Multimodal tumor-agnostic ctDNA analysis for minimal residual disease detection and risk stratification in ovarian cancer: results from the MITO16a/MaNGO-OV2 trial

Abstract

Background

Advanced-stage epithelial ovarian cancer (EOC) remains a therapeutic challenge due to high relapse rates and limited survival, while standard post-surgical parameters such as residual tumor (RT) incompletely capture minimal residual disease (MRD) and offer limited insight into tumor evolution. To address this gap, we investigated whether a multimodal, tumor-agnostic analysis of circulating tumor DNA (ctDNA)—integrating tumor fraction (TF) and genome-wide fragmentomic profiles (PF)—could refine early risk stratification after cytoreductive surgery and enable longitudinal monitoring during therapy.

Materials and methods

A total of 393 plasma samples from 173 patients in the phase IV MITO16a/MaNGO-OV2a trial were analyzed by shallow whole-genome sequencing at three time points: post-surgery/pre-chemotherapy (B1), post-chemotherapy (B2), and at the end of maintenance therapy or upon disease progression during maintenance (B3). Associations with progression-free survival (PFS) and overall survival (OS) were assessed using multivariable Cox models adjusted for clinical covariates.

Results

TF was detectable in 97% of patients at B1, including those classified as optimally debulked, and outperformed established clinical covariates in predicting survival [PFS: hazard ratio (HR) 1.02, P = 0.008; OS: HR 1.04, P = 0.005]. PF provided independent prognostic values (PFS: HR 1.06, P = 0.010; OS: HR 1.10, P = 0.005), and combined TF/PF modeling identified subgroups with distinct survival trajectories beyond clinical predictors (PFS: HR 1.76, P = 0.015; OS: HR 2.06, P = 0.029). Longitudinal copy number profiling revealed dynamic remodeling under treatment pressure, with recurrent 19q13.42 amplification emerging at B2 and B3.

Conclusions

Together, these findings establish multimodal ctDNA profiling as a sensitive, non-invasive strategy for MRD detection and longitudinal surveillance in advanced EOC, refining prognostic assessment beyond clinical and surgical factors while paving the way for precision-guided therapeutic management.

Highlights

• •Multimodal ctDNA detects MRD after surgery in ovarian cancer trial patients.
• •TF predicts survival better than surgical RT.
• •Fragmentomics (PF) independently stratifies post-surgery patient risk.
• •A combined TF/PF model refines early prognosis in the MITO16a/MaNGO-OV2 trial.
• •Longitudinal ctDNA reveals therapy-driven genomic changes, including 19q13.42 amplifications.

Introduction

Epithelial ovarian cancer (EOC) remains one of the most lethal gynecological malignancies, with the majority of patients diagnosed at advanced stages [International Federation of Gynecology and Obstetrics (FIGO) III-IV] and relapsing within 18 months of first-line therapy, despite initial platinum (Pt) sensitivity.¹ The lack of predictive biomarkers and the marked molecular heterogeneity of the disease continue to hinder personalized treatment approaches and long-term disease control.

Maintenance therapies such as poly (ADP-ribose) polymerase inhibitors (PARPi) and anti-angiogenic agents have significantly improved outcomes in selected subgroups of patients: PARPi show maximal efficacy in the presence of homologous recombination deficiency (HRD) or BRCA1/2 mutations, while bevacizumab provides greater benefit in patients with high-risk clinical features.^2,3

Among clinicopathological features, the extent of residual tumor (RT) after cytoreductive surgery remains the most powerful prognostic determinant in advanced stages EOC and is now increasingly recognized as a key factor in guiding maintenance treatment allocation.⁴ Beyond its well-established prognostic value, RT also appears to influence the relative efficacy of available maintenance strategies. Subgroup analyses from ICON-7 and GOG-0218 indicated that patients with macroscopic RT (R2), who represent a high-risk subgroup, derive a greater overall survival (OS) benefit from the addition of bevacizumab to first-line chemotherapy and maintenance compared with the broader trial populations.^5,6 Conversely, trials evaluating PARPi, such as PRIMA and PAOLA-1, have consistently shown that patients achieving no visible disease (R0) or minimal RT (R1) derive the most pronounced and durable benefit from PARPi maintenance, although efficacy is maintained across residual disease subgroups.^2,7

To date, RT assessment relies on intraoperative judgment and lacks standardization. Furthermore, widely used markers such as CA-125^8,9 or Pt-free interval remain suboptimal due to limited sensitivity and predictive value.¹⁰ Similarly, although BRCA1/2 status and HRD testing have advanced patient selection for PARPi,^2,11,12 these molecular features are insufficient to identify patients at high risk of early relapse or to guide timely treatment adaptation.

