Early and accurate detection of ovarian cancer by profiling cell-free mitochondrial DNA fragmentomics in multi-ethnic population.

A total of 1,837 female participants were recruited from four medical centers: Xijing Hospital affiliated to the Fourth Military Medical University (September 2022- December 2024), Tangdu Hospital affiliated to the Fourth Military Medical University (January 2024 - December 2024), Fudan University Shanghai Cancer Center (March 2024- December 2024) and Tumor Hospital affiliated to Xinjiang Medical University (August 2024- December 2024). The eligibility criteria required that women should be between 20 and 80 years old, have a confirmed menstrual status (natural menopause ≥ 12 months or surgical menopause), and have undergone pelvic imaging within three months prior to their enrollment. Participants were assigned to one of four groups based on the following criteria: (1) OC group: surgically confirmed primary epithelial OC, with no prior anti-tumor therapy before surgery. (2) BAM group: presence of ovarian lesions preoperatively assessed as O-RADS 3, 4, or 5 and subsequently confirmed by surgical pathology as ovarian endometriotic cysts, benign ovarian tumors or others. (3) HC group: participants demographically matched to OC patients based on age, menstrual status, and parity, with no clinical, biochemical, or imaging evidence suggestive of any ovarian pathology. (4) Borderline ovarian tumor (BOT) group: surgically confirmed primary epithelial BOT, with no prior anti-tumor therapy administered before surgery. The following exclusion criteria were applied uniformly to all groups: (1) concurrent diagnosis of any other malignancy; (2) presence of acute pelvic inflammatory disease or pelvic abscess; (3) history of blood transfusion within 1 month prior to enrollment; (4) history of ovarian malignancy. This study was approved by the four Hospital Review Boards and all participants provided written informed consent. OC patients were staged according to the 2014 International Federation of Gynecology and Obstetrics (FIGO) classification system. The demographic and pathological characteristics of the participants are presented in Table S1-3 (Additional file 2). For the OC group, all consecutive eligible patients with early‑stage OC (FIGO I–II) who met the inclusion criteria during the study period were enrolled. To balance the representation of disease stages, late‑stage OC (FIGO III–IV) patients were then randomly matched to the early‑stage cases at a 1:1 ratio, using age and menopausal status as matching variables. Healthy controls (HC) were similarly matched to the OC group at a 1:1 ratio based on age and menopausal status. For the BAM group, patients with adnexal masses preoperatively classified as O-RADS 3–5 who underwent surgery were consecutively enrolled during the same period. This sampling strategy ensures that the OD model is trained and validated on a clinically representative spectrum of OC (with balanced early/late stages) and that the BMD model focuses on the most diagnostically challenging benign lesions. After enrollment, the participants were assigned to six cohorts according to hospital source, ethnicity, tumor type, and predefined model-development purposes: the training cohort ( n  = 677), internal validation cohort ( n  = 289), external validation cohort 1 ( n  = 376), external validation cohort 2 ( n  = 290), Chinese minority cohort ( n  = 153), and BOT cohort ( n  = 52). Participants recruited from Xijing Hospital were randomly allocated to the training and internal validation cohorts at a ratio of 7:3, whereas the internal validation cohort consisted of participants from Xijing Hospital and was used for validation of both models. Participants recruited from Tangdu Hospital were assigned to external validation cohort 1, and participants from Fudan University Shanghai Cancer Center were assigned to external validation cohort 2. These two external validation cohorts were used to evaluate the generalizability of the models across independent medical centers. Participants recruited from the Tumor Hospital affiliated to Xinjiang Medical University were assigned to the Chinese minority cohort, which was analyzed separately to assess model robustness across different ethnic backgrounds. The BOT cohort was not used for model training; instead, it was reserved for exploratory evaluation of model behavior in borderline tumors. Plasma sample collection was performed prior to the initiation of any treatment. Peripheral blood samples (5 mL for each) from all participants were collected into EDTA tubes and centrifuged within 2 h of venipuncture using a standard two-step protocol to obtain plasma. After separation, plasma aliquots were immediately stored at -80℃ and subjected to only one freeze-thaw cycle. The QIAamp Circulating Nucleic Acid kit (Qiagen, USA) was utilized for cfDNA extraction from the plasma within 3 months according to the manufacturer’s instructions. The cfDNA was quantified using Qubit 4.0 (Thermo Fisher, USA). Subsequently, the whole-genome sequencing (WGS) libraries were constructed using plasma cfDNA (10–20 ng for each sample) with the NEB ultra v2 kit (NEB, US). The number of PCR cycles during library construction was 12. Then, the WGS libraries were hybridized with homemade biotinylated ccf-mtDNA capture probes covering the complete mitochondrial genome (rCRS, NC_012920.1 ) with 1× tiling density (no regional bias), as previously described [ 32 – 35 ]. Finally, all captured ccf-mtDNA libraries were sequenced by the Illumina Hiseq Novaseq 6000 platform using paired-end runs with 2 × 150 cycles (PE 150). To minimize bias resulting from systematic errors in the experiment, experiments were performed in multiple batches, with each batch containing samples from different disease types. All experiments were carried out by two individuals in a centralized laboratory, utilizing the same instruments across all samples. Totally, based on our sequencing and analysis methods, it will take 5–7 working days to accomplish one processing pipeline (including cell-free DNA extraction, library construction, next generation sequencing and data analysis) and thus obtain the fragmentomic data of ccf-mtDNA, with cost of 30–40 US dollars for each sample. These supplementary information support our approach to be cost-effective. After sequencing, raw reads were trimmed (Fastp software, v.0.20.0), aligned to both rCRS and hg19 (BWA software, v.0.7.1521), and only uniquely rCRS-mapped reads with MAPQ ≥ 20 were retained to exclude nuclear-derived mitochondrial sequence (NUMT). Reads were sorted, PCR duplicates removed (Picard MarkDuplicates, v.1.81), and local realignment performed (Genome Analysis Toolkit 4, v.3.2-2). This workflow produced a high quality ccf-mtDNA dataset with mean coverage 3,979× (Additional file 2: Table S4), following our group’s protocols for hepatocellular carcinoma and colorectal cancer [ 32 , 33 ]. Moreover, to assess potential batch effects, principal component analysis (PCA) was performed on the coverage profiles or fragment features across all samples, and no significant batch-driven clustering was observed. The fragment size of ccf-mtDNA was determined from mitochondrial genome alignments using Picard Tools (v.1.119). Fragments were subsequently categorized as short or long based on the sample-specific