Chemotherapy-driven gut microbiota remodeling in ovarian cancer: a prospective longitudinal study

Abstract

Background

The gut microbiome shapes chemotherapy efficacy and outcomes in several cancers, but evidence in ovarian cancer (OC) remains limited and largely cross-sectional. Despite high initial response rates, long-term relapse in OC remains frequent, while conventional markers capture only short-term therapeutic sensitivity. Whether longitudinal gut-microbiome trajectories during chemotherapy are associated with long-term recurrence remains unknown.

Methods

Within the prospective SOCFCP cohort (N = 91; 13 recurrences), 100 serial fecal samples from a 33-patient sub-cohort were analyzed by 16S rRNA sequencing across the early, middle and late chemotherapy phases. Microbial successional trajectories and their association with recurrence were assessed by linear mixed-effects modeling, multivariable MaAsLin3 and repeated-measures correlation (rmcorr) networks, alongside stratified and covariate-adjusted sensitivity analyses and patient-level bootstrap assessment. The cumulative severe-toxicity–recurrence relationship was estimated by Firth penalized-likelihood regression, suited to sparse, separation-prone events.

Results

Microbial α-diversity rose progressively across the chemotherapy course (Shannon time effect p = 0.002), consistent with ecological succession, with higher turnover among peripheral than in core taxa (p = 0.005). Cumulative severe toxicity was not associated with recurrence (Firth OR = 0.99, 95% CI 0.68–1.39). Crucially, recurrent patients exhibited a progressive depletion of Fusicatenibacter (recurrence × time coefficient = −5.35, q < 0.001) that persisted across all sensitivity analyses—stratified, medication-adjusted, antibiotic-depleted and clinically-adjusted models (coefficient −4.60 to −5.66, all q < 0.001). PICRUSt2-based functional inference identified recurrence-associated differences in predicted de novo nucleotide-biosynthesis and cell-wall-assembly pathway abundance (q < 0.05). A bootstrap-supported co-variation network further linked specific taxa, notably Escherichia-Shigella and Roseburia, to these recurrence-associated pathways.

Conclusion

Chemotherapy-driven gut-microbiome remodeling, in particular the recurrence-associated depletion of Fusicatenibacter, was associated with long-term OC relapse, whereas cumulative severe toxicity showed no significant association with recurrence. These longitudinal microbial dynamics support a candidate non-invasive marker that warrants external validation, and provide a hypothesis-generating rationale for testing whether targeting specific predicted bacterial functional pathways can modulate the host anti-tumor milieu.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12967-026-08731-9.

Background

Ovarian cancer (OC) remains the most lethal gynecological malignancy [1]. Frontline primary cytoreductive surgery followed by platinum-based chemotherapy achieves initial remission in roughly 80% of patients, yet more than 70% relapse within five years [2]. This contrast between a strong initial response and frequent late relapse highlights a critical clinical challenge: conventional prognosticators such as cancer antigen 125 (CA125) track short-term therapeutic sensitivity but do not forecast long-term recurrence [3]. The relationship between acute chemotherapy-associated severe adverse events (SAEs) and long-term oncologic outcomes is equally unsettled [4]. It remains unclear whether acute tissue toxicity and eventual relapse share a common mechanism or instead represent two parallel, distinct consequences of treatment. These observations underscore the need for dynamic markers that prospectively reflect long-term recurrence risk rather than transient treatment response.

The gut microbiome has emerged as a modulator of chemotherapy efficacy and host anti-tumor immunity across several cancers, acting through short-chain fatty acid producers and other taxa that shape systemic immunity [5], intestinal barrier integrity [6], and the circulating metabolic milieu [7, 8]. In OC, however, this evidence is sparse and almost entirely cross-sectional [9–13]. Such single-timepoint snapshots are fundamentally limited for causal inference, as they cannot distinguish whether a dysbiotic signature is a pre-existing high-risk ecological substrate of relapse or a passive passenger phenomenon of advancing disease and acute drug toxicity. Furthermore, OC chemotherapy is a prolonged and cyclical process in which treatment exposure and cumulative pharmacologic stress fluctuate continuously; static designs are therefore vulnerable to unmeasured confounding [14] and cannot capture the structured ecological succession that the gut community undergoes under repeated chemotherapeutic pressure. Prospective, longitudinal tracking is required to move beyond these constraints.

This highlights two pivotal, unanswered scientific inquiries: First, what is the longitudinal successional trajectory of the gut microbiome across cyclic chemotherapy? Second, and more importantly, is this trajectory associated with long-term recurrence in a way that still holds after acute treatment toxicity is taken into account?

