A multiplex biomarker assay improves the diagnostic performance of HE4 and CA125 in ovarian tumor patients.

Around 700 Swedish women are diagnosed with ovarian cancer or borderline tumors every year. Symptoms are few and non-specific in the early stages, causing delays in diagnosis and treatment. While patients with borderline tumors have an excellent prognosis, with a five-year survival rate of 97%, the prognosis is poor in ovarian cancer patients. Half will die within five years of diagnosis [ 1 , 2 ]. Ovarian cancer is predominantly in the form of epithelial tumors (90%); the remaining 10% comprise germ cell and sex-cord stromal tumors. The main morphological subtypes in epithelial ovarian cancer (EOC) are high-grade serous (HGSC) (70%), endometrioid (EC) (10%), clear cell (CCC) (10%), mucinous (MC) (3%) and low grade serous cancer (LGSC) (<5%) [ 3 ]. These subtypes differ in origin and behavior, and respond very differently to oncological treatment [ 4 , 5 ]. Despite advances in surgical and oncological treatment, little improvement has been seen in long-term survival in EOC [ 6 , 7 ]. The majority of patients are diagnosed in late stages. In order to improve survival, the patients must be diagnosed earlier, when the disease is still curable. A screening method for ovarian cancer, for use in the general population, has been sought for decades. Two large-scale prospective population studies, the PLCO and UKCTOCS trials, were unable to show a significant decrease in ovarian cancer mortality from screening with the plasma protein biomarker CA125 and / or transvaginal ultrasound [ 8 , 9 ]. Apart from screening, a way to earlier ovarian cancer diagnosis is to improve the risk assessment when a patient presents with an adnexal mass. Patients with an estimated high risk of malignancy should be referred to the proper level of care without unnecessary delay. The multivariate Risk of Malignancy Index (RMI) algorithm (incorporating CA125, ultrasound score and menopause status) has been in clinical use since the 1990s [ 10 ]. The use of RMI requires ultrasound competence, which is not always available at the primary care level. In 2009, Moore et al. [ 11 ] introduced the Risk of Ovarian Malignancy Algorithm (ROMA) (CA125, HE4 and menopause status) dispensing with the need for ultrasound evaluation [ 11 ]. In their study comparing ROMA and RMI in 2010, Moore et al. [ 12 ] found better performance for ROMA compared to RMI, although these findings have been questioned by subsequent studies [ 12 – 15 ]. Both algorithms have reduced sensitivity and specificity in early stages of EOC when they would be of most diagnostic value [ 16 ]. Karlsen et al 2015 [ 17 ] introduced a modified version of the ROMA, the Copenhagen Index (CPH-I), substituting menopause status for age. The CPH-I, ROMA and RMI had comparable performance in a multicenter study [ 17 ]. However, ultrasound-based models have been found superior for the preoperative assessment of an adnexal mass, provided there is access to good quality ultrasonography [ 18 ]. Many research groups, including ours, have evaluated a range of other biomarkers and combinations of biomarkers for their potential use in ovarian cancer [ 19 , 20 ]. CA125 continues to stand out as the single-best biomarker [ 21 , 22 ] and considerable research has been focused on the search for additional biomarkers to improve the performance of CA125 alone [ 23 ]. Lately, researchers have turned to the rapidly evolving field of proteomics in the search for new candidate biomarkers, using new techniques for high throughput multiplex analysis in large-scale protein studies [ 24 , 25 ]. In this study we analyzed the Olink ® Oncology II and Inflammation panels (in total 177 unique protein biomarkers) in 180 women with benign tumor, borderline tumor, early (stage I) or late (stage II-IV) EOC, with the aim of searching for new candidate biomarkers with the potential to improve the performance of HE4 and CA125 for discrimination between benign disease and EOC. We tested the individual biomarkers in three-biomarker combinations with HE4, CA125 and age. Only the addition of ITGAV improved the reference model of HE4, CA125 and age to a significant level. In order to test whether a multiplex biomarker model could further improve performance of the reference model, we combined cross-validation with LASSO regression. A 6-biomarker model (HE4, CA125, CXCL1, ITGAV, CEACAM1, IL-10RB and age) was found to be the best model for discrimination between benign tumors and EOC including borderline tumors.

