Evaluation of a multimodal ctDNA-based assay for detection of aggressive cancers lacking standard screening tests

Abstract

Aim: Cancers lacking standard screening (LSS) options account for approximately 70% of cancer-related deaths due to late-stage diagnosis. Circulating tumor DNA (ctDNA) is a promising biomarker for multi-cancer early detection. We previously developed SPOT-MAS, a multimodal ctDNA-based assay analyzing methylation and fragmentomic profiles, effective in detecting common cancers (breast, colorectal, liver, lung and gastric). This study extends the analysis to five LSS cancers: endometrial, esophageal, head and neck, ovarian and pancreatic.

Methods: SPOT-MAS was applied to profile cfDNA methylation and fragmentomic patterns in 739 healthy individuals and 135 LSS cancer patients.

Results: We identified 347 differentially methylated regions and observed genome-wide hypomethylation across all five LSS cancers. Esophageal and head and neck cancers showed an enrichment of short cfDNA fragments (<150 bp). Eleven 4-mer end motifs were consistently altered in cfDNA fragments across all LSS cancers. Many significant signatures were consistent with previous observations in common cancers. Notably, SPOT-MAS achieved 96.2% specificity and 74.8% overall sensitivity, with a lower sensitivity of 60.7% in early-stage cancers.

Conclusion: This proof-of-concept study demonstrates that SPOT-MAS a non-invasive test trained on five common cancer types, could detect a number of LSS cancer cases, potentially complementing existing screening programs.

Plain Language Summary

Many cancers do not have standard tests, so they are often found too late, which leads to about 70% of cancer deaths. We've created a blood test that can help find cancer early. This test has already worked well for common cancers like breast and lung cancer, and now we're testing it on five harder-to-detect cancers: endometrial, esophageal, head and neck, ovarian and pancreatic cancers. In our study, we tested our blood test on 739 healthy people and 135 patients with these difficult cancers. Our method correctly identified healthy people 96.2% of the time and found cancer cases 74.8% of the time. This new test could help with screening for types of cancer that do not have good tests right now.

Plain language summary

Article highlights.

• Approximately 70% of cancer deaths result from late-stage diagnoses in cancers lacking standard-of-care screening (LSS).

• The SPOT-MAS test is a ctDNA-based assay designed to detect five common cancers (breast, lung, colorectal, gastric and liver) and has been extended to five additional LSS cancers: endometrial, esophageal, head and neck, ovarian and pancreatic.

• The assay analyzed cfDNA methylation and fragmentomic patterns in 135 LSS cancer patients and 739 healthy individuals.

• It identified 347 differentially methylated regions and specific fragmentomic alterations.

• SPOT-MAS achieved 96.2% specificity and 74.8% overall sensitivity, with 60.7% sensitivity for early-stage detection.

• This study suggests that SPOT-MAS could serve as a non-invasive method for detecting LSS cancers.

1. Introduction

Cancer remains a major global public health challenge, driving the need for continuous advancements in research, prevention and treatment strategies, particularly given its increasing incidence and diverse impact across different regions [1]. To address the burden of cancer-related mortality, the U.S. Preventive Services Task Force (USPSTF) has issued screening recommendations for lung, prostate, colorectal, breast and cervical cancers. These screening programs have significantly contributed to increased patient survival rates and improved quality of life, ultimately leading to reduced mortality rates for these specific cancer types [2–6]. However, more than 60% of newly diagnosed malignancies lack standard-of-care (SOC) screening (LSS), accounting for over 70% of cancer-related deaths annually [7]. Among the LSS cancers, pancreatic, ovarian, endometrial, head and neck and esophageal cancers exhibit the highest mortality rates in Southeast Asia [8,9]. Despite these elevated mortality rates, single-organ screening programs have proven ineffective in detecting these cancers, largely due to their low prevalence. For instance, approximately 1,000 screening samples are needed to identify a single case of esophageal cancer [10]. Given this low prevalence, single-organ screening tests must achieve exceptionally high sensitivity and specificity to be effective in community-wide screening [10].