Driven by this clinical question, we leveraged the prospective SOCFCP cohort (N = 91), integrating systematic longitudinal capture of chemotherapy delivery, cumulative SAEs, and recurrence outcomes with serial 16S rRNA profiling in a representative sub-cohort across the early, middle, and late chemotherapy phases. We analyzed these data using linear mixed-effects regression, multivariable MaAsLin3 [30], and repeated-measures correlation networks. Because recurrence events were sparse, the toxicity–recurrence relationship was estimated via Firth penalized-likelihood regression, and every signal was examined through stratified, medication-adjusted, and bootstrap sensitivity analyses. This comprehensive framework allowed us to delineate the chemotherapy-driven ecological succession of the gut microbiome, demonstrating that cumulative acute toxicity was not associated with recurrence. Instead, we identified a recurrence-associated depletion trajectory of Fusicatenibacter that remained consistent across all sensitivity analyses, coupled with a bootstrap-supported network linking specific commensals to predicted functional profiles. Collectively, these findings position longitudinal microbial dynamics as a candidate non-invasive marker warranting external validation, providing a framework for hypotheses on microbiome-informed strategies targeting implicated taxa and predicted functional pathways to modulate the host anti-tumor milieu.

Methods

Study design and participant enrollment

This longitudinal observational study was nested within the prospective Shanghai Ovarian Cancer and Family Care cohort (SOCFCP, NCT06118307) [15]. Patients with histologically confirmed epithelial ovarian cancer (diagnosed between August 2023 and December 2024) who underwent primary debulking surgery (PDS) followed by standard adjuvant paclitaxel-carboplatin (TC) chemotherapy were enrolled. To limit baseline microbial confounding, patients with prior neoadjuvant therapy, a concurrent active malignancy, or exposure to antibiotics, probiotics, or corticosteroids within three months before enrollment were excluded.

Clinical phenotyping and endpoint

Chemotherapy-associated adverse events were graded prospectively according to CTCAE v5.0, and grade ≥ 3 events were defined as severe adverse events (SAEs). Dynamic clinical indices, including CA125 decline kinetics, the systemic immune-inflammation index (SII), and the prognostic nutritional index (PNI), were recorded throughout treatment (definitions and calculation formulas in Supplementary Table S1 and S2). The primary endpoint was progression-free survival (PFS), defined as the interval from surgery to the first RECIST 1.1–confirmed recurrence or death from any cause.

Microbiome profiling and predicted functional profiling

Fecal samples were collected during the inpatient admission for each cycle, either on the morning of the infusion day or within the next 1–2 days, in the early (cycles 1–2), middle (cycles 3–4), and late (cycles 5–6) phases. The 16S rRNA V3–V4 region (primers 338F/806 R) was amplified and sequenced on the Illumina MiSeq platform, denoised into amplicon sequence variants using DADA2, and taxonomically classified against SILVA 138. All samples were processed in a single sequencing run and a single bioinformatic pipeline. The 33 sequenced patients provided 101 serial samples; one sample taken on the first postoperative day, which could not be assigned to a chemotherapy cycle, was excluded from every analysis, leaving 100. Predicted pathway abundances were inferred from taxonomic profiles using PICRUSt2 [16] and the MetaCyc database and were interpreted as predicted functional profiles rather than direct measurements of microbial metabolic activity. Taxa were classified as core (prevalence > 50% and mean relative abundance > 0.001) or peripheral (prevalence < 30%).

Statistical analysis

Baseline characteristics were compared using the Student’s t-test, Mann–Whitney U test, or chi-square test as appropriate. Longitudinal cycle-wise SAE trends were evaluated via the Cochran–Armitage test.

To characterize community succession, α-diversity (Shannon, Chao1, and Simpson indices) was compared across the three phases using the Wilcoxon signed-rank test (BH-adjusted), and community structure (Jaccard and Bray-Curtis distances) was assessed by PERMANOVA constrained within subjects (adonis2, vegan; 999 permutations). These phase-wise comparisons used one representative sample per patient per phase; all subsequent longitudinal analyses used the full 100 samples. To identify clinical drivers of diversity shifts, z-scored indices were fitted by linear mixed-effects (LME) regression (lme4/lmerTest) with age, stage, days from surgery, CA125 decline rate, and SAE occurrence as fixed effects incorporating a patient random intercept.