median size. The count of fragments covering each of the 16,569 mtDNA sites was computed. The fragment size distribution (FSD) score was defined as ratio of the coverage depth of long fragments to that of short fragments and was corrected for GC content via LOESS regression (span = 0.75, degree = 1). Then, the FSD score was standardized into a z-score relative to the sample mean. Subsequently, a genome-wide fragmentation profile was generated based on the standardized scores across all mtDNA positions [ 33 , 34 ]. The median reference profile was constructed by taking the median of each fragmentomic feature independently across 30 healthy controls (HCs) from our prior study (Ref. 26). These 30 HCs were not part of the current study’s HC cohort and were excluded from all training/validation to prevent data leakage. All sequencing protocols were identical between the reference HCs and current samples [ 35 ]. Peaks were defined in the fragmentation profile with a minimum width of 5 bp. New peaks were identified as those positioned more than 20 bp from the corresponding peaks in the median HC profile. The mitochondrial genome was partitioned into 255 windows, and the areas of all positive and negative peaks (baseline FSD = 0) were quantified within each window. Subsequently, the Euclidean distance from each fragmentation profile to the median HC profile within each window was computed using the Scipy software (v.1.6.2). In parallel, paired-end reads were processed into fragments using BEDTools (v.2.26.0), with the 5’ end of the light strand defined as the start of Read 1 and the heavy strand end as its complement. The proportion of each 5’ end base and 4-mer motif was then determined. The 5’ end base preference was evaluated by normalizing the observed proportions against the base composition of the mitochondrial reference genome( NC_012920.1 ), with the following base frequencies: adenine (A) 0.2782, thymine (T) 0.2782, cytosine (C) 0.2218, guanine (G) 0.2218. Finally, the diversity of 4-mer motifs was quantified via normalized Shannon entropy, defined the motif diversity score (MDS) with the following equation: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{MDS}=\sum\limits_{\mathrm{i}=1}^{\mathrm{256}}-\mathrm{Pi}\mathrm{*log(}\mathrm{Pi}\mathrm{)/}\mathrm{log}256$$\end{document} , where Pi denotes motif frequency. The theoretical range of MDS is 0 to 1 [ 35 ]. Based on ccf-mtDNA fragmentomic features, we constructed two random forest models: the OC Detection (OD) model and the Benign vs. Malignant Differential diagnosis (BMD) model in the training cohort. Both models were developed by the random forest (RF) algorithm as implemented in the “randomForest” R package (v.4.6–14), using 500 trees and a random selection of six features at each split. The ten-fold cross validation was used to evaluate the model performance in the training cohort. The OD score and BMD score for each patient were computed, respectively. The performance of these two models was subsequently evaluated using an internal and four external validation cohorts. Serum CA-125 and HE4 levels were quantified for all participants. Venous blood samples (5 mL for each) were collected in serum separator tubes prior to any treatment or intervention. CA-125 and HE4 concentrations were determined using electrochemiluminescence immunoassays (Roche Diagnostics, Mannheim, Germany). All assays were conducted by technicians blinded to clinical diagnoses. Quality control procedures compassed daily calibration and inclusion of manufacturer-provided controls (low/medium/high concentrations) with inter-assay coefficients of variation < 5% for both biomarkers. To systematically evaluate the diagnostic performance of these serum biomarkers, we employed three distinct analytical strategies: (i) Binary rule (Clinical Thresholds): Based on established clinical guidelines, patients were categorized as biomarker-positive or -negative. The cut-off for CA-125 was set at 35 U/mL [ 36 ]. For HE4, we applied menopause-stratified thresholds: 68.5 pmol/L for pre-menopausal women and 114.9 pmol/L for post-menopausal women [ 37 ]. Patients were defined as “biomarker-positive” if at least one marker exceeded its corresponding threshold, and “biomarker-negative” if both were below these cut-offs. (ii) Logistic regression (continuous CA125 & HE4): A logistic regression (LR) model was constructed using continuous serum CA-125 and HE4 concentrations as independent variables to calculate a composite probability score, offering a quantitative approach to diagnostic prediction. (iii) Risk of Ovarian Malignancy Algorithm (ROMA) Index: The ROMA index was calculated using the validated algorithm incorporating both CA‑125 and HE4 levels, adjusted for menopausal status. Specifically, the predictive index (PI) was first computed as follows: for premenopausal women, PI = -12.0 + 2.38×ln(HE4) + 0.0626×ln(CA125); for postmenopausal women, PI = -8.09 + 1.04×ln(HE4) + 0.732×ln(CA125), where HE4 is expressed in pmol/L and CA125 in U/mL. The ROMA value (%) was then derived as ROMA = exp(PI) / [1 + exp(PI)]×100. To reflect real-world clinical practice, the cut-off values were defined strictly according to the established clinical thresholds: a ROMA value ≥ 11.4% indicated high risk for premenopausal women, and ≥ 29.9% indicated high risk for postmenopausal women. For the binary rule and ROMA index, these pre‑defined thresholds were used without further modification. Finally, we performed a comparative analysis to assess the diagnostic efficacy of these three strategies, identifying the approach that yielded the highest predictive accuracy for distinguishing OC from HC and for distinguishing OC from BAM. To maximize diagnostic performance, we developed a combined logistic regression model integrating CA125, HE4, and the BMD score. This unified model leverages the three predictors to generate a single composite probability, with the classification threshold calibrated to fix the sensitivity at 95%. Specifically, a patient is classified as positive if the predicted probability exceeds this threshold. Decision curve analysis was performed using the dca() function from the dcurves package. Net benefit (NB) was calculated across a clinically relevant probability threshold range of 1% to 50%. Statistical analyses were conducted using GraphPad Prism v.9.5.1 (GraphPad Software, USA). The Friedman test was utilized to examine the differences in fragment size distribution across multiple groups, whereas the Wilcoxon signed-rank test was applied for comparison between two groups. Additionally, comparisons between two groups were performed using the Wilcoxon rank-sum test, while multiple group comparisons were performed using the Kruskal-Wallis H test, Bonferroni’s correction, and Dunn’s multiple comparisons test. Spearman’s rank correlation coefficient was calculated to assess the correlation between two groups. All P values are two-tailed, and statistical significance was set at P  < 0.05. Statistical performance metrics, including sensitivity and specificity et al. were calculated as as previously described [ 32 – 34 ]. Detailed statistical methods are provided in the corresponding figure legends. Sample size was estimated a priori using MedCalc v19.4, assuming an expected AUC of 0.95, 90% statistical power, alpha error = 0.05 and beta error = 0.10.