This unmet clinical need underscores the critical importance of accurately assessing minimal residual disease (MRD) after surgery—a decisive window for tailoring therapy and improving patient outcomes.¹³

Liquid biopsy (LB), and in particular cell-free DNA (cfDNA) analysis, offers a promising, non-invasive strategy to monitor tumor burden and molecular evolution in real time. In EOC, where serial tissue biopsies are rarely feasible, circulating tumor DNA (ctDNA)—the tumor-derived fraction of cfDNA—can reveal both point mutations and somatic copy number variations (CNVs), providing a dynamic readout of MRD.¹⁴ This is particularly relevant to high grade serous ovarian cancer (HGSOC), the most genomically unstable and prevalent EOC subtype.^15,16 Recent advances have extended ctDNA analyses beyond genetic alterations to include fragmentomics, the systematic characterization of cfDNA fragmentation patterns, which are largely influenced by chromatin structure, nucleosome positioning, DNA methylation, and nuclease activity. ctDNA often exhibits distinctive fragment size distributions, enriched for shorter fragments [<150 base pair (bp)] compared with cfDNA from normal cells, as well as specific end motifs and coverage patterns that reflect nucleosome footprints and epigenetic features. This fragmentomic signature provides molecular insights and enables improved detection and quantification of ctDNA, even when tumor burden is low.^17,18

Recent advances in LB have enabled highly sensitive, blood-based assays for MRD detection. Among these, shallow whole-genome sequencing (sWGS) has significantly broadened the scope of ctDNA profiling by supporting comprehensive, genome-wide analyses without requiring prior knowledge of tumor-specific alterations (named as tumor-agnostic approach). Unlike tumor-informed methods, tumor-agnostic strategies can simultaneously provide estimates of different ctDNA metrics, including tumor fraction (TF), CNV burden, and fragmentomic signatures, thereby enhancing sensitivity for MRD detection beyond conventional clinical and molecular predictors and informing therapeutic decisions.¹⁹

Building on our previous observation that increases in TF can anticipate clinical relapse by up to 18 months,¹⁶ we hypothesized that in the immediate post-surgical setting, combining TF with fragmentomic profile (PF) (i.e. multimodal approach) could refine MRD detection and improve early risk stratification. To test this, we conducted a translational study nested within the MITO16a/MaNGO-OV2 trial,²⁰ assessing the prognostic relevance of multimodal ctDNA analysis in patients with advanced EOC treated with Pt-based chemotherapy and bevacizumab maintenance. The primary aim of this study was to determine whether integrating TF and PF can enhance post-operative MRD assessment and support personalized therapy adaptation. In addition, by carrying out longitudinal profiling of ctDNA-derived CNVs, we sought to explore clonal dynamics evolution during and after treatment, thereby providing new insights into the mechanisms of tumor evolution and resistance.

Materials and methods

Study approval and cohort selection

Among 389 patients enrolled in the MITO16a/MaNGO-OV2a trial (EudraCT 2012-003043-29), 173 with stage IIIB-IV EOC and available longitudinal plasma samples were included. All participants received first-line treatment with carboplatin, paclitaxel, and bevacizumab, followed by maintenance bevacizumab after primary debulking surgery.²⁰ Details are reported in the Supplementary Results section, available at https://doi.org/10.1016/j.esmoop.2026.106087. The study was conducted according to the Declaration of Helsinki and International Council on Harmonization and Good Clinical Practice guidelines. Ethical approval was granted by the Ethics Committee of the Istituto Nazionale Tumori, IRCCS, Fondazione G. Pascale, Naples (Ref. No. 187/16E, 10 October 2016). All patients or their legal representatives provided written informed consent.

Study design and translational aims

Figure 1 illustrates the overall study design and analytical framework. A total of 393 plasma samples were collected at three time points: baseline (B1, post-surgery and pre-chemotherapy; n = 172), post-chemotherapy (B2; n = 120), and during maintenance therapy (B3, n = 101), which included patients who either relapsed during bevacizumab maintenance (B3 relapsed, sampled at relapse) or remained progression-free until the end of therapy (B3 responders, sampled at the cycle 22, including six cycles of chemotherapy) (Supplementary File 1, available at https://doi.org/10.1016/j.esmoop.2026.106087). To detect and profile ctDNA at multiple time points, sWGS was carried out on each plasma sample to derive three key cfDNA-based biomarkers: (i) TF, (ii) PF, and (iii) the landscape of CNVs. While TF and PF are quantitative and analyzed in combination, CNV profiles are qualitative in nature and are described separately.

Figure 1.