Recurrence-associated taxa and predicted pathway abundances were then modeled utilizing multivariable MaAsLin3 on total-sum-scaled (TSS), log-transformed abundances. The model included recurrence status, continuous time (days from surgery), collapsed stage, and a recurrence × time interaction term to capture dynamic trajectories, with continuous predictors standardized. To avoid imputation-induced bias, samples missing covariates were removed by list-wise deletion without imputation, and the false discovery rate (FDR) was strictly controlled at q < 0.1 for taxa and tightened to q < 0.05 for predicted pathways.

Acute toxicity was related to recurrence in two steps. First, per-cycle SAE incidence was modeled using a binomial generalized linear mixed model (lme4) with a patient random intercept. Second, because all 13 recurrence events occurred exclusively in advanced-stage (FIGO III/IV) disease—producing quasi-complete separation—and because overall events were sparse, conventional unpenalized models yielded unstable estimates. Therefore, the association between cumulative severe toxicity (the number of cycles with a grade ≥ 3 SAE) and recurrence was re-estimated using Firth penalized-likelihood logistic regression (logistf), adjusting for age.

To test for residual confounding, the principal microbiome signals were re-examined under four conditions: (i) restriction exclusively to advanced-stage (FIGO III/IV) patients; (ii) adjustment for peri-chemotherapy systemic antibiotic, probiotic, and corticosteroid exposure within 7 days of sampling, entered as time-varying covariates; (iii) an antibiotic-depleted subset; and (iv) joint adjustment for baseline BMI, chemotherapy compliance, and surgical residual disease (R0 status). Temporal co-variation among taxa, predicted pathways, and host indices was assessed via repeated-measures correlation (rmcorr). Candidate nodes were strictly restricted to significant MaAsLin3 features reformatted into a patient × cycle matrix; to ensure stable covariation, evaluated edges required non-zero variance, at least 10 non-missing paired observations, coverage across at least 2 cycles, and at least 5 patients, followed by BH adjustment. Finally, network stability was assessed via a patient-level non-parametric bootstrap (B = 1000): in each iteration, patients rather than observations were resampled with replacement to preserve within-subject structure, the full filtering pipeline was re-applied, and rmcorr was recomputed and BH-adjusted. Each edge’s selection probability (the fraction of iterations significant at q < 0.05 with a consistent sign) was recorded, and only edges with a probability ≥ 0.80 were retained as bootstrap-supported. All analyses were performed in R (v4.5.2).

Results

Longitudinal clinical phenotypes and toxicity trajectories

The final analysis cohort comprised 91 patients with a median follow-up of 611 days and a recurrence rate of 14.3% (13 events; Supplementary Figure S1). All recurrences arose in advanced-stage (FIGO III/IV) disease (100% vs 70.5% of non-recurrent patients, p = 0.033), whereas age, the surgical R0 rate, BRCA status, pre-surgery CA125, the systemic immune-nutritional indices, and the CA125 elimination kinetics were comparable between groups (Table 1, Supplementary Table S3).

Table 1.

Clinical characteristics of ovarian cancer patients stratified by recurrence status