As shown in Fig.  1 , a total of 1,837 female participants were recruited from four hospitals and analyzed across six cohorts, including training cohort ( n  = 677), internal validation cohort ( n  = 289), external validation cohort 1 and 2 ( n  = 376 and 290, respectively), Chinese minority cohort ( n  = 153), BOT cohort ( n  = 52). Detailed clinicopathological characteristics (such as age, menopausal status, parity, FIGO stage, et al.) of HC, BAM patients and OC patients were summarized in Table S1 (Additional file 2). Each cohort showed comparable distributions of key characteristics (e.g., age, menopausal status) across the HC, BAM, and OC groups (all P  > 0.05). The ethnic distribution of Chinese minority participants was detailed in Table S2 (Additional file 2). Demographic and clinicopathological characteristics of patients with BOT were shown in Table S3(Additional file 2). PCA based on ccf-mtDNA fragmentomic features showed no obvious batch effect across HC, BAM and OC groups (Additional file 1: Fig. S1). Based on ccf-mtDNA fragmentomic features, we developed two machine learning models: OD model and BMD model. The performance of both models underwent extensive evaluation. Fig. 1 Study design. A total of 1,837 participants were enrolled from four hospitals. Participants including 384 °C patients, 188 BAM patients, 394 HC and 52 BOT were enrolled from Xijing Hospital affiliated to the Fourth Military Medical University. Among them, OC patients, BAM patients and HC were randomly divided into training and internal validation cohorts at the ratio of 7:3. Four cohorts were used to evaluate the performance, including external validation cohort 1 (participants were enrolled from Fudan University Shanghai Cancer Center): 150 °C, 70 BAM, 156 HC; external validation cohort 2 (participants were enrolled from Tangdu Hospital affiliated to the Fourth Military Medical University): 119 °C, 51 BAM, 120 HC; ethnic minority cohort (participants were enrolled from Tumor Hospital Affiliated to Xinjiang Medical University): 63 C, 34 BAM, 56 HC; BOT cohort comprised 52 patients with BOT who were enrolled from Xijing Hospital. A standard procedure was used for all experiments (including plasma collection, cfDNA extraction, library construction and capture-based mtDNA NGS) and analyses of ccf-mtDNA fragmentomic features (including fragment size, fragmentation profile, 5′ end base preference, 5′ end base motifs, and 5′ end MDS). Then, random forest algorithm was used to construct the OC detection model and Benign vs. Malignant Differential diagnosis model in the training cohort. The performance (such as AUC, sensitivity, specificity) of models were evaluated in four cohorts. HC, healthy control; BOT, borderline ovarian cancer; BAM, benign adnexal masses; OC, ovarian cancer; cfDNA, cell-free DNA; ccf-nDNA, circulating cell-free nDNA; ccf-mtDNA, circulating cell-free mtDNA; MDS, motif diversity score; OD, ovarian cancer detection; BMD, Benign vs. Malignant Differential diagnosis; AUC, area under curve; PC, principal component Study design. A total of 1,837 participants were enrolled from four hospitals. Participants including 384 °C patients, 188 BAM patients, 394 HC and 52 BOT were enrolled from Xijing Hospital affiliated to the Fourth Military Medical University. Among them, OC patients, BAM patients and HC were randomly divided into training and internal validation cohorts at the ratio of 7:3. Four cohorts were used to evaluate the performance, including external validation cohort 1 (participants were enrolled from Fudan University Shanghai Cancer Center): 150 °C, 70 BAM, 156 HC; external validation cohort 2 (participants were enrolled from Tangdu Hospital affiliated to the Fourth Military Medical University): 119 °C, 51 BAM, 120 HC; ethnic minority cohort (participants were enrolled from Tumor Hospital Affiliated to Xinjiang Medical University): 63 C, 34 BAM, 56 HC; BOT cohort comprised 52 patients with BOT who were enrolled from Xijing Hospital. A standard procedure was used for all experiments (including plasma collection, cfDNA extraction, library construction and capture-based mtDNA NGS) and analyses of ccf-mtDNA fragmentomic features (including fragment size, fragmentation profile, 5′ end base preference, 5′ end base motifs, and 5′ end MDS). Then, random forest algorithm was used to construct the OC detection model and Benign vs. Malignant Differential diagnosis model in the training cohort. The performance (such as AUC, sensitivity, specificity) of models were evaluated in four cohorts. HC, healthy control; BOT, borderline ovarian cancer; BAM, benign adnexal masses; OC, ovarian cancer; cfDNA, cell-free DNA; ccf-nDNA, circulating cell-free nDNA; ccf-mtDNA, circulating cell-free mtDNA; MDS, motif diversity score; OD, ovarian cancer detection; BMD, Benign vs. Malignant Differential diagnosis; AUC, area under curve; PC, principal component To address the critical unmet need for non-invasive EOC detection, we systematically analyzed the ccf-mtDNA fragmentomic features of HC (n = 276), BAM patients (n = 132), EOC patients (n = 142), and late-stage OC (LOC) patients (n = 127). We hypothesized that mitochondrial genomic instability may generate stage-specific molecular signatures undetectable by conventional diagnostics. As shown in Fig.  2 A-B, the ccf-mtDNA fragment size increased gradually from HC and BAM to EOC and LOC. Significantly higher proportions of long fragments were found in the BAM, EOC and LOC groups compared to the HC group (all P   0.05). Then, we further observed heterogeneity in the ccf-mtDNA fragmentation profiles of both EOC and LOC groups. Notably, the correlation between ccf-mtDNA fragmentation profiles and the median HC profile progressively weakened from HC, BAM to EOC and LOC groups (all P  < 0.05, Fig.  2 C-D). This progression occurred concurrently with a continuous increase of the number of total peaks and new peaks (all P  < 0.05, Fig.  2 E-F). In addition, significant difference was found in 5’ end base preference among four groups. Detailed, in comparison to HC or BAM group, both EOC and LOC groups exhibited a significant underrepresentation of C-ends but a significant overrepresentation of T-ends (all P   0.05, Fig.  2 H). Fig. 2 Comparative analyses of ccf-mtDNA fragmentomic features among different cohorts. Comparison of ( A ) fragment size distribution (10–400 bp), ( B ) proportion of fragments ≤ 100 bp and > 100 bp, ( C ) ccf-mtDNA fragmentation profiles, ( D ) the correlation to ccf-mtDNA fragmentation profile of median HC, ( E ) the number of total peaks, ( F ) the number of new peaks, and ( G ) 5′ end base preference, ( H ) 5′ end MDS among HC ( n  = 276), patients with BAM ( n  = 132), patients with EOC ( n  = 142), LOC ( n  = 127) from Xijing Hospital of Fourth Military Medical University. Statistical significance in (A) was determined by the Friedman test. Comparisons among multiple groups in (B) and (D-H) were evaluated using the Kruskal-Wallis H test followed by Dunn’s multiple comparisons test with Bonferroni correction. Correlation in (C) was evaluated using Spearman’s rank correlation coefficient. HC, healthy control; BAM, benign adnexal masses; OC, ovarian cancer; EOC, early-stage ovarian cancer; LOC, late-stage ovarian cancer MDS, motif diversity score; ns, not significant; * P  < 0.05; ** P  < 0.01; *** P  < 0.001. In B and D-H, center line indicates the median, while lower and upper hinges represent the 25th and 75th percentiles, respectively Comparative analyses of ccf-mtDNA fragmentomic features among different cohorts. Comparison of ( A ) fragment size distribution (10–400 bp), ( B ) proportion of fragments ≤ 100 bp and > 100 bp, ( C ) ccf-mtDNA fragmentation profiles, ( D ) the correlation to ccf-mtDNA fragmentation profile of median HC, ( E ) the number of total peaks, ( F ) the number of new peaks, and ( G ) 5′ end base preference, ( H ) 5′ end MDS among HC ( n  = 276), patients with BAM ( n  = 132), patients with EOC ( n  = 142), LOC ( n  = 127) from Xijing Hospital of