

Multimodal analysis of methylomics and fragmentomics in plasma cell-free DNA for multi-cancer early detection and localization

Gene Solutions, Viet Nam
Medical Genetics Institute, Viet Nam
University Medical Center, Viet Nam
MEDIC Medical Center, Viet Nam
National Cancer Hospital, Viet Nam
Hanoi Medical University, Viet Nam
Thu Duc City Hospital, Viet Nam
Department of Biostatistics, Virginia Commonwealth University, School of Medicine, United States
Pham Ngoc Thach University of Medicine, Viet Nam

Oct 11, 2023

https://doi.org/10.7554/eLife.89083.3

Open access
Copyright information

eLife assessment

This study provides insights into the early detection of malignancies with noninvasive methods by developing a framework, which assesses methylation, CNA, and other genomic features. They established a solid model in discriminating malignancies from healthy controls, as well as the ability to distinguish tumor of origin. This important study will demonstrate its practical impacts in the clinic and other researchers of the field.

https://doi.org/10.7554/eLife.89083.3.sa0

Significance of the findings:

Important: Findings that have theoretical or practical implications beyond a single subfield

Landmark
Fundamental
Important
Valuable
Useful

Strength of evidence:

Solid: Methods, data and analyses broadly support the claims with only minor weaknesses

Exceptional
Compelling
Convincing
Solid
Incomplete
Inadequate

During the peer-review process the editor and reviewers write an eLife Assessment that summarises the significance of the findings reported in the article (on a scale ranging from landmark to useful) and the strength of the evidence (on a scale ranging from exceptional to inadequate). Learn more about eLife Assessments

Abstract
Introduction
Results
Discussion
Materials and methods
Data availability
References
Article and author information
Metrics

Abstract

Despite their promise, circulating tumor DNA (ctDNA)-based assays for multi-cancer early detection face challenges in test performance, due mostly to the limited abundance of ctDNA and its inherent variability. To address these challenges, published assays to date demanded a very high-depth sequencing, resulting in an elevated price of test. Herein, we developed a multimodal assay called SPOT-MAS (screening for the presence of tumor by methylation and size) to simultaneously profile methylomics, fragmentomics, copy number, and end motifs in a single workflow using targeted and shallow genome-wide sequencing (~0.55×) of cell-free DNA. We applied SPOT-MAS to 738 non-metastatic patients with breast, colorectal, gastric, lung, and liver cancer, and 1550 healthy controls. We then employed machine learning to extract multiple cancer and tissue-specific signatures for detecting and locating cancer. SPOT-MAS successfully detected the five cancer types with a sensitivity of 72.4% at 97.0% specificity. The sensitivities for detecting early-stage cancers were 73.9% and 62.3% for stages I and II, respectively, increasing to 88.3% for non-metastatic stage IIIA. For tumor-of-origin, our assay achieved an accuracy of 0.7. Our study demonstrates comparable performance to other ctDNA-based assays while requiring significantly lower sequencing depth, making it economically feasible for population-wide screening.

Introduction

The incidence of cancer-related morbidity and mortality is rapidly increasing globally, and accounted for nearly one fifth of all deaths in 2020 (Sung et al., 2021). High-cost treatment is a significant financial burden for cancer patients, with almost 286 billion dollars in 2021 and an increase of 8.2% – 581 billion dollars in 2030. In Vietnam, GLOBOCAN 2020 reported over 182,500 newly diagnosed cases and 122,690 cancer-related deaths (Sung et al., 2021). Among these, liver (14.5%), lung (14.4%), breast (11.8%), gastric (9.8%), and colorectal cancer (9%) are the five most common types. Up to 80% of cancer patients in Vietnam were diagnosed at stage III or stage IV, resulting in a high rate of 1 year mortality (25%) and a low 5-year survival rate compared to other countries (Pham et al., 2019). Diagnostic delays are associated with a lower chance of survival, greater treatment-associated problems, and higher costs (Hawkes, 2019). Cancer detection at earlier stages can improve the opportunity to control cancer progression, increase the patient survival rate, and lower medical expenses (Kakushadze et al., 2017).

Although currently guided screening tests have each been shown to provide better treatment outcomes and reduce cancer mortality, some of them are invasive, thus having low accessibility. Importantly, most of them are single cancer screening tests, which may result in high false positive rates when used sequentially (Sasieni et al., 2023). Multi-cancer early detection (MCED) tests can potentially overcome these challenges by simultaneously detecting multiple cancer types from a single test (Liu et al., 2020). Liquid biopsy, an emerging non-invasive approach for MCED, can capture a wide range of tumor features, including cell-free DNA (cfDNA), circulating tumor DNA (ctDNA), exosomes, proteins, mRNA, and metabolites (Li et al., 2018; Nguyen et al., 2022b). Among them, ctDNA has become a promising biomarker for detecting early-stage cancers because it is a carrier of genetic and epigenetic modifications from cancer-derived DNA (Gao et al., 2022). Indeed, ctDNA detection has demonstrated several advantages in non-invasive diagnostic, prognostic, and monitoring of cancer patients during and after treatment (Pascual et al., 2022; Nguyen et al., 2020). Furthermore, ctDNA carrying tumor-specific alterations could be used to identify the corresponding unknown primary cancer and tumor localization.

In recent years, there has been considerable interest in exploring the potential of ctDNA alterations for early detection of cancer (Nguyen et al., 2020; Moser et al., 2023). One such approach is the PanSeer test, which uses 477 differentially methylated regions (DMRs) in ctDNA to detect five different types of cancer up to 4 years prior to conventional diagnosis (Im et al., 2021). The DELFI assay employs a genome-wide analysis of ctDNA fragment profiles to increase sensitivity in early detection (Zhou et al., 2022b). Recently, the Galleri test has emerged as a multi-cancer detection assay that analyzes more than 100,000 methylation regions in the genome to detect over 50 cancer types and localize the tumor site (Cohen et al., 2018).

Despite their great potential, there remain several challenges that these assays must solve to deliver accessible and reliable clinical adoption for the large population, including the low fraction of ctDNA in the blood of early-stage cancer patients, the heterogeneity of ctDNA signatures from diverse cancer types, subtypes and stages (Moser et al., 2023), and the high sequencing depth required. To address these challenges, recent studies have focused on multi-analyte approach – combining genomic and nongenomic features such as methylomics and fragmentomics to increase the detection of ctDNA and accuracy for tissue of origin (TOO) identification (Moser et al., 2023; Im et al., 2021; Zhou et al., 2022b). Advances in multimodal analysis approaches have led to the development of powerful screening tests that enable high sensitivity and cost-effectiveness. For example, CancerSEEK uses a combined approach of protein biomarkers and genetic alterations to detect and locate the presence of eight types of cancers (Cohen et al., 2018). In this assay, cancer-associated serum proteins play a complementary role in tumor localization as cfDNA mutations are not tissue specific. However, detecting both protein and genetic biomarkers are time-consuming and costly. Thus, the development of future MCED tests should endeavor to deliver a screening approach with high sensitivity, specificity, and TOO identification at cost-effective price to provide better clinical outcomes and treatment opportunities for all cancer patients.

In an effort to address the challenges of early cancer detection, we have developed a multimodal approach called SPOT-MAS (screening for the presence of tumor by DNA methylation and size). This assay was previously applied to cohorts of colorectal (Nguyen et al., 2022a) and breast cancer patients (Pham et al., 2023) and demonstrated ability for early detection of these cancers at high sensitivity across different cancer stages and patient age groups. In this study, we aimed to expand our multimodal approach, SPOT-MAS, to comprehensively analyze methylomics, fragmentomics, DNA copy number, and end motifs (EMs) of cfDNA and evaluate its utility to simultaneously detecting and locating cancer from a single screening test. As proof of concept, we used 2288 participants, including 738 non-metastatic patients and 1550 healthy controls, to train and fully validate this approach on five commonly diagnosed cancers, including breast, gastric, lung, colorectal, and liver cancer. Our findings demonstrate that the multimodal approach of SPOT-MAS enables profiling of multiple ctDNA signatures across the entire genome at low sequencing depth to detect five different cancer types in their early stages. Beyond detecting the presence of cancer signals, our assay was able to predict the tumor location, which is important for clinicians to fast-track the follow-up diagnostic and guide necessary treatment. Thus, SPOT-MAS has the potential to become a universal, simple, and cost-effective approach for early multi-cancer detection in a large population.

Results

Clinical characteristics of cancer and healthy participants

This study recruited 738 patients with five common cancer types, including breast cancer (n=223), CRC (n=159), gastric cancer (n=98), liver cancer (n=122), lung cancer (n=136), and 1550 healthy participants (Supplementary file 1, Table S1). Cancer patients were diagnosed by either imaging and/or histology analysis, depending on cancer type. All cancer patients were treatment-naïve at the time of blood collection. Healthy participants had no history of cancer at the time of sample collection and remained cancer-free at the 6- and 12-month follow-ups. Cancer patients and healthy participants were randomly assigned to the discovery and validation cohorts (Table 1 and Supplementary file 1, Table S2). The discovery cohort was used to profile multiple cancer- and tissue-specific signatures and to construct machine learning algorithm while the validation cohort was used solely to external evaluation of the performance of machine learning models.

Table 1

Summary of clinical features of 738 cancer patients and 1550 healthy controls in discovery and validation cohorts.

Clinical features		Discovery cohort (N=1575)					Validation cohort (N=713)
		Cancer (N=499)		Healthy (N=1076)		p-Value (cancer vs healthy)	Cancer (N=239)		Healthy (N=474)		p-Value (cancer vs healthy)
		N	Percentage	N	Percentage	p-Value (cancer vs healthy)	N	Percentage	N	Percentage	p-Value (cancer vs healthy)
Gender	Female	279	55.9%	599	55.7%	0.9281*	126	52.72%	270	56.1%	0.2818*
Gender	Male	220	44.1%	477	44.3%	0.9281*	113	47.28%	204	43.9%	0.2818*
Age	Median	58		47		<0.0001^†	59		48		<0.0001^†
	Min	25		18			28		19
	Max	97		84			92		85
Stage	I	52	10.4%				23	9.6%			0.4947*
	II	169	33.9%				69	28.9%
	IIIA	150	30.1%				77	32.2%
	Non-metastasis with unknown staging information	128	25.7%				70	29.3%

*

p-Values from Chi-square test.
†

p-Values from Mann-Whitney test.

The discovery cohort comprised of 499 cancer patients (156 breast, 106 CRC, 67 gastric, 77 liver, and 93 lung, Supplementary file 1, Table S1) and 1076 healthy participants. The cancer group had a median age of 58 (range 25–97, Table 1) and consisted of 279 females and 220 males. The discovery healthy group consisted of 599 females and 477 males, with a median age of 47 (range 18–84, Table 1). In the discovery cohort, gender ratios were similar between cancer and healthy control groups, whereas cancer patients were older than controls (p<0.0001, Mann-Whitney test, Table 1). Of the cancer patients, 10.4% were at stage I, 33.9% were at stage II, and 30.1% were at non-metastatic stage IIIA. Staging information was not available for 25.7% of cancer patients, who were confirmed by specialized clinicians to have non-metastatic tumors (Table 1).

The validation cohort consisted of 239 cancer patients (67 breast, 53 CRC, 31 gastric, 45 liver, and 43 lung, Supplementary file 1, Table S1) and 474 healthy participants (Table 1). Consistent with the discovery cohort, the gender distribution was comparable between the cancer and healthy control groups, and the cancer group was older than the control group, with a median age of 59 and 48 years of age, respectively (p<0.0001, Mann-Whitney test, Table 1). The percentage of cancer patients with each stage was similar to that of the discovery cohort, with 9.6% at stage I, 28.9% at stage II, and 32.2% at stage IIIA. Staging information was unavailable for 29.3% of non-metastatic cancer patients (Table 1).

The multimodal SPOT-MAS assay for multi-cancer and TOO detection

In our recent study of SPOT-MAS, we have demonstrated that the integration of ctDNA methylation and fragmentomic features can significantly improve the early detection of colorectal cancer (Nguyen et al., 2022a) and breast cancer (Pham et al., 2023). Here, we expanded the breadth of ctDNA analyses by adding two sets of features including DNA copy number and EM into SPOT-MAS to maximize cancer detection rate and identify TOO. Briefly, a novel and cost-effective workflow of SPOT-MAS was developed involving three main steps (Figure 1). In step 1, cfDNA was isolated from peripheral blood and subjected to bisulfite conversion and adapter ligation to create a single whole-genome bisulfite library of cfDNA. From this library, in step 2, a hybridization reaction was performed to collect the target capture fraction (450 cancer-specific regions), then the whole-genome fraction was retrieved by collecting the ‘flow-through’ and hybridizing with probes specific for adapter sequences of DNA library. Both the target capture fraction and whole-genome fraction were sequenced to the depth of ~52× and 0.55×, respectively (Supplementary file 1, Table S3). Data pre-processing was performed to generate five different sets of cfDNA features, including methylation changes at target regions (TM), genome-wide methylation (GWM), fragment length patterns (FLEN), copy number aberrations (CNA), and EM. In step 3, these features were used as inputs for a two-stage model to obtain prediction outcomes. Stage 1 of our model comprised of a stacking ensemble machine learning model for binary classification of cancer versus healthy. Then, the samples predicted as cancer were passed to stage 2 where graph convolutional neural network (GCNN) was adopted to predict TOO (Figure 1).

Figure 1

Download asset Open asset

Workflow of SPOT-MAS (screening for the presence of tumor by methylation and size) assay for multi-cancer detection and localization.

There are three main steps in the SPOT-MAS assay. First, cell-free DNA (cfDNA) is isolated from peripheral blood, then treated with bisulfite conversion and adapter ligation to make whole-genome bisulfite cfDNA library. Second, whole-genome bisulfite cfDNA library is subjected to hybridization by probes specific for 450 target regions to collect the target capture fraction. The whole-genome fraction was retrieved by collecting the ‘flow-through’ and hybridized with probes specific for adapter sequences of DNA library. Both the target capture and whole-genome fractions were subjected to massive parallel sequencing and the resulting data were pre-processed into five different features of cfDNA: target methylation (TM), genome-wide methylation (GWM), fragment length profile (FLEN), DNA copy number (CNA), and end motif (EM). Finally, machine learning models and graph convolutional neural networks are adopted for classification of cancer status and identification tissue of origin.