Characteristic	All patients	Non-recurrence	Recurrence	P
	N = 91	n = 78	n = 13
Age (Years)	56.5(11.5)	56.1(11.4)	58.8(12.1)	0.452
BMI (kg/m2)	21.9 (3.1)	21.8 (3.2)	22.1 (2.5)	0.757
FIGO Stage:				0.033
Early (I–II)	23(25.3%)	23(29.5%)	0(0.00%)
Late (III–IV)	68(74.7%)	55(70.5%)	13(100%)
R0 resection rate				0.356
R0	56(61.5%)	50(64.1%)	6(46.2%)
Non-R0	35(38.5%)	28(35.9%)	7(53.8%)
Bevacizumab Use				>0.999
No	67(73.6%)	57(73.1%)	10(76.9%)
Use	24(26.4%)	21(26.9%)	3(23.1%)
BRCA mutation status				0.503
Mutation	25(27.5%)	23(29.5%)	2(15.4%)
Wild-type/unknown	66(72.5%)	55(70.5%)	11(84.6%)
Preoperative baseline inflammatory/nutritional indices
NLR	4.15[2.47;6.34]	3.82[2.44;6.07]	5.88[3.11;7.43]	0.259
PLR	238[172;334]	240[174;322]	228[144;419]	0.932
SII	1155[729;2183]	1139[752;1832]	1249[624;2750]	0.606
LMR	2.90(1.37)	2.94(1.33)	2.69(1.62)	0.62
PNI	48.2[43.9;50.8]	48.6[44.6;50.8]	44.0[37.5;49.5]	0.125
HALP	20.0[13.7;32.0]	20.5[15.2;32.3]	19.9[9.01;23.3]	0.373
Surgery to chemo interval (Days)	15.0[12.0;21.0]	14.5[12.0;21.0]	17.0[14.0;20.0]	0.577
Severe toxicity (Grade ≥ 3) during chemotherapy				0.611
No	9(9.89%)	7(8.97%)	2(15.4%)
Yes	82(90.1%)	71(91.0%)	11(84.6%)
Chemo Compliance				0.142
Compliant	71(78.0%)	62(79.5%)	9(69.2%)
Mild Delay	13(14.3%)	9(11.5%)	4(30.8%)
Severe Delay	7(7.69%)	7(8.97%)	0(0.00%)
Total chemo duration (Days)	110[106;120]	110[106;120]	110[106;121]	0.729
Baseline CA125 (Pre-surgery)	337[121;1165]	334[94.5;1170]	539[282;1165]	0.286
CA125 normalization during chemo				>0.999
No	2(2.20%)	2(2.56%)	0(0.00%)
Yes	89(97.8%)	76(97.4%)	13(100%)
CA125 reduction % (at end-chemo)	0.98[0.91;0.99]	0.98[0.89;0.99]	0.98[0.95;0.99]	0.247
Time to CA125 normalization (Days from surgery)	45.0[34.0;65.0]	42.0[34.0;63.2]	60.5[50.2;71.0]	0.166
CA125 half-life (Days)	19.8[14.9;30.2]	19.7[14.5;31.1]	20.7[16.6;25.6]	0.878
Modeled KELIM slope (k)	0.03[0.01;0.04]	0.03[0.01;0.04]	0.03[0.02;0.04]	0.578
Myelosuppression (G3/4) during chemotherapy				0.189
No	11(12.1%)	8(10.3%)	3(23.1%)
Yes	80(87.9%)	70(89.7%)	10(76.9%)
Liver dysfunction (G3/4) during chemotherapy				0.466
No	87(95.6%)	75(96.2%)	12(92.3%)
Yes	4(4.40%)	3(3.85%)	1(7.69%)
Renal dysfunction (G3/4) during chemotherapy: No	91(100%)	78(100%)	13(100%)	-
Coagulation disorder (G3/4) during chemotherapy: No	91(100%)	78(100%)	13(100%)	-
Electrolyte imbalance (G3/4) during chemotherapy				>0.999
No	90(98.9%)	77(98.7%)	13(100%)
Yes	1(1.10%)	1(1.28%)	0(0.00%)
Metabolic disorder (G3/4) during chemotherapy				0.374
No	88(96.7%)	76(97.4%)	12(92.3%)
Yes	3(3.30%)	2(2.56%)	1(7.69%)

Note: Continuous variables are presented as mean (standard deviation) or median (interquartile range); categorical variables are presented as n (%). p-values were calculated using Student’s t-test, Mann–Whitney U test, or chi-square test, as appropriate. Body mass index was calculated from weight measured after primary debulking surgery and before the first chemotherapy cycle. Abbreviations: FIGO, International Federation of Gynecology and Obstetrics; R0, complete resection; BRCA, breast cancer gene; CA125, cancer antigen 125; HE4, human epididymis protein 4; KELIM, CA125 elimination kinetics; NLR, neutrophil-to-lymphocyte ratio; PLR, platelet-to-lymphocyte ratio; SII, systemic immune-inflammation index; LMR, lymphocyte-to-monocyte ratio; PNI, prognostic nutritional index; HALP, hemoglobin, albumin, lymphocyte, and platelet; BMI, body mass index

SAEs occurred in 90.1% of the cohort (82/91), predominantly as myelosuppression, and their per-cycle incidence rose steeply over the treatment course (Cochran–Armitage p < 0.001; Fig. 1C). However, this toxicity trajectory showed no significant association with long-term oncologic outcomes. Although per-cycle SAE incidence increased with consecutive cycles (OR = 1.32 per cycle, 95% CI 1.17–1.48, p < 0.001), it did not differ between recurrent and non-recurrent patients (group OR = 0.99, 95% CI 0.43–2.30, p = 0.984; group × cycle p = 0.439). Crucially, Firth penalized-likelihood regression—appropriate for the sparse events and stage-driven quasi-complete separation—detected no significant association between cumulative grade ≥ 3 toxicity and recurrence (OR = 0.99, 95% CI 0.68–1.39, p = 0.934), with near-identical mean cumulative grade ≥ 3 cycles between the two groups (2.38 vs 2.49; Fig. 1D, Supplementary Table S4).