Fourth Military Medical University. Statistical significance in (A) was determined by the Friedman test. Comparisons among multiple groups in (B) and (D-H) were evaluated using the Kruskal-Wallis H test followed by Dunn’s multiple comparisons test with Bonferroni correction. Correlation in (C) was evaluated using Spearman’s rank correlation coefficient. HC, healthy control; BAM, benign adnexal masses; OC, ovarian cancer; EOC, early-stage ovarian cancer; LOC, late-stage ovarian cancer MDS, motif diversity score; ns, not significant; * P  < 0.05; ** P  < 0.01; *** P  < 0.001. In B and D-H, center line indicates the median, while lower and upper hinges represent the 25th and 75th percentiles, respectively We further analyzed the differences in ccf-mtDNA fragmentomic features among different subgroups. As shown in Fig. S2-5 (Additional file 1), no significant differences in ccf-mtDNA fragment size, fragmentation profiles, 5’ end base preference and 5’ end MDS were detected among subgroups stratified by age, menopausal state, and parity (all P  > 0.05). Collectively, these results demonstrated that the fragmentomic features of ccf-mtDNA were independent of physiological changes related to age, menopause, and parity. Further analysis revealed differences in the proportions of long/short fragments and 5’ T-end base preference between high-grade serous ovarian carcinoma (HGSOC) and other epithelial ovarian cancer subtypes (including endometrioid, clear cell, mucinous and low-grade serous ovarian carcinomas, additional file 1: Fig. S5). These results reveal aberrant ccf-mtDNA fragmentomic features in OC patients, supporting its potential as a novel biomarker for detection. To tackle the urgent need for OC detection, we constructed a model using the ccf-mtDNA fragmentomic features to distinguish OC patients ( n  = 269) from HC ( n  = 276) in the training cohort. The top 50 features ranked by mean importance in the OD model are listed in Table S5 (Additional file 2). The OD score, representing the probability of OC detection, was calculated. As shown in Fig. S6A-B (Additional file 1), OD scores were significantly higher in OC patients than in HC. The OD model demonstrated robust diagnostic power, achieving an AUC of 0.987. Given the low real-world incidence of OC, a clinically practical detection model should prioritize high specificity to maximize the positive predictive value. At specificities of 99%, 95%, and 90%, the OD model achieved sensitivities of 77.32%, 90.71%, and 95.17%, respectively (Additional file 1: Fig. S6C, additional file 2: Table S6). For the intended use in high-risk or symptomatic individuals, a specificity of 95% (cut-off value: 0.535) was selected to balance robust sensitivity (90.71%) with a clinically acceptable false-positive rate. To assess generalizability and reduce the risk of overfitting, the model was evaluated in an internal validation cohort. As shown in Fig.  3 A, OD scores remained significantly elevated in OC patients compared to HC ( P  < 0.001). Furthermore, as shown in Fig.  3 A, the model exhibited excellent performance at the cut-off value of 0.535, with an AUC of 0.984 (95% CI: 0.973–0.995), a sensitivity of 92.17% (95% CI: 85.79–95.83%), and a specificity of 94.07% (95% CI: 88.26 to 97.10%). To assess the generalizability and real-world potential of the model, we evaluated the OD model in two external validation cohorts. In external validation cohort 1 and 2, OD scores were significantly higher in OC patients than in HC, with exceptional performance (AUC = 0.983 and 0.979, respectively; Fig.  3 B-C). Importantly, at the predefined cut-off value of 0.535, the OD model achieved high specificities (94.87% and 91.07%, respectively) and high sensitivities (90.00% and 91.60%, respectively) in both cohorts (Fig.  3 B-C). To directly assess the cross-population generalizability of OD model, we validated it in an independent cohort of 153 individuals from Uyghur, Kazakh, Hui, Mongol and Tajik ethnic groups. In this cohort, the model maintained strong performance, achieving an AUC of 0.982 (95% CI: 0.965–0.999), with a sensitivity higher than 90% (Fig.  3 D). These results confirms that the OD model is both robust and generalizable, highlighting its considerable potential for clinical application (Additional file 2: Table S7). Furthermore, detailed performance at ≥ 99% specificity was also provided in Table S8 (Additional file 2). Fig. 3 Validation of the OD model for distinguishing OC from HC in four validation cohorts. ( A - D left column) Comparison of OD score between HCs and OC patients in four cohorts. ( A - D middle column) The ROC curves evaluating the performance of OD model in distinguishing patients with OC from HC. ( A - D right column) The specificity and sensitivity of OD model in distinguishing patients with OC from HC. Statistical significance for the comparisons in the left column was determined using the Wilcoxon rank-sum test. The 95% confidence intervals for sensitivity and specificity in the right column were calculated using the exact binomial method. OC, ovarian cancer; OD, OC detection; HC, healthy control; ROC, receiver operating characteristic; AUC, area under the curve; ns, not significant; *** P  < 0.001. In left column, center line indicates the median, while lower and upper hinges represent the 25th and 75th percentiles, respectively. In right column, the error bars represent the 95% confidence interval. Chinese minority cohort comprised Uyghur, Kazakh, Hui, Mongol, and Tajik individuals Validation of the OD model for distinguishing OC from HC in four validation cohorts. ( A - D left column) Comparison of OD score between HCs and OC patients in four cohorts. ( A - D middle column) The ROC curves evaluating the performance of OD model in distinguishing patients with OC from HC. ( A - D right column) The specificity and sensitivity of OD model in distinguishing patients with OC from HC. Statistical significance for the comparisons in the left column was determined using the Wilcoxon rank-sum test. The 95% confidence intervals for sensitivity and specificity in the right column were calculated using the exact binomial method. OC, ovarian cancer; OD, OC detection; HC, healthy control; ROC, receiver operating characteristic; AUC, area under the curve; ns, not significant; *** P  < 0.001. In left column, center line indicates the median, while lower and upper hinges represent the 25th and 75th percentiles, respectively. In right column, the error bars represent the 95% confidence interval. Chinese minority cohort comprised Uyghur, Kazakh, Hui, Mongol, and Tajik individuals We further benchmarked the OD model against three established serum biomarker-based diagnostic strategies: the binary rule, logistic regression (LR) using continuous CA125 and HE4 levels, and the ROMA Index (Additional file 2: Table S9). Among these, the LR model (using a training-derived cut-off of 0.433) achieved the highest specificity (94.4%; 95% CI: 91.8%-96.3%) and sensitivity (84.3%; 95% CI: 80.6%-87.5%). Conversely, the binary rule and ROMA Index demonstrated comparable sensitivities (~ 90%) but lower specificities (73.7% and 76.4%, respectively). ROC analysis revealed that the OD model (AUC = 0.982) significantly outperformed the LR model based on CA125 and HE4 (AUC = 0.937; additional file 1: Fig. S7). These findings demonstrate that the OD model provides superior diagnostic accuracy, effectively balancing sensitivity and specificity compared with serum-based algorithms. Subgroup analyses were performed in the combined validation cohort to assess the robustness of the OD model across clinically relevant patient strata. As our primary focus, we first analyzed the detection performance across different stages of OC. As shown in Fig.  4 A, no significant difference of OD score was found among the OC patients at different FIGO stages. When detecting the OC patients stratified by FIGO stage, the model exhibited consistently high accuracy across all stages, with AUC values ranging from 0.973 