Study design, longitudinal sampling, and analytical framework of the MITO16a/MaNGO-OV2 translational study. Schematic overview of the clinical timeline, patient cohorts, and sampling strategy in the MITO16a/MaNGO-OV2 translational study. Among 389 patients enrolled in the parent phase IV trial, 173 were included in the translational cohort. Plasma samples were obtained at three predefined time points: baseline (B1, post-surgery and pre-chemotherapy; n = 172), post-chemotherapy (B2; n = 120), and post-bevacizumab maintenance therapy (B3; n = 101). At B3, patients were stratified according to clinical outcome into those who relapsed during maintenance therapy (B3-relapsed) and those who completed maintenance without recurrence (B3-responders). All plasma samples underwent sWGS for the analysis of TF, genome-wide cfDNA PF, and somatic CNVs. The translational study addressed two primary objectives: aim 1, to assess TF and PF at B1, individually or in combination, for MRD detection and early prognostic stratification. Aim 2, to carry out longitudinal CNV profiling across B1, B2, and B3 to monitor tumor evolution. Representative genomic events, including 19q13.42 amplification, are shown as examples. Partially generated with BioRender. Beva, bevacizumab; cfDNA, cell-free DNA; CNV, copy number variation; CT, chemotherapy; ctDNA, circulating tumor DNA; EOC, epithelial ovarian cancer; MRD, minimal residual disease; PF, fragmentomic profile; Pt, platinum; sWGS, shallow whole-genome sequencing; TF, tumor fraction.

The primary aim was to assess whether post-operative ctDNA analysis could serve as a marker of MRD, enabling the identification of patients with distinct outcomes. Accordingly, aim 1 focused on evaluating the prognostic value of TF and PF—individually and in combination—in B1 post-surgical plasma samples collected before Pt-based chemotherapy, thereby minimizing potential treatment-related confounding effects. The secondary aim (aim 2) was to investigate longitudinal tumor genome dynamics under therapeutic pressure by profiling CNVs in matched plasma samples collected at B1, B2, and B3.

Sample processing and sWGS analysis

Peripheral blood was collected at three time points: baseline (B1, post-surgery/pre-chemotherapy), post-chemotherapy (B2), and during bevacizumab maintenance (B3). Details on plasma cfDNA extraction purification and library preparation are reported in the Supplementary Methods, available at https://doi.org/10.1016/j.esmoop.2026.106087, and as previously described.¹⁶

Statistical analyses

Baseline characteristics were summarized as absolute numbers and percentages for categorical variables, and as mean (±standard deviation) or median [interquartile range (IQR)] for continuous variables, and were compared with the overall MITO16a/MaNGO-OV2a cohort. Associations between TF, PF, and clinical features were evaluated using non-parametric tests. The prognostic value of ctDNA-derived metrics for progression-free survival (PFS) and OS was assessed using Cox proportional hazards models in univariate and multivariable settings, adjusted for key clinical covariates. Optimal cut-off values for TF and PF were identified using an outcome-oriented approach based on PFS. Candidate thresholds were evaluated within Cox proportional hazards models, and the cut-offs yielding the strongest separation in PFS, as reflected by the hazard ratio (HR), were selected. These thresholds were subsequently applied in Kaplan–Meier survival analyses. Given the exploratory nature of this study, cut-off selection was not intended for definitive inference but to support risk stratification and hypothesis generation. The two complementary variables, TF and PF, were combined into a composite variable comprising three categories: group 0 (low TF/low PF), group 1 (low TF/high PF or high TF/low PF), and group 2 (high TF/high PF). Survival differences among groups were evaluated using the log-rank test. Discordant TF/PF combinations were aggregated into a single intermediate group to ensure adequate event numbers and robust statistical comparisons.

Additional details on the study methods are provided in the Supplementary Methods, available at https://doi.org/10.1016/j.esmoop.2026.106087.

Results

Patient cohort

A subset of 173 patients from the 389 enrolled in the multicenter, phase IV MITO16a/MaNGO-OV2a clinical trial were selected for this translational study. This subgroup, defined as the translational cohort, was chosen based on the availability of longitudinal plasma samples (REMARK diagram, Supplementary Figure S1, available at https://doi.org/10.1016/j.esmoop.2026.106087). Baseline clinical and pathological characteristics were comparable between the translational cohort and the full trial population. As summarized in Table 1, the median age at diagnosis was comparable between the translational cohort and the full study population (58.5 versus 59.2 years), as were the distribution of Eastern Cooperative Oncology Group (ECOG) performance status, FIGO stage, and RT following cytoreduction surgery. FIGO stage IIIC was the most frequent substage (73% versus 69%), and HGSOC was the predominant histological subtype (86% versus 84%). Kaplan–Meier survival analysis demonstrated overlapping PFS and OS curves between the two groups confirming the representativeness of the translational cohort for real-world advanced EOC (Supplementary Figure S2, available at https://doi.org/10.1016/j.esmoop.2026.106087).

Table 1.

