Early detection of gastric cancer with the assistance of cell-free DNA fragmentomics Authors: Xiaoli Huang, Zhengxiang Han, Ying Zhang, Hongmei Wang, Xuqi Li, Chenxi Bai, Ying Pan, Jingwen Huang, Jiaojiao Xie, Xiaohan Qian, Xuan Wang, Tao Ding, Jinpeng Zhang, Shanshan Yang, Xinyue Hong, Qifan Jing, Gang Li, Xiaohua Li Published: 2026 ABSTRACT AbstractBackgroundGastric cancer remains one of the most prevalent malignancies globally. As early-stage gastric cancer is typically asymptomatic or presents with non-specific symptoms, most patients are diagnosed at advanced stages, leading to poor survival outcomes. Effective early detection strategies are important for reducing gastric cancer-related mortality. In this study, we developed a non-invasive assay utilizing cell-free DNA to distinguish patients with early-stage gastric cancer from healthy individuals.ResultsWe performed low-depth whole genome sequencing to profile cell-free DNA and extracted three distinct features: fragment size patterns, coverage at transcription factor binding sites, and methylation-based profiles. These features were integrated via machine learning to construct a stacked ensemble model. The study included a training cohort (108 gastric cancer patients and 108 healthy controls), a temporally independent validation cohort (79 patients and 79 healthy controls), and an external validation cohort recruited from two independent centers (136 patients and 136 healthy controls). The ensemble model demonstrated robust performance, achieving area under the curve values of 0.986, 0.978, and 0.967 in the training, validation, and external cohorts, respectively. Specificity and sensitivity were 98.1% and 89.8% in the training cohort, 97.5% and 87.6% in the validation cohort, and 96.3% and 87.5% in the external cohort. Notably, the sensitivity for detecting stage I gastric cancer exceeded 85% across all cohorts.ConclusionsBy integrating multi-dimensional cell-free DNA fragmentomic features, this assay provides accurate, non-invasive detection of gastric cancer, particularly at early stages. While its performance was high, the specificity reported here may be overestimated due to the use of a strictly screened healthy control group. Nevertheless, this fragmentomic-based approach represents a promising tool to complement existing screening strategies, potentially improving early diagnosis rates.Supplementary InformationThe online version contains supplementary material available at 10.1186/s13148-026-02150-9.Keywords: Gastric cancer, Early detection, Cell-free DNA, Fragmentomics Introduction Gastric cancer is a leading cause of cancer-related death globally [1], with patient survival depending on the stage at diagnosis. While the 5-years survival rate for early-stage gastric cancer exceeds 90% [2], the lack of symptoms often delays diagnosis until advanced stages, resulting in poor outcomes. Consequently, early detection is critical to reducing mortality. Current gastric cancer screening relies primarily on endoscopic surveillance. However, its invasive nature and requirements limit patient compliance and widespread application [3]. Non-invasive tumor marker assays (e.g., CEA, CA19-9) offer an alternative to endoscopy but are hampered by low sensitivity, rendering them ineffective for early-stage screening [4]. This limitation has spurred the development of liquid biopsy. Liquid biopsy utilizes various biomarkers, including extracellular vesicles, circulating RNA, and cell-free DNA (cfDNA) [5]. Among these, cfDNA has shown particular promise. It consists of DNA fragments released into the bloodstream during cellular processes such as apoptosis and necrosis [6]. cfDNA retains genetic and epigenetic signatures from its tissue of origin, including mutations and methylation patterns. Specifically, the fraction of cfDNA derived from tumor cells, known as circulating tumor DNA (ctDNA), harbors tumor-specific alterations that are critical for early cancer detection. Despite the potential of cfDNA, current approaches have limitations. Recent studies utilizing cfDNA methylation, such as those by Lei et al. [7] and Qi et al. [8], have demonstrated the feasibility for gastric cancer detection. However, these methods have historically faced challenges in achieving high sensitivity for Stage I disease while maintaining the high specificity required for screening. Consequently, there remains a critical need for a robust, non-invasive method capable of accurately detecting gastric cancer at its earliest, most treatable stages. cfDNA fragmentomics has been developed rapidly in recent years. Fragmentomics involves multiple features, including fragment size, end motifs, and fragment coverage [9]. Genetic and epigenetic variations alter the chromatin construction and gene expression of cancer cells, resulting in changes to cfDNA fragmentomics [10]. For example, ctDNA tends to have shorter fragment sizes and specific end motifs. Fragmentomics approaches have been applied to early cancer detection with high performance [11,12]. In this study, we developed a multi-feature liquid biopsy assay based on cfDNA fragmentomics to detect gastric cancer at an early stage. Using low-depth whole-genome sequencing (WGS), we profiled three distinct cfDNA features: fragment size pattern (FSP), coverage at transcription factor binding sites (CTBS), and methylation-based profiles (MBP). These features were integrated into an ensemble machine learning model to distinguish gastric cancer patients from healthy individuals. The model was trained and initially validated using cohorts from a single center, followed by independent verification in an external cohort recruited from two additional centers. This study presents a rigorously validated, high-performance assay that holds promise as a complement to standard screening tests for the early detection of gastric cancer. Methods and materials Study population and samples The study was designed with three independent cohorts: a training cohort, an internal validation cohort, and an external validation cohort (Fig. 1). Both training and internal validation cohorts were enrolled at the Fourth Affiliated Hospital of Nanjing Medical University but were recruited during distinct time periods. The training cohort recruited participants from June 1st, 2023, to March 31st, 2024. Following quality control (QC), this cohort included 108 patients with pathologically confirmed gastric cancer and 108 healthy participants. The internal validation cohort was subsequently enrolled from April 1st, 2024, to December 31st, 2024. Initially, 160 participants were enrolled; after excluding cases that failed QC, 79 gastric cancer patients and 79 healthy individuals were included in the final analysis. Overview of datasets and workflow of model construction To assess generalizability, an external validation cohort was recruited from two independent centers between May 22, 2024, and May 21, 2025. External cohort 1 consisted of 132 participants enrolled at The Affiliated Hospital of Xuzhou Medical University, and External cohort 2 consisted of 140 participants enrolled at The First Affiliated Hospital of Xi’an Jiaotong University. Both sub-cohorts maintained a balanced design with equal numbers of gastric cancer patients and healthy controls. Data from these two centers were pooled to form the final external validation cohort (n = 272) for the primary performance evaluation. All participants provided informed consent. Peripheral blood samples (~ 10 ml) were collected from each participant prior to other screening tests, such as gastroscopy, CT scans, and tumor marker assays. Plasma was isolated from blood sample within 4 h of collection. cfDNA was then extracted within 72 h and stored at -80 ℃ for subsequent sequencing. All procedures were conducted at the respective hospitals. Sample preparations followed standardized protocols on an automated liquid handling platform to ensuring consistency. Cohort inclusion and exclusion criteria Eligible participants met the following inclusion criteria: (1) age between 18 and 80 years; (2) no prior cancer history; (3) plasma samples passing QC; and (4) provision of informed consent. Participants were classified into the gastric cancer group or healthy group based on specific criteria. The gastric cancer group included individuals diagnosed with gastric cancer via histopathological confirmation without concomitant malignancies. Disease staging was determined according to the American Joint Committee on Cancer (AJCC) Cancer Staging Manual (8th Edition) [13]. The healthy group included individuals who showed no signs of cancers based on routine physical examinations (blood tests, ultrasound and CT) and were confirmed negative for gastric cancer by gastroscopy. Exclusion criteria for the gastric cancer group were: (1) diagnosis of stage IV gastric cancer; (2) withdrawal of informed consent; or (3) sequencing data failing QC. Exclusion criteria for the healthy group were: (1) significant chronic diseases of other systems, such as severe cardiovascular diseases, uncontrolled diabetes, hypertension, or infectious diseases; (2) diagnosis of active gastric inflammation or benign gastric diseases (e.g., gastric ulcer, atrophic gastritis); (3) abnormal tumor marker results within the past year; (4) lack of informed consent; or (5) sequencing data failing QC. Sample collection and sequencing library preparations Blood samples were centrifuged at 16,000× at 4 °C for 10 min to collect plasma. Circulating cfDNA was isolated from 2 to 4 mL of plasma using the QIAamp Circulating Nucleic Acid Kit (Qiagen), following the manufacturer’s protocol. The concentration of cfDNA was quantified using Qubit dsDNA HS Assay Kit (Thermo Fisher Scientifc). Sequencing libraries for wete prepared after cfDNA isolation and quantification. The KAPA Hyper Prep Kit (KAPA Biosystems) was used for library construction according to the manufacturer’s instructions. Briefly, 5–10 ng of cfDNA was end-repaired and A-tailed, followed by adapter ligation. Library construction was automated on the Hamilton Microlab STAR automated liquid handling platform (Hamilton Company). Libraries were quantified by qPCR using the KAPA SYBR FAST qPCR Master Mix (KAPA Biosystems). Low‑coverage whole‑genome sequencing and alignment Whole-genome libraries were used for DNA nanoball generation and sequenced using 150-bp paired-end runs on the DNBSEQ-T7 platforms (MGI), targeting 8× coverage per genome. Sequencing data underwent quality controlled, including trimming with Trimmomatic. QC criteria were applied based on standardized protocols: samples were retained only if the total cfDNA input was ≥ 4 ng, the library cyclization rate was ≥ 7.5%, the peak insert size fell within 150–200 