to 0.986. Specifically OC in patients with FIGO I ( n  = 120), the OD model demonstrated an AUC of 0.982 and sensitivity of 91.67% when distinguishing OC from HC (Fig.  4 B-C). Remarkably, further analysis revealed no significant differences in OD scores across all histopathological subtypes (all P  > 0.05, Fig.  4 D). Critically, as shown in Fig.  4 E-F, high accuracy was maintained in HGSOC ( n  = 290; AUC = 0.982, sensitivity = 91.72%) and other subtypes: mucinous carcinoma ( n  = 48; AUC = 0.980, sensitivity = 87.50%), clear cell carcinoma ( n  = 48; AUC = 0.983, sensitivity = 91.67%), endometrioid carcinoma ( n  = 35; AUC = 0.980, sensitivity = 91.43%), and low-grade serous ovarian carcinoma (LGSOC, n  = 26; AUC = 0.976, sensitivity = 88.46%)(Additional file 2: Table S10). Moreover, the model consistently maintained high performance across all subgroups, including different menopausal status, parity and with or without ascites (Additional file 1: Fig. S8). These results collectively demonstrate that the OD model maintains robust and consistent performance across all major clinical and pathological subgroups. Fig. 4 Subgroup analyses of OD model. Comparison of OD score ( A ) among FIGO stage Ⅰ, Ⅱ, Ⅲ, Ⅳ and ( D ) among different histological subtypes. The ROC curves evaluating the performance of OD model in distinguishing OC patients at ( B ) each FIGO stage (I, II, III, IV), and ( E ) OC patients with different histological type from HC. ( C , F ) Sensitivity of OD model in different OC subgroups. Statistical significance in ( A ) and ( D ) was evaluated using the Kruskal-Wallis H test followed by Dunn’s multiple comparisons test with Bonferroni correction.The 95% confidence intervals in ( C ) and ( F ) were calculated using the exact binomial method. OC, ovarian cancer; OD, OC detection; HC, healthy control; ROC, receiver operating characteristic; AUC, area under the curve; HGSOC, high-grade serous ovarian carcinoma; MC mucinous carcinoma; CCC, clear cell carcinoma; EC, endometrioid carcinoma; LGSOC, low-grade serous ovarian carcinomas; ns, not significant; *** P  < 0.001. In A and D, center line indicates the median, while lower and upper hinges represent the 25th and 75th percentiles, respectively. In C and F, the error bars represent the 95% confidence interval Subgroup analyses of OD model. Comparison of OD score ( A ) among FIGO stage Ⅰ, Ⅱ, Ⅲ, Ⅳ and ( D ) among different histological subtypes. The ROC curves evaluating the performance of OD model in distinguishing OC patients at ( B ) each FIGO stage (I, II, III, IV), and ( E ) OC patients with different histological type from HC. ( C , F ) Sensitivity of OD model in different OC subgroups. Statistical significance in ( A ) and ( D ) was evaluated using the Kruskal-Wallis H test followed by Dunn’s multiple comparisons test with Bonferroni correction.The 95% confidence intervals in ( C ) and ( F ) were calculated using the exact binomial method. OC, ovarian cancer; OD, OC detection; HC, healthy control; ROC, receiver operating characteristic; AUC, area under the curve; HGSOC, high-grade serous ovarian carcinoma; MC mucinous carcinoma; CCC, clear cell carcinoma; EC, endometrioid carcinoma; LGSOC, low-grade serous ovarian carcinomas; ns, not significant; *** P  < 0.001. In A and D, center line indicates the median, while lower and upper hinges represent the 25th and 75th percentiles, respectively. In C and F, the error bars represent the 95% confidence interval Accurately differentiating between malignant OC and BAM continues to pose a significant challenge, critically impacting treatment planning for high-risk or symptomatic patients. A noninvasive diagnostic tool could optimize referral pathways and reduce unnecessary interventions, especially in young or frail individuals. To address this need, we developed a BMD model using a random forest algorithm in the training cohort ( n  = 399) with BAM patients serving as controls. The feature importance scores are summarized in Table S11 (Additional file 2). The BMD model generated a malignancy probability score (BMD score), with significantly higher scores observed in OC patients compared to those with BAM (AUC = 0.982; additional file 1: Fig. S9A-B). In clinical practice, when evaluating high-risk or symptomatic populations, high sensitivity is paramount to rule out malignancy. Therefore, a moderate specificity may be acceptable if the positive predictive value remains sufficient to justify further workup. As shown in Fig. S9C (Additional file 1 ), the BMD model achieved specificities of 80.45% and 84.96% at sensitivity thresholds of 99% and 95%, respectively(Additional file 2: Table S12). To balance the need to minimize missed malignancies against the risk of excessive false positives, we set a sensitivity of 95% with a corresponding cut-off value of 0.599. Across three validation cohorts (one internal, two external), BMD scores were consistently elevated in the OC group compared to the BAM group. The model demonstrated robust diagnostic performance, with AUCs of 0.980, 0.965, and 0.961; sensitivities of 88.70%, 86.00%, and 89.92%; and specificities of 92.86%, 91.43%, and 90.20%, respectively. Furthermore, the model also performed robustly, achieving an AUC of 0.963 (95% CI: 0.925-1.000) and sensitivity exceeding 95% in Chinese minority cohort (Additional file 2: Table S13). These results confirm the generalizability and clinical utility of the BMD model across diverse patient populations(Fig.  5 ). Fig. 5 Performance of the BMD model across all validation cohorts. ( A - D left column) Comparison of BMD score between BAM patients and OC patients in four cohorts. ( A - D middle column) The ROC curves evaluating the performance of BMD model in distinguishing patients with OC from BAM. ( A - D right column) The specificity and sensitivity of BMD model in distinguishing patients with OC from BAM. Statistical significance for the comparisons in the left column was determined using the Wilcoxon rank-sum test. The 95% confidence intervals in the right column were calculated using the exact binomial method. BAM, benign adnexal masses; OC, ovarian cancer; BMD, Benign vs. Malignant Differential diagnosis; ROC, receiver operating characteristic; AUC, area under the curve; ns, not significant; *** P  < 0.001. In left column, center line indicates the median, while lower and upper hinges represent the 25th and 75th percentiles, respectively. In right column, the error bars represent the 95% confidence interval Performance of the BMD model across all validation cohorts. ( A - D left column) Comparison of BMD score between BAM patients and OC patients in four cohorts. ( A - D middle column) The ROC curves evaluating the performance of BMD model in distinguishing patients with OC from BAM. ( A - D right column) The specificity and sensitivity of BMD model in distinguishing patients with OC from BAM. Statistical significance for the comparisons in the left column was determined using the Wilcoxon rank-sum test. The 95% confidence intervals in the right column were calculated using the exact binomial method. BAM, benign adnexal masses; OC, ovarian cancer; BMD, Benign vs. Malignant Differential diagnosis; ROC, receiver operating characteristic; AUC, area under the curve; ns, not significant; *** P  < 0.001. In left column, center line indicates the median, while lower and upper hinges represent the 25th and 75th percentiles, respectively. In right column, the error bars represent the 95% confidence interval The current O-RADS classification system is inherently constrained by its dependence on sonographic morphology, which may not adequately capture the underlying tumour biology. The limitation can result in diagnostic misclassification, wherein the overclassification of BAM leads to unnecessary invasive procedures, while the underclassification of OC risks delayed therapeutic intervention. To address this diagnosis dilemma, we validated the BMD model as an adjunct tool for the preoperative reclassification of ultrasound-determined