Fig. 1.

Longitudinal clinical phenotypes and toxicity trajectories during chemotherapy. (A) Overview of the prospective study design and longitudinal fecal sampling across early, middle, and late chemotherapy phases. (B) Kinetics of short-term tumor elimination, illustrated by the KELIM slope (left) and CA125 decline rate (right) between non-recurrence and recurrence groups. (C) Sankey diagram of per-patient severe adverse event (SAE) grade trajectories (CTCAE v5.0) across cycles 1–6 and the final recurrence outcome. Vertical axes denote cycles, and the right-most axis the recurrence outcome; patients are stratified by their highest SAE grade per cycle, and ribbon width is proportional to the number of patients transitioning between grades, illustrating the cycle-wise rise in severe toxicity and its distribution across recurrence status. (D) Grade ≥ 3 severe adverse event (SAE) trajectories show no significant association with long-term recurrence (full cohort, n = 91). Per-cycle incidence of grade ≥ 3 SAEs is shown for non-recurrence (blue) versus recurrence (orange) patients

Chemotherapy-driven diversity succession and community restructuring

The sequenced 33-patient sub-cohort was clinically comparable to the remaining unsequenced patients across key baseline characteristics (Supplementary Table S5). Within this sub-cohort, microbial α-diversity (Chao1 and Shannon indices) increased progressively from the early to the mid-to-late phases (p < 0.05, Fig. 2A, Supplementary Figure S2). Concurrently, principal-coordinates analysis of Jaccard (R² = 0.033, p = 0.001) and Bray-Curtis (R² = 0.030, p = 0.001) revealed a pronounced longitudinal drift in community structure (Fig. 2C, Supplementary Figure S3). This structural shift occurred unevenly across ecological niches. Specifically, the peripheral microbiota experienced a significantly greater divergence from its early-phase baseline compared with the core microbiota during both the early-to-middle (p = 0.007) and early-to-late (p = 0.005) transitions (Fig. 2B). The temporal gain in α-diversity persisted after adjustment for clinical covariates (Shannon estimate = 0.008, p = 0.002; Chao1 estimate = 0.008, p < 0.001; Table 2), and within advanced-stage patients (β = 0.007, p = 0.020; Supplementary Table S6), and was not significantly associated with SAE occurrence (p > 0.05).

Fig. 2.

Longitudinal evolution of gut microbiota diversity and community structure during chemotherapy. (A) Longitudinal shifts in α-diversity (Chao1, Shannon, and Simpson indices) across the three predefined chemotherapy stages. Pairwise comparisons were evaluated using the wilcoxon signed-rank test (* p < 0.05; ns, not significant). (B) Temporal divergence of the core and peripheral sub-communities from their respective early-phase baselines. The y-axis denotes structural divergence, with higher values indicating greater compositional shifts. p-values reflect the significant differences in volatility between core and peripheral microbiota across the early-to-middle and early-to-late transitions. (C) Principal coordinates analysis (PCoA) based on Jaccard (left) and Bray-Curtis (right) distance matrices, depicting significant longitudinal drift in community structure. (D) Linear mixed-effects (LME) estimates relating Z-score-standardized α-diversity (Shannon and Chao1) to the host’s CA125 decline rate

Table 2.

Longitudinal associations between clinical factors and gut microbiota α-diversity indices

	Shannon		Chao1		Simpson
Term	Standardized estimate (95% CI)	p value	Standardized estimate (95% CI)	p value	Standardized estimate (95% CI)	p value
Age	−0.008(−0.031,0.016)	0.506	−0.011 (−0.039, 0.017)	0.436	−0.013 (−0.039, 0.013)	0.304
Tumor stage	−0.412(−1.163,0.340)	0.274	−0.177 (−1.005, 0.650)	0.667	−0.140 (−0.961, 0.681)	0.731
Time	0.008(0.003,0.013)	0.002	0.008 (0.004, 0.012)	<0.001	0.003 (−0.002, 0.009)	0.242
CA125 Decline rate	1.880(0.607,3.154)	0.005	1.269 (0.105, 2.434)	0.033	0.466 (−0.949, 1.881)	0.512
Severe adverse events	0.075(−0.295,0.446)	0.688	−0.020 (−0.308, 0.268)	0.891	−0.007 (−0.430, 0.415)	0.972

Note: Standardized estimates were derived from linear mixed-effects (LME) models adjusted for age, tumor stage, postoperative time (days), CA125 decline rate, and SAEs (grade ≥ 3). α-diversity indices were Z-score transformed prior to modeling. A patient-specific random intercept was included to account for intra-subject temporal autocorrelation