Identification of DMRs in cancer patients from target capture fraction

DNA methylation is an important epigenetic signature responsible for major changes in regulating expression of cancer-associated genes by impacting the binding of transcription factors to regulatory sites and the structure of chromatin (Yin et al., 2017; Buitrago et al., 2021). Of the 450 target regions associated with cancer that were selected from public data (Chen et al., 2020; Nguyen et al., 2021; Phan et al., 2022), 402 regions were identified as DMRs in cancer patients when compared to healthy participants from the discovery cohort (Wilcoxon rank-sum test, p-values <0.05, Figure 2A and Supplementary file 1, Table S4). Of those, 339 (84.3%) regions were identified as hypermethylated (log₂ fold change [logFC] >0), and 63 (15.7%) regions as hypomethylated in cancer samples (logFC <0, Figure 2A). We next examined the genomic location of the 402 DMRs and found 100, 108, 107, and 87 DMRs that were mapped to promoter, exon, intron, and intergenic regions, respectively (Figure 2B). To understand the relationship between the differences in methylation regions and biological pathways, we performed pathway enrichment analysis using g:Profiler on hypermethylated DMRs. We detected 36 enriched pathways, including 14 from Kyoto Encyclopedia of Genes and Genomes (KEGG) and 22 from WikiPathway (WP) (Figure 2C and Supplementary file 1, Table S5). These significant pathways were known to regulate tumorigenesis of breast, gastric, hepatocellular, and colorectal cancer. Therefore, the methylation changes in the targeted regions, particularly the hypermethylated DMRs, mostly occur early in tumorigenesis and are crucial for distinguishing early-stage cancer patients from healthy individuals.

Figure 2

Download asset Open asset

Analysis of targeted methylation in cell-free DNA (cfDNA).

(A) Volcano plot shows log₂ fold change (logFC) and significance (-log₁₀ Benjamini-Hochberg adjusted p-value from Wilcoxon rank-sum test) of 450 target regions when comparing 499 cancer patients and 1076 healthy controls in the discovery cohort. There are 402 DMRs (p-value <0.05), color-coded by genomic locations. (B) Number of differentially methylated regions (DMRs) in the four genomic locations. (C) Kyoto Encyclopedia of Genes and Genomes (KEGG) and WikiPathway (WP) pathway enrichment analysis using g:Profiler for genes associated with the DMRs. A total of 36 pathways are enriched, suggesting a link between differences in methylation regions and tumorigenesis.

GWM changes in cfDNA of cancer patients

In addition to site-specific hypermethylation, hypomethylation is a significant genome-wide change that has been identified in many types of cancers (Das and Singal, 2004; Ehrlich, 2002; Hoffmann and Schulz, 2005). To investigate the methylation changes at genome-wide level, bisulfite sequencing reads from the whole-genome fraction were mapped to the human genome, split into bins of 1 Mb (2734 bins across the genome), and the reads from each bin were used to calculate methylation ratio. As expected, we observed a left-ward shift in the distribution of methylation ratio in cancer samples compared to healthy controls, indicating global hypomethylation in the cancer genome (p<0.0001, two-sample Kolmogorov-Smirnov test, Figure 3A). Of these bins, we identified 1715 (62.7%) bins as significantly hypomethylated in cancer, located across 22 autosomes of the genome (Figure 3B, Wilcoxon rank-sum test with Benjamini-Hochberg adjusting p-value <0.05). In contrast, there were only 10 bins identified as hypermethylated and mapped to chromosome 1, 2, 3, 5, 6, 7, and 12 in the cancer genome (Figure 3B). Therefore, our data confirmed the widespread hypomethylation across the genome and this would potentially serve to distinguish cancer patients from healthy controls.

Figure 3

Download asset Open asset

Genome-wide methylation changes in cell-free DNA (cfDNA) of cancer patients.

(A) Density plot showing the distribution of genome-wide methylation ratio for all cancer patients (red curve, n=499) and healthy participants (blue curve, n=1076). The left-ward shift in cancer samples indicates global hypomethylation in the cancer genome (p<0.0001, two-sample Kolmogorov-Smirnov test). (B) Log₂ fold change of methylation ratio between cancer patients and healthy participants in each bin across 22 chromosomes. Each dot indicates a bin, identified as hypermethylated (red), hypomethylated (blue), or no significant change in methylation (gray).

Increase DNA CNAs in cfDNA of cancer patients

Somatic CNAs in the cancer genome are associated with the initiation and progression of numerous cancers by altering transcriptional levels of both oncogenes and tumor suppressor genes (Shao et al., 2019). Recent studies have shown that CNAs detection could identify and quantify the fraction of ctDNA in plasma cfDNA (Baldacchino and Grech, 2020; Knuutila et al., 1999; Dereli-Öz et al., 2011). To examine CNAs at genome-wide scale, we used 1 Mb bin to determine the percentage of bins that showed significant copy number gains or losses between cancer and control group. We identified 729 bins (27.1%) with a significant gain and 976 bins (36.3%) with a significant loss in copy number across 22 chromosomes of the cancer genome (Benjamini-Hochberg adjusting p-value <0.05, Wilcoxon rank-sum test, Figure 4A). We noted that chromosome 8 had the highest proportion of bins with CNA gains, while chromosome 22 showed the highest proportion of bins with CNA losses (Figure 4B).

Figure 4 with 1 supplement see all

Download asset Open asset

Analysis of copy number aberration (CNA) in cell-free DNA (cfDNA).

(A) Log₂ fold change of DNA copy number in each bin across 22 autosomes between 499 cancer patients and 1076 healthy participants in the discovery cohort. Each dot represents a bin identified as gain (red), loss (blue), or no change (gray) in copy number. (B) Proportions of different CNA bins in each autosomes.

It is thought that the abnormal hypomethylation at genome-wide level is linked with somatic CNA, resulting in genome instability, which is an important tumorigenic event (Brennan and Flanagan, 2012; Zhang et al., 2020; Chan et al., 2013). Indeed, our data showed a significant increase in levels of CNA in hypomethylated bins compared to bins with unchanged methylation (p=0.024, Figure 4—figure supplement 1). Consistently, bins with CNA gains showed significant decreases in methylation as compared to those with CNA losses or unchanged CNA (p<0.01, Figure 4—figure supplement 1). In summary, SPOT-MAS enables comprehensive profiling of both global differences in methylation and somatic CNA as individual feature types, as well as exploring their functional links during cancer initiation and development, rendering them ideal biomarkers for cancer detection.

Fragment length analysis captured patterns of ctDNA in plasma

Several studies have shown that the fragmentation pattern of cfDNA is a non-random event mediated by apoptotic-dependent caspases and ctDNA fragments tend to be shorter than non-cancer cfDNA (Underhill et al., 2016; Mouliere et al., 2018; Lo et al., 2021; Cristiano et al., 2019; Nguyen et al., 2023b). One novel technical aspect of SPOT-MAS is the use of bisulfite sequencing data not only for methylation but also for fragment length analysis. Certain studies showed evidence of DNA degradation followed bisulfite treatment, possibly due to high temperature and low pH conditions of the bisulfite conversion procedure, while other showed that bisulfite sequencing affects large genomic DNA but not small size cfDNA (Raizis et al., 1995; Tanaka and Okamoto, 2007; Kint et al., 2018; Ehrich et al., 2007). Therefore, to demonstrate the use of bisulfite-treated cfDNA for fragment length analysis, we randomly selected three healthy controls and nine cancer samples to perform pair-wise comparison between bisulfite and non-bisulfite sequencing results. We observed a strong correlation between fragment length profile of non-bisulfite and bisulfite sequencing (Pearson’s correlation, R² >0.9, p<0.0001, Figure 5—figure supplement 1) for all 12 tested samples, indicating the feasibility of using bisulfite sequencing data for cfDNA fragment length analysis. Indeed, the fragment size distributions of bisulfite-treated cfDNA in both cancer patients and control subjects showed a peak at 167 bp (Figure 5A), corresponding to the length of DNA wrapped around histone (~ 147 bp) plus linker regions (~ 2 × 10 bp), which was in good agreement with previous studies using non-bisulfite cfDNA (Cristiano et al., 2019; Jiang et al., 2020). Importantly, our results showed that cfDNA of cancer patients was more fragmented than that of healthy participants, with a higher frequency of fragments ≤150 bp and a lower frequency of fragment >150 bp (Figure 5A).

Figure 5 with 1 supplement see all

Download asset Open asset

Analysis of fragment length patterns of circulating tumor DNA (ctDNA) in plasma.

(A) Density plot of fragment length between cancer patients (red, n=499) and healthy participants (blue, n=1076) in the discovery cohort. Inset corresponds to an x-axis expansion of short fragment (<150 bp). (B) Ratio of short to long fragments across 22 autosomes. Each dot indicates a mean ratio for each bin in cancer patients (red) and healthy participants (blue).

To examine whether the fragment length variation in cancer-derived cfDNA and non-cancer cfDNA could be position-dependent (Cristiano et al., 2019), we calculated the ratios of short (≤150 bp) to long fragments (>150 bp) across the genome in cancer patients and healthy controls. The mean ratio of short to long fragments in cancer patients was 0.29 (range 0.28–0.42), which was higher than the mean ratio of 0.27 (range 0.26–0.39) for healthy controls (Figure 5B). The changes of mean ratio were across 22 autosomes of the genome. Our results indicate that the SPOT-MAS technology can effectively capture differences in fragmentation patterns between cancer and healthy participants across the entire genome, making them potential biomarkers for the detection of ctDNA in plasma.

Profile of 4-mer EMs reflecting differences between cancer and healthy cfDNA

Associated with differences in fragment length is the differences in the DNA motifs at the end of each fragment as the consequences of differential cleavage between DNA in cancer cells and normal cells during apoptosis (Jiang et al., 2020; Jin et al., 2021). Here, we calculated the frequencies of 256 4-mer EMs of cfDNA fragments and compared them between cancer patients and healthy participants. Consistent with the fragment length features, we also confirmed the correlation of EM frequency between bisulfite and non-bisulfite sequencing results of 12 randomly selected samples, suggesting that EM profiles were reserved in bisulfite-treated cfDNA (Figure 5—figure supplement 1). Of the 256 4-mer EMs, we detected 78 motifs with increased frequencies and 106 motifs with decreased frequencies between cancer and healthy controls (Figure 6A and Supplementary file 1, Table S6).

Figure 6

Download asset Open asset

Differences in 4-mer end motif between cancer and healthy cell-free DNA (cfDNA).

(A) Heatmap shows log₂ fold change of 256 4-mer end motifs in cancer patients (n=499) compared to healthy controls (n=1076). (B) Box plots showing the top 10 motifs with significant differences in frequency between cancer patients (red) and healthy controls (blue) using Wilcoxon rank-sum test with Bonferroni-adjusted p-value <0.0001.

Interestingly, EMs beginning with cytosine (C) exhibited the highest number of EMs with significant changes of frequency in cancer samples (Figure 6A). Figure 6B shows the top 10 EMs exhibiting significant differences. Specifically, the frequencies of five motifs (CAAA, TAGA, CAGA, CAAG, and CAAT) were found to be significantly increased, while the frequencies of another five motifs (CGCT, CGCC, CGCA, GCCT, and CGTT) were significantly decreased in cancer patients (Figure 6B). Therefore, the differences in EM frequency identified by SPOT-MAS between cancer patients and healthy participants may serve as a promising target for the identification of ctDNA.

SPOT-MAS assay combining different features of cfDNA to enhance the accuracy of cancer detection

In order to increase the sensitivity of early cancer detection while avoiding the high cost of deep sequencing, a screening test should survey a wide range of ctDNA signatures (Moser et al., 2023). Therefore, we utilized multiple ctDNA signatures to construct classification models for distinguishing cancer patients from healthy individuals. To expand the feature space, we generated four additional features based on fragment length, including short, long, total fragment count, and short-to-long ratio, resulting in nine input feature groups (Figure 7A). For each feature group, we tested three different algorithms, including random forest (RF), logistic regression (LR), or extreme gradient boosting (XGB), to tune hyperparameters and select the optimal algorithms (Figure 7A). To evaluate the performance of these single-feature models, we performed 20-fold cross-validation on the discovery dataset and calculated ‘area under the curve’ (AUC) of the ‘receiver operating characteristic’ (ROC) curve. Among the nine features, EM-based model showed the highest AUC of 0.90 (95% CI: 0.89–0.92, Figure 7B) while the SHORT-based model had the lowest AUC of 0.71 (95% CI: 0.69–0.74, Figure 7B).

Figure 7 with 2 supplements see all

Download asset Open asset

Model construction and performance validation for SPOT-MAS (screening for the presence of tumor by methylation and size).

(A) Two-model construction strategies for cancer detection. (B, C) Receiver operating characteristic (ROC) curves comparing the performance of single-feature models, and two combination models (concatenate and ensemble stacking) in the discovery (B) and validation cohorts (C). (D, E) Bar charts showing the specificity and sensitivity of single-feature models and two combination models (concatenate and ensemble stacking) in the discovery (D) and validation cohorts (E). (F, G) Dot plots showing the sensitivity of SPOT-MAS assay in detection of five different cancer types in the discovery (F) and validation cohorts (G). The points and error bars represent the sensitivity and 95% confidence intervals. Feature abbreviations as follows: TM – target methylation density, GWM – genome-wide methylation density, CNA – copy number aberration, EM – 4-mer end motif, FLEN – fragment length distribution, LONG – long fragment count, SHORT – short fragment count, TOTAL – all fragment count, RATIO – ratio of short/long fragment.

To assess whether combining features could improve classification, we used two strategies to construct multi-feature models. In the first strategy, all nine feature groups were concatenated into a single data frame before being fed into the RF, LR, or XGB algorithms. Of the three algorithms, the XGB model exhibited the best performance with an AUC of 0.88 (95% CI: 0.87–0.90, Figure 7B). However, this AUC is still lower than that of the EM-based model (0.88 vs 0.90, Figure 7B). In the second strategy, we constructed an ensemble stacking model using LR to combine the prediction results of the single-feature models. We conducted an exhaustive search approach to evaluate the performance of 511 possible combinations. The stacking ensemble model based on combining eight features, including TM, GW, CNA, FLEN, LONG, TOTAL, RATIO, and EM, exhibited the best performance and outperformed the single-feature models (Supplementary file 1, Table S7), with an AUC of 0.93 (95% CI: 0.92–0.95, Figure 7B and Figure 7—figure supplement 1). In the independent validation cohort, we obtained similar results, where the ensemble model also outperformed single-feature models, with an AUC of 0.95 (95% CI: 0.93–0.96, Figure 7C).