bp, and the mean deduplicated sequencing depth was ≥ 3X. The Picard toolkit (http://broadinstitute.github.io/picard/) was used to remove PCR duplicates. Qualified reads were aligned to the human reference genome (GRCh37/UCSC hg19) using the Burrows-Wheeler Aligner (BWA). Insert fragments with lengths less than 100 bp or greater than 220 bp were filtered out. This filtering step was implemented to minimize technical noise from short artifacts (< 100 bp) and to exclude high-molecular-weight genomic DNA contamination (> 220 bp), focusing the analysis on the informative mono-nucleosomal region. To eliminate the potential impact of sequencing depth variations, all plasma samples were down-sampled to a uniform coverage of 5×, except for samples with an initial raw depth of less than 5×, which were retained at their original depth. Fragment size pattern (FSP) We analyzed the ratio of short to long fragments and the fragment size distribution. The procedure for calculating the short-to-long ratio was adapted from Mathios et al. [14]. The genome was divided into non-overlapping 5 Mb bins. A total of 541 bins with an average GC content ≥ 0.3 and an average mapping rate ≥ 0.9 were retained. For each bin, the counts of short (100–150 bp) and long (151–220 bp) fragments were calculated. The 150 bp cutoff was selected to distinguish tumor-derived cfDNA, which is typically shorter, from the background non-tumor cfDNA, which peaks around 167 bp [11,14]. The results were standardized to z-scores, and the ratio of short fragments to long fragments was calculated for each bin. Coverage at transcription factor binding sites (CTBS). Coverage at transcription factor binding sites was analyzed following the method described by Doebley et al. [15]. Transcription factors (TFs) with more than 10,000 highly mappable sites were selected from GTRD database (v21.12) [16]. A subset of 854 TFs with known binding sites in the CIS-BP database (v2.00) [17] was selected. For each TF, the 10,000 mappable sites with the highest peak counts were retained to analyze coverage patterns. To profile the coverage, fragments between 100 and 220 bp within a window of -5k bp to +5k bp relative to each TF binding site were extracted. A GC-bias weight was assigned to each read. Each site was divided into 15-bp bins, and the weights of all reads with midpoints falling within each bin were summed to generate GC-corrected midpoint coverage profiles. These profiles were averaged and normalized for each TF. A Savitzky-Golay filter (150 bp window, polynomial order = 3) was applied to smooth the coverage curves. From the generated coverage profiles, three quantified features were extracted for each of the 854 selected TFs: (1) central coverage, calculated as the mean depth at the central binding site (specifically, the − 30 bp to + 30 bp window); (2) flanking mean coverage, calculated within ± 1 kb of the site center; and (3) oscillation amplitude, estimated by Fast Fourier Transform (FFT) to quantify the strength of nucleosome positioning signals. Methylation-based profile (MBP) The MBP consists of 21 features designed to infer methylation status from cleavage patterns, adapted from the method by Zhou et al [18]. Theoretically, unmethylated cytosines are more prone to enzymatic cleavage than methylated ones. The analysis involves two genomic contexts. First, for all CpG sites in Alu regions, we calculated the fragment ratios for eight specific patterns where the 5’ end breakpoint is either CGN or NCG (N = A, T, C, G). Additionally, the aggregate fragment count ratio of CGN to NCG (CGN/NCG) was computed, resulting in nine features. Second, expanding beyond Alu regions, we examined genome-wide CGCG sites. Fragment ratios were calculated for 10 patterns where the 5’ end breakpoints correspond to specific cleavage combinations: CGC and NCG (defined as the C1 set), and CGN and GCG (defined as the C2 set). Alongside these, the ratios of CGC/NCG and CGN/GCG were calculated, contributing 12 additional features. These analyses were combined to create the comprehensive set of 21 features, which were standardized using z-scores prior to model construction. Machine learning and cross-validation analyses A machine learning classifier was developed to distinguish gastric cancer samples and healthy samples. A two-level stacked ensemble classifier was developed using the H2O AutoML framework [19]. At Level 0, base models were trained for each feature type (FSP, CTBS, MBP) using six distinct algorithms: (1) XGBoost, which optimizes gradient boosting decision trees; (2) Gradient Boosting Machine (GBM), which iteratively refines residual errors; (3) Generalized Linear Model (GLM), which provides interpretable probability scores; (4) Distributed Random Forest (DRF), which reduces overfitting via bagging; (5) Extremely Randomized Trees (XRT), which minimizes variance through random split selection; and (6) Deep Learning (DL), which captures non-linear patterns using multi-layer neural networks. To prevent overfitting, 5-fold cross-validation was employed during base model training. The predictions generated from the out-of-fold cross-validation holds were used as input features for the next level. At Level 1, a Generalized Linear Model (GLM) meta-learner with non-negative constraints was trained to optimally combine these out-of-fold predictions into a final ensemble score. This