surgical candidates, to directly align management decisions with true biological risk. As shown in Fig. S10A (Additional file 1), C and E, BMD scores were consistently higher in OC group compared to BAM across different O-RADS categories (3, 4, and 5). Within the O-RADS 3 subgroup, which poses diagnostic challenges for clinical management, the BMD model effectively differentiated malignant tumors from BAM, achieving an AUC of 0.943 with 80.00% sensitivity and 93.41% specificity (Additional file 1: Fig. S10B). This robust diagnostic performance was maintained in the O-RADS 4 subgroup (AUC = 0.965; sensitivity = 88.60%; specificity = 90.10%) and the O-RADS 5 subgroup (AUC = 0.936; sensitivity = 90.26%; specificity = 84.21%), demonstrating its consistent utility across all risk strata defined by ultrasound-based classification (Additional file 1: Fig. S10D and F, additional file 2: Table S14). Remarkably, the moderate specificity observed in the O-RADS 5 subgroup stemmed from the higher malignant potential of BAMs within this category, making them challenging to distinguish from malignant tumors based on the BMD score. Totally, our data revealed that the BMD model avoided 91.00% (192/211) of unnecessary invasive surgeries in BAM patients. To further validate the clinical utility and generalizability of the BMD model in guiding management decisions, we quantified its potential to avoid unnecessary treatment, defined as the therapeutic de-escalation rate, across key clinical subgroups of patients with BAM. Our analysis demonstrated consistently high therapeutic de-escalation rates in all predefined BAM subgroups (Additional file 1: Fig. S11). Specifically, the model would have facilitated the avoidance of unnecessary surgeries in 89.57% (103/115) of patients with benign tumor, 91.84% (45/49) of those with endometriosis. This robust performance was maintained irrespective of menopausal status, with de-escalation rates of 91.53% (138/152) in postmenopausal and 90.79% (54/59) in premenopausal women. Furthermore, the model showed uniformly high efficacy across groups with varying parity. These results collectively demonstrate that the potential of the BMD model to guide therapeutic de-escalation is robust and consistent across diverse patient demographics and pathological subtypes. To establish the clinical value of our newly developed BMD model, we conducted a direct comparison against serum biomarkers, the current clinical standard for diagnosing adnexal masses, to determine if BMD model provides a meaningful diagnostic advantage. We first evaluated the performance of the BMD model in subgroups of patients stratified by serum biomarker status (positive vs. negative). As shown in Fig.  6 A, no significant difference of BMD score was found between two groups ( P  > 0.05). Furthermore, the BMD model consistently differentiated OC from BAM with high accuracy in both biomarker-positive ( n  = 397; AUC = 0.967, sensitivity = 88.92%) and biomarker-negative ( n  = 50; AUC = 0.967, sensitivity = 92.00%) subgroups (Fig.  6 B-C). We systematically compared the BMD model with three mainstream serum biomarker strategies, including binary rule, the LR (CA125&HE4) model and ROMA Index using standardized cut-offs to eliminate selection bias (Additional file 2: Table S15). Given the primary clinical need to avoid missed diagnoses, the LR (CA125&HE4) model with high sensitivity was used to further analysis. ROC analysis confirmed the BMD model’s superior diagnostic accuracy, with a substantially higher AUC (0.967 vs. 0.879; Fig.  6 D). Critically, at the clinically critical 95% sensitivity, the BMD model achieved a 2.27-fold higher specificity (84.93% vs. 37.4%), greatly reducing unnecessary surgeries for benign lesions. To optimize biomarker integration, we constructed a logistic regression model incorporating BMD score, CA125 and HE4 as continuous variables in the training cohort (Additional file 2: Table S16). As shown in the Fig.  6 E, combined model achieved 95.70% specificity at 91.00% sensitivity. Decision curve analysis further validated that both the BMD and combined models yielded significantly higher clinical net benefits than serum biomarkers across all relevant thresholds (Fig.  6 F). Fig. 6 Comparison of performance between serum biomarkers and the BMD model. ( A ) Comparison of BMD score between OC patients with positive or negative serum biomarkers. ( B ) The ROC curves evaluating the performance of BMD model in distinguishing biomarker-positive or biomarker-negative OC patients from BAM. ( C ) The sensitivity of BMD model in detecting biomarker-positive or biomarker-negative OC patients. ( D ) The ROC curves comparing the diagnostic performance of BMD model and serum biomarkers (combined via logistic regression) in distinguishing OC from BAM. ( E ) Sensitivity and specificity of three distinct diagnostic models: the BMD model, serum biomarkers (combined CA125&HE4 via LR), and the combined model (BMD&CA125&HE4 via LR). ( F ) Decision curve analysis (DCA) showing the net benefit of the BMD model, serum biomarkers (combined CA125&HE4 via LR), and the combined model (BMD&CA125&HE4 via LR) at various threshold probabilities. Statistical significance in (A) was determined using the Wilcoxon rank-sum test. The 95% confidence intervals in (C) were calculated using the exact binomial method. BAM, benign adnexal masses; OC, ovarian cancer; BMD, Benign vs. Malignant Differential diagnosis; AUC, area under the curve; LR, logistic regression; ns, not significant. In A, center line indicates the median, while lower and upper hinges represent maximum and minimum. In C, the error bars represent the 95% confidence interval Comparison of performance between serum biomarkers and the BMD model. ( A ) Comparison of BMD score between OC patients with positive or negative serum biomarkers. ( B ) The ROC curves evaluating the performance of BMD model in distinguishing biomarker-positive or biomarker-negative OC patients from BAM. ( C ) The sensitivity of BMD model in detecting biomarker-positive or biomarker-negative OC patients. ( D ) The ROC curves comparing the diagnostic performance of BMD model and serum biomarkers (combined via logistic regression) in distinguishing OC from BAM. ( E ) Sensitivity and specificity of three distinct diagnostic models: the BMD model, serum biomarkers (combined CA125&HE4 via LR), and the combined model (BMD&CA125&HE4 via LR). ( F ) Decision curve analysis (DCA) showing the net benefit of the BMD model, serum biomarkers (combined CA125&HE4 via LR), and the combined model (BMD&CA125&HE4 via LR) at various threshold probabilities. Statistical significance in (A) was determined using the Wilcoxon rank-sum test. The 95% confidence intervals in (C) were calculated using the exact binomial method. BAM, benign adnexal masses; OC, ovarian cancer; BMD, Benign vs. Malignant Differential diagnosis; AUC, area under the curve; LR, logistic regression; ns, not significant. In A, center line indicates the median, while lower and upper hinges represent maximum and minimum. In C, the error bars represent the 95% confidence interval Due to their histopathological similarities with both BAM and OC, BOT present significant diagnostic and therapeutic challenges. Approximately 10–20% of BOT are misclassified during initial evaluation, frequently leading to inappropriate conservative management and subsequent unnecessary repeated surgeries [ 38 ]. Accurate preoperative diagnosis of BOT is critical to guide definitive surgical planning. Therefore, we evaluated the ability of the BMD model to detect patients with BOT. As shown in Fig.  7 , the BMD model demonstrated significantly higher sensitivity and positive predictive value (PPV) than serum biomarkers (combined CA125&HE4 via logistic regression) for identifying BOT patients (73.08% vs. 26.92% and 66.67% vs.16.20%). These results indicate that the BMD model achieves superior sensitivity and positive predictive value over serum biomarkers, enabling more reliable preoperative