Translational study cohort description

	Translational study population	MITO16a/MaNGO-OV2a population	P value
	(n = 173)	(n = 398)
Median age (IQR), years	58.5 (49.9-65.8)	59.2 (49.9-66.5)	0.770
Age category, n (%)			0.660
<65 years	124 (72)	278 (70)	—
≥65 years	49 (28)	120 (30)	—
ECOG performance status, n (%)			0.627
0	135 (78)	315 (79)	—
1	34 (20)	69 (17)	—
2	4 (2)	14 (4)	—
Residual disease, n (%)			0.597
None	68 (40)	153 (38)	—
≤1 cm	36 (21)	72 (18)	—
>1 cm/not operated	69 (39)	173 (44)	—
FIGO stage, n (%)			0.667
IIIB	14 (8)	36 (9)	—
IIIC	126 (73)	275 (69)	—
IV	33 (19)	87 (22)	—
Tumor histology, n (%)			0.684
High-grade serous	148 (86)	333 (84)	—
Low-grade serous	4 (2)	13 (3)	—
Endometrioid	7 (4)	9 (2)	—
Clear cell	4 (2)	11 (3)	—
Mucinous	2 (1)	3 (1)	—
Mixed	0 (0)	4 (1)	—
Other	8 (5)	25 (6)	—

Summary of the main clinical, pathological, and demographic characteristics of patients included in the translational study cohort (n = 173) compared with the overall population enrolled in the MITO16a/MaNGO-OV2a clinical trial (n = 398). Statistical comparisons were carried out using the Wilcoxon rank-sum test for continuous variables, Pearson’s chi-square test for categorical variables, and Fisher’s exact test with simulated P values for categories with small sample sizes.

ECOG, Eastern Cooperative Oncology Group; FIGO, International Federation of Gynecology and Obstetrics; IQR, interquartile range.

Tumor fraction reflects disease burden and residual tumor after cytoreductive surgery

At time point B1, TF was quantifiable in 167 of 172 plasma samples (97%) with a median value of 6.1% (IQR 4.3%-10.5%), consistent with the platform’s validated limit¹⁶ (Supplementary File 1, available at https://doi.org/10.1016/j.esmoop.2026.106087). To evaluate its translational relevance, we examined associations between B1 TF and prognostic variables including age, FIGO stage (III versus IV), ECOG performance status, histological subtype, BRCA mutation, and RT after primary cytoreduction.

As shown in Supplementary Figure S3 and Table S1, available at https://doi.org/10.1016/j.esmoop.2026.106087, TF was significantly higher in FIGO stage IV versus stage III disease (median TF 8.71% versus 5.73%, P = 0.028), reflecting increased tumor burden. No significant associations emerged with ECOG performance status, histotype, BRCA status, or age.

Importantly, TF correlated with RT, the most important post-operative prognostic factor. Patients with macroscopic residual disease (R2) had higher TF (median 6.83%, IQR 4.63-13.83) compared with those with minimal (R1) or no residual disease (R0) (P = 0.0094). TF values were comparable in patients with R0 (median 5.52%, IQR 4.06-6.63) and those with R1 (median 5.77%, IQR 5.65-8.65), suggesting TF may uncover microscopic disease not captured by surgical assessment. These findings support TF as a minimally invasive biomarker of post-surgical tumor burden, offering prognostic insight that complements—or may surpass—conventional surgical metrics.

Tumor fraction as an independent prognostic biomarker following cytoreductive surgery

To evaluate the prognostic relevance of ctDNA before chemotherapy, we analyzed TF levels at time point B1, assessed both as a continuous and a dichotomous variable in relation to PFS and OS in univariate and multivariate analysis.

Kaplan–Meier analysis (Figure 2A and B) showed that patients with TF < 15.08% (n = 144) had significantly longer PFS than those with TF ≥15.08% [n = 23; median PFS 22.7 versus 15.9 months; HR 2.56, 95% confidence interval (CI) 1.22-5.38, P = 0.010], while no significant OS association was observed. When modeled continuously, TF was strongly associated with both PFS (HR 1.04, 95% CI 1.02-1.06, P < 0.001) and OS (HR 1.04, 95% CI 1.02-1.07, P < 0.001) (Supplementary Table S2, available at https://doi.org/10.1016/j.esmoop.2026.106087).

Figure 2.

Prognostic impact of plasma biomarkers on survival outcomes. Kaplan–Meier survival curves (log-rank test) illustrating the association of dichotomized TF and plasma PF with progression-free survival (panels A, C, E) and overall survival (panels B, D, F). (A and B) Patients stratified by TF levels into high (>15.08%, blue) and low (≤15.08%, red). (C and D) Patients stratified by PF levels into high (>0.1659, blue) and low (≤0.1659, red). (E and F) Combined TF/PF multimodal score stratification: group 0 (red): low TF and low PF; group 1 (green): discordant (low TF/high PF or high TF/low PF); group 2 (blue): high TF and high PF. PF, fragmentomic profile; TF, tumor fraction.

Multivariable Cox regression adjusting for RT status, ECOG performance status, age, and FIGO stage confirmed TF as an independent prognostic factor. In this model, TF remained predictive of both PFS (HR 1.02, 95% CI 1.01-1.05, P = 0.008) and OS (HR 1.04, 95% CI 1.01-1.06, P = 0.005) as a continuous variable. Dichotomized TF showed a borderline association with PFS (P = 0.054), and did not stratify OS, suggesting that the continuous modeling provided greater prognostic granularity. These findings establish TF at B1 as an independent, clinically informative biomarker measurable at a critical post-surgical, pre-chemotherapeutic time point.