In order to ensure cost-effectiveness and minimize psychological impact of cancer screening tests in a large population, high specificity is a crucial requirement. Accordingly, we established the cutoff value for each constructed model based on a minimum specificity threshold of 95%. Of the nine single-feature models, EM and GWM models exhibited the highest sensitivities, at 59.5% and 60.9%, respectively. The stacking ensemble model achieved a sensitivity of 73.8% and a specificity of 95.1% with a cutoff value of 0.546 in the discovery cohort (Figure 7D), and a mean sensitivity of 72.4% and a specificity of 97.0% in the validation cohort (Figure 7E). Stratification of samples by cancer types revealed that the ensemble model performed most accurately in predicting liver cancer (89.6% sensitivity), followed by CRC (82.1% sensitivity), lung cancer (78.5% sensitivity), and gastric cancer (71.6% sensitivity) (Figure 7F, Supplementary file 1, Table S8). Breast cancer had the lowest detection rate of 58.3% (91/156 patients). Importantly, the performance of our ensemble model remained consistent in the validation cohort, with liver cancer again showing the highest sensitivity (91.1%), followed by lung cancer (83.7%), CRC (83.0%), gastric cancer (61.3%), and breast cancer (49.3%) (Figure 7G, Supplementary file 1, Table S8).

Influence of clinical features on model prediction

Upon stratifying our dataset by gender, we found that there was no significant difference in the prediction of healthy status between males and females (Figure 7—figure supplement 2). However, in the case of cancer prediction, our model demonstrated higher accuracy in males than females in both the discovery and validation cohorts (Figure 7—figure supplement 2). Notably, when breast cancer samples were removed from our analysis, there was no difference in the detection rates between male and female patients (Figure 7—figure supplement 2), suggesting that the observed gender bias may be attributed to the high proportion of breast cancer patients (all females) in our cohort, who exhibited the lowest detection rate among the five cancer groups.

We next evaluated the potential confounding effect of age on our prediction model by examining the correlation between the model prediction scores and the participants’ ages. The results revealed no significant correlation, suggesting that age differences are unlikely to affect the accuracy of our model (Figure 7—figure supplement 2). With regard to cancer burden (i.e. tumor size), our model performed better for cancers with higher burden, as reflected by the higher cancer scores assigned to these cases (Figure 7—figure supplement 2). Specifically, patients with tumor diameter ≥3.5 cm were more likely to be detected than those with a diameter <3.5 cm (Figure 7—figure supplement 2). Similarly, cancer stages also influence the performance of our stacking ensemble model, showing increasing detection accuracy as the stages get more advanced. In the discovery cohort, the model’s accuracy was highest for stage IIIA cancers, with an AUC of 0.95 (95% CI 0.93–0.97), and lowest for stage I cancer, with an AUC of 0.90 (95% CI 0.86–0.95) (Figure 7—figure supplement 2). Consistently, our model performance was lower with an AUC of 0.94 (95% CI 0.89–0.98) and 0.93 (95% CI 0.90–0.96) for stage I and II cancer, respectively, increasing to 0.98 (95% CI 0.97–0.99) for stage IIIA in the validation cohort (Figure 7—figure supplement 2). These results demonstrated that our ensemble model can detect cancers at all stages found in our cohorts, despite a slightly lower performance in early stages (I and II) compared to non-metastatic stage (IIIA).

SPOT-MAS enables prediction of cancer types

The ability to predict the tissue origin of ctDNA is critical for early cancer detection as this can guide subsequent diagnostic tests and treatment. Previous studies have attempted to use either fragment length or methylation landscapes to achieve this goal (Liu et al., 2020; Cristiano et al., 2019; Klein et al., 2021). In this study, we demonstrated the ability of SPOT-MAS to identify the TOO using low-depth bisulfite sequencing to generate multiple sets of cfDNA features. We first concatenated the nine sets of cfDNA features into a single data frame and focused our analysis on 499 cancer patients with five cancer types in the discovery cohort. We then constructed an RF and two neural network models (convolutional neural network and GCNN) to predict the TOO and used 10-fold cross-validation to estimate and compare the performance of these models (Figure 8A and Figure 8—figure supplement 1). The GCNN was chosen due to its superior performance and stability (Figure 8—figure supplement 1 and Table S9).

Figure 8 with 2 supplements see all

Download asset Open asset

The performance of SPOT-MAS (screening for the presence of tumor by methylation and size) assay in prediction of the tissue of origin.

(A) Model construction strategy to predict tissue of origin by combining nine sets of cell-free DNA (cfDNA) features using graph convolutional neural networks. (B) Heatmap shows feature important scores of five cancer types. (C) Bar chart indicates the contribution of important features for classifying five different cancers. (D) Three dimensions graph represents the classification of five cancer types. (E, F) Cross-tables show agreement between the prediction (x-axis) and the reference (y-axis) to predict tissue of origin in the discovery cohort (E) and validation cohort (F).

We then used the GNNExplainer tool to measure the importance of different cfDNA features. Our results showed that breast cancer had the highest number of features with an important score >0.9 (497 features), while lung cancer had the lowest number of important features (126 features) (Figure 8B). Colorectal, gastric, and liver cancers had 363, 309, and 204 important features, respectively (Figure 8B and Supplementary file 1, Table S10). GWM and CNA were the most important features for differentiating breast, CRC, lung, gastric, and liver cancer from other cancer types, while the EM had the lowest contribution to distinguish cancer types (Figure 8C). Visualization of the 3D GCNN showed that this set of discriminative features could segregate the five different cancer types (Figure 8D), highlighting the benefits of a multimodal approach for predicting TOO.

The median accuracy for TOO identification among the five cancer types by the GCNN-based multi-feature model was 0.73 (range 0.54–0.87) in the discovery cohort (Figure 8E). The accuracy in the discovery cohort was highest for breast (0.87) and liver cancer (0.82) and lowest for gastric cancer (0.54). In the validation cohort, we obtained a slightly lower accuracy with a median of 0.70 (range 0.55–0.78). The accuracies for individual cancer types were 0.78 for breast, 0.76 for liver, 0.66 for colorectal, 0.63 for lung, and 0.55 for gastric cancer (Figure 8F). Among the five cancer types, breast cancer showed the highest TOO accuracy, possibly due to the highest number of important features detected by the model. In contrast, CRC and gastric cancer exhibited the lowest TOO accuracy with high misprediction rates between these two cancer types (0.11 and 0.19 for CRC versus gastric and gastric versus CRC, respectively). Together, our study highlights the benefits of integrating multimodal analysis with the GCNN model to capture the broad landscape of tissue-specific markers in different cancer types.

Discussion

In an era marked by a global rise in cancer-related morbidity and mortality, the development of liquid biopsy screening tests that can detect and localize cancer at an early stage holds tremendous potential to revolutionize cancer diagnosis and therapy. However, the low amount of ctDNA fragments in plasma samples of patients with early-stage cancer as well as the molecular heterogeneity of different cancer types are known as the major challenges for liquid biopsy-based multi-cancer detection assays. Thus, sequencing at high depth coverages is required to capture enough informative cancer DNA fragments in the finite plasma sample to achieve early cancer detection. In support to this notion, many groups (Liu et al., 2020; Cohen et al., 2018; Cristiano et al., 2019; Stackpole et al., 2022) have developed assays that exploited high depth coverage of sequencing to detect ctDNA fragments in plasma of early-stage cancer patients. However, this strategy might not be cost-effective and feasible for population wide screening in developing countries. Alternatively, we argued that increasing breadth of ctDNA analysis could maximize the ability to detect ctDNA fragments with heterogeneous genetic and epigenetic changes at shallow sequencing depth, thus improving the sensitivity for multi-cancer detection. To demonstrate the feasibility of this approach, we built a stacking ensemble model to combine nine different ctDNA signatures and demonstrated its superior performance on cancer detection in comparison to single-feature models (Figure 7B and C). SPOT-MAS achieved a sensitivity of 72.4% at a specificity of 97.0% for detecting five common cancer types using shallow depth sequencing. Furthermore, it can predict the TOO with an accuracy of 70%.

Previous studies have reported that methylation changes at target regions could be exploited for detecting ctDNA in plasma of patients with early-stage cancer (Xu et al., 2017; Luo et al., 2020). Consistently, in TM analysis, out of 450 TM regions chosen from previous publications (Chen et al., 2020; Nguyen et al., 2021), we identified 402 regions as significant differentially methylated regions (DMRs) in cancer patients (Figure 2A). These DMRs were enriched for regulatory regions of well-known cancer-related gene families such as PAX family genes, TBX family genes, FOX family genes, and HOX family genes, and some have previously been reported as biomarkers for non-invasive cancer diagnosis, such as SEPT9 and SHOX2 (Ilse et al., 2014; Warren et al., 2011). In addition to the targeted hypermethylation regions, our study also showed widespread hypomethylation patterns across 22 autosomes of cancer patients (Figure 3), a hallmark of cancer (Jones et al., 2019). In addition to methylation alterations, recent studies have revealed that the DNA copy number, fragmentomics profile (Cristiano et al., 2019), and EM profile (Jiang et al., 2020) at genome-wide scales have been shown as useful features for healthy-cancer classification. Therefore, we propose that the combination of these markers might provide added value to increase the performance of liquid biopsy assays. We demonstrated that the same bisulfite sequencing data could be used to identify somatic CNA (Figure 4), cancer-associated fragment length (Figure 5), and EMs (Figure 6), highlighting the advantage of SPOT-MAS in capturing the broad landscape of ctDNA signatures without high-cost deep sequencing. For cancer-associated fragment length, we pre-processed this data into five different feature tables to better reflect the information embedded within the data. Overall, we integrated multiple features of ctDNA including methylation, fragment length, EM, and copy number changes into a multi-cancer detection model and demonstrated that this approach could distinguish healthy individuals with patients from five popular cancer types. This strategy enables increased breadth of ctDNA analysis at shallow sequencing depth to overcome the limitation of low amount of ctDNA fragments in plasma samples as well as molecular heterogeneity of cancers.

The involvement and orthogonal links of the above features in the transcriptional regulation of cancer-associated genes during carcinogenesis prompted us to examine whether the combination of multiple cancer-specific signatures in cfDNA could improve the efficiency of cancer detection (Ulz et al., 2019; Ivanov et al., 2015). We first determined the performance of models constructed using individual type of cfDNA features. Next, by performing exhaustive searches for all possible combinations of single-feature models, we identified that the stacking ensemble of seven features could achieve the AUC of 0.95 (95% CI: 0.93–0.96, Figure 8C and Figure 7—figure supplement 1), which is superior to all single-feature models. Moreover, this study showed that the feature of EM achieved the highest performance among the five examined ctDNA signatures in discriminating cancer from healthy controls (Figure 8—figure supplement 2). Importantly, we found that combining EM with other ctDNA signatures in a stack model could further improve the sensitivity for detecting cancer samples, with significant improvement for lung cancer patients (Figure 8—figure supplement 2). These findings highlighted that the multimodal analysis of multiple ctDNA signatures by SPOT-MAS could increase the breadth of ctDNA feature analysis, thus enhancing the detection sensitivity while maintaining the low cost of sample preparation and sequencing. Among the five cancer types, breast cancer showed the lowest detection rate of 58.3% and 49.3% in the discovery and validation cohort, respectively. Variations in detection rates among different cancer types have been previously reported (Liu et al., 2020; Cohen et al., 2018; Klein et al., 2021). Consistently, it has been reported that the detection of breast cancer, particularly in early stages, is challenging due to the low levels of ctDNA shedding and heterogeneity of molecular subtypes of breast tumors (Liu et al., 2020). In contrast, we obtained the highest detection rate for liver cancer patients with the sensitivity of 89.6% and 91.1% in the discovery and validation cohort, respectively. Our finding is in good agreement with the literature showing that liver tumors shed high amounts of ctDNA (Caggiano et al., 2021). Despite a slightly higher AUC value in the validation cohort compared to the discovery cohort, no significant differences in AUC values were observed between the two cohorts at CV of 10 or 50 (p=0.1277, DeLong’s test). This result demonstrated the advantage of a multimodal approach to enhance ctDNA detection in plasma. We also conducted a survey of liquid biopsy assays to put our SPOT-MAS into the context of current state-of-the-art in the field. Supplementary file 1, Table S11 showed that SPOT-MAS is using the lowest sequencing depth approach (with a depth coverage of ~0.55×) and making up for this by integrating the greatest number of cfDNA features to achieve comparable performance to other assays.

For TOO identification, our results showed that the GCNN performed the best among the models tested (Figure 8—figure supplement 1 and Figure 8). GCNN has the ability to explore the similarity and mutual representation among samples, therefore achieving great success in multi-class classification tasks (Yin et al., 2022; Huang and Chung, 2022). Unlike the reference-based deconvolution approaches (Moss et al., 2018; Loyfer et al., 2023), our GCNN approach is independent of a reference methylation atlas, which was developed from tissue or cell type-specific methylation markers and thus may introduce bias due to discordance between the methylomes of tissue gDNA and plasma cfDNA (Moser et al., 2023; Zhou et al., 2022a). Although the methylation changes were reported as most predictive for TOO in previous studies (Moss et al., 2018; Loyfer et al., 2023), our results showed the contribution of each of the nine features for TOO identification (Figure 8C). In addition to GWM, fragment ratio (RATIO) and CNA are the major contributors to the discrimination of different tissue types. This finding provided additional evidence that the multimodal approach capturing the breadth of tissue-specific signatures could improve the accuracy of TOO identification (Liu et al., 2020). Our GCNN model achieved an accuracy of 0.70 for TOO prediction in validation cohort. This was comparable to the performance of CancerLocator, which was based on a probabilistic distribution model of tissue-specific methylation markers (Kang et al., 2017). Recently, Liu et al., 2020, developed a methylation atlas-based method, which achieved a higher accuracy of 93% for locating 50 types of cancer. However, this approach is based on deep genome-wide sequencing with high depth coverage of 30× (Supplementary file 1, Table S11), thus might not be a cost-effective approach for cancer screening in large populations, especially in low-income countries.