detection of BOT. Fig. 7 Performance of the BMD model in detecting patients with BOT. ( A ) Comparison of BMD scores among patients with BAM, BOT and OC. The dashed line indicates the optimal cut-off value (0.599). ( B ) Sensitivity and ( C ) positive predictive value of BMD model and serum biomarkers (combined CA125&HE4 via LR) in detecting BOT. Statistical significance in (A) was evaluated using the Kruskal-Wallis H test followed by Dunn’s multiple comparisons test with Bonferroni correction. The 95% confidence intervals in (B) and (C) were calculated using the exact binomial method. BAM, benign adnexal masses; BOT, borderline ovarian cancer; OC, ovarian cancer; BMD, Benign vs. Malignant Differential diagnosis; *** P  < 0.001. In A, center line indicates the median, while lower and upper hinges represent the maximum and minimum values. In B and C, the error bars represent the 95% confidence interval Performance of the BMD model in detecting patients with BOT. ( A ) Comparison of BMD scores among patients with BAM, BOT and OC. The dashed line indicates the optimal cut-off value (0.599). ( B ) Sensitivity and ( C ) positive predictive value of BMD model and serum biomarkers (combined CA125&HE4 via LR) in detecting BOT. Statistical significance in (A) was evaluated using the Kruskal-Wallis H test followed by Dunn’s multiple comparisons test with Bonferroni correction. The 95% confidence intervals in (B) and (C) were calculated using the exact binomial method. BAM, benign adnexal masses; BOT, borderline ovarian cancer; OC, ovarian cancer; BMD, Benign vs. Malignant Differential diagnosis; *** P  < 0.001. In A, center line indicates the median, while lower and upper hinges represent the maximum and minimum values. In B and C, the error bars represent the 95% confidence interval

Ovarian cancer (OC) represents a formidable global health challenge, with over 300,000 new cases and approximately 200,000 deaths annually worldwide [ 1 , 2 ]. OC is the most lethal malignancy of the female reproductive system, a stark statistic largely attributable to its insidious onset and the absence of distinct early symptoms [ 3 ].Consequently, more than 70% of patients are diagnosed at an advanced stage (FIGO Stage III or IV), where the 5-year survival rate plummets to below 30%, compared to over 90% when the disease is detected early (Stage I) [ 4 ].The profound disparity in survival underscores the critical importance of developing effective strategies for early detection. Furthermore, the accurate differentiation of malignant ovarian tumors from benign adnexal masses (BAM) is a pivotal yet persistent clinical dilemma [ 5 – 7 ]. Each year, a vast number of women undergo surgery for suspicious adnexal masses, yet the majority of these procedures reveal benign pathology. The high rate of unnecessary interventions not only imposes significant physical, psychological, and economic burdens on patients but also strains healthcare resources. Therefore, a reliable non-invasive tool for both OC early detection and precise benign-malignant differential diagnosis is urgently needed to improve survival outcomes and reduce overtreatment. Current approaches for OC detection and differential diagnosis primarily rely on pelvic ultrasound and serum biomarker assays, most notably Cancer Antigen 125 (CA-125) and Human Epididymis Protein 4 (HE4) [ 4 – 8 ]. Pelvic ultrasound is employed for morphological assessment, while the Ovarian-Adnexal Reporting and Data System (O-RADS) has been widely adopted in ultrasound assessment for providing a standard malignant risk stratification system. Despite this advancement, the interpretation of ultrasound findings remains limited by insufficient specificity in distinguishing benign from malignant lesions, particularly pronounced for indeterminate and suspicious lesions classified as O-RADS 3 and 4 [ 9 ]. Serum biomarkers also have well-documented limitations. CA-125 can be elevated in various benign conditions such as endometriosis, leading to false positives [ 10 , 11 ]. Conversely, it is not elevated in a significant proportion of early-stage OC (EOC) cases, resulting in false negatives. Although HE4 improves specificity, its sensitivity for early-stage disease remains suboptimal. The combination of these tools into multivariate indices (e.g., the Risk of Malignancy Index) has improved performance but still falls short of the accuracy required for reliable risk stratification in clinically relevant settings, including high-risk or symptomatic individuals [ 12 ]. The limitations of existing modalities highlight a significant unmet clinical need for novel, highly specific, and sensitive molecular biomarkers that can complement or surpass current standards of care. As a minimally invasive technique, liquid biopsy interrogates circulating tumor-derived components, with circulating tumor DNA (ctDNA) and cell-free DNA (cfDNA) representing the most commonly used analytes [ 13 , 14 ]. There is growing evidence that this approach holds significant potential to transform the clinical management of OC, especially by enabling non-invasive monitoring for early detection. However, technical hurdles remain. Methylation analysis, for example, is fundamentally limited by the low concentration of extracted ctDNA, a problem especially exacerbated in EOC [ 15 – 17 ]. Beyond the issue of concentration, the lack of specificity (since some methylated markers are not unique to OC) and the poor reproducibility of methylation assays raise further concerns [ 18 , 19 ]. Although ctDNA mutation analysis can achieve high specificity, its sensitivity for early-stage (I-II) cancer is often suboptimal and highly variable [ 20 – 22 ]. As demonstrated in the study by Phallen et al., ctDNA analysis has demonstrated a 68% detection rate in EOC [ 23 ], while Abbosh C. et al. have reported that the detection rate of ctDNA varies not only across different pathological types of lung cancer but also within the same pathological type [ 24 ].The emerging field of cfDNA fragmentomics presents a promising alternative [ 25 – 27 ]. A recent study has demonstrated that circulating cell-free nuclear DNA(ccf-nDNA) fragmentomics combined with serum biomarkers can non-invasively detect EOC and distinguish OC from benign adnexal masses (BAM) [ 25 ]. However, the low sequencing coverage poses a challenge to the reproducibility of cfDNA fragmentation profiles, an issue that must be addressed to enable their clinical translation. Beyond ccf-nDNA, plasma contains abundant circulating cell-free mitochondrial DNA (ccf-mtDNA) [ 28 ]. Unlike nuclear DNA, mtDNA exhibits a higher copy number per cell, increased susceptibility to oxidative damage, and distinct patterns of fragmentation, positioning it as a potentially source of cancer-derived signals in plasma [ 28 , 29 ]. We previously established a cost-efficient and adaptable capture-based sequencing method for comprehensive profiling of ccf-mtDNA fragmentomics and demonstrated that ccf-mtDNA carried tumor-specific fragmentomic features, suggesting a potential link to mitochondrial genomic instability and malignancy-associated nucleolytic cleavage patterns [ 30 , 31 ]. Critically, ccf-mtDNA fragmentomics exhibits superior diagnostic accuracy for early-stage hepatocellular, colorectal, and non-small cell lung cancers compared to conventional biomarkers [ 32 – 34 ]. However, the clinical utility of ccf-mtDNA fragmentomics in early detection and differential diagnosis of OC has not been systematically explored. In this multicenter study, we leveraged plasma ccf-mtDNA fragmentomics to develop two machine-learning models for OC management: an OC Detection (OD) model to distinguish OC patients from healthy controls (HC), and a benign vs. malignant differential diagnosis (BMD) model to discriminate OC from BAM. We rigorously evaluated the models’ performance in internal, external, and Chinese minority validation cohorts, specifically testing their generalizability across populations with distinct genetic and lifestyle backgrounds.