Current screening strategies for LSS cancers often rely on imaging diagnostics or protein markers, yet these methods have significant limitations in both sensitivity and specificity, leading to late-stage diagnoses and poor outcomes for LSS cancers. For instance, while CT scans for pancreatic cancer achieve a sensitivity of 81.4%, their specificity is only 43% [11]. Ovarian cancer screening with CA125 has a sensitivity of 50% in early stages and 80% in advanced stages, while transvaginal ultrasound (TVUS) detects 59% to 65% of Stage I cases [12,13]. In a cohort of 10,000 women, annual CA125 testing or biannual TVUS led to 300 to 350 unnecessary recalls, causing potential distress [14]. Similarly, CA125 for endometrial cancer detection shows limited effectiveness, with 52.63% sensitivity and 80% specificity [15]. For head and neck cancer, oral examination every 3 years achieved a sensitivity of 81.5% and a specificity of 84.8%, respectively, with a positive predictive value of 39.6% [16,17]. Flexible endoscopy with biopsy is the primary method for diagnosing esophageal cancer, with a sensitivity and specificity of 100% and 84%, respectively, but only for suspicious esophageal lesions [17].

To address the limitations of current screening methods, recent research has focused on developing liquid biopsy (LB) assays to detect multiple cancer types in a single test [10]. These LB assays analyze circulating tumor DNA (ctDNA) signatures, such as methylation levels and fragmentomic profiles, to enable multicancer early detection (MCED) without invasive procedures [18]. Several studies have explored ctDNA signatures as a biomarker for distinguishing LSS cancers from healthy individuals, particularly through the use of targeted methylation panels [8,9,19]. Recent studies suggest that integrating multiple ctDNA features could enhance the efficiency of early cancer detection [18]. Burgener et al. (2021) used multimodal ctDNA profiling, including mutations, methylation and fragment length, to differentiate 30 head and neck squamous cell carcinoma patients from 20 matched healthy controls [20]. However, comprehensive profiling of multiple ctDNA signatures across the five lethal LSS cancers remains underexplored.

We previously developed a multimodal ctDNA assay, Screening for the Presence of Tumor by Methylation And Size (SPOT-MAS), to profile methylation and fragmentomic signatures for multi-cancer early detection [21,22]. SPOT-MAS demonstrated its ability to capture ctDNA signatures from five common cancer types (breast, lung, liver, colorectal and gastric cancer) and has successfully facilitated early-stage cancer detection in a retrospective validation study [21,22]. In this study, we aim to evaluate the performance of SPOT-MAS in detecting LSS cancers. We hypothesize that SPOT-MAS can identify both distinct and shared ctDNA signals across five LSS cancer types, enabling effective differentiation from healthy individuals.

2. Materials & methods

2.1. Patient enrollment

This study enrolled 739 healthy volunteers, and 135 patients diagnosed with one of five cancer types for which there are no standard screening tests: endometrial cancer (n = 9), esophageal cancer (n = 29), head and neck cancer (n = 29), ovarian cancer (n = 38) and pancreatic cancer (n = 30). Cancer staging was conducted in accordance with the guidelines set by the American Joint Committee on Cancer and the International Union for Cancer Control (version VIII). Recruitment took place across 15 hospitals from May 2023 to July 2024. All cancer patients had not undergone any treatment at the time of blood sample collection.

Healthy participants were recruited from May 2023 to July 2023 during their annual health check-ups at the same hospitals where LSS cancer patients were enrolled. At the time of enrollment, these participants were verified to have no history of cancer and their cancer-free status was closely monitored through a 12-month follow-up.

The study received approval from the Ethics Committee of the Medical Genetics Institute in Ho Chi Minh City, Vietnam. Written informed consent was obtained from each participant.

2.2. Isolation of cfDNA

Blood sample (10 ml) 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 (2,000 × g for 10 min and then 16,000 × g for 10 min). The plasma fraction was aliquoted for long-term storage at -80°C. Cell free DNA (cfDNA) was extracted from 1 ml plasma aliquots using the MagMAX Cell-Free DNA Isolation kit (ThermoFisher, USA), according to the manufacturer's instructions. Extracted cfDNA was quantified by the QuantiFluor dsDNA system (Promega, USA).