For an effective screening test, careful consideration of disease prevalence, cancer in this context, is imperative. Given the low prevalence of cancers, even a small proportion of false-positive test results arising from reduced assay specificity, if extrapolated to a national population, could significantly escalate the need for confirmatory imaging and biopsy procedures for benign abnormalities detected during screening. Thus, false positives can have substantial implications for both healthcare resources and patient well-being. Conversely, a screening test with high sensitivity ensures that most cancer cases are detected and minimizes delays in diagnosis. To address potential limitations posed by low sensitivity in cancer screening tests, we suggest that current liquid biopsy tests should be employed as a complementary approach to existing diagnostic methods to enhance cancer detection rates. To be used a stand-alone test, further work is required to improve its performance, with a particular emphasis on improving sensitivity while preserving high specificity.

There are several limitations in our study. First, despite using a large dataset of 738 cancer samples, there was an unequal distribution of samples among cancer types, with breast cancer accounting for 30.2% (223/738, Supplementary file 1, Table S2) of the total samples and gastric cancer having a much smaller representation (13.3%, Supplementary file 1, Table S2). As a result, our models may have been influenced by this imbalance, potentially introducing bias in the training and evaluation process. Therefore, future studies should consider incorporating more samples to better estimate the overall performance of the SPOT-MAS test. Second, tumor staging information was not available for 26.8% of cancer patients (198/738) in our study. For patients with unavailable staging information, their initial imaging examinations were conducted at the study hospitals. However, subsequent tests and surgical procedures were performed at a different hospital, as per the patients’ preferences. Consequently, the original study hospitals lacked access to comprehensive tumor staging data. To address this limitation, the metastasis status of these patients was obtained via communication channels between the clinicians at the study hospitals and those at the surgery hospitals. This enabled the retrieval of limited information, adhering to an established data-sharing agreement between the two institutions. To maintain the robustness of our analysis, patients diagnosed with metastatic cancer or those with indeterminate metastatic status were subsequently excluded from the study. Therefore, all cancer patients recruited in this study were confirmed to have non-metastatic tumors. Third, the cancer patients in both the discovery and validation cohort were older than the healthy participants. Age differences could be a confounding variable of methylation and could affect the model performance (Yusipov et al., 2020; Field et al., 2018). However, we observed no significant association between the participants’ age and model prediction scores (Figure 7—figure supplement 2). Fourth, the ability of SPOT-MAS to differentiate cancer patients from those with benign lesions has not been examined in this study. Fifth, this study only focused on the top 5 common cancer types, thus the current version of SPOT-MAS might misidentify cancer patients of other types, resulting in lower sensitivity to real-world application. At each research sites, blood samples from both cancer patients and healthy subjects were collected in Streck Cell-Free DNA BCT tubes and subsequently transported to a central laboratory located in Medical Genetics Institute for cfDNA isolation, library preparation, and sequencing. In a recent publication (Nguyen et al., 2023a), we have investigated the impact of logistic time and hemolysis rates of blood samples collected from different clinical sites on cfDNA concentration and sequencing quality. We did not observe any noticeable impact of such variations on cfDNA concentrations or sequencing library yields. However, future analytical validation studies using a larger sample size are required to evaluate the impact of variation in sampling technique across different clinical sites on the robustness or accuracy of assay results. Lastly, this was a retrospective cohort study and may be biased by the nature of this study design. In an interim 6-month report of a prospective study named K-DETEK, we were encouraged by the preliminary data demonstrated the ability of SPOT-MAS to detect cancer patients who exhibited no symptoms at the time of testing (Nguyen et al., 2023a). Despite these promising results, the performance of SPOT-MAS as an early cancer screening test remains to be fully validated in a large, multi-center prospective study with 1–2 years of follow-up.

In conclusion, we have developed the SPOT-MAS assay to comprehensively profile methylomic, fragmentomic, CNAs, and motif end signatures of plasma cfDNA. Our large-scale case-control study demonstrated that SPOT-MAS, with its unique combination of multimodal analysis of cfDNA signatures and innovative machine-learning algorithms, can detect and localize multiple types of cancer with high accuracy at a low-cost sequencing. These findings provided important supporting evidence for the incorporation of SPOT-MAS into clinical settings as a complementary cancer screening method for at-risk populations.

Materials and methods

Patient enrollment

Request a detailed protocol

This study recruited 738 cancer patients (223 breast cancer, 159 CRC, 122 liver cancer, 136 lung cancer, 98 gastric cancer) and 1550 healthy subjects. All cancer patients were confirmed to have one of the five cancers analyzed in this study. Cancer patients were confirmed to have cancer by abnormal imaging examination and subsequent tissue biopsy confirmation of malignancy. Cancer stages were determined by the TNM (tumor, node, metastasis) system classification according to the American Joint Committee on Cancer and the International Union for Cancer Control. Our study only recruited cancer patients with non-systemic metastatic stages (stages I–IIIA) in which cancer is localized to the primary sites and has not spread to other organs. We excluded patients who were diagnosed with metastatic stage IIIB and IV cancer. All healthy subjects were confirmed to have no history of cancer at the time of enrollment. They were followed up at 6 months and 1 year after enrollment to ensure that they did not develop cancer. Study subjects were recruited from the University of Medicine and Pharmacy, Thu Duc City Hospital, Medic Medical Center, Medical Genetics Institute in Ho Chi Minh City, Vietnam, National Cancer Hospital and Hanoi Medical University in Hanoi from May 2019 to December 2022.

Written informed consent was obtained from each participant in accordance with the Declaration of Helsinki. This study was approved by the Ethics Committee of the Medic Medical Center, University of Medicine and Pharmacy and Medical Genetics Institute, Ho Chi Minh City, Vietnam. All cancer patients were treatment-naïve at the time of blood sample collection.

Isolation of cfDNA

Request a detailed protocol

10 mL of blood was collected from each participant in a Cell-Free DNA BCT tube (Streck, USA). Plasma was collected from blood samples after centrifugation with two rounds (2000×g for 10 min and then 16,000×g for 10 min). The plasma fraction was aliquoted for long-term storage at –80°C. cfDNA was extracted from 1 mL plasma aliquots using the MagMAX Cell-Free DNA Isolation kit (Thermo Fisher, USA), according to the manufacturer’s instructions. Extracted cfDNA was quantified by the QuantiFluor dsDNA system (Promega, USA).

Bisulfite conversion and library preparation

Request a detailed protocol

According to the manufacturer’s instructions, bisulfite conversion and cfDNA purification were prepared by EZ DNA Methylation-Gold Kit (Zymo Research, D5006, USA). DNA library was prepared from bisulfite-converted DNA samples using xGen Methyl-Seq DNA Library Prep Kit (Integrated DNA Technologies, 10009824, USA) with Adaptase technology, according to the manufacturer’s instructions. The QuantiFlour dsDNA system (Promega, USA) was used to analyze the concentration of DNA.

Target region capture, whole-genome hybridization, and sequencing

Request a detailed protocol

DNA from library products were pooled equally, hybridized and captured using The XGen hybridization and wash kit (Integrated DNA Technologies, 1072281, USA), together with our customized panel of xGen Lockdown Probes including 450 regions across 18,000 CpG sites (Integrated DNA Technologies, USA). The construction of panel was built as previously described (Nguyen et al., 2022a; Chen et al., 2020; Nguyen et al., 2021). After hybridization, the flow-through product was concentrated using SpeedVac (N-Biotek, NB-503CIR, Korea) at 65°C. The samples were then added with the hybridization master mixture (hybridization buffer, hybridization enhancer, and H₂O) and denatured. Biotinylated P5 and P7 probes (P5-biotin: /5Biosg/AATGATACGGCGACCACCGA, P7-biotin: /5Biosg/CAAGCAGA AGACGGCATACGAGAT) on streptavidin magnetic beads (Invitrogen, CA, USA) were hybridized with the single-stranded DNA. The captured DNA products were amplified by a PCR with free P5 and P7 primers (P5 primer: AATGATACGGCGACCACCGA, P7 primer: CAAGCAGAAGACGGCATACGA). The concentrations of DNA libraries were determined using the QuantiFluor dsDNA system (Promega, USA). Both target and flow-through fraction were sequenced on the DNBSEQ-G400 DNA system (MGI Tech, Shenzhen, China) with 100 bp paired-end reads at a sequencing depth of 20 million reads per fraction. Data was demultiplexed by bcl2fastq (Illumina, CA, USA). FASTQ files were then examined using FastQC v. 0.11.9 and MultiQC v. 1.12.

TM analysis

Request a detailed protocol

All paired-end reads were processed by Trimmomatic v 0.32 with the option HEADCROP. The trimmed reads were then aligned by Bismark v. 0.22.3. Deduplication and sorting of BAM files were conducted using Samtools v. 1.15. Reads falling into our 450 target regions were filtered using Bedtools v. 2.28. Methylation calling was performed using Bismark methylation extractor (Nguyen et al., 2022a). Briefly, methylation ratio was measured for each target region:

M e t h y l a t i o n r a t i o = \frac{m e t h y l a t e d c y t o \sin e (C)}{m e t h y l a t e d C + u n m e t h y l a t e d C}

Methylation fold change from cancer to control was calculated for each target region. For analyzing differential methylated regions, significance level was set at p≤0.05, corresponding to a -log10 adjusted p-value ≥1.301 (Benjamini-Hochberg correction).

GWM analysis

Request a detailed protocol

The integrated bioinformatics pipeline Methy-Pipe was used to analyze GWM. We carried out the trimming step using Trimmomatic, removing adapter sequence and low-quality bases at fragment ends (Nguyen et al., 2022a). The methylation ratio for each bin was calculated as follows.

M e t h y l a t i o n r a t i o = \frac{m e t h y l a t e d c y t o \sin e (C)}{m e t h y l a t e d C + u n m e t h y l a t e d C}

Mean methylation ratio was calculated for each bin and subsequently used to plot GWM density curves. To identify bins with significant methylation changes between cancer and control group, methylation ratio in each bin of cancer samples were compared with corresponding values in control samples using Wilcoxon rank-sum test. Bins with adjusted p-value (Benjamini-Hochberg correction)≤0.05 were considered significant. Those with logFC (cancer vs control)>0 were categorized as hypermethylated bins. Those with logFC (cancer vs control)<0 were categorized as hypomethylated bins.

CNA analysis

Request a detailed protocol

CNA analysis was performed using the R-package QDNAseq (Mouliere et al., 2018). We also used 1 Mb segmentation strategy to analyze CNA. We excluded bins that felt into the low mappability and Duke blacklist regions (Ehrlich, 2002). The number of reads mapped to each bin was measured by the function ‘binReadCounts’, and GC-content correction was conducted by the functions ‘estimateCorrection’ and ‘correctBins’. The final CNA feature was derived by bin-wise normalizing and outlier smoothing with the functions ‘normalizeBins’ and ‘smoothOutlierBins’. This process resulted in a feature vector of a length of 2691 bins.

To identify significant DNA gain or loss between cancer and control group, CNA values in each bin of cancer samples were compared with corresponding values in control samples using Wilcoxon rank-sum test. Bins with adjusted p-value (Benjamini-Hochberg correction)≤0.05 were considered significant. Those with logFC (cancer vs control)>0 were categorized as significant increase. Those with logFC (cancer vs control)<0 were categorized as significant decrease.

Fragment length analysis (FLEN, SHORT, LONG, TOTAL, RATIO)

Request a detailed protocol

We used an in-house python script to convert the .bsalign files into BAM files and collected the fragment length from 100 to 250 bp, resulting in 151 possible fragment lengths for further analysis. The fragment frequency in each length (%) was measured by getting the proportion of reads with that length to the total read count in the range of 100–250 bp. Fragment length (bp) against fragment frequency (%) was plotted to obtain a FLEN distribution curve.

We divided the whole genome into 588 non-overlapping bins of 5 Mb (5 million bases) long and then extracted the read counts regarding these bins. Short fragments have lengths from 100 to 150 bp and long fragments have lengths from 151 to 250 bp. The ratio of short and long fragments was calculated by dividing the number of each fragment. All the short, long, and total read counts for each sample in 588 bins were normalized using z-score normalization. The short, long, and total normalized read counts and short/long ratios were chosen as features analyzed (SHORT, LONG, TOTAL, RATIO).

EM analysis

Request a detailed protocol

Adaptase technology (Integrated DNA Technologies, USA) was used during library preparation to ligate adapters to ssDNA fragments in a template-independent reaction (Jiang et al., 2020). This step involved adding a random tail to the 5’ end of reverse reads. Although median length of the tail was 8 bp and thus allowed trimming to obtain information for other analysis, the random-length tails did not allow exact determination of the 5’ end of the reverse reads. Therefore, EM features were determined based on the genomic coordinate of the 5’ end of the forward reads. We determined the first 4-mer sequence based on the human reference genome hg19. In 256 possible 4-mer motifs, the frequency of each motif was calculated by dividing the number of reads carrying that motif by the total number of reads, generating an EM feature vector of a length of 256 for each sample.

Construction of machine learning models

Request a detailed protocol

All samples in the discovery cohort were used for model training to classify if a sample is cancerous or not. For every feature type (TM, GWM, CNA, FLEN, SHORT, LONG, TOTAL, RATIO, and EM), three machine learning algorithms, including LR, RF, and XGB, were applied. By using the ‘GridSearchCV’ function in the scikit-learn (v.1.0.2), model hyperparameters with the best performance were chosen with ‘CV’ parameter (cross-validation) set to 5. The best hyperparameters for each algorithm were found using function ‘best_params_’ implemented in GridSearchCV. Subsequently, feature selection was performed for each algorithm as follows: (1) for LR, the ‘penalty’ parameter with ‘l1’ (LASSO regression), ‘l2’ (Ridge regression), and ‘none’ (no penalty) were examined to select the setting with the best performance; (2) for RF and XGB, a ‘SelectFromModel’ function with the ‘threshold’ was set at 0.0001 to get all features. Then, the three algorithms (LR, RF, XGB) trained with the best hyperparameters and selected features were validated using k-fold cross validation approach on the dataset of training cohort with k-fold set to 20-fold, and ‘scoring’ parameter set to ‘roc_auc’. This split the data into 20 groups, in which 19 groups were model-fitted and the remaining group was tested, which resulted in 20 ‘roc_auc’ scores. The average of these scores was used to obtain the prediction performance of each model. The model with the highest ‘roc_auc’ average score was chosen (either LR, RF, or XGB). Ensembled models were constructed by combining probability scores of nine single-feature base models (TM, GWM, CNA, FLEN, SHORT, LONG, TOTAL, RATIO, EM) with different combination using LR, resulting in one probability score for every sample. An extensive search was performed to evaluate the performance of all possible combinations (n=511) and the combination with highest AUC was selected as the final model. The model cutoff was set at the threshold specificity of >95%. This combination model performance was evaluated on an independent validation dataset to examine the model classification power.