architecture ensures that the meta-learner is not exposed to the same data used to train the base models, thereby maximizing generalizability. The prediction scores from the selected intermediate models were averaged to generate the prediction score for each sample. The prediction score ranged between 0 and 1 and a higher score indicated a higher probability of cancer. Finally, feature importance was evaluated using SHAP (SHapley Additive exPlanations) values to identify the main contributors to model predictions. Statistical analysis All subsequent statistical analyses were conducted in R (v4.1.2). Gene Ontology (GO) enrichment analysis was performed using the clusterProfiler package, with significance defined as an adjusted FDR < 0.05. Receiver operating characteristic (ROC) curves were plotted and the AUC was calculated using the pROC package (v1.18.0). To statistically compare the AUCs of different models (e.g., ensemble vs. single-feature models), the DeLong test was employed. Performance metrics including sensitivity [TP/(TP + FN)], specificity [TN/(TN + FP)], and accuracy [(TP + TN)/(TP + FP+TN + FN)] were computed. The corresponding 95% confidence intervals (CIs) for these proportions were calculated using the Wilson score interval method. Comparisons of continuous variables between two independent groups were conducted using the Wilcoxon rank-sum test. To evaluate the trend of prediction scores across ordered groups (e.g., Stage I, II, III), the Jonckheere-Terpstra test was applied. A two-sided p-value of < 0.05 was considered statistically significant. Results Plasma samples contributed from each participant were collected to perform low depth whole-genome sequencing (WGS). Samples, which failed quality control, were excluded. In the training dataset, three features, including fragment size pattern (FSP), methylation-based profile (MBP) and coverage at transcription factor binding sites (CTBS), were extracted to construct base models through machine learning. Base models were selected and integrated to form the final model. The final model was locked and evaluated on the validation dataset. Study design and cohort information The study workflow is illustrated in Fig. 1. A total of 646 participants were enrolled across three independent cohorts. The clinical characteristics of all participants are summarized in Table 1. The training cohort included 108 gastric cancer patients and 108 healthy individuals. The groups were well-balanced in terms of age (mean: 62.3 vs. 60.9 years) and gender (54 males and 54 females in both groups). The cancer patients were evenly distributed across Stages I, II, and III, with 36 patients (33.3%) in each stage. The majority of tumors were poorly differentiated (58.3%, 63/108), followed by moderately (37.0%, 40/108) and well-differentiated (4.7%, 5/108) cases. Table 1Clinical characteristics of patients in training cohort, validation cohort, and external validation cohortsPatient characteristicTraining data setValidation data setExternal validationHealthy(n = 108)Cancer(n = 108)Healthy(n = 79)Cancer(n = 79)Healthy(n = 136)Cancer(n = 136)GenderFemale54 (50.0%)54 (50.0%)54 (68.4%)6 (7.6%)76 (55.9%)30 (22.1%)Male54 (50.0%)54 (50.0%)25 (31.6%)73 (92.4%)60 (44.1%)106 (77.9%)Agemean (min, max)60.9 (36, 76)62.3 (45, 78)60.7 (40, 77)63.0 (45, 79)60.0 (19, 75)63.2 (47, 79)StageI/36 (33.3%)/34 (43.0%)/92 (67.6%)II/36 (33.3%)/22 (27.9%)/22 (16.2%)III/36 (33.3%)/23 (29.1%)/22 (16.2%)DifferentiationWell/5 (4.7%)/3 (3.8%)/11 (8.1%)Moderately/40 (37.0%)/25 (31.6%)/51 (37.5%)Poorly/63 (58.3%)/51 (64.6%)/74 (54.4%)Differences in age between healthy and cancer groups were assessed using the Wilcoxon rank-sum test. No significant age differences were observed in the Training (p = 0.93) and Validation (p = 0.14) cohorts. The External Validation cohort showed a statistical difference (p = 0.009); however, correlation analysis (Figure S1B) confirmed that age does not influence prediction scores Clinical characteristics of patients in training cohort, validation cohort, and external validation cohorts Differences in age between healthy and cancer groups were assessed using the Wilcoxon rank-sum test. No significant age differences were observed in the Training (p = 0.93) and Validation (p = 0.14) cohorts. The External Validation cohort showed a statistical difference (p = 0.009); however, correlation analysis (Figure S1B) confirmed that age does not influence prediction scores The internal validation cohort comprised 79 gastric cancer patients and 79 healthy individuals. While age was comparable between groups, the cancer group showed a higher proportion of males (92.4%, 73/79 vs. 31.6%, 25/79). Notably, early-stage disease was prevalent, with 43.0% (34/79) of patients diagnosed at Stage I. Similar to the training cohort, most tumors were poorly differentiated (64.6%, 51/79). The external validation cohort (n = 272) was pooled from two independent centers. It included 136 gastric cancer patients and 136 healthy controls. The cancer group was predominantly male (77.9%, 106/136) with a mean age of 63.2 years. Importantly, this cohort was heavily enriched for early-stage disease, with Stage I patients accounting for 67.6% (92/136) of the cancer group. Tumor differentiation was predominantly poor (54.4%, 74/136) or moderate (37.5%, 51/136). cfDNA fragmentomic features of gastric cancer