2.3. Multimodal analysis of methylation & fragmentomics features of cfDNA by SPOT-MAS protocol

The SPOT-MAS assay employs a multimodal approach to capture multiple signatures of ctDNA to enhance detection sensitivity. As such, it generates a high-dimensional feature dataset, for which machine learning algorithms are required to efficiently recognize complex patterns, combine multiple features and generalize them to make accurate predictions for the presence of ctDNA [22,23]. Briefly, the SPOT-MAS workflow involves three main steps:

In step 1, a minimum of 2 ng of cfDNA isolated from peripheral blood was subjected to bisulfite conversion and adapter ligation to create a single whole-genome bisulfide library of cfDNA.

In step 2, a hybridization reaction was performed from this library to collect the target capture fraction (450 cancer-specific regions). The whole-genome fraction was then retrieved by collecting the ‘flow-through’ and hybridizing it with probes specific to the adapter sequences of the DNA library. Both the target capture fraction and the whole-genome fraction were sequenced to depths of approximately 52X and 0.55X, respectively, on the DNBSEQ-G400 DNA sequencing system (MGI Tech, Shenzhen, China), generating sequencing data with 100-bp paired-end reads at a depth of 20 million reads per fraction (target capture and genome-wide capture). The sequencing data was demultiplexed to obtain FASTQ files. All FASTQ files were then quality-checked using FastQC v. 0.11.9 and MultiQC v. 1.12. Data pre-processing was performed to generate four different sets of cfDNA features, including methylation changes at target regions (TM), genome-wide methylation (GWM), fragment length patterns and end motifs (EM) as follows:

2.4. Targeted & genome-wide methylation analysis

Bisulfite sequencing reads were mapped to the 450 target regions or 2,734 bins of 1 Mb (1 million bases) across the human genome. The methylation calling was performed using Bismark methylation extractor (16). Methylation ratio was measured for each target region (TM) or for each bin (GWM):

$$\text{Methylation ratio}=\frac{\mathit{methylated}\mathit{cyto}sin\mathit{e}\left(\mathit{C}\right)}{\mathit{methylated}\mathit{C}+\mathit{unmethylated}\mathit{C}}$$

The methylation ratio in each target region or bin of cancer samples was compared with corresponding values in control samples using the Wilcoxon rank sum test to identify regions or bins with significant methylation changes between the cancer and control group. Target regions or bins with adjusted p-value (Benjamini-Hochberg correction) ≤ 0.05 were considered significant. Those with log₂ fold change (cancer vs control) >0 were categorized as hypermethylated regions or bins. Those with log₂ fold change (cancer vs control) <0 were categorized as hypomethylated regions or bins.

2.5. Fragment length analysis

The preprocessing of data and subsequent calculations were conducted as detailed in a previous study [22]. To avoid potential contamination from index and adapter sequences, only fragments between 100 and 250 bp were included in the fragment length analysis. Briefly, read pairs with fragment lengths ranging from 100 bp to 250 bp were selected. For each fragment length, 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 to 250 bp. Fragment length (bp) against fragment frequency (%) was plotted to obtain a distribution curve.

2.6. End motif analysis

Data preprocessing and calculations followed methods outlined in an earlier study [22]. Specifically, the initial 4-nucleotide sequence was identified based on the human reference genome hg19. Among the 256 possible 4-nucleotide motifs, the frequency of each motif was calculated by dividing the count of reads containing that motif by the total number of reads, resulting in an EM feature vector of length 256 for each sample.

In step 3, these features were used as inputs for a machine learning algorithm to obtain prediction outcomes. A binary classification model of SPOT-MAS, as described in an earlier study [22], was utilized to predict cancer probabilities for all samples in this study using the data generated as described above. The model's performance was evaluated by calculating sensitivity and specificity.

2.7. Statistical analysis

The Wilcoxon rank-sum test (Mann-Whitney test) was employed to identify statistically significant differences between cancer and control groups. The Benjamini-Hochberg correction was applied to adjust p-values for multiple comparisons, with a significance threshold set at α ≤ 0.05. All statistical analyses were performed using custom Python scripts and packages, including scikit-learn and matplotlib.