In addition to the stacking ensemble, another combinatory strategy was examined. Instead of combining nine base models, we generated a single dataframe consisting of raw data of all nine features. The model hyperparameters tuning and features selecting were followed the same strategy as described above. After choosing the best algorithm, the model performance was also evaluated using the same external validation dataset.

Construction of models for TOO

Strategy 1: RF model

Request a detailed protocol

A single data frame of nine features in discovery cohort was used to train the RF to classify five cancer types. By using the ‘GridSearchCV’ function in the scikit-learn (v.1.0.2), model hyperparameters with the best performance were chosen with ‘CV’ parameter (cross-validation) set to 3 and ‘class_weight’ parameter set to ‘balanced’. The best hyperparameters were found by function ‘best_params_’. Then, the model was validated using k-fold cross-validation approach on the training cohort with k-fold set to 10-fold and its performance was evaluated on the validation cohort.

Strategy 2: DNN model

Request a detailed protocol

Backpropagation trained the H₂O deep neural network (DNN) (multi-layer feedforward artificial neural network) (H₂O package, version 3.36.1.2) with stochastic gradient descent. The random grid search was selected as previously described (Nguyen et al., 2022a).

Strategy 3: GCNN model

Request a detailed protocol

The model training utilized an input graph formed from a discovery dataset and a validation dataset as transudative setting (Chen et al., 2020) comprising patients diagnosed with five types of cancer: breast, colorectal (CRC), gastric, liver, and lung. The discovery dataset contains a set of sample-label pairs $J = {(X_{i}, Y_{i}) | i = 1, \dots, N}$ , where $X_{i}$ represents the $i$ th sample and $Y_{i}$ represents $i$ th label, and $N$ is the number of sample-label pairs. For each $X_{i}$ in the discovery dataset, a node’s feature vector $f = {F_{0}, \dots, F_{d}} \in R^{d}$ is constructed by combining groups of features, where $F_{i}$ is the $i$ th feature, $d$ is the number of features. The same procedure was applied for the independent validation dataset. To construct an interaction graph between cancer nodes, we employed the k-nearest neighbors’ algorithm. An interaction graph defined as $G = (V, E)$ , where $V = {X_{i} | i = 1, \dots, N}$ is a node set formed by the discovery samples, and $E = {e_{i j}}$ is an edge set, where $e_{i j}$ denotes an edge. Given $N$ nodes in the node set, that is $|V| = N$ , a graph topology $A \in R^{N \times N}$ is defined by:

A_{i j} = {\begin{cases} 1, e_{i j} \in E a n d d_{i j} < δ \\ 0, o t h e r w i s e \end{cases}

where $d_{i j}$ is the Euclidean distance of node $i$ and $j$ , and $δ$ is set to 0.8.

In accordance with Nguyen et al., 2021, a GCNN was constructed for the purpose of TOO classification. The network comprised three message-passing layers, each with a hidden size of 44 and a head number of 4. TOO classification was approached as a node classification problem, wherein the model assigned each node to one of five cancer types: breast, colorectal, gastric, liver, or lung cancer. Focal loss was employed for multi-class classification optimization and the Adam optimizer was utilized for gradient-based optimization. A 10-fold cross-validation approach was implemented on the discovery dataset; nine groups were used for model training and one group for evaluation. The optimal model was selected based on its ability to achieve the highest accuracy on the validation set during 10-fold cross-validation. This model was subsequently applied to an independent validation dataset consisting of 239 cancer patients across five cancer types to obtain the performance of TOO classification.

Given the predictions of trained model and the graph topology, we estimated the feature importance score by the GNN Explainer. The feature was considered important if it satisfied:

F_{i} > δ_{f}

where $F_{i}$ is the important score of $i$ th feature estimated by the GNN Explainer, $δ_{f}$ is the chosen cutoff and was set to 0.9.

Statistical analysis

Request a detailed protocol

This study used either the Wilcoxon rank-sum test or t-test to find statistically significant differences between cancer and control. The Kolmogorov-Smirnov test was used to decide whether two cohorts have the same statistical distribution. The Benjamini-Hochberg correction was used to correct p-value for multiple comparisons (with a corrected p-value cutoff α≤0.05). DeLong’s test was used to compare the differences between AUCs. All statistical analyses were performed using R (4.1.0) packages, including ggplot2, pROC, and caret. 95% confident interval (95% CI) was presented in a bracket next to a value accordingly.

Data availability

All draw data supporting the results in this paper are available at https://doi.org/10.6084/m9.figshare.24145392.v1. The code used in the study is publicly available at https://github.com/GS-ECD-Research/Multimodal_analysis_for_ECD, (copy archived at Doan, 2023). Sequencing data (fastqc files) from individuals are only available to interested researchers upon individual requests due to ethical restrictions imposed by The Ethic Committee of University of Medicine and Pharmacy at Ho Chi Minh City, Vietnam. To access the data, please contact the senior author Dr Le Son Tran, Medical Genetic Institute (MGI); Address: 186-188 Nguyen Duy Duong, District 10, Ho Chi Minh city, Vietnam; Email: leson1808@gmail.com. Interested researchers when requesting access to the sequencing data need to submit a data sharing agreement. The interested researchers agree to be bound by the terms and conditions in the agreement. The Director of Medical Genetics Institute will assess the data sharing agreement.

The following data sets were generated

1. Nguyen C
(2023) figshare
ELIFE pan-cancer data.

https://doi.org/10.6084/m9.figshare.24145392.v1

References

1. Baldacchino S
2. Grech G
(2020) Somatic copy number aberrations in metastatic patients: The promise of liquid biopsies
Seminars in Cancer Biology 60:302–310.

https://doi.org/10.1016/j.semcancer.2019.12.014
- PubMed
- Google Scholar
1. Brennan K
2. Flanagan JM
(2012) Is there a link between Genome-Wide Hypomethylation in Blood and Cancer Risk?
Cancer Prevention Research 5:1345–1357.

https://doi.org/10.1158/1940-6207.CAPR-12-0316
- Google Scholar
1. Buitrago D
2. Labrador M
3. Arcon JP
4. Lema R
5. Flores O
6. Esteve-Codina A
7. Blanc J
8. Villegas N
9. Bellido D
10. Gut M
11. Dans PD
12. Heath SC
13. Gut IG
14. Brun Heath I
15. Orozco M
(2021) Impact of DNA methylation on 3D genome structure
Nature Communications 12:3243.

https://doi.org/10.1038/s41467-021-23142-8
- PubMed
- Google Scholar
1. Caggiano C
2. Celona B
3. Garton F
4. Mefford J
5. Black BL
6. Henderson R
7. Lomen-Hoerth C
8. Dahl A
9. Zaitlen N
(2021) Comprehensive cell type decomposition of circulating cell-free DNA with CelFiE
Nature Communications 12:2717.

https://doi.org/10.1038/s41467-021-22901-x
- PubMed
- Google Scholar
1. Chan KCA
2. Jiang P
3. Chan CWM
4. Sun K
5. Wong J
6. Hui EP
7. Chan SL
8. Chan WC
9. Hui DSC
10. Ng SSM
11. Chan HLY
12. Wong CSC
13. Ma BBY
14. Chan ATC
15. Lai PBS
16. Sun H
17. Chiu RWK
18. Lo YMD
(2013) Noninvasive detection of cancer-associated genome-wide hypomethylation and copy number aberrations by plasma DNA bisulfite sequencing
PNAS 110:18761–18768.

https://doi.org/10.1073/pnas.1313995110
- Google Scholar
1. Chen X
2. Gole J
3. Gore A
4. He Q
5. Lu M
6. Min J
7. Yuan Z
8. Yang X
9. Jiang Y
10. Zhang T
11. Suo C
12. Li X
13. Cheng L
14. Zhang Z
15. Niu H
16. Li Z
17. Xie Z
18. Shi H
19. Zhang X
20. Fan M
21. Wang X
22. Yang Y
23. Dang J
24. McConnell C
25. Zhang J
26. Wang J
27. Yu S
28. Ye W
29. Gao Y
30. Zhang K
31. Liu R
32. Jin L
(2020) Non-invasive early detection of cancer four years before conventional diagnosis using a blood test
Nature Communications 11:17316.

https://doi.org/10.1038/s41467-020-17316-z
- Google Scholar
1. Cohen JD
2. Li L
3. Wang Y
4. Thoburn C
5. Afsari B
6. Danilova L
7. Douville C
8. Javed AA
9. Wong F
10. Mattox A
11. Hruban RH
12. Wolfgang CL
13. Goggins MG
14. Dal Molin M
15. Wang TL
16. Roden R
17. Klein AP
18. Ptak J
19. Dobbyn L
20. Schaefer J
21. Silliman N
22. Popoli M
23. Vogelstein JT
24. Browne JD
25. Schoen RE
26. Brand RE
27. Tie J
28. Gibbs P
29. Wong HL
30. Mansfield AS
31. Jen J
32. Hanash SM
33. Falconi M
34. Allen PJ
35. Zhou S
36. Bettegowda C
37. Diaz LA
38. Tomasetti C
39. Kinzler KW
40. Vogelstein B
41. Lennon AM
42. Papadopoulos N
(2018) Detection and localization of surgically resectable cancers with a multi-analyte blood test
Science 359:926–930.

https://doi.org/10.1126/science.aar3247
- PubMed
- Google Scholar
1. Cristiano S
2. Leal A
3. Phallen J
4. Fiksel J
5. Adleff V
6. Bruhm DC
7. Jensen SØ
8. Medina JE
9. Hruban C
10. White JR
11. Palsgrove DN
12. Niknafs N
13. Anagnostou V
14. Forde P
15. Naidoo J
16. Marrone K
17. Brahmer J
18. Woodward BD
19. Husain H
20. van Rooijen KL
21. Ørntoft M-BW
22. Madsen AH
23. van de Velde CJH
24. Verheij M
25. Cats A
26. Punt CJA
27. Vink GR
28. van Grieken NCT
29. Koopman M
30. Fijneman RJA
31. Johansen JS
32. Nielsen HJ
33. Meijer GA
34. Andersen CL
35. Scharpf RB
36. Velculescu VE
(2019) Genome-wide cell-free DNA fragmentation in patients with cancer
Nature 570:385–389.

https://doi.org/10.1038/s41586-019-1272-6
- PubMed
- Google Scholar
1. Das PM
2. Singal R
(2004) DNA methylation and cancer
Journal of Clinical Oncology 22:4632–4642.

https://doi.org/10.1200/JCO.2004.07.151
- PubMed
- Google Scholar
(2011) Studies of genomic copy number changes in human cancers reveal signatures of DNA replication stress
Molecular Oncology 5:308–314.

https://doi.org/10.1016/j.molonc.2011.05.002
- PubMed
- Google Scholar
Software
1. Doan NNT
(2023) Multimodal analysis of Methylomics and Fragmentomics in plasma cell-free DNA for multi-cancer early detection and localization, version swh:1:rev:1ac97f7c7c7e80ae212cee5609da1bb881bef796
Software Heritage.

https://archive.softwareheritage.org/swh:1:dir:c98b788dc01cbb3df4c25714e60880e79198b602;origin=https://github.com/GS-ECD-Research/Multimodal_analysis_for_ECD;visit=swh:1:snp:f9e2fe409d49b72b58573c5b2300c17cead514ff;anchor=swh:1:rev:1ac97f7c7c7e80ae212cee5609da1bb881bef796
1. Ehrich M
2. Zoll S
3. Sur S
4. van den Boom D
(2007) A new method for accurate assessment of DNA quality after bisulfite treatment
Nucleic Acids Research 35:e29.

https://doi.org/10.1093/nar/gkl1134
- PubMed
- Google Scholar
1. Ehrlich M
(2002) DNA methylation in cancer: too much, but also too little
Oncogene 21:5400–5413.

https://doi.org/10.1038/sj.onc.1205651
- PubMed
- Google Scholar
1. Field AE
2. Robertson NA
3. Wang T
4. Havas A
5. Ideker T
6. Adams PD
(2018) DNA methylation clocks in aging: categories, causes, and consequences
Molecular Cell 71:882–895.

https://doi.org/10.1016/j.molcel.2018.08.008
- PubMed
- Google Scholar
1. Gao Q
2. Zeng Q
3. Wang Z
4. Li C
5. Xu Y
6. Cui P
7. Zhu X
8. Lu H
9. Wang G
10. Cai S
11. Wang J
12. Fan J
(2022) Circulating cell-free DNA for cancer early detection
Innovation 3:100259.

https://doi.org/10.1016/j.xinn.2022.100259
- PubMed
- Google Scholar
1. Hawkes N
(2019) Cancer survival data emphasise importance of early diagnosis
BMJ 364:l408.

https://doi.org/10.1136/bmj.l408
- Google Scholar
1. Hoffmann MJ
2. Schulz WA
(2005) Causes and consequences of DNA hypomethylation in human cancer
Biochemistry and Cell Biology = Biochimie et Biologie Cellulaire 83:296–321.

https://doi.org/10.1139/o05-036
- PubMed
- Google Scholar
1. Huang Y
2. Chung ACS
(2022) Disease prediction with edge-variational graph convolutional networks
Medical Image Analysis 77:102375.

https://doi.org/10.1016/j.media.2022.102375
- PubMed
- Google Scholar
(2014)
Analysis of SHOX2 Methylation as an Aid to Cytology in Lung Cancer Diagnosis

Cancer Genomics & Proteomics 11:251–258.
- Google Scholar
1. Im YR
2. Tsui DWY
3. Diaz LA
4. Wan JCM
(2021) Next-Generation Liquid Biopsies: embracing data science in Oncology
Trends in Cancer 7:283–292.

https://doi.org/10.1016/j.trecan.2020.11.001
- PubMed
- Google Scholar
1. Ivanov M
2. Baranova A
3. Butler T
4. Spellman P
5. Mileyko V
(2015) Non-random fragmentation patterns in circulating cell-free DNA reflect epigenetic regulation
BMC Genomics 16:S1.

