A two-step strategy for identification of plasma protein biomarkers for endometrial and ovarian cancer

Clinical material and ethics approval

Plasma protein measurements

The abundance of 441 unique protein in plasma were analyzed using the Olink Multiplex assays CVD II, CVD III, INF I, ONC II and NEU I (http://www.olink.com) and quantified by real-time PCR using the Fluidigm BioMark™ HD real-time PCR platform as described earlier [12]. Briefly, for each protein a unique pair of oligonucleotide-labeled antibody probes bind to the targeted protein, and if the two probes are in close proximity a PCR target sequence is formed by a proximity-dependent DNA polymerization event and the resulting sequence is subsequently detected and quantified using standard real-time PCR. Data is then normalized and transformed using internal extension controls and inter-plate controls, to adjust for intra- and inter-run variation as described earlier [12]. The final assay read-out is given in Normalized Protein eXpression (NPX), which is an arbitrary unit on log2-scale where a high value corresponds to a higher protein expression. Each PEA measurement has a lower detection limit (LOD) calculated based on negative controls that are included in each run, and measurements below LOD were removed from further analysis. All assay characteristics including detection limits and measurements of assay performance and validations are available from the manufacturers webpage (http://www.olink.com). The analyses were based on 1 μL of plasma for each panel of 92 assays. To avoid batch effects, samples from the different disease entities and cohorts, including benign tumors, were randomized across assay plates. Each plate included internal controls, as described previously used to adjust for technical variation and/or sample irregularities.

Development of focused multiplex PEA panels

The focused PEA assays were designed to be compatible with the Fluidigm 192 × 24 Integrated Fluidic Circuit (IFC) which can measure up to 24 assays on 192 samples simultaneously. Hence, 21 protein marker assays and three internal controls for run QC (Incubation Control, Extension Control and Detection control) can be analyzed per IFC. The main difference between the focused assays and the discovery assays is that the focused assays allow higher PEA probe concentration, which in turn means that higher levels of antigen can be measured without reaching the hook in the measuring range (i.e. too much antigen will decrease signals in a homogenous immunoassay format). This modification allowed inclusion of PEA assays from the CVD III panel, where samples are normally diluted 1:100 before analysis, into the focused assays used to analyze undiluted samples, making the differences in protein concentration to be measured more than 7 logs.

Statistical methods

All calculations were carried out in R version 3.2.3 (R core team) [13]. Individual protein levels were also normalized by plate and sampling round using the MDimNormn-package [14]. This was done separately for the discovery and replication cohorts.

Significance levels for comparison of protein levels between cases and controls in the discovery cohort were calculated using the two-sided rank-based Spearman test (Wilcoxon). From the entire discovery-cohort, the top-ranking proteins for each of the two cancers was identified based on the p value in the comparison of benign tumors to cancer samples. 15 proteins were selected for association with OC and 16 with EC out of which 4 overlapped. Lasso and Elastic-Net Regularized Generalized Linear Models (’glmnet’ package in R) was used to fit multivariate models for each cancer using the selected proteins. These models were then evaluated based on the best point and their specificity at a fixed sensitivity of 0.95. This was repeated for each of the tumor/control combination investigated.

A few proteins overlapped between the two cancers, and a total of 27 proteins were selected to be characterized in the replication cohort using focused PEA-panels.

Model performance was evaluated by randomly splitting the observations in the replication cohort into a training set (75%) and a test set (25%). A model was then built using the training set and with the same proteins as selected from the discovery-cohort. The multivariate models were retrained in the replication cohorts using the ‘lm’ function in R. The reason for re-building the model in the replication data set is that the scaling of the NPX-values can differ between the multiplex PEA panels. This model was then used to predict the response variable in the test set. The random-split into training and test set will often result in different performance due to the limited sample-size. To accommodate for this, a cross-validation schema was applied and the process was repeated 100 times and model prediction errors on the training and test set were recorded. Sensitivity and specificity from the replication were reported as mean ± 1 SD of the 100 runs.

Discovery step analyses

The protein biomarker panels were identified using a two-step study design (Fig. 1). First, the levels of 441 individual proteins was measured in pretreatment plasma collected at time of diagnosis from women with OC or EC and compared to benign tumors and population controls (Table 1). No difference was found in the mean age of women with benign tumors and OC (stage I–IV) or between healthy controls and women with OC (Table 1). EC is diagnosed at a higher age than OC, which is reflected in the age difference seen between women with EC and those with benign tumors (Table 1).

The distribution of differences in protein abundance (NPX-values) between women with OC or EC and benign tumors is shown for all proteins in Fig. 2. Proteins with a significant (Bonferroni-corrected significance threshold) difference in NPX values between benign tumors and OC (Fig. 2a–c) and between OC and EC (Fig. 2d) are labelled by name. The clinically used biomarker CA125 showed the largest differences in NPX value between OC cases and benign tumors. Nominal p-values for the comparisons of OC or EC with benign tumors and population controls are given for each protein in Additional file 1: Table 1.

At the highest AUC (best point analysis) the 15-protein selected for OC could distinguish women with benign tumors from those with OC stage I–IV with a sensitivity = 0.79 and specificity = 0.85 (AUC = 0.86), from OC stage I–II with a sensitivity = 0.68 and specificity = 0.73 (AUC = 0.72) and from OC stage III–IV with a sensitivity = 0.93 and specificity = 0.93 (AUC = 0.95) (Table 4, Fig. 3). At a minimum sensitivity of 0.95, as a cutoff value for the biomarker panel to be used as a preoperative diagnostic test, the specificity to distinguish between benign ovarian tumors and OC stage III–IV was 0.54 (Table 4).