A recurrence-associated depletion trajectory of Fusicatenibacter

Genus-level multivariable modeling (MaAsLin3) of the longitudinal dataset singled out Fusicatenibacter as the primary recurrence-associated taxon (Fig. 3). Recurrent patients entered treatment with a higher baseline prevalence of this genus (recurrence main effect Coef = 14.94, q < 0.001), but experienced a progressive depletion over the treatment course (recurrence × time interaction Coef = −5.35, q < 0.001). This effect was specific, given that Citrobacter declined with time regardless of outcome (Coef = −1.17, q = 0.049), and replacing recurrence with SAEs as the outcome yielded no significant genus- or ASV-level signal. Four sensitivity analyses confirmed the consistency of this Fusicatenibacter depletion trajectory: the signal remained highly significant when confined exclusively to advanced-stage patients (Coef = −5.66), when peri-chemotherapy systemic antibiotic, probiotic, and corticosteroid exposure was adjusted for (Coef = −5.16), in an antibiotic-depleted subset (Coef = −5.35), and after joint adjustment for baseline BMI, chemotherapy compliance, and surgical residual disease (Coef = −4.60). Across all scenarios, the depletion trajectory maintained high statistical significance (all q < 0.001; Supplementary Table S6, Supplementary Figure S4).

Fig. 3.

Longitudinal taxonomic trajectories associated with long-term recurrence. Multivariable MaAsLin3 modeling at the genus level, adjusted for age, tumor stage, and intra-subject correlation. (left) dot plot illustrating the effect sizes (β coefficients) for recurrence status (main effect) and the recurrence × time interaction. Positive values denote higher baseline prevalence/abundance in the recurrence group or an ascending temporal trend. Circles indicate abundance associations, while triangles denote prevalence associations. (right) Heatmap displaying FDR-adjusted significance (q values) across all covariates for both prevalence and abundance models. Cell background color indicates the effect direction (red: positive; blue: negative)

Predicted functional profiles and a bootstrap-supported co-variation network

PICRUSt2-based functional inference (mean weighted NSTI = 0.10) identified 162 pathways with significant temporal dynamics at q < 0.05 (Fig. 4A). The leading recurrence-associated predicted pathways converged on de novo nucleotide biosynthesis (adenosine and deoxyadenosine) and cell-wall assembly (peptidoglycan), alongside dTDP-L-rhamnose biosynthesis, one-carbon metabolism, tRNA charging, and serine/glycine biosynthesis. These predicted functional signals remained consistent across advanced-stage restriction, medication-adjusted, and antibiotic-depleted sensitivity analyses (Supplementary Table S6).

Fig. 4.

Temporal co-variation network of predicted microbial functional profiles and host indices. (A) Recurrence-associated predicted pathway abundances inferred by PICRUSt2 and modeled using MaAsLin3 at q < 0.05; bars show MaAsLin3 coefficients (β) for recurrence-associated predicted pathway signals. (B) Tripartite longitudinal co-variation network constructed using repeated-measures correlation (rmcorr) and assessed by patient-level non-parametric bootstrap resampling (B = 1000). Solid edges indicate bootstrap-supported correlations with selection probability ≥ 0.80; the dashed predicted pathway–host edge (dTDP-L-rhamnose–prognostic nutritional index, selection probability 0.62) is exploratory

To link these predicted pathway shifts to specific taxa and host status, a longitudinal co-variation network was constructed. Patient-level bootstrap assessment (B = 1000) supported eight microbe–pathway couplings (selection probability ≥ 0.80). Specifically, Escherichia-Shigella showed bootstrap-supported negative co-variation with predicted peptidoglycan synthesis (r = −0.680, selection probability 1.00), adenosine biosynthesis, and tRNA charging, whereas Roseburia showed the reciprocal positive couplings. Additionally, Bacteroides closely tracked predicted dTDP-L-rhamnose biosynthesis (r = 0.416, selection probability 0.84; Fig. 4B, Supplementary Table S7). Bridging these predicted pathway signals to the host, predicted bacterial dTDP-L-rhamnose biosynthesis co-varied positively with the prognostic nutritional index (r = 0.399), though this correlation exhibited moderate bootstrap support (selection probability 0.62).