Following low-pass WGS, we performed a comparative analysis of cfDNA fragmentomic features between gastric cancer patients and healthy controls. The distribution of fragment sizes exhibited clear differences between the two groups. As illustrated in Fig. 2A, the ratio of short fragments (100–150 bp) to long fragments (151–220 bp) across the genome was significantly higher in the gastric cancer group compared to the healthy group, indicating an enrichment of shorter cfDNA fragments in patients with cancer. Analysis of cleavage-inferred methylation patterns revealed distinct epigenetic alterations. The CGN/NCG ratio at fragment 5’-ends, a proxy for global methylation levels, was notably lower in the gastric cancer group compared to healthy controls (p = 0.046, Cohen’s d = 0.27; Fig. 2B). This observation is consistent with the global genomic hypomethylation frequently reported in gastric cancer [20]. Cell-free DNA (cfDNA) features of cancer and healthy samples. A The ratio of short fragments (100–150 bp) to long fragments (151–220 bp) distribution across the genome, B Methylation-based profile (MBP) (CGN/NCG ratio) of cancer and healthy samples. Statistical significance was determined by independent samples t-test (p = 0.04), with a Cohen’s d of 0.27, C Heatmap of coverage at transcription factor binding sites (CTBS) in cancer and healthy samples. Rows represent individual samples, and columns represent standardized coverage scores for selected transcription factors, D Gene Ontology (GO) enrichment analysis of biological processes (BP) for transcription factors showing significant differential coverage between gastric cancer and healthy groups. The y-axis lists the top enriched terms. The size of the dot represents the number of genes involved (Count), and the color scale indicates the statistical significance (adjusted p-value). We further examined nucleosome coverage patterns around transcription factor binding sites. Heatmap visualization demonstrated that cancer and healthy samples clustered separately based on their CTBS profiles (Fig. 2C). GO analysis of TFs revealed that these TFs were primarily enriched in developmental processes, such as pattern specification process, embryonic organ development, and cell fate commitment (Fig. 2D). Furthermore, KEGG pathway analysis highlighted enrichment in transcriptional misregulation in cancer and signaling pathways regulating pluripotency of stem cells (Figure S6). These findings support that the assay captures the chromatin footprint of cellular dedifferentiation and oncogenic reprogramming that are frequently associated with gastric carcinogenesis. Statistical significance of AUC differences was assessed using the DeLong test (p < 0.05 for ensemble vs. single-feature models). Comparison of prediction scores between groups was performed using the two-sided Wilcoxon rank-sum test (p < 0.001). Overall model performance We employed machine learning to construct a classifier capable of distinguishing gastric cancer patients and healthy individuals based on three cfDNA features: FSP, CTBS and MBP. Even when using single cfDNA features, the base models demonstrated robust discriminatory ability, with AUC values ranging from 0.87 to 0.96 (Fig. 3A-B). Specifically, in the validation cohort, the CTBS model achieved the highest individual AUC (0.938), followed by FSP (0.928) and MBP (0.904), suggesting that CTBS was the primary contributor to predictive power. Integrating these three features into an ensemble model further improved performance. DeLong tests confirmed that the ensemble model significantly outperformed the best single-feature models in both the training (ensemble vs. MBP, p = 0.039) and validation cohorts (ensemble vs. CTBS, p = 0.019). The ensemble model achieved AUCs of 0.986 and 0.978 in the training and validation cohorts, respectively (Fig. 3A-B). Key features contributing to the model’s performance were identified via SHAP analysis (Figure S5). Model performance in the training cohort and validation cohort The ensemble model generated a prediction score for each sample, ranging from 0 to 1, where a higher score indicated a high probability of gastric cancer. As expected, prediction scores in the gastric cancer group were significantly higher than those in the healthy group across all cohorts (p < 0.001, Fig. 3C). To establish a decision rule, we selected a cutoff value of 0.605 based on the training cohort data. This threshold was chosen to target a high specificity (≥ 98%) suitable for screening purpose, yielding a specificity of 98.1% and a sensitivity of 88.9% in the training set (Fig. 3D; Table S1). This cutoff was then locked and applied to all subsequent validation cohorts. In the validation cohort, specificity and sensitivity were 97.5% and 88.6% respectively (Fig. 3D; Table S1). It’s worth noting that this high specificity was achieved against a highly selected healthy control population and may be overestimated for real-world screening settings where benign gastric conditions are prevalent. Additionally, down-sampling analysis demonstrated that the model maintains robust performance (AUC > 0.95) even at ultra-low coverage (1×–2×), highlighting its potential for cost-effective screening (Figure S4). Given the potential imbalance in age and sex distribution among the cohorts, we evaluated their impact on prediction scores. Analysis revealed