3. Results

3.1. Clinical characteristics of cancer & healthy participants

In this study, we recruited 739 healthy individuals and 135 patients diagnosed with one of five LSS cancer types: endometrial cancer (9), esophageal cancer (29), head and neck cancer (29), ovarian cancer (38) and pancreatic cancer (30). Participants were recruited across 15 hospitals between May 2023 and July 2024 (Supplementary Table S1).

The 739 healthy individuals underwent annual health check-ups and were monitored for a period of 12 months to ensure their cancer-free status. All 135 patients with cancer had their diagnoses confirmed through histological analysis.

Of the 135 cancer patients included in this study, 68 were male and 69 were female, with a median age of 60 years (range: 15–85 years) (Table 1). In comparison, the healthy control group had a median age of 47 years (range: 18–90 years) and consisted of 424 females and 315 males. The cancer patients were significantly older than the control group (p < 0.0001, Mann-Whitney test, Table 1). The gender distribution between the cancer and control groups was comparable. Notably, all patients with ovarian and endometrial cancers were female, while the majority of patients with esophageal and head and neck cancers were male. The gender distribution of pancreatic cancer cases was balanced, with 50% of patients being female and 50% male. Moreover, the majority of cancer patients (79.3%) were diagnosed at metastatic stages (stage III and IV, Table 1) while 20.7% had early-stage tumors (stage I and II, Table 1).

Table 1.

Summary of clinical information of 135 cancer patients and 739 healthy subjects (total N = 874).

Clinical features		Healthy control (N = 739)		All cancers (N = 135)		p-value (All cancer vs Healthy control)	Endometrial (N = 9)		Esophageal (N = 29)		Head and neck (N = 29)		Ovarian (N = 38)		Pancreatic (N = 30)
		N	Percentage	N	Percentage		N	Percentage	N	Percentage	N	Percentage	N	Percentage	N	Percentage
Gender	Female	424	57.4%	67	49.6%	Chi-Square test p-value = 0.095	9	100.0%	1	3.4%	4	13.8%	38	100.0%	15	50.0%
	Male	315	42.6%	68	50.4%		0	0.0%	28	96.6%	25	86.2%	0	0.0%	15	50.0%
Age	Median	47		60		Mann-Whitney test p-value < 0.0001	52		60		63		52		59
	Min	18		15			30		44		39		15		24
	Max	90		85			67		74		82		85		81
Stage	Early stage (I-II)			28	20.7%		2	22.2%	6	20.7%	7	24.1%	7	18.4%	6	20%
	Metastasis stage (III-IV)			107	79.3%		7	77.8%	23	79.3%	22	75.9%	31	81.6%	24	80.0%

3.2. Distinct methylation & fragmentomic alterations in cell-free DNA of LSS cancers

We employed the SPOT-MAS multimodal analysis workflow to analyze distinct ctDNA signatures in the plasma of patients with LSS cancers. This workflow integrates two sequencing approaches: deep target sequencing and shallow genome-wide sequencing [22]. For the target sequencing fraction, we profiled methylation changes across 450 selected target regions, which were chosen based on their critical roles in the transcriptional regulation of cancer-associated genes. From these 450 regions, 347 differentially methylated regions (DMRs) were identified (Wilcoxon rank-sum test, p-values <0.05, Figure 1A & Supplementary Table S2) when comparing all LSS cancer groups to the healthy controls. We identified 323, 188, 153 and 2 DMRs in head and neck cancer, pancreatic cancer, esophageal cancer and ovarian cancer, respectively, with no DMRs observed in endometrial cancer (Figure 1A).

Figure 1.

Methylation and fragmentomic profiles of cancers lacking standard screening (LSS) tests.

(A) The number of differentially methylated regions (DMRs) profiled in each cancer type and their overlapping. (B) LSS cancers exhibited a consistent pattern of genome-wide hypomethylation. (C) Cell-free DNA fragment length distribution in each LSS cancer type compared with healthy samples. (D) The number of end motifs (EMs) significantly different between LSS cancers and control samples, along with their overlap.

In addition to site-specific DMRs, genome-wide hypomethylation is a significant and pervasive alteration associated with various cancers. To investigate methylation changes at a genome-wide level, bisulfite sequencing reads from the whole-genome fraction were mapped to the human genome, split into 2,734 bins of 1 Mb each and used to calculate the methylation ratio for each bin. Compared with healthy controls, a consistent leftward shift in the methylation ratio distribution was observed across all cancer groups, indicating widespread hypomethylation in the cell-free DNA of LSS cancer genomes (Figure 1B).

