Development and validation for prognostic nomogram of epithelial ovarian cancer recurrence based on circulating tumor cells and epithelial–mesenchymal transition

We aimed to determine the prognosis value of circulating tumor cells (CTCs) undergoing epithelial–mesenchymal transition in epithelial ovarian cancer (EOC) recurrence. We used CanPatrol CTC-enrichment technique to detect CTCs from blood samples and classify subpopulations into epithelial, mesenchymal, and hybrids. To construct nomogram, prognostic factors were selected by Cox regression analysis. Risk stratification was performed through Kaplan–Meier analysis among the training group (n = 114) and validation group (n = 38). By regression screening, both CTC counts (HR 1.187; 95% CI 1.098–1.752; p = 0.012) and M-CTC (HR 1.098; 95% CI 1.047–1.320; p = 0.009) were demonstrated as independent factors for recurrence. Other variables including pathological grade, FIGO stage, lymph node metastasis, ascites, and CA-125 were also selected (p < 0.005) to construct nomogram. The C-index of internal and external validation for nomogram was 0.913 and 0.874. We found significant predictive values for the nomogram with/without CTCs (AUC 0.8705 and 0.8097). Taking CTC counts and M-CTC into separation, the values were 0.8075 and 0.8262. Finally, survival curves of risk stratification based on CTC counts (p = 0.0241), M-CTC (p = 0.0107), and the nomogram (p = 0.0021) were drawn with significant differences. In conclusion, CTCs could serve as a novel factor for EOC prognosis. Nomogram model constructed by CTCs and other clinical parameters could predict EOC recurrence and perform risk stratification for clinical decision-making. Trial registration Chinese Clinical Trial Registry, ChiCTR-DDD-16009601, October 25, 2016.

www.nature.com/scientificreports/ Univariable and multivariate regression analysis of training group. Figure 1 showed that patients suffered cancer recurrence had higher CTC counts and M-CTC percentage (p-value < 0.05). To further determine the independent predictive indexes, univariate and multivariate analyses were performed (  (Table 3).  Fig. 2A), while a nomogram without CTC counts and M-CTC percentage were also constructed for comparison (Fig. 2B). In the training group, the C-index values of 1000 sample bootstraps were 0.913 and 0.832 for the nomograms with and without CTCs. When applied to the validation cohort, the C-index values were 0.874 and 0.782, respectively, which showed a significant prognosis value of discrimination in both cohorts for the nomogram with CTC counts and M-CTC percentage. Further risk stratification in EOC patients calibration curves manifested that the probability of predicted 1-year and 2-year recurrence rate in nomogram was well consistent between the predicted outcome of cancer recurrence and actual observation in the training group (Fig. 3A,B). Moreover, in the external validation group, the calibration curves also illustrated good validation between predicted and observed 1-and 2-year recurrence proportions (Fig. 3C,D). The discrimination and calibration validation of the external group certificated that nomogram models in this study were comparatively accurate enough to predict the recurrence probability of patients with EOC.  4A). For the nomogram with/without CTC counts and M-CTC percentage, the AUCs were 0.8705 and 0.8097 (Fig. 4B). Meanwhile, as illustrated in Fig. 4C,D, the discriminatory values of CTC counts and M-CTC percentage were significant among ovarian cancer patients, with the log-rank p-value of 0.0241 and 0.0107, respectively. When stratified by CTC counts, patients with CTCs ≥ 9 and 5 ≤ CTCs < 9 were associated with a 1.98-fold increase (95% CI 1.04-2.47) and 1.24-fold increase (95% CI 1.07-2.29) of recurrence rate, comparing to those with CTCs < 5, while patients with M-CTC percentage ≥ 0.3 and 0.1 ≤ M-CTC < 0.3 were associated with a 2.10-fold increase (95% CI 1.54-2.66) and 1.43-fold increase (95% CI 1.14-2.53) of recurrence rate, comparing to those with M-CTC < 0.1. Moreover, the patients were then divided into three risk groups (low-, intermediate-and high-risk groups) based on the nomogram-predicted recurrence probabilities. For the nomogram without CTCs, when compared with the low-risk group, the high-risk and intermediate-risk groups were associated with a 2.37-fold