https://doi.org/10.1186/1471-2164-16-S13-S1
- PubMed
- Google Scholar
1. Jiang P
2. Sun K
3. Peng W
4. Cheng SH
5. Ni M
6. Yeung PC
7. Heung MMS
8. Xie T
9. Shang H
10. Zhou Z
11. Chan RWY
12. Wong J
13. Wong VWS
14. Poon LC
15. Leung TY
16. Lam WKJ
17. Chan JYK
18. Chan HLY
19. Chan KCA
20. Chiu RWK
21. Lo YMD
(2020) Plasma dna end-motif profiling as a fragmentomic marker in cancer, pregnancy, and transplantation
Cancer Discovery 10:664–673.

https://doi.org/10.1158/2159-8290.CD-19-0622
- PubMed
- Google Scholar
1. Jin C
2. Liu X
3. Zheng W
4. Su L
5. Liu Y
6. Guo X
7. Gu X
8. Li H
9. Xu B
10. Wang G
11. Yu J
12. Zhang Q
13. Bao D
14. Wan S
15. Xu F
16. Lai X
17. Liu J
18. Xing J
(2021) Characterization of fragment sizes, copy number aberrations and 4-mer end motifs in cell-free DNA of hepatocellular carcinoma for enhanced liquid biopsy-based cancer detection
Molecular Oncology 15:2377–2389.

https://doi.org/10.1002/1878-0261.13041
- PubMed
- Google Scholar
(2019) Epigenetic therapy in immune-oncology
Nature Reviews. Cancer 19:151–161.

https://doi.org/10.1038/s41568-019-0109-9
- PubMed
- Google Scholar
(2017) Estimating cost savings from early cancer diagnosis
Data 2:30.

https://doi.org/10.3390/data2030030
- Google Scholar
1. Kang S
2. Li Q
3. Chen Q
4. Zhou Y
5. Park S
6. Lee G
7. Grimes B
8. Krysan K
9. Yu M
10. Wang W
11. Alber F
12. Sun F
13. Dubinett SM
14. Li W
15. Zhou XJ
(2017) CancerLocator: non-invasive cancer diagnosis and tissue-of-origin prediction using methylation profiles of cell-free DNA
Genome Biology 18:53.

https://doi.org/10.1186/s13059-017-1191-5
- PubMed
- Google Scholar
(2018) Evaluation of bisulfite kits for DNA methylation profiling in terms of DNA fragmentation and DNA recovery using digital PCR
PLOS ONE 13:e0199091.

https://doi.org/10.1371/journal.pone.0199091
- PubMed
- Google Scholar
1. Klein EA
2. Richards D
3. Cohn A
4. Tummala M
5. Lapham R
6. Cosgrove D
7. Chung G
8. Clement J
9. Gao J
10. Hunkapiller N
11. Jamshidi A
12. Kurtzman KN
13. Seiden MV
14. Swanton C
15. Liu MC
(2021) Clinical validation of a targeted methylation-based multi-cancer early detection test using an independent validation set
Annals of Oncology 32:1167–1177.

https://doi.org/10.1016/j.annonc.2021.05.806
- PubMed
- Google Scholar
1. Knuutila S
2. Aalto Y
3. Autio K
4. Björkqvist AM
5. El-Rifai W
6. Hemmer S
7. Huhta T
8. Kettunen E
9. Kiuru-Kuhlefelt S
10. Larramendy ML
11. Lushnikova T
12. Monni O
13. Pere H
14. Tapper J
15. Tarkkanen M
16. Varis A
17. Wasenius VM
18. Wolf M
19. Zhu Y
(1999) DNA copy number losses in human neoplasms
The American Journal of Pathology 155:683–694.

https://doi.org/10.1016/S0002-9440(10)65166-8
- PubMed
- Google Scholar
1. Li J
2. Han X
3. Yu X
4. Xu Z
5. Yang G
6. Liu B
7. Xiu P
(2018) Clinical applications of liquid biopsy as prognostic and predictive biomarkers in hepatocellular carcinoma: circulating tumor cells and circulating tumor DNA
Journal of Experimental & Clinical Cancer Research 37:213.

https://doi.org/10.1186/s13046-018-0893-1
- PubMed
- Google Scholar
1. Liu MC
2. Oxnard GR
3. Klein EA
4. Swanton C
5. Seiden MV
6. CCGA Consortium
(2020) Sensitive and specific multi-cancer detection and localization using methylation signatures in cell-free DNA
Annals of Oncology 31:745–759.

https://doi.org/10.1016/j.annonc.2020.02.011
- PubMed
- Google Scholar
1. Lo YMD
2. Han DSC
3. Jiang P
4. Chiu RWK
(2021) Epigenetics, fragmentomics, and topology of cell-free DNA in liquid biopsies
Science 372:eaaw3616.

https://doi.org/10.1126/science.aaw3616
- PubMed
- Google Scholar
1. Loyfer N
2. Magenheim J
3. Peretz A
4. Cann G
5. Bredno J
6. Klochendler A
7. Fox-Fisher I
8. Shabi-Porat S
9. Hecht M
10. Pelet T
11. Moss J
12. Drawshy Z
13. Amini H
14. Moradi P
15. Nagaraju S
16. Bauman D
17. Shveiky D
18. Porat S
19. Dior U
20. Rivkin G
21. Or O
22. Hirshoren N
23. Carmon E
24. Pikarsky A
25. Khalaileh A
26. Zamir G
27. Grinbaum R
28. Abu Gazala M
29. Mizrahi I
30. Shussman N
31. Korach A
32. Wald O
33. Izhar U
34. Erez E
35. Yutkin V
36. Samet Y
37. Rotnemer Golinkin D
38. Spalding KL
39. Druid H
40. Arner P
41. Shapiro AMJ
42. Grompe M
43. Aravanis A
44. Venn O
45. Jamshidi A
46. Shemer R
47. Dor Y
48. Glaser B
49. Kaplan T
(2023) A DNA methylation atlas of normal human cell types
Nature 613:355–364.

https://doi.org/10.1038/s41586-022-05580-6
- PubMed
- Google Scholar
1. Luo H
2. Zhao Q
3. Wei W
4. Zheng L
5. Yi S
6. Li G
7. Wang W
8. Sheng H
9. Pu H
10. Mo H
11. Zuo Z
12. Liu Z
13. Li C
14. Xie C
15. Zeng Z
16. Li W
17. Hao X
18. Liu Y
19. Cao S
20. Liu W
21. Gibson S
22. Zhang K
23. Xu G
24. Xu R-H
(2020) Circulating tumor DNA methylation profiles enable early diagnosis, prognosis prediction, and screening for colorectal cancer
Science Translational Medicine 12:524.

https://doi.org/10.1126/scitranslmed.aax7533
- PubMed
- Google Scholar
(2023) Bridging biological cfDNA features and machine learning approaches
Trends in Genetics 39:285–307.

https://doi.org/10.1016/j.tig.2023.01.004
- PubMed
- Google Scholar
1. Moss J
2. Magenheim J
3. Neiman D
4. Zemmour H
5. Loyfer N
6. Korach A
7. Samet Y
8. Maoz M
9. Druid H
10. Arner P
11. Fu K-Y
12. Kiss E
13. Spalding KL
14. Landesberg G
15. Zick A
16. Grinshpun A
17. Shapiro AMJ
18. Grompe M
19. Wittenberg AD
20. Glaser B
21. Shemer R
22. Kaplan T
23. Dor Y
(2018) Comprehensive human cell-type methylation atlas reveals origins of circulating cell-free DNA in health and disease
Nature Communications 9:5068.

https://doi.org/10.1038/s41467-018-07466-6
- PubMed
- Google Scholar
1. Mouliere F
2. Chandrananda D
3. Piskorz AM
4. Moore EK
5. Morris J
6. Ahlborn LB
7. Mair R
8. Goranova T
9. Marass F
10. Heider K
11. Wan JCM
12. Supernat A
13. Hudecova I
14. Gounaris I
15. Ros S
16. Jimenez-Linan M
17. Garcia-Corbacho J
18. Patel K
19. Østrup O
20. Murphy S
21. Eldridge MD
22. Gale D
23. Stewart GD
24. Burge J
25. Cooper WN
26. van der Heijden MS
27. Massie CE
28. Watts C
29. Corrie P
30. Pacey S
31. Brindle KM
32. Baird RD
33. Mau-Sørensen M
34. Parkinson CA
35. Smith CG
36. Brenton JD
37. Rosenfeld N
(2018) Enhanced detection of circulating tumor DNA by fragment size analysis
Science Translational Medicine 10:466.

https://doi.org/10.1126/scitranslmed.aat4921
- PubMed
- Google Scholar
1. Nguyen HT
2. Tran DH
3. Ngo QD
4. Pham H-AT
5. Tran T-T
6. Tran V-U
7. Pham T-VN
8. Le TK
9. Le N-AT
10. Nguyen NM
11. Vo BT
12. Nguyen LT
13. Nguyen T-CV
14. Bui QTN
15. Nguyen H-N
16. Luong BA
17. Le LGH
18. Do DM
19. Do T-TT
20. Hoang AV
21. Dinh KT
22. Phan M-D
23. Tran LS
24. Giang H
25. Nguyen H-N
(2020) Evaluation of a liquid biopsy protocol using ultra-deep massive parallel sequencing for detecting and quantifying circulation tumor DNA in colorectal cancer patients
Cancer Investigation 38:85–93.

https://doi.org/10.1080/07357907.2020.1713350
- PubMed
- Google Scholar
1. Nguyen H-N
2. Cao N-PT
3. Van Nguyen T-C
4. Le KND
5. Nguyen DT
6. Nguyen Q-TT
7. Nguyen T-HT
8. Van Nguyen C
9. Le HT
10. Nguyen M-LT
11. Nguyen TV
12. Tran VU
13. Luong BA
14. Le LGH
15. Ho QC
16. Pham H-AT
17. Vo BT
18. Nguyen LT
19. Dang A-TH
20. Nguyen SD
21. Do DM
22. Do T-TT
23. Hoang AV
24. Dinh KT
25. Phan M-D
26. Giang H
27. Tran LS
(2021) Liquid biopsy uncovers distinct patterns of DNA methylation and copy number changes in NSCLC patients with different EGFR-TKI resistant mutations
Scientific Reports 11:16436.

https://doi.org/10.1038/s41598-021-95985-6
- PubMed
- Google Scholar
1. Nguyen HT
2. Khoa Huynh LA
3. Nguyen TV
4. Tran DH
5. Thu Tran TT
6. Khang Le ND
7. Le N-AT
8. Pham T-VN
9. Le M-T
10. Quynh Pham TM
11. Nguyen TH
12. Van Nguyen TC
13. Nguyen TD
14. Tran Nguyen BQ
15. Phan M-D
16. Giang H
17. Tran LS
(2022a) Multimodal analysis of ctDNA methylation and fragmentomic profiles enhances detection of nonmetastatic colorectal cancer
Future Oncology 18:3895–3912.

https://doi.org/10.2217/fon-2022-1041
- PubMed
- Google Scholar
1. Nguyen H-T
2. Luong BA
3. Tran D-H
4. Nguyen T-H
5. Ngo QD
6. Le LGH
7. Ho QC
8. Nguyen H-HT
9. Nguyen CM
10. Tran VU
11. Pham TVN
12. Le MT
13. Le NAT
14. Le TK
15. Nguyen TL
16. Pham H-AT
17. Le HT
18. Duong HDT
19. Hoang AV
20. Nguyen HB
21. Truong Dinh K
22. Phan M-D
23. Nguyen H-N
24. Do T-TT
25. Giang H
26. Tran LS
27. Tran DT
(2022b) Ultra-Deep Sequencing of Plasma-Circulating DNA for the Detection of Tumor- Derived Mutations in Patients with Nonmetastatic Colorectal Cancer
Cancer Investigation 40:354–365.

https://doi.org/10.1080/07357907.2021.2017951
- PubMed
- Google Scholar
1. Nguyen THH
2. Lu Y-T
3. Le VH
4. Bui VQ
5. Nguyen LH
6. Pham NH
7. Phan TH
8. Nguyen HT
9. Tran VS
10. Bui CV
11. Vo VK
12. Nguyen PTN
13. Dang HHP
14. Pham VD
15. Cao VT
16. Nguyen TD
17. Nguyen LHD
18. Phan NM
19. Nguyen TH
20. Nguyen VTC
21. Pham TMQ
22. Tran VU
23. Le MP
24. Vo DH
25. Tran TMT
26. Nguyen MN
27. Nguyen TT
28. Tieu BL
29. Nguyen HTP
30. Truong DYA
31. Cao CTT
32. Nguyen VT
33. Le TLQ
34. Luong TLA
35. Doan TKP
36. Dao TT
37. Phan CD
38. Nguyen TX
39. Pham NT
40. Nguyen BT
41. Pham TTT
42. Le HL
43. Truong CT
44. Jasmine TX
45. Le MC
46. Phan VB
47. Truong QB
48. Tran THL
49. Huynh MT
50. Tran TQ
51. Nguyen ST
52. Tran V
53. Tran VK
54. Nguyen HN
55. Nguyen DS
56. Nguyen TQT
57. Phan TV
58. Do TT-T
59. Truong D-K
60. Tang HS
61. Phan MD
62. Giang H
63. Nguyen HN
64. Tran LS
(2023a) Clinical validation of a ctDNA-Based Assay for Multi-Cancer Detection: An Interim Report from a Vietnamese Longitudinal Prospective Cohort Study of 2795 Participants
Cancer Investigation 41:1–17.

https://doi.org/10.1080/07357907.2023.2173773
- PubMed
- Google Scholar
1. Nguyen V-C
2. Nguyen TH
3. Phan TH
4. Tran T-HT
5. Pham TTT
6. Ho TD
7. Nguyen HHT
8. Duong M-L
9. Nguyen CM
10. Nguyen Q-TB
11. Bach H-PT
12. Kim V-V
13. Pham T-A
14. Nguyen BT
15. Nguyen TNV
16. Huynh LAK
17. Tran VU
18. Tran TTT
19. Nguyen TD
20. Phu DTB
21. Phan BHH
22. Nguyen Q-TT
23. Truong D-K
24. Do T-TT
25. Nguyen H-N
26. Phan M-D
27. Giang H
28. Tran LS
(2023b) Fragment length profiles of cancer mutations enhance detection of circulating tumor DNA in patients with early-stage hepatocellular carcinoma
BMC Cancer 23:233.