It has been shown that ctDNA shed by cancer cells tends to be shorter than that from normal cells [24,25]. However, the fragment length patterns vary among different cancer types. Consistent with previous studies, our results revealed varied fragment length patterns among the five LSS cancer types. Specifically, esophageal and head and neck cancers were characterized by an enrichment of short cfDNA fragments (<150 bp), whereas endometrial, ovarian and pancreatic cancers exhibited a reduced proportion of these short fragments (Figure 1C).

The 4-mer end motif of cfDNA fragments has been identified as a significant signature for distinguishing ctDNA from various cancer types compared with healthy cells [26,27]. In our study, we identified 219, 204, 127, 71 and 22 significant end motifs in pancreatic cancer, ovarian cancer, esophageal cancer, head and neck cancer and endometrial cancer (Figure 1D & Supplementary Table S3). Moreover, we identified 11 significant end motifs overlapping among the five LSS cancer types (Figure 1D). These motifs – AGTG, CGCA, TGAA, TGAC, TGAT, TGCA, TGCT, TGGT, TTCA, TTGA and TTGT (Supplementary Figure S1) – could serve as shared features for detecting ctDNA from LSS cancers.

By applying a multimodal analysis approach, we identified significant changes in methylation patterns, fragment length profiles and motif ends in plasma cfDNA from patients with five lethal cancers that currently lack SOC screening options.

3.3. Classification of LSS cancers using a multi modal ctDNA machine learning model

Following the identification of significant features between each LSS cancer type and healthy individuals, we next investigated whether these signatures overlap with those from our previous analysis of the five most common cancer types (breast, colorectal, lung, liver and gastric cancers). Target regions with methylation alterations consistent with the five common cancers (Figure 2A, blue bar) were predominantly found in head and neck cancer (312 shared DMRs), esophageal cancer (144 shared DMRs) and pancreatic cancer (181 shared DMRs). Notably, we identified over 1,000 bins with overlapping methylation patterns seen in the five common cancers (Figure 2A, green bar) across four LSS cancer types-esophageal, head and neck, ovarian and pancreatic cancers-while no such patterns were observed in endometrial cancer (Supplementary Table S4). Although fewer in number, shared fragment length changes (Figure 2A, orange bar) and end motifs (Figure 2A, yellow bar) were present across all five LSS cancer types (Supplementary Table S4). Consistent with our previous findings in the five common cancer types, we observed predominant CNA gains on chromosome 8 and losses on chromosome 4 (Supplementary Figure S2A). Additionally, we detected CNA losses on chromosome 13 and gains on chromosomes 18, 19 and 22 in LSS cancers, which were not observed in the analysis of the five common cancers (Supplementary Figure S2B). When stratifying the data by each LSS cancer type, we identified varying numbers of CNA bins (ranging from 8 to 213) that exhibited similar patterns to the common cancer types, with the exception of endometrial cancer (gray bar, Figure 2). These results indicate that ctDNA from the five LSS cancers exhibits overlapping signatures with those from the five common cancer types, suggesting that these signatures could serve as universal ctDNA biomarkers for detecting mutiple cancer types. This prompted us to further evaluate whether a machine learning based model previously trained on shared signatures across five common cancer types could accurately distinguish LSS cancer samples from healthy controls.

Figure 2.

Overlapping significant ctDNA signatures between the five LSS cancers and the top five common cancers, and detection accuracy by a multimodal machine learning algorithm. (A) Number of features overlapping with the top five common cancer types. (B) Accuracy of the multimodal machine learning algorithm in distinguishing between these LSS cancers and healthy individuals.

Our trained model demonstrated a specificity of 96.2% and an overall sensitivity of 74.8% with sensitivities of 66.7% for endometrial cancer, 69.0% for head and neck cancer, 73.7% for ovarian cancer, 79.3% for esophageal cancer and 80.0% for pancreatic cancer. Sensitivity was 60.7% for early-stage cancers (stage I and II), increasing to 78.5% for metastatic stages (stage III and IV) (Table 2). These findings demonstrated that our previously trained machine learning model can effectively differentiate LSS cancer samples from healthy individuals.