Out of the 180 patient samples, eight samples did not pass internal quality control in the PEA analyses ( www.olink.com ) and were excluded from statistical analyses. These samples comprised two borderline tumors and six advanced stage EOC cases. The analyses below include 172 patients. Non-hierarchical clustering analysis was performed for the whole patient cohort and for serous tumors alone. Heat maps indicating protein expression levels for each patient are shown in the S1 Fig . No clustering could be observed visually. Principal component analysis did not segregate the patients into groups according to protein expression levels ( S2 Fig ). Out of the 177 biomarkers analyzed, a statistically significant difference in NPX levels between benign tumors and cancer was found for eight proteins (using a conservative cut-off p < 0.001, p-values adjusted with False Discovery Rate (FDR)). HE4 (WFDC2) and CA125 (MUC16) were highest ranked. Most of the proteins were up-regulated in cancer patients although lower NPX levels were seen for two proteins, ITGAV and DNER, in cancer patients ( Table 2 ). Box plots for the six proteins with the lowest p-values are shown in S3 Fig . Table 3 shows the AUC values for discriminating benign tumors from cancer for the individual proteins. 34 proteins had AUC > 0.7. HE4 (WFDC2) ranked highest with AUC 0.830 (95% CI 0.739–0.921). ROC curves for the six proteins with the highest AUC values are depicted in Fig 1 . Table 4 shows the AUC values, sensitivities (95% specificity) and specificities (95% sensitivity) for the reference model with HE4, CA125 and age (AUC 0.838 (0.752–0.924) and for the reference model with the addition of each one of the remaining 32 proteins with AUC > 0.7. ITGAV was the only biomarker to significantly improve the diagnostic performance of the reference model, to AUC 0.874 (0.799–0.949) (p = 0.045). Sensitivities and specificities are low with wide confidence intervals, calling for caution when interpreting results. Benign tumors vs. cancer. Benign tumors vs. cancer, p 0.7. Benign tumors vs. cancer. a Comparing the reference model and the reference model with an added biomarker. Cross-validation and LASSO regression were used to select multi-biomarker combinations for logistic regression models. A six-biomarker model including HE4, CA125, CEACAM1, CTSV, CXCL6, S100A4 and age was found to be the best model for discriminating between benign tumors and EOC with AUC 0.921 (0.863–0.979), sensitivity 0.897 / specificity 0.889 at best point cut-off (p = 0.025) ( Table 5 and Fig 2 ). Benign tumors vs. cancer. Benign tumors vs. cancer. a Comparing the reference model and the reference model with additional marker combinations. A statistically significant difference in NPX levels between benign tumors and borderline tumors + cancer was found for only two proteins, HE4 (WFDC2) and CA125 (MUC16) (conservative cut-off p < 0.001, p-values adjusted with False Discovery Rate (FDR)) ( Table 6 ). Box plots for the six proteins with the lowest p-values are shown in S4 Fig . Table 7 shows the AUC values for discriminating benign tumors from borderline tumors + cancer for the individual proteins. Fifteen proteins had AUC > 0.7. HE4 (WFDC2) ranked highest with AUC 0.767 (0.672–0.861). ROC curves for the six proteins with the highest AUC values are depicted in Fig 3 . Table 8 shows the AUC values, sensitivities and specificities for the reference model with HE4, CA125 and age (AUC 0.770 (0.674–0.865)) and for the reference model with the addition of each one of the remaining 13 proteins with AUC > 0.7. Again, only the addition of ITGAV would significantly increase the diagnostic performance of the reference model, to AUC 0.818 (0.737–0.900) (p<0.05). For both models the sensitivities and specificities were low and confidence intervals wide, indicating statistical uncertainty. Benign tumors vs. borderline + cancer. Benign tumors vs. borderline + cancer, p 0.7. Benign tumors vs. borderline + cancer. a Comparing the reference model and the reference model with an added biomarker. Multi-marker models were developed using cross-validation and LASSO regression. A six-biomarker model (HE4, CA125, CXCL1, ITGAV, CEACAM1, IL-10RB and age) was the best model for discrimination between benign tumors and borderline tumors + cancer, with AUC 0.868 and sensitivity 0.86 / specificity 0.82 at best point cut-off (p = 0.016) ( Table 9 and Fig 4 ). Benign tumors versus borderline + cancer. Benign tumors vs. borderline + cancer. a Comparing reference model and model with added biomarkers.

HE4 was the best performing biomarker for discrimination of benign tumors versus EOC including borderline tumors in our study. ITGAV was the only individual biomarker found to improve the diagnostic performance of HE4, CA125 and age. Using LASSO regression, a multiplex model including 6 biomarkers (HE4, CA125, ITGAV, CXCL1, CEACAM1, IL-10RB) and age had the highest diagnostic accuracy for discrimination between benign ovarian tumors and EOC including borderline tumors. We find that the addition of other known carcinogenesis-related biomarkers in multiple marker combinations has potential to improve the performance of the established markers HE4 and CA125.