Fragmentation-based LB profiling stratifies risk after debulking surgery

We first evaluated whether global cfDNA fragmentomics could serve as a marker for untargeted ctDNA detection. As shown in Supplementary Figures S4 and S5, available at https://doi.org/10.1016/j.esmoop.2026.106087, cfDNA from MITO16a/MaNGO-OV2 was more fragmented than in healthy controls, supporting the feasibility of sWGS for untargeted ctDNA detection. We then calculated the ratio of short (100-150 bp) versus long (151-220 bp) fragments (i.e. PF) and evaluated its prognostic relevance.

At the B1, PF was associated with clinical outcomes (Figure 2C and D). Using a threshold of 0.1659, patients were stratified into high-risk (PF > 0.1659, n = 35) and low-risk (PF ≤ 0.1659, n = 132) groups. Kaplan–Meier analysis showed higher PF linked to shorter PFS (HR 1.97, 95% CI 1.07-3.65, P = 0.030), with a non-significant trend for OS (HR 2.67, 95% CI 0.93-7.65, P = 0.067) (Supplementary Table S3, available at https://doi.org/10.1016/j.esmoop.2026.106087).

As a continuous variable, PF showed stronger associations (Supplementary Table S3, available at https://doi.org/10.1016/j.esmoop.2026.106087). In univariate analysis, higher PF was significantly associated with shorter PFS (HR 1.07, P = 0.002) and OS (HR 1.10, P = 0.002). These associations remained significant in multivariate Cox regression adjusting for the clinical covariates (PFS: HR 1.06, P = 0.010; OS: HR 1.10, P = 0.005). Notably, PF lost significance when dichotomized (Supplementary Table S3, available at https://doi.org/10.1016/j.esmoop.2026.106087). Collectively, these findings support continuous PF as a non-invasive, independent prognostic biomarker in the immediate post-operative setting.

Integrated multimodal ctDNA profiling enhances early prognostic stratification

To improve prognostic precision beyond single-biomarker approaches, we evaluated whether integrating TF and PF improved ctDNA analysis. A moderate positive correlation (r² = 0.356, P < 0.001; Supplementary Figure S6, available at https://doi.org/10.1016/j.esmoop.2026.106087) confirmed these metrics capture partially distinct tumor signals, supporting their combined use.

Using defined thresholds (TF ≥ 15.08%; PF > 0.1659) patients were stratified into three risk categories: (i) group 0 (both markers below cut-offs), (ii) group 1 (one marker elevated), and (iii) group 2 (both markers elevated). Kaplan–Meier curves (Figure 2E and F) illustrate clear prognostic separation.

In univariate analysis (Table 2), group 1 had shorter PFS (17.7 versus 23.0 months) and OS (38.5 versus 47.6 months) compared with group 0, with increased risk of relapse (PFS: HR 1.76, 95% CI 1.12-2.79, P = 0.015; OS: HR 2.06, 95% CI 1.08-3.95, P = 0.029). Group 2 patients had the poorest prognosis, with a median PFS of 13.3 months and OS of 27.4 months (PFS: HR 3.58, 95% CI 2.03-6.31, P < 0.001; OS: HR 4.06, 95% CI 1.90-8.69, P < 0.001).

Table 2.

Association of combined dichotomized TF and PF with PFS and OS in univariate and multivariate analysis

TF and PF B1 dichotomous variables
(TF cut-off = 15.08%, PF cut-off = 0.1659)
Univariate model
Variables	PFS		OS
	HR (95% CI)	P value	HR (95% CI)	P value
TF + PF_group 1 versus group 0	1.76 (1.12-2.79)	0.015	2.06 (1.08-3.95)	0.029
TF + PF_group 2 versus group 0	3.58 (2.03-6.31)	<0.001	4.06 (1.9-8.69)	<0.001

Multivariate model
Variables	PFS		OS
	HR (95% CI)	P value	HR (95% CI)	P value
TF + PF_group 1 versus group 0	1.62 (1.02-2.58)	0.041	2.02 (1.03-3.96)	0.040
TF + PF_group 2 versus group 0	2.46 (1.37-4.42)	0.003	2.88 (1.29-6.44)	0.010
RT (none versus present)	1.35 (0.92-1.99)	0.123	1.56 (0.85-2.85)	0.151
ECOG PS (0 versus 1_2)	1.74 (1.14-2.67)	0.011	1.78 (0.98-3.26)	0.060
Stage (III versus IV)	2.24 (1.44-3.48)	<0.001	1.43 (0.77-2.67)	0.262
Age (<65 years versus ≥65 years)	0.98 (0.65-1.46)	0.907	0.86 (0.47-1.58)	0.635

Univariate and multivariate model reporting the association in terms of PFS and OS with patients’ group classification on the basis of combined dichotomized TF and PF variables. To define the HRs, patients of group 1 (low TF/high PF or high TF/low PF) and group 2 (high TF/high PF) were both compared with patients belonging to group 0 (low TF/low PF). The multivariate analysis was adjusted for RT, performance status (ECOG PS), tumor stage, and age. HRs are presented with 95% CIs and corresponding P values. Bold values indicate statistically significant P-values.