Table 2.

SPOTMAS performance on lacking standard screening cancer types.

	All cases		Early stage (I-II)		Metastasis stage (III-IV)
	N	Accuracy	N	Accuracy	N	Accuracy
Endometrial	9	66.7%	2	0.0%	7	85.7%
Esophageal	29	79.3%	6	66.7%	23	82.6%
Head and neck	29	69.0%	7	57.1%	22	72.7%
Ovarian	38	73.7%	7	71.4%	31	74.2%
Pancreatic	30	80.0%	6	66.7%	24	83.3%
All cancers	135	74.8%	28	60.7%	107	78.5%
Control	739	96.2%

4. Discussion

Cancers without recommended screening programs account for over 60% of newly diagnosed cases and contribute to approximately 70% of cancer-related deaths annually [7]. These LSS cancers often lack noticeable symptoms and are frequently diagnosed at advanced stages, leading to poor prognoses [13]. Early detection is thus clinically significant, as it can improve 5-year survival rates [13]. MCED tests offer a promising approach by aggregating the prevalence of multiple rare cancers and complementing existing single-cancer screening programs [10]. In the K-DETEK study involving 9,057 asymptomatic participants [21], we demonstrated that the SPOT-MAS assay could prospectively detect cancers that lack SOC screening, even though it was initially trained on samples from common cancers such as breast, gastric, colorectal, liver and lung cancer. This finding suggests that the SPOT-MAS machine learning model may capture ctDNA signatures shared across multiple cancer types, enabling the detection of more cancer types than the five it was trained. In this study, using the SPOT-MAS workflow, we identified multiple significant methylation,fragmentomics changes and motif ends, in plasma cfDNA from five LSS cancers, which overlapped with the top five common cancer types from our previous case-control studies [14]. Our findings are consistent with recent studies showing that ovarian and pancreatic cancers exhibit distinct methylation and fragment length patterns compared with healthy individuals, supporting their potential as universal ctDNA signatures for early detection [19,20,28].

Our trained model demonstrated a specificity of 96.2% and an overall sensitivity of 74.8%, with a sensitivity of 60.7% for early-stage cancers (stage I and II) and 78.5% for metastatic stages (stage III and IV). Among early-stage cancers, esophageal and pancreatic cancers showed the highest sensitivity at 66.7% (Figure 2B), while head and neck cancer had the lowest sensitivity at 57.1% (Figure 2B). For metastatic stages, despite the limited sample size, 6 out of 7 endometrial cancers (85.7%) were correctly detected, whereas head and neck cancer had the lowest detection rate with a sensitivity of 72.7%. These findings indicate that our previously trained machine learning model can effectively differentiate LSS cancer samples from those of healthy individuals. Since we observed multiple overlapping signatures between LSS cancers and the top 5 common cancers, this limits the ability of our current model to accurately distinguish LSS cancers from the cancers it was trained on. We believe that future studies will require an increase in sample size for each cancer type to enable pairwise feature selection and retraining of our multiclass model, to achieve higher accuracy in TOO identification. To assess the potential confounding effect of age differences on the performance of our model and features, we conducted a correction analysis. Consistent with our previous study, we did not observe any significant correlation between age and model performance (Supplementary Figure S3A & B) or ctDNA features, including methylation (Supplementary Figure S3C), fragment length (Supplementary Figure S3D and motif ends (Supplementary Figure S3E).

The SPOT-MAS test employs a cost-effective approach with low sequencing depth (∼0.55 × coverage) for genome-wide sequencing while maintaining comparable overall sensitivity by integrating the maximum number of cfDNA features to achieve performance similar to other assays. In contrast, the CancerSeek test demonstrated strong performance in ovarian cancer through the use of genetic mutations and protein markers, but it required extensive wet lab operations, particularly to differentiate tumor-derived mutations from clonal hematopoiesis of indeterminate potential (CHIP) [29]. The Galleri test, another non-invasive method capable of detecting over 50 cancer types, still exhibited poor performance in stage I cancers across 12 predefined types, despite utilizing a significantly higher sequencing depth and it lacked specific data for individual cancers [30] (Supplementary Table S5).