increase (95% CI 1.28-4.83) and 1.48-fold increase (95% CI 1.17-2.64) in the risk of recurrence, with the p-value of 0.0021 (Fig. 4E). Separately, the log-rank p-values were 0.0386 between high and intermediate risk groups, while 0.0930 between intermediate-risk and low-risk groups. Meanwhile, in the CTCs based nomogram, the high-risk and intermediate-risk groups were associated with a 3.14-fold increase (95% CI 1.16-4.50) and 1.86-fold increase (95% CI 1.70-3.96) in the risk of recurrence, with the p-value of 0.0002 (Fig. 4F). Separately, the log-rank p-values were 0.0292 between high-risk and intermediate-risk groups, while 0.0491 between intermediate-risk and low-risk groups. www.nature.com/scientificreports/

Discussion
The clinical value of CTCs is constantly growing, as they could serve precision-medicine-based treatment of EOC patients by stratifying those with potential high recurrence risk. In this prospective study, we developed and validated a novel nomogram based on CTCs and other clinicopathological variables to categorize EOC patients concerning tumor recurrence. We also found that the presence of CTC subpopulations, especially the M-CTC percentage is associated with ovarian cancer recurrence. To our knowledge, this is the very first recurrence risk stratification developed for EOC patients especially refer to CTCs undergoing EMT.
Increasing evidence indicated that CTCs count is an independent predictor for prognosis in various solid carcinoma, including breast cancer, prostate cancer, and hepatocellular cancer. The breast cancer studies have demonstrated that patients with CTCs < 5 per 7.5 mL blood would suffer shorter PFS (2.1 months vs 7.0 months, p < 0.001) 19,20 . In prostate cancer, CTCs count is considered as an independent predictor of the overall survival rate among castration-resistant prostate cancer patients (p < 0.05) 6 . However, regarding ovarian cancer, whether CTCs detection was associated with prognosis remains controversial 10,21 . Judson et al. 21 characterized CTCs by immunomagnetic beads conjugated to epithelial markers followed with the microscopic evaluation refer to specific cytoplasmic staining and did not find a significant correlation between CTCs and prognosis. In contrast, www.nature.com/scientificreports/ Poveda et al. 10 analyzed CTCs using the CellSearch system and concluded that elevated CTCs could impart unfavorable prognoses of ovarian cancer patients. Differences in isolation and characterization technique in previous studies make it difficult to combine conclusions in agreement 22 . So, the standardization of CTCs detection techniques is of great importance. In our study, we revealed that CTCs count was an independent prognosis factor for ovarian cancer recurrence through both univariable and multivariable analyses using the CanPatrol CTC-enrichment technique System. The high sensitivity of the CanPatrol technique might be attributed to a simple filter-based separation method that might reduce CTC loss caused by the complicated washing and centrifugation process 23 . Meanwhile, the routine approach of the Cellsearch System used in previous studies might fail to detect CTCs undergoing EMT, since it only isolates CTCs by the only tumor epithelial cell expression of EpCAM 11,23 and not mesenchymal ones without epithelial markers. Thus, we used the CanPatrol CTC-enrichment technique System to detect aggressive CTCs subpopulations that might have undergone EMT through various target sequences, including EpCAM, CD45, CK8/18/19, vimentin, and Twist 5 . For hepatocellular carcinoma, a previous study concluded that M-CTC percentage ≥ 2% before the operation was a novel predictor for early recurrence with the AUC 0.75 (95% CI 0.66-0.84) 8 , which was partly consistent with our finding that ovarian cancer patients with M-CTC percentage ≥ 0.3 and 0.1 ≤ M-CTC < 0.3 were associated with a 2.10-fold increase and 1.43-fold increase of recurrence rate, when compared to those with M-CTC < 0.1. However, regarding the results of univariable regression analysis, E-CTC percentage and hybrid-CTC percentage were not considered as independent prognostic factors for OS of EOC patients (p-value ≥ 0.05). To the best of our knowledge, this is the first study to reveal the considerable clinical value of both CTC counts and M-CTC percentage in ovarian cancer prognosis.