Discussion

This prospective longitudinal study reframes the gut microbiome in OC from a static snapshot to a dynamic trajectory. Tracking the microbial community throughout frontline chemotherapy showed that its longitudinal ecological remodeling is associated with long-term relapse. Crucially, a Firth penalized-likelihood analysis detected no significant association between cumulative severe toxicity and recurrence (OR = 0.99, 95% CI 0.68–1.39). Because the microbiome trajectory lacks a significant statistical relationship with acute treatment toxicity, this microbial remodeling appears to represent a distinct dimension of patient risk, offering a time-resolved prognostic perspective that conventional short-term markers cannot provide.

This dynamic view also clarified the nature of the host response. Under sustained chemotherapy, the gut ecosystem did not simply degrade; α-diversity rose progressively through the mid-to-late phases, and the peripheral microbiota diverged from its early-phase state far more than the core. This pattern of core stability and peripheral variation points to a structured reassembly rather than indiscriminate dysbiosis, implying that the host retains considerable ecological resilience even under repeated cytotoxic pressure. We read this diversity gain as ecological succession, or partial recovery, which most likely reflects the combined influence of hospitalization, postoperative recovery, and other treatment-phase factors rather than a direct effect of chemotherapy alone [17]. Notably, an initial apparent link between diversity gain and rapid CA125 decline vanished and reversed sign when restricted exclusively to advanced-stage (FIGO III/IV) patients (Shannon p = 0.894). Thus, we regard the diversity expansion as a stage-sensitive ecological phenomenon during treatment, rather than a direct readout of tumor response.

Within this broader ecological context, genus-level multivariable modeling identified a recurrence-specific trajectory for Fusicatenibacter, a commensal producer of anti-inflammatory short-chain fatty acids (SCFAs) [18]. It was not deficient at baseline in patients who later relapsed; its baseline prevalence was, if anything, higher. The difference lay in maintenance rather than in starting point: these patients lost the genus progressively over treatment. The prognostic signal emerged from its progressive depletion over the treatment course, capturing a vulnerability only detectable through serial sampling. This depletion remained directionally identical and highly significant (q < 0.001) after restriction to advanced-stage patients, adjustment for peri-chemotherapy medication, exclusion of antibiotic-exposed samples, and joint adjustment for baseline BMI, compliance, and residual disease. This supports the Fusicatenibacter trajectory as a consistent correlate of recurrence rather than a fragile or confounded one. Because SCFA producers help sustain the anti-inflammatory tone and barrier function of the colonic mucosa [19, 20], their progressive loss could plausibly erode this protection, prompting us to evaluate the accompanying shifts in predicted functional potential.

Functional inference (cohort-wide mean weighted NSTI = 0.10) identified a coherent pattern of recurrence-associated predicted functional shifts. At a tightened false-discovery rate (q < 0.05), these shifts converged on de novo nucleotide biosynthesis, dTDP-L-rhamnose generation, one-carbon metabolism, and peptidoglycan cell-wall assembly. Their convergence on related circuits points to a focused predicted pathway pattern and offers a mechanistic framework for future testing. Specifically, we propose two hypotheses: first, that nucleotide and one-carbon precursors from the gut could reach the circulation via the gut-liver axis, supplying substrates for residual chemoresistant cells [21, 22]; and second, that heightened peptidoglycan turnover—as chemotherapy weakens the mucosal barrier [23, 24]—could favor the translocation of pathogen-associated molecular patterns (PAMPs), fostering low-grade systemic inflammation and anti-tumor immune exhaustion [25, 26]. Bridging to the host, the predicted abundance of bacterial dTDP-L-rhamnose biosynthesis positively tracked the host prognostic nutritional index (r = 0.399). However, given its moderate bootstrap support in our patient-level bootstrap network (selection probability 0.62), this specific host link is strictly offered as an exploratory observation. Like the other predicted functional inferences, it rests on computational prediction rather than direct measurement and requires multi-omic confirmation [27].

Taken together, these findings outline a clinical opportunity. Because a recurrence-associated trajectory was already detectable during mid-to-late treatment, serial fecal profiling could offer a non-invasive way to monitor emerging risk while patients remain under active care, the window in which intervention is most feasible. Furthermore, the specificity of the taxonomic and predicted functional signals suggests that indiscriminate broad-spectrum probiotics are poorly suited to reverse a defined ecological pattern [28]. This points instead toward precision approaches, such as evaluating strategies targeting the implicated predicted pathways or limiting PAMP translocation with mucosal-protective strategies. The natural next steps are external validation in independent cohorts with a pre-specified prediction model, and direct mechanistic testing in germ-free and metabolomic systems [29].