CI, confidence interval; ECOG PS, Eastern Cooperative Oncology Group performance status; HR, hazard ratio; OS, overall survival; PF, fragmentomic profile; PFS, progression-free survival; RT, residual tumor; TF, tumor fraction.

Multivariate models confirmed group 1 retained intermediate risk (PFS: HR 1.62, P = 0.041; OS: HR 2.02, P = 0.040), and group 2 consistently demonstrated the highest risk, with significantly shorter survival (PFS: HR 2.46, 95% CI 1.37-4.42, P = 0.003; OS: HR 2.88, 95% CI 1.29-6.44, P = 0.010).

Overall, in the post-surgical setting (B1), in models incorporating TF and PF as covariates, the prognostic contribution of RT appeared attenuated when compared with TF and PF, consistent with a stronger prognostic effect of TF and PF on both PFS and OS. At the post-chemotherapy (B2) time point, prognostic significance was reduced, with only TF retaining modest OS association and PF showing no impact (Supplementary Table S4, available at https://doi.org/10.1016/j.esmoop.2026.106087). Longitudinal analysis across the three time points evidenced marked inter-patient heterogeneity, with no clear or consistent divergence in TF and PF trajectories between responders and relapsed patients, although reduced sample availability at later time points may have limited the power to detect more subtle associations with clinical outcome (Supplementary Figure S7, available at https://doi.org/10.1016/j.esmoop.2026.106087).

Altogether, these results support multimodal ctDNA analysis as a robust and independent prognostic strategy to refine early risk stratification after surgery and before chemotherapy.

Prognostic significance of tumor fraction and fragmentomic profiles in the post-chemotherapy setting

In response to the clinical need for prognostic biomarkers at the end of first-line chemotherapy, we evaluated the prognostic relevance of TF and PF in the post-chemotherapy setting. At the B2 time point, both biomarkers were analyzed exclusively as continuous variables to minimize potential biases associated with cut-offs defined in the B1 setting, where patient selection was restricted to therapy responders, and could have artificially inflated prognostic associations. In univariate analyses, neither TF nor PF was associated with PFS or OS. Similarly, in multivariate models assessing the two biomarkers separately, neither TF nor PF showed an independent association with PFS. In contrast, TF retained independent prognostic significance for OS in the B2 setting (HR 1.11, 95% CI 1.01-1.22, P = 0.039), whereas PF was not significantly associated with outcome in any multivariate model (Supplementary Table S4, available at https://doi.org/10.1016/j.esmoop.2026.106087). Notably, when TF and PF were simultaneously included as covariates in the multivariate model, PF was not significant, while TF preserved its independent prognostic value for OS after adjustment for PF (HR 1.05, 95% CI 0.98-1.12, P = 0.048) (Supplementary Table S4, available at https://doi.org/10.1016/j.esmoop.2026.106087). Overall, these findings indicate that, in the post-chemotherapy setting, TF provides a modest yet independent contribution to prognostic stratification, with its association with OS consistently maintained across multivariate models, including those adjusted for PF. In contrast, PF did not demonstrate prognostic relevance at this setting.

Treatment-associated genomic remodeling and recurrent 19q13.42 amplification in longitudinal ctDNA profiles

To investigate dynamic genomic changes, we analyzed 386 plasma samples from 173 patients. A heatmap of CNVs across time points (Supplementary Figure S8, available at https://doi.org/10.1016/j.esmoop.2026.106087) highlights the evolving ctDNA landscape.

We initially questioned whether the prognostic role of TF and PF reflects underlying genomic differences. At B1, the hierarchical clustering depicted in Figure 3A identified two major groups: cluster C1, enriched for patients in TF/PF-defined groups 1-2 (18/36, 50%), and displaying relative CNV homogeneity, and C2, marked by heterogeneity and subdivided into C2a and C2b, the latter largely composed of group 0 patients (88%, Supplementary File 2, available at https://doi.org/10.1016/j.esmoop.2026.106087). Amplification of the 19q13.42 was predominantly observed in C2b and often absent in C1, suggesting a selective role in genomic stability. Additional alterations at 10q24 and 10q26 were confined to C2, consistent with divergent tumor evolution.

Figure 3.