A single cohort-design was used for biomarker discovery, with the aim of validation in a larger subsequent cohort of patients in case of positive findings. Peripheral blood samples were obtained preoperatively from 180 women with an adnexal mass admitted for surgery at the Department of Obstetrics and Gynecology, Skåne University Hospital Lund, Sweden 2005 to 2012. Blood was collected in citrate tubes, centrifuged, and then the plasma was stored at −20°C until it was analyzed. All diagnoses were verified by histopathologic examination. The histological type and stage of the disease according to the International Federation of Gynecology and Obstetrics (FIGO) were available in all malignant cases. The patient cohort included 30 cases of benign adnexal mass, 28 cases with borderline tumors, 25 early EOC cases (FIGO stage I) and 97 advanced EOC cases (FIGO stage II-IV) ( Table 1 ). The frozen plasma samples were shipped to Olink Proteomics AB, Uppsala, Sweden, for analyses. a Information on age /date of surgery not available in one patient with a benign tumor. Proteins were measured using the Olink® Oncology II and Inflammation panels (Olink Proteomics AB, Uppsala, Sweden) according to the manufacturer's instructions. The biomarkers included in each panel are listed in the supporting information, in the S1 and S2 Files. The Proximity Extension Assay (PEA) technology used for the Olink protocol has been well described [ 26 ] and enables 92 analytes to be analyzed simultaneously, using 1 μL of each sample. Pairs of oligonucleotide-labeled antibody probes bind to their targeted protein, and if the two probes are brought into close proximity the oligonucleotides will hybridize in a pair-wise manner. The addition of a DNA polymerase leads to a proximity-dependent DNA polymerization event, generating a unique PCR target sequence. The resulting DNA sequence is subsequently detected and quantified using a microfluidic real-time PCR instrument (Biomark HD, Fluidigm). The final assay read-out is presented in Normalized Protein eXpression (NPX) values, which is an arbitrary unit on a log2-scale where a high value corresponds to a higher protein expression. The NPX values are relative and not comparable between different proteins. Analyses were performed by biomedical technicians at Olink Proteomics AB in Uppsala, Sweden. Disease status of the patients was unknown to the technicians performing the analyses. Samples were randomized across the plates and run in duplicates. Data was quality controlled and normalized using an internal extension control and an inter-plate control, to adjust for intra- and inter-run variation. All assay validation data (detection limits, intra- and inter-assay precision data, etc.) are available on the manufacturer's website ( www.olink.com ). Hierarchical clustering analysis and principal component analysis were performed to search for clusters of proteins associated with the different tumor categories. Patients were subsequently categorized into benign tumors versus cancer, or benign tumors versus borderline tumors and cancer, and differences in protein expression between groups were analyzed with a Student’s t-test with a p-value < 0.001 indicating a statistically significant difference; p-values were adjusted for multiple comparisons using the False Discovery Rate (FDR). Each biomarker was used as a continuous variable in univariate logistic regression models, with the binary outcome benign tumors versus cancer or benign tumors versus borderline tumors and cancer. Receiver Operator Curves (ROC) were constructed and the Area under the Curve (AUC) was calculated with 95% confidence intervals using the non-parametric bootstrap procedure. In order to evaluate the biomarkers’ potential to improve the performance of the ROMA and CPH-I algorithms, a multivariate logistic regression model including the biomarkers HE4 and CA125 and age was constructed to serve as a reference model. Each of the biomarkers with AUC > 0.7 was added in turn to the reference model. The classification accuracy of each model was evaluated with the AUC, and the AUC for each model was compared to the AUC of the reference model using DeLong’s method. A p-value < 0.05 for differences in AUC was considered statistically significant. For each model, the sensitivity corresponding to a specificity of 0.95, and specificity corresponding to sensitivity of 0.95 was calculated. We wished to test whether a multiplex biomarker model could further improve diagnostic performance of the reference model. In order to select which combination of biomarkers to include in a final logistic regression model in addition to HE4, and CA-125 and age, a combination of cross-validation and LASSO-regression was employed. To start with the data was randomly split suing a 50/50 split into a training and test set. In the training set the shrinkage parameter (λ) was estimated using k-fold cross-validation. The estimated shrinkage parameter λ CV was then used in the test set in order to perform variable selection. The selected variables and the absolute value of the coefficients were saved. This process was then repeated 10 times. Next, the variables were ordered by the number of times they were selected and the sum of its estimated coefficients. The lowest ranked variable was removed and the entire process was repeated until a final model was selected. The final models were estimated with logistic regression. Receiver Operator Curves (ROC) were constructed and Area under Curve (AUC) calculated with 95% confidence intervals using the non-parametric bootstrap procedure. The AUC for each model was compared to the AUC of the reference model using DeLong’s method. A p-value < 0.05 for differences in AUC was considered statistically significant. All statistical analyses were carried out using R v 4.0.0 (R Core Team (2018). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. URL http://www.R-project.org/ ). Written informed consent was obtained from all study participants. Ethical approval was granted by the Ethical Review Board at the Faculty of Medicine, Lund University, Sweden. Dnr 495 2016 (amendment to Dnr 558–2004 and 94–2006).

a) All patients b) Serous tumors only. (TIF) Click here for additional data file. a) All patients b) Serous tumors only. (TIF) Click here for additional data file. Benign tumors, early and late stage EOC. (TIFF) Click here for additional data file. Benign tumors, borderline tumors, early and late stage EOC. (TIF) Click here for additional data file. (DOCX) Click here for additional data file. (DOCX) Click here for additional data file. (DOCX) Click here for additional data file. (DOCX) Click here for additional data file.