no significant difference in scores between males and females within either the healthy or cancer groups (healthy: p = 0.27, cancer: p = 0.59; Figure S1A). Similarly, no significant correlation was observed between age and prediction scores (healthy R2 = 0.001, cancer R2 = 0.006; Figure S1B). These findings indicate that age and sex had negligible influence on the model’s output. Model performance in the cancer subgroups Pathological stage is a critical determinant of patient management and prognosis. We observed a positive correlation between prediction scores and disease stage across all three cohorts (Fig. 4A). Similarly, detection sensitivity increased with advancing stage (Fig. 4B). In the training cohort, sensitivity for stage I, II, and III was 86.1% (31/36), 88.9% (32/36), and 94.4% (34/36), respectively. These results were successfully replicated in the validation cohort, which showed sensitivities of 85.3% (29/34) for stage I, 86.4% (19/22) for stage II and 95.7% (22/23) for stage III. The model performance in the cancer subgroups. A Prediction scores of gastric cancer samples at stage I-III in the training cohort and validation cohort. B Sensitivity of gastric cancer samples at stage I-III in the training cohort and validation cohort. C Prediction scores of gastric cancer samples with different differentiation degrees in the training cohort and validation cohort. D Sensitivity of gastric cancer samples with different differentiation degrees in the training cohort and validation cohort. Differentiation grade serves as a predictive factor of tumor aggressiveness. We analyzed the relationship between prediction scores and histological differentiation in the training and validation cohorts. As shown in Fig. 4C, prediction scores were lowest in well-differentiated tumors and highest in poorly differentiated tumors. Correspondingly, sensitivity improved with decreasing differentiation (Fig. 4D). In the training cohort, sensitivities were 80.0% (4/5), 87.5% (35/40), and 92.1% (58/63) for well, moderately and poorly differentiated tumors, respectively. In the validation cohort, the model achieved sensitivities of 66.6% (2/3) for well-differentiated, 88.0% (22/25) for moderately differentiated, and 90.2% (46/51) for poorly differentiated cases. Model performance in the external validation cohort To independently assess the generalizability of our assay, we evaluated the model in the pooled external validation cohort (n = 272), which comprised participants from two independent centers. The model exhibited robust performance, achieving an AUC of 0.967, with a specificity of 96.3% and a sensitivity of 87.5% (Fig. 5A-B). Consistent results were observed across the individual centers: External Cohort 1 showed an AUC of 0.966 (95.5% specificity, 87.9% sensitivity), and External Cohort 2 showed an AUC of 0.969 (97.1% specificity, 87.9% sensitivity) (Figure S2-3). Model performance in the external cohort. A Receiver operating characteristic (ROC) curves and area under the ROC curve (AUROC) values of base model (fragment size pattern [FSP], methylation-based profile [MBP], coverage at transcription factor binding sites [CTBS]) and ensemble model (total features). B Prediction scores of gastric cancer and healthy samples. C Prediction scores and sensitivity of gastric cancer samples at stage I-III. D Prediction scores and sensitivity of gastric cancer samples with different differentiation degrees. Statistical significance of AUC differences was assessed using the DeLong test. Comparison of prediction scores between groups was performed using the two-sided Wilcoxon rank-sum test. We further analyzed model performance in subgroups of the external cohort, which mirrored the trends observed in the training and validation sets. Both predicted scores and sensitivity increased with advancing disease stage (Fig. 5C). Importantly, over two-thirds of patients (67.6%) in this external validation cohort were diagnosed with stage I disease, the model maintained a high sensitivity of 85.9% (79/92) for Stage I cancers. Regarding differentiation, sensitivities were 81.8% (9/11), 84.3% (43/51), and 90.5% (67/74) for well, moderately and poorly differentiated tumors, respectively (Fig. 5D). These findings from the multi-center cohort confirmed the model’s efficacy for the gastric cancer early detection. Discussion Gastric cancer remains a significant global health burden. While early detection is pivotal for improving survival outcomes, traditional endoscopic screening is limited by its invasive and low compliance, and conventional tumor markers (e.g., CEA, CA19-9) lack the sensitivity required for early-stage disease [4]. In this study, we developed and validated a non-invasive assay based on multi-dimensional cfDNA fragmentomics. Our approach demonstrated robust performance in distinguishing early-stage gastric cancer from healthy controls, achieving high sensitivity (> 85%) for Stage I disease across training, internal validation, and external multi-center cohorts. The prediction scores derived from our model exhibited a strong positive correlation with disease stage and differentiation grade, suggesting that the assay quantitatively captures the extent of tumor burden and biological malignancy. Liquid biopsy has emerged as a promising alternative for cancer screening. While previous studies exploring long noncoding RNA [21] and microRNAs [22–24] have shown potential, RNA-based biomarkers are often limited by stability issues in plasma. In contrast, cfDNA offers a more stable analyte. Although cfDNA methylation has been widely investigated [7,25], existing assays have historically struggled to balance high specificity with sufficient sensitivity for Stage I gastric cancer. Ren et al.. reported sensitivities of only 44% and 59% for stage I and II diseases, respectively [25]. Similarly, Lei et al. achieved a stage I sensitivity of 33.3–57.1% using targeted methylation sequencing [7]. By leveraging fragmentomic features (FSP, CTBS, MBP) via low-pass WGS, our assay overcame these limitations, achieving a Stage I sensitivity exceeding 85% while maintaining high specificity (> 96%). The superior performance of our assay is likely attributable to the integration of multiple fragmentomic dimensions. Specifically, the altered CTBS patterns observed in cancer patients likely reflect the widespread chromatin remodeling characteristic of gastric carcinogenesis. Our functional enrichment analyses provide a biological explanation for these patterns. GO analysis revealed enrichment in developmental pathways such as pattern specification and cell fate commitment, while KEGG analysis highlighted transcriptional misregulation in cancer and signaling pathways regulating pluripotency of stem cells. Together, these findings suggest that our assay captures the chromatin footprint of cellular dedifferentiation, a process where tumor cells regress to a stem-like state. Specifically, we identified altered coverage at binding sites for CDX1, a key regulator of intestinal metaplasia and cancer stemness [26], and POU5F1 (OCT4), a master regulator of pluripotency whose overexpression is linked to advanced stage and poor differentiation in gastric cancer [27]. The detection of these signals implies that the model leverages the chromatin remodeling signatures associated with oncogenic reprogramming. Furthermore, the progressive increase in model scores from stage I to III supports the biological validity of these features, indicating that the magnitude of fragmentomic alteration scales with disease burden. Additionally, the integration of MBP inferred from cleavage patterns provided robust epigenetic signals without the need for bisulfite conversion, which can degrade DNA and reduce sensitivity. When compared to similar WGS-based approaches, our assay demonstrates distinct advantages. For example, Yu et al.. reported a multi-dimensional liquid biopsy method with a specificity of 89.8% [12]. In contrast, our assay demonstrated a superior specificity of 96.3% in the external validation cohort, a critical metric for reducing false positives in population screening. Additionally, unlike methods relying on complex antibody enrichment like cfMeDIP-seq [8], our streamlined low-pass WGS workflow reduces technical variability and cost, facilitating scalable clinical application. The successful validation in a multi-center external cohort further underscores the robustness and generalizability of our model. In terms of clinical utility, it is important to interpret the Positive Predictive Value (PPV) in the context of disease prevalence. While our case-control design yielded high PPVs, in real-world screening settings targeting high-risk populations, the prevalence of gastric cancer typically ranges from 0.8% in community-based programs [28] to 2.7% in hospital-based cohorts [29]. Under these scenarios, the estimated PPV of our assay would range from approximately 19% to 40% (based on 87.5% sensitivity and 96.3% specificity). Although lower than in the study cohort, this represents a substantial enrichment (up to 40-fold) compared to the baseline prevalence. Therefore, we envision this assay functioning primarily as a triage tool to prioritize high-risk individuals for endoscopy, thereby optimizing healthcare resource allocation and improving the diagnostic yield of invasive procedures. Several limitations of this study should be noted. First, although we employed gastroscopy as the gold standard to ensure healthy controls were cancer-free at baseline, long-term follow-up was not performed to rule out future cancer development. A second limitation is the nature of our control cohorts, which consisted of strictly screened, endoscopy-negative individuals, creating a highly selected reference population free of active gastric inflammation or benign pathology. While this rigorous design was essential to establish a clean diagnostic baseline, it represents an idealized clinical scenario. Consequently, the exceptionally high specificities reported in this study likely represent an overestimation of the assay’s performance in real-world screening settings and create unrealistic expectations regarding clinical utility. In a true screening population, benign conditions (e.g., atrophic gastritis, gastric ulcers) are highly prevalent and may share overlapping fragmentomic features with malignancy. Indeed, our preliminary data indicate a decline in specificity when tested against these conditions. Therefore, future validation in large-scale, independent cohorts that include a spectrum of benign gastric diseases is imperative to determine the true clinical specificity. Third, the model was constructed and validated based on the Chinese population, which may limit its