Unsupervised clustering of genomic alterations in serial plasma samples. Heatmap summarizing genomic alterations identified by sWGS sequencing in plasma samples collected at three clinical time points: B1 (A), B2 (B), and B3 (C). Each column represents a specific cytogenetic region, with amplifications (red) and deletions (blue) annotated by z-scores. In the upper part of the figure, the row labeled ‘TF’ indicates patients classified as ‘high’ (blue) or ‘low’ (yellow) based on a TF cut-off of 15.08%; ‘PF’ indicates patients classified as ‘high’ (blue) or ‘low’ (yellow) based on a PF cut-off of 0.1659. ‘Category’ indicates patients classified based on the different cut-offs of TF and PF in group 0, 1, and 2. "Status" distinguishes between non-responder from responder patients. The row asterisks (∗) denote cytobands significantly altered between the high and low TF subgroups. PF, fragmentomic profile; sWGS, shallow whole-genome sequencing; TF, tumor fraction.

Following Pt-based chemotherapy (B2, Figure 3B) two clusters (L1 and L2) emerged. Amplification of 19q13.42 was present in 91% of assessable cases, indicating widespread clonal selection. L1 was further characterized by co-deletions at 18q11.2 and 18q12.3, absent in L2 except one case.

In the longitudinal analysis (Supplementary Figure S8, available at https://doi.org/10.1016/j.esmoop.2026.106087), no definitive CNV patterns clearly distinguished responders by non-responders, reflecting the overall heterogeneity of patient trajectories (details are reported in the Supplementary Results section, available at https://doi.org/10.1016/j.esmoop.2026.106087). The 19q13.42 amplification, recurrent across all time points and relatively enriched in responders, encompasses key cancer-associated genes such as ZNF family transcription factors and the miR-515 cluster, suggesting potential roles in clonal expansion and therapeutic adaptation. Similarly, 3q26.2 gains are observed across all time points. Persistent deletions at 12q24.32, 10q26.11, and 1p36.13 are detected across all time points, independent of response, indicating stable genomic alterations during disease progression. Additionally, a subset of B3 responders (∼20%) exhibits a gain at 14q11.2, a region including BCL2L2, PABPN1, and EFS, potentially reflecting selective clonal expansion driven by pro-survival and transcriptional regulatory pathways under maintenance therapy.

Discussion

To our knowledge, this is the first study conducted within a randomized, multicenter phase IV clinical trial to demonstrate that multimodal ctDNA analysis provides superior prognostic resolution compared with conventional metrics in advanced EOC. Although carried out before widespread PARPi use, the translational arm of the MITO16a/MaNGO-OV2 trial provided a unique opportunity to assess the clinical relevance of post-surgical ctDNA, leveraging centralized sample processing and curated clinical annotation.

Two principal findings emerge from our study, each corresponding to a distinct clinical time point.

At the post-surgical, pre-chemotherapy time point, TF—reflecting the proportion of ctDNA in plasma—demonstrated robust prognostic performance, and outperformed RT in predicting survival.²¹ Notably, TF retained prognostic significance even among optimally debulked patients, indicating that ctDNA-based MRD assessment captures microscopic residual disease not detectable by surgical evaluation alone. This early post-operative time point was intentionally selected to interrogate residual disease before exposure to Pt-based chemotherapy, thereby providing a direct readout of surgical completeness. In this context, our findings extend the current MRD paradigm by showing that a tumor-agnostic, multimodal ctDNA approach can identify biologically and clinically relevant high-risk subgroups even within surgically favorable populations.

In contrast, at the B2 time point, the prognostic landscape differed substantially. In this setting, TF retained a modest yet independent association with OS, whereas the prognostic value of PF was no longer evident. These findings suggest that, following systemic therapy, ctDNA burden remains informative of treatment-resistant residual disease, while PF may be attenuated or biologically altered, potentially reflecting therapy-induced effects on cfDNA fragmentation patterns.

The post-adjuvant chemotherapy setting has been extensively investigated for MRD assessment. In this context, Knisely et al. reported an association between post-treatment ctDNA positivity, residual disease at second-look laparoscopy, and poorer outcomes ctDNA positivity after first-line therapy was associated with residual disease at second-look laparoscopy and poorer clinical outcomes.²² Although these findings may appear to differ with our observations, such discrepancies should be interpreted in light of the limited sample size and the tumor-informed methodological framework adopted in this study, in contrast to the tumor-agnostic approach employed here. These methodological differences may, at least in part, account for the observed variations in prognostic performance.

Our study extends this paradigm by showing that ctDNA can identify high-risk patients after surgery, when therapeutic decisions remain potentially modifiable. Importantly, the extent of RT has historically guided the risk assessment and is now emerging as a potential determinant of maintenance therapy allocation. Evidence from pivotal phase III trials has indicated that patients with macroscopic disease (R2), representing a high-risk subgroup, may derive greater benefit from bevacizumab-based maintenance.^5,23 Conversely, clinical trials of PARPi have shown that patients with no visible disease (R0) or minimal RT (R1) consistently experience the most pronounced and durable benefit from PARPi while efficacy is maintained across all residual disease subgroups.⁷ Although these findings require further validation in prospective studies, they underscore the clinical relevance of accurately assessing RT. Consistent and reproducible measurement of residual disease at the time of surgery could therefore be pivotal to refining patient stratification and optimizing maintenance treatment strategies in the future.