Several limitations qualify these conclusions. The number of recurrences was small (13 events); although Firth penalized estimation, stratified sensitivity analyses, stricter false-discovery control, and patient-level bootstrapping were used to reduce small-sample and multiple-testing concerns, the findings remain associations rather than causal effects. Because all recurrences occurred in advanced-stage (FIGO III/IV) disease, the microbial associations are strictly observed within this subgroup and cannot be fully separated from residual disease severity. Although peri-chemotherapy antibiotic, probiotic, and corticosteroid exposure was addressed in sensitivity analyses, dietary intake and gastrointestinal symptoms remain important unmeasured confounders. These variables must be acknowledged as limitations, as continuous outpatient dietary monitoring was unfeasible. The absence of a treatment-naïve baseline also limits the distinction between pre-existing microbial differences and treatment-phase changes. Finally, the predicted functional findings were based on computational inference and require multi-omic confirmation, and the bootstrap analysis assesses internal stability rather than external generalizability, making prospective validation in independent cohorts essential.

Conclusion

By tracking the gut microbiota longitudinally across frontline chemotherapy, this study shows that its dynamic remodeling over treatment is associated with long-term ovarian cancer recurrence, a trajectory that exhibited no significant statistical relationship with acute treatment toxicity. The recurrence-associated signal was defined not by a baseline deficit but by the progressive depletion of the short-chain-fatty-acid producer Fusicatenibacter, together with shifts in predicted functional profiles whose exploratory link to a host immune-nutritional index suggests possible microbe-host crosstalk. These findings support longitudinal microbial dynamics as a candidate non-invasive marker of recurrence risk that warrants external validation, and provide a hypothesis-generating rationale for precision strategies that target implicated taxa and predicted functional pathways to modulate the anti-tumor milieu.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Abbreviations

OC

Ovarian cancer

SAEs

Severe adverse events

SOCFCP

Shanghai Ovarian Cancer and Family Care cohort

LME

Linear mixed-effects

OR

Odds ratio

CI

Confidence interval

rmcorr

Repeated-measures correlation

PDS

Primary debulking surgery

TC

Paclitaxel and carboplatin

CTCAE

Common Terminology Criteria for Adverse Events

SII

Systemic immune-inflammation index

PNI

Prognostic nutritional index

PFS

Progression-free survival

RECIST

Response Evaluation Criteria in Solid Tumors

ASVs

Amplicon sequence variants

GLMM

Generalized linear mixed model

TSS

Total-sum-scaled

FDR

False discovery rate

FIGO

International Federation of Gynecology and Obstetrics

BRCA

Breast cancer gene

NLR

Neutrophil-to-lymphocyte ratio

PLR

Platelet-to-lymphocyte ratio

CA125

Cancer antigen 125

KELIM

CA125 elimination rate constant / CA125 elimination kinetics

PCoA

Principal Coordinates Analysis

SCFAs

Short-chain fatty acids

PAMPs

Pathogen-associated molecular patterns

LMR

Lymphocyte-to-monocyte ratio

HALP

Hemoglobin, albumin, lymphocyte, and platelet

HE4

Human epididymis protein 4

Author contributions

Conceptualization: W.S., Z.L. and Y.W.; Clinical sample collection: N.L., W.S., R.B., and X.L.; Cohort maintenance: N.L., W.S., S.C., Y.Z., Y.Z., H.D., Y.L., and Z.L.; Formal analysis: W.S. and N.L.; Writing—original draft preparation: W.S., N.L., Z.L. and Y.W.; Writing—review and editing: W.S., Z.L. and Y.W.; Visualization: W.S., R.B. and X.L.; Supervision: Z.L. and Y.W. All authors read and approved the final manuscript.

Funding

This study was supported by grants from the Shanghai First Maternity and Infant Hospital Science Project (2023B09) and Shanghai Public Health Research Special Project (2025GKQ26). The funders played no role in study design, data collection, data analysis, or manuscript preparation. No other financial support was received.

Data availability

The raw sequence data reported in this paper have been deposited in the Genome Sequence Archive (Genomics, Proteomics & Bioinformatics 2025) in National Genomics Data Center (Nucleic Acids Res 2025), China National Center for Bioinformation / Beijing Institute of Genomics, Chinese Academy of Sciences (GSA: CRA047203) that are publicly accessible at https://ngdc.cncb.ac.cn/gsa.

Declarations

Ethics approval and consent to participate

This study was conducted in accordance with the Declaration of Helsinki, and the protocol was reviewed and approved by the Scientific and Ethical Committee of the Shanghai First Maternity and Infant Hospital affiliated with Tongji University (Approval No. KS23167). Written informed consent to participate was obtained from all individual patients included in the study.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.