Moreover, we aimed to develop a predictive nomogram to help facilitate the risk triage of ovarian cancer recurrence. Besides the presence of CTCs, we also selected several routinely collected risk factors including pathological grade, FIGO stage, lymph node metastasis, ascites, and CA-125 to construct the nomogram in training group [24][25][26] . The clinical relevance of our nomogram was demonstrated by its internal and external validation with the C-index of 0.913 and 0.874, which indicated that our model included in CTCs could provide a more reliable predictive evaluation for ovarian cancer recurrence than previous studies 26,27 .
Nevertheless, we further performed risk stratification of EOC patients based on CTC counts, M-CTC percentage, and points derived from the nomogram. All the risk stratification was well validated by survival analysis (p < 0.05) with the AUC higher than 0.75 as well. According to risk stratification, especially by the nomogram, we could carry out individualized and targeted treatment to improve the prognosis of ovarian cancer.
However, there are also some limitations of our study. Firstly, the prospective study enrolled a relatively small sample size of 152 EOC patients in a single-center, which might limit the accuracy of results. To overcome this problem, additional multi-center studies with a larger sample size would be of great importance to further validate our results. Second, detection efficiency might be biased since the CanPatrol system is a filtration-based system, allowing small CTCs to easily cross the barrier. Thus, other CTCs collection techniques might also be used to improve detection efficiency in future studies.
In conclusion, CTCs, especially those undergoing EMT hold promise prognostic value as minimally-invasive biomarkers for ovarian cancer recurrence. By the advanced CanPatrol CTC-enrichment technique, our study evaluated both CTC counts and M-CTC percentage to clarify their clinical value. The prognostic nomogram based on CTCs and EMT could support clinical decision-making and provide cues for early intervention among EOC patients.

Methods
Study design and patients. We enrolled 181 patients with pathologically diagnosed EOC who underwent surgery at the Department of Obstetrics and Gynecology, Renji Hospital Affiliated to Shanghai Jiaotong University School of Medicine between June 2017 to October 2019. The criteria for inclusion in this study were: (1) newly diagnosis EOC confirmed by pathological biopsy; (2) no coexisting cancers or prior cancers within 5 years; (3) no preoperative treatment, including neoadjuvant chemotherapy or radiotherapy. The exclusion criteria were as follows: (1) lost to follow-up (n = 9); (2) without detailed clinical, laboratory, imaging, and treatment data (n = 8); (3) underwent other treatments, such as radiotherapy or immunotherapy (n = 5); (4) without consent to use medical information for the research purpose (n = 4), and (5) with status not allowing the treatment of operation followed by chemotherapy (n = 3). As a result, 152 patients were assessed in the analysis (Fig. 5). Moreover, we also involved 30 patients with benign gynecologic diseases at our institution as negative controls.
In order to achieve optimal tumor debulking, the operation for all involved patients was aimed at maximal ovarian tumor resection without visible residual tumor. The surgery was followed by standardized paclitaxel and platinum chemotherapy. All patients were followed up until September 1st, 2020. This study was approved by the Ethics Committee of Renji Hospital Affiliated to Shanghai Jiaotong University School of Medicine and all involved subjects provided informed consent for use of their information for research purposes. All experiments were performed in accordance with relevant guidelines and regulations.