This study focuses on the critical unmet needs for non-invasive early detection of OC and accurate differential diagnosis between BAM and OC. Our analysis revealed distinct plasma ccf-mtDNA fragmentomic features as novel biomarkers and developed two models (OC vs. HC and OC vs. BAM). The robust performance of these models was rigorously confirmed in all validation cohorts comprising the Han and other ethnic populations. Characteristic cfDNA fragmentomic signatures associated with distinct pathological states provide a theoretical basis for exploring tumorigenesis and differentiating diseases [ 39 , 40 ]. In this study, we identified significant differences in plasma ccf-mtDNA fragmentomics among HC, BAM and OC, including fragment size distribution, fragmentation profiles, 5’ end motif preferences, and 5’ end MDS. These alterations may reflect changes in nuclease activity or dysregulation of mtDNA-binding proteins. Contrary to the shortening of ccf-nDNA fragments commonly seen in cancer, and even the shortening of ccf-mtDNA we previously observed in colorectal cancer [ 33 ], OC patients exhibited elevated proportions of long ccf-mtDNA fragments. This divergent fragmentation pattern strongly suggests tumor-type-specific mechanisms of cfDNA release and processing. While the anatomical proximity of OC to highly vascularized peritoneal and omental surfaces may facilitate continuous nucleic acid shedding into the circulation, this distinct fragmentomic elongation is likely driven by profound alterations in underlying molecular and cellular mechanisms. First, this may reflect a shift in tumor cell death pathways. While physiological apoptosis generates precisely cleaved short DNA fragments, the rapid proliferation and subsequent hypoxia within OC masses frequently trigger tumor necrosis, leading to the haphazard release of incompletely digested, longer DNA fragments [ 41 ]. Second, OC is notorious for its robust generation of extracellular vesicles (EVs) [ 42 ]. It is highly plausible that a substantial fraction of tumor-derived mtDNA is encapsulated within these EVs, which physically shield the DNA from complete degradation by circulating nucleases (such as DNASE1) in the plasma. Thus, the enriched long ccf-mtDNA signature in OC likely represents a composite readout of tumor necrosis and vesicular compartmentalization. Future studies should precisely elucidate the cellular origins and these molecular mechanisms underlying cfDNA release in malignancies. Such insights may facilitate the development of novel liquid biopsy platforms with improved diagnostic accuracy and clinical utility. Early detection of OC patients can significantly delay disease progression and reduce mortality. However, serum biomarkers like CA125 and HE4 demonstrate limited diagnostic performance. We successfully established the OD model based on ccf-mtDNA fragmentomics, which achieved a sensitivity of 85.87% at 95% specificity, significantly outperforming serum biomarkers [ 8 , 43 ]. The OD model also demonstrated high diagnostic accuracy across four validation cohorts, with AUC values of 0.981, 0.994, 0.995, 0.994, underscoring its robustness and generalizability. Subgroup analyses confirmed that the model’s consistent performance across different tumor stages, with excellent sensitivity in EOC, demonstrating its suitability for early detection. Importantly, the OD model is intended for opportunistic testing in clinically enriched populations, such as individuals with hereditary risk, persistent symptoms, abnormal biomarkers, or suspicious imaging findings, rather than for population-wide screening of average-risk women. In these target populations, the expected prevalence of OC and the pretest probability are higher than those in the general population, making the 95% specificity more clinically actionable [ 45 , 46 ]. The high sensitivity of the OD model may help rule out OC when the test result is negative, whereas positive results should prompt further evaluation using ultrasound, serum biomarkers, and clinical assessment rather than being interpreted as a stand-alone diagnostic result. Compared to previously reported approaches for OC detection based on DNA methylation and nuclear genomics, this model demonstrated superior performance [ 25 , 44 ]. To our knowledge, this represents the first study to explore the application of ccf-mtDNA fragmentomics features for early detection of OC. The clinical differential diagnosis between benign and malignant adnexal masses relies primarily on serum biomarkers and imaging such as ultrasound [ 47 , 48 ]. However, the efficacy of ultrasound is often limited by the biological heterogeneity of tumors. This diagnostic uncertainty frequently results in suboptimal management decisions, such as unnecessary invasive surgeries for benign cases or inadequate initial staging for malignancies.To address these challenges, we developed the BMD model by leveraging unique fragmentomic profiles that distinguish BAM from OC. The model demonstrated strong discriminatory power, achieving an AUC of 0.982 for OC detection, with a specificity of 90% at 95% sensitivity. It consistently maintained sensitivities above 85% across all validation cohorts. The BMD model is intended for preoperative triage of patients with O-RADS 3–5 lesions, a clinically enriched setting in which the prevalence of malignancy is substantially higher than that in the general population. In our center, the estimated malignancy rate among patients with O-RADS 3–5 lesions was approximately 50%, consistent with the reported risk range for O-RADS 3–5 categories [ 9 ]. In this context, the clinical goal is to rule out malignancy with high confidence, thereby reducing unnecessary surgeries while ensuring timely referral for oncological management. Adnexal masses categorized as O-RADS 3 or 4 on ultrasound imaging often present diagnostic uncertainty due to nonspecific or overlapping imaging features, frequently leading to clinical misclassification and inconsistent management. The BMD model exhibited superior performance in these challenging subgroups, achieving AUC of 0.981 and 0.975 for O-RADS 3 and 4 lesions, respectively. By serving as a precision adjunct to ultrasound, the model definitively resolves the diagnosis ambiguity of O-RADS 3–4 lesions, allowing for more confident preoperative triage. This capability is crucial for reducing unnecessary extensive surgeries in benign cases while ensuring that patients with malignancy receive immediate referral to gynecologic oncologists. Our findings suggest that ccf-mtDNA fragmentomics can serve as a molecular “second opinion” for indeterminate adnexal masses. Furthermore, the BMD model showed promising performance in identifying BOT, achieving sensitivity of 73.08%. It offered a reliable tool for the prediction and treatment dilemma of BOT in clinical practice. It also outperformed conventional serum biomarkers in accuracy of differential diagnosis and maintained robust diagnostic performance even among serum biomarker-negative OC patients, highlighting its broad applicability. By combining the low cost of routine ultrasound with the diagnostic power of ccf-mtDNA fragmentomics, this model establishes a cost-effective platform that is readily scalable for clinical application in preoperative decision-making. To comprehensively evaluate the clinical applicability of the OD and BMD models given the variations in genetic background, lifestyle, and environmental exposure factors among different ethnic minorities, our study additionally included a multi-ethnic cohort from Tumor Hospital Affiliated to Xinjiang Medical University, encompassing Hui, Uyghur, Mongolian, and other ethnic groups. In this cohort, the OD model demonstrated strong discriminatory performance, with an AUC of 0.994 and a sensitivity of 92.96%. The BMD model also exhibited excellent results, achieving an AUC of 0.963 and a sensitivity of 96.83%. The high predictive accuracy of both models in multi-ethnic population not only validates their cross-ethnic robustness and generalizability but also supports their future clinical translation to broader populations. Collectively, our findings define a plasma ccf-mtDNA fragmentomic signature unique to OC and demonstrate its translational potential. The development and rigorous validation of the OD and BMD models across diverse populations provide robust evidence that ccf-mtDNA fragmentomics can address two critical, unmet clinical needs: early detection and differential diagnosis. Notably, the OD model showed no statistically significant difference in prediction scores across FIGO stages (Fig.  4 A). While nuclear ctDNA signals typically increase with tumor burden, ccf‑mtDNA fragmentomics primarily reflect systemic mitochondrial stress rather than direct tumor shedding. As demonstrated in our previous studies on hepatocellular carcinoma and colorectal cancer [ 32 , 33 ], ccf-mtDNA alterations are already elevated in early-stage disease and do not exhibit a strong stage-dependent increase. However, the reduced sensitivity observed for FIGO stage II is likely attributable to the small sample size of this subgroup, which limits the stability of performance estimates. Larger multi-center studies specifically targeting stage II patients are warranted to further validate our findings. Furthermore, our work provides new perspectives on tumor cfDNA biology and introduces a novel liquid biopsy approach grounded in mtDNA fragmentation profiling. Future research should validate these models in larger, prospective cohorts and more clinical scenario, including distinct entities like BOT and explore the utility of ccf-mtDNA profiles as dynamic biomarkers for monitoring treatment response and detecting recurrence. Ultimately, multi-center validation and technical standardization will be essential to translate this methodology from bench to bedside, with the long-term goal of reducing the incidence and mortality of OC via proactive, evidence-based intervention strategies.

In summary, we identify a plasma ccf-mtDNA fragmentomic signature unique to ovarian cancer. This signature underpins two clinical models that significantly improve early detection and differential diagnosis from benign masses across diverse populations in targeted clinical populations. Our findings reveal a distinct pattern of mtDNA fragmentation in OC biology and establish a novel framework for liquid biopsy. Prospective multi-center trials are the critical next step toward clinical translation.

Supplementary Material 1. Supplementary Material 1. Supplementary Material 2. Supplementary Material 2.