https://doi.org/10.1186/s12885-023-10681-0
- PubMed
- Google Scholar
1. Pascual J
2. Attard G
3. Bidard F-C
4. Curigliano G
5. De Mattos-Arruda L
6. Diehn M
7. Italiano A
8. Lindberg J
9. Merker JD
10. Montagut C
11. Normanno N
12. Pantel K
13. Pentheroudakis G
14. Popat S
15. Reis-Filho JS
16. Tie J
17. Seoane J
18. Tarazona N
19. Yoshino T
20. Turner NC
(2022) ESMO recommendations on the use of circulating tumour DNA assays for patients with cancer: a report from the ESMO Precision Medicine Working Group
Annals of Oncology 33:750–768.

https://doi.org/10.1016/j.annonc.2022.05.520
- PubMed
- Google Scholar
1. Pham T
2. Bui L
3. Kim G
4. Hoang D
5. Tran T
6. Hoang M
(2019) Cancers in vietnam-burden and control efforts: a narrative scoping review
Cancer Control 26:1073274819863802.

https://doi.org/10.1177/1073274819863802
- PubMed
- Google Scholar
1. Pham TMQ
2. Phan TH
3. Jasmine TX
4. Tran TTT
5. Huynh LAK
6. Vo TL
7. Nai THT
8. Tran TT
9. Truong MH
10. Tran NC
11. Nguyen VTC
12. Nguyen TH
13. Nguyen THH
14. Le NDK
15. Nguyen TD
16. Nguyen DS
17. Truong DK
18. Do TTT
19. Phan M-D
20. Giang H
21. Nguyen H-N
22. Tran LS
(2023) Multimodal analysis of genome-wide methylation, copy number aberrations, and end motif signatures enhances detection of early-stage breast cancer
Frontiers in Oncology 13:1127086.

https://doi.org/10.3389/fonc.2023.1127086
- PubMed
- Google Scholar
1. Phan TH
2. Chi Nguyen VT
3. Thi Pham TT
4. Nguyen V-C
5. Ho TD
6. Quynh Pham TM
7. Tran T-H
8. Nguyen TD
9. Khang Le ND
10. Nguyen T-H
11. Duong M-L
12. Bach H-PT
13. Kim V-V
14. Pham T-A
15. Nguyen BT
16. Vo Nguyen TN
17. Nguyen TD
18. Bieu Phu DT
19. Huu Phan BH
20. Nguyen D-S
21. Truong D-K
22. Do T-TT
23. Giang H
24. Nguyen H-N
25. Phan M-D
26. Tran LS
(2022) Circulating DNA methylation profile improves the accuracy of serum biomarkers for the detection of nonmetastatic hepatocellular carcinoma
Future Oncology 18:4399–4413.

https://doi.org/10.2217/fon-2022-1218
- PubMed
- Google Scholar
(1995) A bisulfite method of 5-methylcytosine mapping that minimizes template degradation
Analytical Biochemistry 226:161–166.

https://doi.org/10.1006/abio.1995.1204
- PubMed
- Google Scholar
1. Sasieni P
2. Smittenaar R
3. Hubbell E
4. Broggio J
5. Neal RD
6. Swanton C
(2023) Modelled mortality benefits of multi-cancer early detection screening in England
British Journal of Cancer 129:72–80.

https://doi.org/10.1038/s41416-023-02243-9
- PubMed
- Google Scholar
1. Shao X
2. Lv N
3. Liao J
4. Long J
5. Xue R
6. Ai N
7. Xu D
8. Fan X
(2019) Copy number variation is highly correlated with differential gene expression: a pan-cancer study
BMC Medical Genetics 20:175.

https://doi.org/10.1186/s12881-019-0909-5
- PubMed
- Google Scholar
1. Stackpole ML
2. Zeng W
3. Li S
4. Liu C-C
5. Zhou Y
6. He S
7. Yeh A
8. Wang Z
9. Sun F
10. Li Q
11. Yuan Z
12. Yildirim A
13. Chen P-J
14. Winograd P
15. Tran B
16. Lee Y-T
17. Li PS
18. Noor Z
19. Yokomizo M
20. Ahuja P
21. Zhu Y
22. Tseng H-R
23. Tomlinson JS
24. Garon E
25. French S
26. Magyar CE
27. Dry S
28. Lajonchere C
29. Geschwind D
30. Choi G
31. Saab S
32. Alber F
33. Wong WH
34. Dubinett SM
35. Aberle DR
36. Agopian V
37. Han S-HB
38. Ni X
39. Li W
40. Zhou XJ
(2022) Cost-effective methylome sequencing of cell-free DNA for accurately detecting and locating cancer
Nature Communications 13:5566.

https://doi.org/10.1038/s41467-022-32995-6
- PubMed
- Google Scholar
1. Sung H
2. Ferlay J
3. Siegel RL
4. Laversanne M
5. Soerjomataram I
6. Jemal A
7. Bray F
(2021) Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries
CA 71:209–249.

https://doi.org/10.3322/caac.21660
- PubMed
- Google Scholar
1. Tanaka K
2. Okamoto A
(2007) Degradation of DNA by bisulfite treatment
Bioorganic & Medicinal Chemistry Letters 17:1912–1915.

https://doi.org/10.1016/j.bmcl.2007.01.040
- PubMed
- Google Scholar
1. Ulz P
2. Perakis S
3. Zhou Q
4. Moser T
5. Belic J
6. Lazzeri I
7. Wölfler A
8. Zebisch A
9. Gerger A
10. Pristauz G
11. Petru E
12. White B
13. Roberts CES
14. John JS
15. Schimek MG
16. Geigl JB
17. Bauernhofer T
18. Sill H
19. Bock C
20. Heitzer E
21. Speicher MR
(2019) Inference of transcription factor binding from cell-free DNA enables tumor subtype prediction and early detection
Nature Communications 10:4666.

https://doi.org/10.1038/s41467-019-12714-4
- PubMed
- Google Scholar
1. Underhill HR
2. Kitzman JO
3. Hellwig S
4. Welker NC
5. Daza R
6. Baker DN
7. Gligorich KM
8. Rostomily RC
9. Bronner MP
10. Shendure J
(2016) Fragment Length of Circulating Tumor DNA
PLOS Genetics 12:e1006162.

https://doi.org/10.1371/journal.pgen.1006162
- PubMed
- Google Scholar
1. Warren JD
2. Xiong W
3. Bunker AM
4. Vaughn CP
5. Furtado LV
6. Roberts WL
7. Fang JC
8. Samowitz WS
9. Heichman KA
(2011) Septin 9 methylated DNA is a sensitive and specific blood test for colorectal cancer
BMC Medicine 9:133.

https://doi.org/10.1186/1741-7015-9-133
- PubMed
- Google Scholar
1. Xu R-H
2. Wei W
3. Krawczyk M
4. Wang W
5. Luo H
6. Flagg K
7. Yi S
8. Shi W
9. Quan Q
10. Li K
11. Zheng L
12. Zhang H
13. Caughey BA
14. Zhao Q
15. Hou J
16. Zhang R
17. Xu Y
18. Cai H
19. Li G
20. Hou R
21. Zhong Z
22. Lin D
23. Fu X
24. Zhu J
25. Duan Y
26. Yu M
27. Ying B
28. Zhang W
29. Wang J
30. Zhang E
31. Zhang C
32. Li O
33. Guo R
34. Carter H
35. Zhu J-K
36. Hao X
37. Zhang K
(2017) Circulating tumour DNA methylation markers for diagnosis and prognosis of hepatocellular carcinoma
Nature Materials 16:1155–1161.

https://doi.org/10.1038/nmat4997
- PubMed
- Google Scholar
1. Yin Y
2. Morgunova E
3. Jolma A
4. Kaasinen E
5. Sahu B
6. Khund-Sayeed S
7. Das PK
8. Kivioja T
9. Dave K
10. Zhong F
11. Nitta KR
12. Taipale M
13. Popov A
14. Ginno PA
15. Domcke S
16. Yan J
17. Schübeler D
18. Vinson C
19. Taipale J
(2017) Impact of cytosine methylation on DNA binding specificities of human transcription factors
Science 356:eaaj2239.

https://doi.org/10.1126/science.aaj2239
- PubMed
- Google Scholar
1. Yin C
2. Cao Y
3. Sun P
4. Zhang H
5. Li Z
6. Xu Y
7. Sun H
(2022) Molecular subtyping of cancer based on robust graph neural network and multi-omics data integration
Frontiers in Genetics 13:884028.

https://doi.org/10.3389/fgene.2022.884028
- PubMed
- Google Scholar
1. Yusipov I
2. Bacalini MG
3. Kalyakulina A
4. Krivonosov M
5. Pirazzini C
6. Gensous N
7. Ravaioli F
8. Milazzo M
9. Giuliani C
10. Vedunova M
11. Fiorito G
12. Gagliardi A
13. Polidoro S
14. Garagnani P
15. Ivanchenko M
16. Franceschi C
(2020) Age-related DNA methylation changes are sex-specific: a comprehensive assessment
Aging 12:24057–24080.

https://doi.org/10.18632/aging.202251
- PubMed
- Google Scholar
1. Zhang W
2. Klinkebiel D
3. Barger CJ
4. Pandey S
5. Guda C
6. Miller A
7. Akers SN
8. Odunsi K
9. Karpf AR
(2020) Global DNA Hypomethylation in Epithelial Ovarian Cancer: Passive Demethylation and Association with Genomic Instability
Cancers 12:764.

https://doi.org/10.3390/cancers12030764
- PubMed
- Google Scholar
1. Zhou X
2. Cheng Z
3. Dong M
4. Liu Q
5. Yang W
6. Liu M
7. Tian J
8. Cheng W
(2022a) Tumor fractions deciphered from circulating cell-free DNA methylation for cancer early diagnosis
Nature Communications 13:7694.

https://doi.org/10.1038/s41467-022-35320-3
- PubMed
- Google Scholar
1. Zhou Q
2. Kang G
3. Jiang P
4. Qiao R
5. Lam WKJ
6. Yu SCY
7. Ma M-JL
8. Ji L
9. Cheng SH
10. Gai W
11. Peng W
12. Shang H
13. Chan RWY
14. Chan SL
15. Wong GLH
16. Hiraki LT
17. Volpi S
18. Wong VWS
19. Wong J
20. Chiu RWK
21. Chan KCA
22. Lo YMD
(2022b) Epigenetic analysis of cell-free DNA by fragmentomic profiling
PNAS 119:44.

https://doi.org/10.1073/pnas.2209852119
- Google Scholar

Article and author information

Author details

Van Thien Chi Nguyen
1. Gene Solutions, Ho Chi Minh City, Viet Nam
2. Medical Genetics Institute, Ho Chi Minh City, Viet Nam
Contribution
Formal analysis

Contributed equally with
Trong Hieu Nguyen

Competing interests
VTCN is affiliated with Gene Solutions. The author has no other competing interests to declare

"This ORCID iD identifies the author of this article:" 0000-0002-4449-9216
Trong Hieu Nguyen
1. Gene Solutions, Ho Chi Minh City, Viet Nam
2. Medical Genetics Institute, Ho Chi Minh City, Viet Nam
Contribution
Conceptualization

Contributed equally with
Van Thien Chi Nguyen

Competing interests
HTN is affiliated with Gene Solutions. The author has no other competing interests to declare
Nhu Nhat Tan Doan
1. Gene Solutions, Ho Chi Minh City, Viet Nam
2. Medical Genetics Institute, Ho Chi Minh City, Viet Nam
Contribution
Formal analysis

Competing interests
NNTD is affiliated with Gene Solutions. The author has no other competing interests to declare
Thi Mong Quynh Pham
1. Gene Solutions, Ho Chi Minh City, Viet Nam
2. Medical Genetics Institute, Ho Chi Minh City, Viet Nam
Contribution
Formal analysis

Competing interests
TMQP is affiliated with Gene Solutions. The author has no other competing interests to declare
Giang Thi Huong Nguyen
1. Gene Solutions, Ho Chi Minh City, Viet Nam
2. Medical Genetics Institute, Ho Chi Minh City, Viet Nam
Contribution
Writing – original draft, Writing – review and editing

Competing interests
GTHN is affiliated with Gene Solutions. The author has no other competing interests to declare

"This ORCID iD identifies the author of this article:" 0000-0001-5156-4904
Thanh Dat Nguyen
1. Gene Solutions, Ho Chi Minh City, Viet Nam
2. Medical Genetics Institute, Ho Chi Minh City, Viet Nam
Contribution
Formal analysis

Competing interests
TDN is affiliated with Gene Solutions. The author has no other competing interests to declare
Thuy Thi Thu Tran
1. Gene Solutions, Ho Chi Minh City, Viet Nam
2. Medical Genetics Institute, Ho Chi Minh City, Viet Nam
Contribution
Writing – original draft, Writing – review and editing

Competing interests
TTTT is affiliated with Gene Solutions. The author has no other competing interests to declare
Duy Long Vo

University Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Conceptualization

Competing interests
No competing interests declared
Thanh Hai Phan

MEDIC Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Conceptualization

Competing interests
No competing interests declared
Thanh Xuan Jasmine

MEDIC Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Conceptualization

Competing interests
No competing interests declared
Van Chu Nguyen
1. National Cancer Hospital, Hanoi, Viet Nam
2. Hanoi Medical University, Hanoi, Viet Nam
Contribution
Conceptualization

Competing interests
No competing interests declared
Huu Thinh Nguyen

University Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Conceptualization

Competing interests
No competing interests declared

"This ORCID iD identifies the author of this article:" 0000-0002-0981-7274
Trieu Vu Nguyen

Thu Duc City Hospital, Ho Chi Minh City, Viet Nam

Contribution
Conceptualization

Competing interests
No competing interests declared
Thi Hue Hanh Nguyen
1. Gene Solutions, Ho Chi Minh City, Viet Nam
2. Medical Genetics Institute, Ho Chi Minh City, Viet Nam
Contribution
Formal analysis

Competing interests
THHN is affiliated with Gene Solutions. The author has no other competing interests to declare
Le Anh Khoa Huynh
1. Gene Solutions, Ho Chi Minh City, Viet Nam
2. Department of Biostatistics, Virginia Commonwealth University, School of Medicine, Richmond, United States
Contribution
Formal analysis

Competing interests
LAKH is affiliated with Gene Solutions. The author has no other competing interests to declare
Trung Hieu Tran
1. Gene Solutions, Ho Chi Minh City, Viet Nam
2. Medical Genetics Institute, Ho Chi Minh City, Viet Nam
Contribution
Formal analysis