Clinicopathological data collection. The clinical stage was evaluated according to the International Federation of Gynecology and Obstetrics (FIGO) stage system. Routine blood tests and tumor marker measurements, including carbohydrate antigen-125 (CA-125), carbohydrate antigen-199 (CA-199), carcinoembryonic antigen (CEA), alpha-fetoprotein (AFP), and human epididymis protein 4 (HE4) were conducted within 1 day before surgery. The clinicopathologic variables, including age, Body Mass Index (BMI), tumor size, menopausal status, fertility history, pathological grade, the FIGO stage, lymph node metastasis, ascites, and histological type were reviewed from medical records. Disease-free survival (DFS) was measured from the date of surgery to the Isolation and characterization of CTCs. Peripheral blood samples (5 mL, anticoagulated with EDTA) were collected 1 day before treatment, stored at 2-8℃, and processed within 4 h after sampling 7 . To avoid potential skin cell contamination caused by venipuncture, the first 2 mL of blood was discarded 28 .
In this study, we isolated and characterized CTCs through the CanPatrol system (Fig. 6). Firstly, the blood sample preserved in cell preservation solution was centrifuged for 5 min at a speed of 1850 rpm. After removing the supernatant, the sample was mixed with phosphate buffer saline (PBS) and 4% formaldehyde for 8 minutes 7 . For filtration, we passed the sample through the vacuum filtration system at 0.08 MPa 7 .This system included a filtration tube containing the membrane with 8-μm diameter pores, a vacuum pump, and a manifold vacuum plate with valve settings.

Construction of nomogram.
The dataset was randomly divided into training and validation cohorts. The selection bias refer to the random classification of the two cohorts was adjusted 32 . T-test and Chi-square test were used to analyze the differences of clinicopathologic characteristics between two cohorts for continuous and categorical variables, respectively. The prognostic factors were determined using both univariate and multivariate analyses through Cox's hazards regression model. The nomogram and calibration plots were generated with the "rms" package of R software 18 . Nomogram points, ranging from 1 to 100, were assigned refer to the weights www.nature.com/scientificreports/ for the relative importance of each model covariate determined by the final hazards regression model. In the nomogram, the total score for each patient was evaluated as a weighted sum of the contribution from each risk factor to predict the probability of recurrence at 1 and 2 years.
Validation of nomogram. The predictive ability of the nomogram model was measured by both discrimination and calibration. The discrimination of the nomogram model was evaluated by calibration curves, overlaying the observed probabilities and nomogram-predicted probabilities with 95% confidence interval (95% CI). As a measurement for internal validation, the Harrell's concordance index (C-index) was analyzed using tenfold cross-validation repeated for 20 times 33 . The calibration plots were generated by the "rms" package of R software 18 . We categorized patients into three risk groups of CTC counts, M-CTC percentage, and nomogram, based on the X-tile (Version 3.6.1, Yale University, New Haven, USA), a newly-developed bioinformatic tool to determine optimal cut-off points for survival analysis 34 . The X-tile software could test all possible cut-off points of target quantitative data by Log-rank test and selected the lowest p-value and highest χ2. The EOC patients involved were then divided into three risk groups: good, intermediate, and poor prognosis. The optimal cut-off values were 128 and 251 for the nomogram with CTCs, while 98 and 169 for the nomogram without CTCs. Taking CTCs into separation, the values were 5 and 9 in CTC count, 0.1 and 0.3 in the M-CTC percentage. Kaplan-Meier methods were used to generate the survival curves and the prognostic differences were assessed by Log-rank test. The receiver operating characteristic (ROC) curve analysis was applied to identify the prognosis value of the nomogram according to the Youden index and area under the curve (AUC). All the statistical analyses were conducted by R software Version 4.0.2 (GUI 1.72 Catalina build, https ://www.R-proje ct.org) and graphed using Graph Prism Version 7.0a (GraphPad Software, San Diego, CA, USA). p-value < 0.05 was defined as statistically significant.

Data availability
The data of these findings cannot be shared at this time as the data also forms part of an ongoing study. Requests for data will be considered by the corresponding author after publication of the study.