Secondly, integrating TF with PF improved risk stratification beyond either biomarker alone. PF, independently associated with survival outcomes, provided orthogonal biological information to TF.^24,25 Multimodal classification identified high-risk patients, despite similar clinicopathological features, underscoring the added value of fragmentomic metrics. These findings are clinically relevant in a therapeutic landscape, where both PARPi and bevacizumab are standard options,^11,26 but tools to guide treatment intensity remain limited. Our results suggest that LB-derived biomarkers could enable de-escalation in low-risk patients to reduce toxicity and escalation or closer monitoring in those at higher risk. Post-operative ctDNA profiling carried out in the post-operative setting thus offers a timely, non-invasive strategy for risk-adapted management.

From a biological point of view, TF and PF capture distinct yet complementary features: TF quantifies ctDNA burden and serves as a proxy for tumor burden while PF reflects fragmentomic patterns linked to chromatin accessibility and cell death.^18,24,25,27 In our cohort, PF added prognostic value independently of TF, underscoring the utility of integrating fragmentomic into ctDNA analysis. This approach improves both sensitivity and specificity of MRD detection and supports tumor-agnostic LB strategies.

Longitudinal CNV analysis further illuminated tumor evolution, with recurrent amplification of 19q13.42 and deletions of 10q and 18q emerging under therapy (B2-B3) The 19q13.42 region, enriched in zinc finger proteins and microRNAs (e.g. miR-515 cluster), was detectable at baseline and became prevalent at later time points. Although previously reported in a small retrospective HGSOC cohort,¹⁶ our study provides the first evidence of this alteration as a recurrent, treatment-associated event in a trial setting. Its consistent emergence under therapeutic pressure highlights a potential role in Pt resistance, warranting functional investigation. While longitudinal ctDNA enables a non-invasive view of clonal selection, the large size of 19q13.42 complicates identification of key drivers, requiring functional studies.

Collectively, these findings demonstrate the clinical utility of three plasma-based biomarkers—TF, PF, and CNVs—derived from a single tube, via sWGS. This multimodal approach enables early prognostic stratification and real-time tracking of tumor evolution and emerging resistance. As maintenance therapies become more widespread, dynamic ctDNA profiling represents a critical tool to personalize treatment, detect molecular relapse, and optimize sequencing.²⁸ Finally, our analyses underscore the intrinsic heterogeneity of advanced-stages EOC. Spaghetti plots of TF and PF for individual patients across time points (Supplementary Figure S7, available at https://doi.org/10.1016/j.esmoop.2026.106087) reveal highly variable trajectories, without consistent trends distinguishing responders from relapsed cases. This variability mirrors the heterogeneity observed in CNV patterns across cytobands (Supplementary Figure S8, available at https://doi.org/10.1016/j.esmoop.2026.106087), including recurrent amplifications such as 19q13.42, enriched in ZNF transcription factors and the miR-515 cluster, and the subset of B3 responders (∼20%) exhibiting a gain at 14q11.2, which contains genes such as BCL2L2, PABPN1, and EFS.

Despite strengths—including centralized processing, multicenter enrollment, and detailed clinical annotations—limitations exist.

Firstly, the absence of standardized CNV protocols hampers clinical translation, with variability in handling and analysis limiting reproducibility.

Secondly, the prognostic significance of TF and PF at the B2 and B3 time points was attenuated, likely due to the smaller number of available samples, which reduced statistical power. Furthermore, the absence of specimens from patients with early progression or poor clinical response introduced a selection bias, resulting in a cohort enriched for chemotherapy-sensitive cases. Consequently, the prognostic value of TF and PF at B2 and B3 remains exploratory and should be interpreted with caution, as these markers may not fully reflect the broader patient population.

Finally, the absence of PARPi in MITO16a/MaNGO-OV2 (incorporated into practice only after trial initiation) limits the generalizability.¹¹ To address this, we are validating our findings in the IOlanTHe study (NCT06121401), which integrates plasma ctDNA analysis with modern maintenance strategies. Results will clarify whether TF, PF, and CNV dynamics can inform adaptive treatment and serve as surrogate trial endpoints.

Conclusions

In conclusion, multimodal, tumor-agnostic ctDNA profiling immediately after surgery refines prognostic assessment in advanced EOC beyond conventional metrics such as RT. TF provides a sensitive, objective marker of MRD that may complement or surpass surgical assessment in guiding maintenance treatment choices, particularly between PARPi and bevacizumab. This framework enables risk-adapted therapy, enhances monitoring of tumor evolution, and supports a precision medicine paradigm for patients with advanced stages EOC.