Competing interests
THT is affiliated with Gene Solutions. The author has no other competing interests to declare
Quang Thong Dang

University Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared

"This ORCID iD identifies the author of this article:" 0000-0003-1206-8881
Thuy Nguyen Doan

University Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
Anh Minh Tran

University Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
Viet Hai Nguyen

University Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
Vu Tuan Anh Nguyen

University Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
Le Minh Quoc Ho

University Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
Quang Dat Tran

University Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
Thi Thu Thuy Pham

MEDIC Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
Tan Dat Ho

MEDIC Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
Bao Toan Nguyen

MEDIC Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
Thanh Nhan Vo Nguyen

MEDIC Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
Thanh Dang Nguyen

MEDIC Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Formal analysis

Competing interests
No competing interests declared
Dung Thai Bieu Phu

MEDIC Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
Boi Hoan Huu Phan

MEDIC Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
Thi Loan Vo

MEDIC Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
Thi Huong Thoang Nai

MEDIC Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
Thuy Trang Tran

MEDIC Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
My Hoang Truong

MEDIC Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
Ngan Chau Tran

MEDIC Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
Trung Kien Le

University Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
Thanh Huong Thi Tran
1. National Cancer Hospital, Hanoi, Viet Nam
2. Hanoi Medical University, Hanoi, Viet Nam
Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
Minh Long Duong
1. National Cancer Hospital, Hanoi, Viet Nam
2. Hanoi Medical University, Hanoi, Viet Nam
Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
Hoai Phuong Thi Bach
1. National Cancer Hospital, Hanoi, Viet Nam
2. Hanoi Medical University, Hanoi, Viet Nam
Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
Van Vu Kim
1. National Cancer Hospital, Hanoi, Viet Nam
2. Hanoi Medical University, Hanoi, Viet Nam
Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
The Anh Pham
1. National Cancer Hospital, Hanoi, Viet Nam
2. Hanoi Medical University, Hanoi, Viet Nam
Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
Duc Huy Tran

University Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
Trinh Ngoc An Le

University Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
Truong Vinh Ngoc Pham

University Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
Minh Triet Le

University Medical Center, Ho Chi Minh City, Viet Nam

Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
Dac Ho Vo
1. Gene Solutions, Ho Chi Minh City, Viet Nam
2. Medical Genetics Institute, Ho Chi Minh City, Viet Nam
Contribution
Formal analysis

Competing interests
DHV is affiliated with Gene Solutions. The author has no other competing interests to declare
Thi Minh Thu Tran
1. Gene Solutions, Ho Chi Minh City, Viet Nam
2. Medical Genetics Institute, Ho Chi Minh City, Viet Nam
Contribution
Formal analysis

Competing interests
TMTT is affiliated with Gene Solutions. The author has no other competing interests to declare
Minh Nguyen Nguyen
1. Gene Solutions, Ho Chi Minh City, Viet Nam
2. Medical Genetics Institute, Ho Chi Minh City, Viet Nam
Contribution
Formal analysis

Competing interests
MNN is affiliated with Gene Solutions. The author has no other competing interests to declare
Thi Tuong Vi Van
1. Gene Solutions, Ho Chi Minh City, Viet Nam
2. Medical Genetics Institute, Ho Chi Minh City, Viet Nam
Contribution
Formal analysis

Competing interests
TTVV is affiliated with Gene Solutions. The author has no other competing interests to declare
Anh Nhu Nguyen
1. Gene Solutions, Ho Chi Minh City, Viet Nam
2. Medical Genetics Institute, Ho Chi Minh City, Viet Nam
Contribution
Formal analysis

Competing interests
ANN is affiliated with Gene Solutions. The author has no other competing interests to declare
Thi Trang Tran
1. Gene Solutions, Ho Chi Minh City, Viet Nam
2. Medical Genetics Institute, Ho Chi Minh City, Viet Nam
Contribution
Formal analysis

Competing interests
TTT is affiliated with Gene Solutions. The author has no other competing interests to declare
Vu Uyen Tran
1. Gene Solutions, Ho Chi Minh City, Viet Nam
2. Medical Genetics Institute, Ho Chi Minh City, Viet Nam
Contribution
Formal analysis

Competing interests
VUT is affiliated with Gene Solutions. The author has no other competing interests to declare
Minh Phong Le
1. Gene Solutions, Ho Chi Minh City, Viet Nam
2. Medical Genetics Institute, Ho Chi Minh City, Viet Nam
Contribution
Formal analysis

Competing interests
MPL is affiliated with Gene Solutions. The author has no other competing interests to declare
Thi Thanh Do
1. Gene Solutions, Ho Chi Minh City, Viet Nam
2. Medical Genetics Institute, Ho Chi Minh City, Viet Nam
Contribution
Formal analysis

Competing interests
TTD is affiliated with Gene Solutions. The author has no other competing interests to declare
Thi Van Phan
1. Gene Solutions, Ho Chi Minh City, Viet Nam
2. Medical Genetics Institute, Ho Chi Minh City, Viet Nam
Contribution
Formal analysis

Competing interests
TVP is affiliated with Gene Solutions. The author has no other competing interests to declare
Hong-Dang Luu Nguyen
1. Gene Solutions, Ho Chi Minh City, Viet Nam
2. Medical Genetics Institute, Ho Chi Minh City, Viet Nam
Contribution
Formal analysis

Competing interests
HDN is affiliated with Gene Solutions. The author has no other competing interests to declare
Duy Sinh Nguyen
1. Gene Solutions, Ho Chi Minh City, Viet Nam
2. Medical Genetics Institute, Ho Chi Minh City, Viet Nam
Contribution
Patient consultancy and Screening

Competing interests
DSN holds equity in Gene Solutions.DSN is affiliated with Gene Solutions. The author has no other competing interests to declare
Van Thinh Cao

Pham Ngoc Thach University of Medicine, Ho Chi Minh City, Viet Nam

Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
Thanh-Thuy Thi Do

Medical Genetics Institute, Ho Chi Minh City, Viet Nam

Contribution
Patient consultancy and Screening

Competing interests
No competing interests declared
Dinh Kiet Truong

Medical Genetics Institute, Ho Chi Minh City, Viet Nam

Contribution
Supervision

Competing interests
DKT is affiliated with Gene Solutions. The author has no other competing interests to declare
Hung Sang Tang
1. Gene Solutions, Ho Chi Minh City, Viet Nam
2. Medical Genetics Institute, Ho Chi Minh City, Viet Nam
Contribution
Patient consultancy and Screening

Competing interests
HST is affiliated with Gene Solutions. The author has no other competing interests to declare
Hoa Giang
1. Gene Solutions, Ho Chi Minh City, Viet Nam
2. Medical Genetics Institute, Ho Chi Minh City, Viet Nam
Contribution
Conceptualization

Competing interests
HG holds equity in Gene Solutions. The funder Gene Solutions provided support in the form of salaries for HG who is inventor on the patent application (USPTO 17930705).HG is affiliated with Gene Solutions. The author has no other competing interests to declare
Hoai-Nghia Nguyen
1. Gene Solutions, Ho Chi Minh City, Viet Nam
2. Medical Genetics Institute, Ho Chi Minh City, Viet Nam
Contribution
Conceptualization

Competing interests
HNN holds equity in Gene Solutions. The funder Gene Solutions provided support in the form of salaries for HNN who is inventor on the patent application (USPTO 17930705).HNN is affiliated with Gene Solutions. The author has no other competing interests to declare
Minh-Duy Phan
1. Gene Solutions, Ho Chi Minh City, Viet Nam
2. Medical Genetics Institute, Ho Chi Minh City, Viet Nam
Contribution
Conceptualization, Writing – review and editing

For correspondence
pmduy@yahoo.com

Competing interests
MDP holds equity in Gene Solutions. The funder Gene Solutions provided support in the form of salaries for MDP who is inventor on the patent application (USPTO 17930705).MDP is affiliated with Gene Solutions. The author has no other competing interests to declare
Le Son Tran
1. Gene Solutions, Ho Chi Minh City, Viet Nam
2. Medical Genetics Institute, Ho Chi Minh City, Viet Nam
Contribution
Conceptualization, Writing – original draft, Writing – review and editing

For correspondence
leson1808@gmail.com

Competing interests
LST holds equity in Gene Solutions. The funder Gene Solutions provided support in the form of salaries for LST who is inventor on the patent application (USPTO 17930705).LST is affiliated with Gene Solutions. The author has no other competing interests to declare

"This ORCID iD identifies the author of this article:" 0000-0002-5382-3903

Funding

Gene Solutions

Van Thien Chi Nguyen
Trong Hieu Nguyen
Nhu Nhat Tan Doan
Thi Mong Quynh Pham
Giang Thi Huong Nguyen
Thanh Dat Nguyen
Thuy Thi Thu Tran
Thi Hue Hanh Nguyen
Le Anh Khoa Huynh
Trung Hieu Tran
Dac Ho Vo
Thi Minh Thu Tran
Minh Nguyen Nguyen
Thi Tuong Vi Van
Anh Minh Tran
Thi Trang Tran
Vu Uyen Tran
Minh Phong Le
Thi Thanh Do
Thi Van Phan
Hong-Dang Luu Nguyen
Duy Sinh Nguyen
Hung Sang Tang
Hoa Giang
Hoai-Nghia Nguyen
Minh-Duy Phan
Le Son Tran

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

We thank all participants who agreed to take part in this study, and all the clinics and hospitals who assisted in patient consultation and sample collection.

Ethics

Version history

Preprint posted: May 3, 2023
Sent for peer review: May 14, 2023
Reviewed Preprint version 1: July 25, 2023
Reviewed Preprint version 2: September 20, 2023
Version of Record published: October 11, 2023

Cite all versions

You can cite all versions using the DOI https://doi.org/10.7554/eLife.89083. This DOI represents all versions, and will always resolve to the latest one.

Copyright

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

6,101

views
450

downloads
15

citations

Views, downloads and citations are aggregated across all versions of this paper published by eLife.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Open citations (links to open the citations from this article in various online reference manager services)

Mendeley

Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)

Van Thien Chi Nguyen
Trong Hieu Nguyen
Nhu Nhat Tan Doan
Thi Mong Quynh Pham
Giang Thi Huong Nguyen
Thanh Dat Nguyen
Thuy Thi Thu Tran
Duy Long Vo
Thanh Hai Phan
Thanh Xuan Jasmine
Van Chu Nguyen
Huu Thinh Nguyen
Trieu Vu Nguyen
Thi Hue Hanh Nguyen
Le Anh Khoa Huynh
Trung Hieu Tran
Quang Thong Dang
Thuy Nguyen Doan
Anh Minh Tran
Viet Hai Nguyen
Vu Tuan Anh Nguyen
Le Minh Quoc Ho
Quang Dat Tran
Thi Thu Thuy Pham
Tan Dat Ho
Bao Toan Nguyen
Thanh Nhan Vo Nguyen
Thanh Dang Nguyen
Dung Thai Bieu Phu
Boi Hoan Huu Phan
Thi Loan Vo
Thi Huong Thoang Nai
Thuy Trang Tran
My Hoang Truong
Ngan Chau Tran
Trung Kien Le
Thanh Huong Thi Tran
Minh Long Duong
Hoai Phuong Thi Bach
Van Vu Kim
The Anh Pham
Duc Huy Tran
Trinh Ngoc An Le
Truong Vinh Ngoc Pham
Minh Triet Le
Dac Ho Vo
Thi Minh Thu Tran
Minh Nguyen Nguyen
Thi Tuong Vi Van
Anh Nhu Nguyen
Thi Trang Tran
Vu Uyen Tran
Minh Phong Le
Thi Thanh Do
Thi Van Phan
Hong-Dang Luu Nguyen
Duy Sinh Nguyen
Van Thinh Cao
Thanh-Thuy Thi Do
Dinh Kiet Truong
Hung Sang Tang
Hoa Giang
Hoai-Nghia Nguyen
Minh-Duy Phan
Le Son Tran

(2023)

Multimodal analysis of methylomics and fragmentomics in plasma cell-free DNA for multi-cancer early detection and localization

eLife 12:RP89083.

https://doi.org/10.7554/eLife.89083.3

Categories and tags

Research organism

Human

Share this article

Cite this article

Summary of clinical features of 738 cancer patients and 1550 healthy controls in discovery and validation cohorts.

Workflow of SPOT-MAS (screening for the presence of tumor by methylation and size) assay for multi-cancer detection and localization.

Analysis of targeted methylation in cell-free DNA (cfDNA).

Genome-wide methylation changes in cell-free DNA (cfDNA) of cancer patients.

Analysis of copy number aberration (CNA) in cell-free DNA (cfDNA).

Analysis of fragment length patterns of circulating tumor DNA (ctDNA) in plasma.

Differences in 4-mer end motif between cancer and healthy cell-free DNA (cfDNA).

Model construction and performance validation for SPOT-MAS (screening for the presence of tumor by methylation and size).

The performance of SPOT-MAS (screening for the presence of tumor by methylation and size) assay in prediction of the tissue of origin.

Author details

Van Thien Chi Nguyen

Contribution

Contributed equally with

Competing interests

Trong Hieu Nguyen

Contribution

Contributed equally with

Competing interests

Nhu Nhat Tan Doan

Contribution

Competing interests

Thi Mong Quynh Pham

Contribution

Competing interests

Giang Thi Huong Nguyen

Contribution

Competing interests

Thanh Dat Nguyen

Contribution

Competing interests

Thuy Thi Thu Tran

Contribution

Competing interests

Duy Long Vo

Contribution

Competing interests

Thanh Hai Phan

Contribution

Competing interests

Thanh Xuan Jasmine

Contribution

Competing interests

Van Chu Nguyen

Contribution

Competing interests

Huu Thinh Nguyen

Contribution

Competing interests

Trieu Vu Nguyen

Contribution

Competing interests

Thi Hue Hanh Nguyen

Contribution

Competing interests

Le Anh Khoa Huynh

Contribution

Competing interests

Trung Hieu Tran

Contribution

Competing interests

Quang Thong Dang

Contribution

Competing interests

Thuy Nguyen Doan

Contribution

Competing interests

Anh Minh Tran

Contribution

Competing interests

Viet Hai Nguyen

Contribution

Competing interests

Vu Tuan Anh Nguyen

Contribution

Competing interests

Le Minh Quoc Ho

Contribution

Competing interests