generalizability to other ethnic groups. Additionally, we acknowledge the imbalance in sex distribution, particularly in the validation cohort (higher proportion of males in the cancer group). However, our subgroup analysis demonstrated no significant difference in prediction scores between males and females in either the healthy or cancer groups, suggesting that the model’s performance is not driven by sex bias. Nevertheless, future validation in more balanced cohorts is recommended. In summary, we successfully developed a non-invasive, convenient, and high-performance assay based on cfDNA fragmentomics. By integrating fragment size, motif-based methylation, and transcription factor binding footprints, this assay provides a powerful tool for the early detection of gastric cancer. Conclusions In this study, we developed a liquid biopsy assay utlizing cfDNA fragmentomics and machine learning to achieve accurate early detection of gastric cancer. The assay integrates three distinct fragmentomic features and demonstrated robust performance across internal and external multi-center validation cohorts, particularly for Stage I disease. These findings suggest that this non-invasive approach holds significant promise as a complement to standard screening strategies, potentially improving early diagnosis rates and patient outcomes. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 2 Supplementary Material 3 Supplementary Material 4 Supplementary Material 5 Supplementary Material 6 Supplementary Material 7 Acknowledgements We want to express our gratitude to the participants for their contributions, to the staff and colleagues for their assistance in the research process, and to the researchers for their guidance. Abbreviations cell-free DNA circulating tumor DNA whole genome sequencing fragment size pattern coverage at transcription factor binding sites methylation-based profile copy number variation generalized linear model extreme gradient boosting gradient boosting machines deep random forest extreme randomized trees deep learning area under the curve quantitative polymerase chain reaction transcription factor gene ontology quality control Author contributions Xiaoli Huang, Zhengxiang Han: Methodology, Formal analysis, Funding acquisition, Investigation, Writing – original draft, Visualization. Ying Zhang: Methodology, Formal analysis, Investigation, Writing – original draft, Visualization. Hongmei Wang, Chenxi Bai, Ying Pan, Jingwen Huang, Jiaojiao Xie, Xiaohan Qian: Data curation, Investigation. Xuan Wang, Tao Ding, Jinpeng Zhang, Shanshan Yang, Xinyue Hong, Qifan Jing: Formal analysis, Investigation, Methodology, Writing – review & editing. Xuqi Li: Data curation, Investigation, Funding acquisition. Gang Li: Conceptualization, Funding acquisition, Resources, Supervision, Project administration, Writing – review & editing. Xiaohua Li: Conceptualization, Resources, Supervision, Project administration, Writing – review & editing. All authors have read and approved the final manuscript. Funding This work was supported by Nanjing Medical Development Research Project Fund (YKK23262) to Xiaoli Huang, Pengcheng Talents – Key Medical Talent Training Project (XWRCHT20220068) to Zhengxiang Han, National Nature Science Foundation of China (No. 82470537) and Fundamental Research Funds for the Central Universities in Xi’an Jiaotong University (xtr062025009) to Xuqi Li, Yishan Research Project of Jiangsu Cancer Hospital (No. YSZD202404) and Excellent Talent Program of Jiangsu Cancer Hospital (No.2017YYCJH-08) to Gang Li. Data availability The datasets generated and/or analysed during the current study are available in the Genome Sequence Archive for Human (GSA-Human) repository, at Accession number: HRA018032 (https://ngdc.cncb.ac.cn/gsa-human/s/RM9QnYwc). Additional clinical data and intermediate analysis results supporting the findings of this study are available from the corresponding authors upon reasonable request. Declarations Competing interests Xuan Wang, Tao Ding, Jinpeng Zhang, Shanshan Yang, Xinyue Hong and Qifan Jing are employees of Nanjing Geneseeq Technology Inc. The remaining authors declare no competing interests. Ethical Approval and consent to participate The study received approval from the Ethics Committee of the Fourth Affiliated Hospital of Nanjing Medical University. The study adhered to the Declaration of Helsinki. All participants provided written informed consent before sample collection. Consent for publication Not applicable. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Xiaoli Huang, Zhengxiang Han and Ying Zhang have contributed equally to this work. Contributor Information Gang Li, Email: ligang@njmu.edu.cn. Xiaohua Li, Email: xjyylixiaohua@163.com. Reference List Associated Data Supplementary Materials Supplementary Material 1 Supplementary Material 2 Supplementary Material 3 Supplementary Material 4 Supplementary Material 5 Supplementary Material 6 Supplementary Material 7 Data Availability Statement The datasets generated and/or analysed during the current study are available in the Genome Sequence Archive for Human (GSA-Human) repository, at Accession number: HRA018032 (https://ngdc.cncb.ac.cn/gsa-human/s/RM9QnYwc). Additional clinical data and intermediate analysis results supporting the findings of this study are available from the corresponding authors upon reasonable request.