To our knowledge, this study is one of the first to comprehensively profile the methylation and fragmentomics features of ctDNA in plasma samples from five distinct LSS cancer types, highlighting the potential utility of the multimodal SPOT-MAS assay for detecting these cancers. Despite the promising results, our study has several limitations. Several factors may account for the differences in the number of significant features across cancer types, including biological variations and technical factors such as differing sample sizes. The sample sizes for each cancer type were small, with a low proportion of early-stage samples and a high proportion of metastatic cases. Therefore, further research is needed to confirm the biological differences of ctDNA signature among these LSS cancers. Beleites et al. (2012) (1) recommend a minimum of 25 samples per cancer type for a dataset containing 2500 features to ensure balanced representation. The limited sample size for each LSS cancer type, particularly for early-stage cancers in this study, constrained our ability to retrain the models. The number of early-stage cases was limited due to the lack of standard screening tests for these cancers, leading to fewer early-stage diagnoses during our recruitment process. This may impact the accuracy of our comparisons of the model's performance across different early-stage cancer types. This limitation mirrors a significant clinical challenge: patients with LSS cancers often exhibit no noticeable symptoms and are frequently diagnosed at advanced stages [10]. To address this limitation, we are actively recruiting additional patients with LSS cancers to retrain the machine learning algorithm and improve its performance in detecting these cancers. Furthermore, this study was a retrospective cohort study, which may introduce biases inherent to this study design. In the K-DETEK study, we were encouraged by the data indicating that SPOT-MAS could detect LSS cancer patients who were asymptomatic at the time of testing [21]. Our current study is an exploratory analysis; the efficacy of SPOT-MAS as an early cancer screening tool for LSS cancers remains to be fully validated in a large, multi-center prospective study.

5. Conclusion

Our study revealed shared methylation and fragmentomic signatures in plasma samples from patients with cancers lacking standard-of-care screening. This proof-of-concept study demonstrates that SPOT-MAS, a non-invasive test trained on five common cancer types, could detect a number of LSS cancer cases, potentially complementing existing screening programs. Future studies are required to further prospectively validate its performance in real-world screening setting.

Funding Statement

This work was supported by Gene Solutions (K-Discovery).

Supplemental material

Supplemental data for this article can be accessed at https://doi.org/10.1080/14796694.2024.2413266

Author contributions

CVT Nguyen, DH Vo, THH Nguyen, TTV Van, TH Tran, TH Nguyen, LAK Huynh, TD Nguyen, HDL Nguyen, UV Tran performed formal analysis. N-H Tran, TMT Ha, PTQ Le, XL Truong, LHD N, TQ Hoang, VB Nguyen, VC Le, XC Nguyen, TM Phuong Nguyen, VH Nguyen, NTNH, TNQ Dang, MH Tran, PN Nguyen, N-T Tran performed patient consultancy and screening. CVT Nguyen, DH Vo, THH Nguyen, TTV Van, TH Tran, TH Nguyen, TD Nguyen performed data curation. DS Nguyen, HS Tang, M-D Phan, H Giang, H-N Nguyen, LS Tran performed the methodology. DS Nguyen, HS Tang, M-D Phan, H Giang, H-N Nguyen, LS Tran performed conceptualization. CVT Nguyen, GTH Nguyen, LS Tran performed writing-original draft. CVT Nguyen, GTH Nguyen, M-D Phan, LS Tran performed writing-review and editing.

Financial disclosure

This work was supported by Gene Solutions (K-Discovery).

Competing interests disclosure

The authors including LS Tran, H Giang, M-D Phan, H-N Nguyen and DS Nguyen hold equity in Gene Solutions. H Giang, M-D Phan and LS Tran are inventors on the patent application (USPTO 17930705). We confirm that this does not alter our adherence to journal policies on sharing data and materials.

Ethical conduct of research

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.

Data availability statement

The authors certify that this manuscript reports original clinical trial data. The data will not be made publicly available.

Patient consent statement

Written informed consent was obtained from each participant in accordance with the Declaration of Helsinki.

Permission to reproduce material from other sources

Not applicable.