Comparison	Proteins	Best point				AUC	Minimum specificity of 0.95 (screening)				Minimum sensitivity of 0.95 (diagnostic)
		Sensitivity		Specificity			Sensitivity		Specificity		Sensitivity		Specificity
		Mean	SD	Mean	SD		Mean	SD	Mean	SD	Mean	SD	Mean	SD
Analysis discovery
Benign tumors versus ovarian cancer stage I–II	All	0.68	0.07	0.73	0.08	0.72	0.33	0.11	0.96	0.01	0.96	0.01	0.07	0.08
Benign tumors versus ovarian cancer stage III–IV	All	0.93	0.04	0.93	0.03	0.95	0.86	0.07	0.96	0.01	0.96	0.00	0.54	0.25
Benign tumors versus ovarian cancer stage I–IV	All	0.79	0.04	0.85	0.04	0.86	0.61	0.07	0.96	0.01	0.96	0.01	0.18	0.10
Population ontrols versus ovarian cancer stage I–II	All	0.70	0.07	0.75	0.07	0.74	0.43	0.07	0.96	0.01	0.97	0.01	0.02	0.05
Population controls versus ovarian cancer stage III–IV	All	0.94	0.03	0.96	0.03	0.97	0.92	0.05	0.96	0.01	0.97	0.01	0.47	0.29
Population controls versus ovarian stage I–IV	All	0.75	0.05	0.87	0.05	0.86	0.66	0.06	0.95	0.00	0.96	0.01	0.13	0.10
Benign tumors versus endometrial cancer	All	0.77	0.05	0.79	0.05	0.83	0.45	0.11	0.96	0.01	0.97	0.01	0.09	0.11
Population controls versus endometrial cancer	All	0.69	0.06	0.78	0.06	0.76	0.43	0.07	0.95	0.00	0.97	0.01	0.07	0.07
Ovarian cancer versus endometrial cancer	All	0.84	0.04	0.88	0.04	0.89	0.64	0.17	0.97	0.01	0.96	0.01	0.42	0.16
Replication
Benign tumors versus ovarian cancer stage I–II	All	0.73	0.08	0.74	0.07	0.76	0.29	0.11	0.97	0.01	0.97	0.01	0.19	0.14
Benign tumors versus ovarian cancer stage III–IV	All	0.88	0.03	0.92	0.03	0.92	0.81	0.07	0.96	0.01	0.96	0.01	0.58	0.15
Benign tumors versus ovarian cancer stage I–IV	All	0.83	0.04	0.87	0.04	0.89	0.68	0.10	0.97	0.01	0.96	0.01	0.48	0.12
Population ontrols versus ovarian cancer stage I–II	All	0.62	0.09	0.68	0.11	0.63	0.24	0.12	0.97	0.01	0.97	0.01	0.06	0.10
Population controls versus ovarian cancer stage III–IV	All	0.85	0.03	0.92	0.04	0.89	0.78	0.09	0.97	0.01	0.96	0.01	0.35	0.15
Population controls versus ovarian stage I–IV	All	0.77	0.05	0.85	0.05	0.86	0.62	0.10	0.96	0.01	0.96	0.01	0.21	0.11
Benign tumors versus endometrial cancer	All	0.70	0.06	0.67	0.07	0.72	0.22	0.09	0.97	0.01	0.96	0.01	0.25	0.08
Population controls versus endometrial cancer	All	0.64	0.07	0.72	0.08	0.71	0.24	0.14	0.97	0.01	0.96	0.01	0.10	0.06
Ovarian cancer versus endometrial cancer	All	0.73	0.03	0.77	0.04	0.78	0.36	0.06	0.96	0.00	0.96	0.00	0.11	0.06
Benign tumors versus ovarian cancer stage I–II	CA125/HE4	0.73	0.08	0.73	0.07	0.78	0.38	0.11	0.97	0.01	0.97	0.01	0.30	0.16
Benign tumors versus ovarian cancer stage III–IV	CA125/HE4	0.87	0.03	0.87	0.04	0.90	0.74	0.06	0.97	0.02	0.96	0.01	0.37	0.12
Benign tumors versus ovarian cancer stage I–IV	CA125/HE4	0.79	0.03	0.87	0.04	0.87	0.67	0.06	0.97	0.01	0.96	0.01	0.39	0.10
Population ontrols versus ovarian cancer stage I–II	CA125/HE4	0.59	0.08	0.77	0.10	0.66	0.35	0.11	0.97	0.01	0.97	0.01	0.14	0.07
Population controls versus ovarian cancer stage III–IV	CA125/HE4	0.84	0.03	0.89	0.04	0.86	0.74	0.07	0.97	0.01	0.96	0.00	0.17	0.09
Population controls versus ovarian stage I–IV	CA125/HE4	0.77	0.03	0.88	0.05	0.81	0.65	0.08	0.96	0.00	0.96	0.01	0.14	0.07

The 15-protein panel was also able to distinguish population controls from OC stage I–IV with a sensitivity = 0.75 and specificity = 0.87 (AUC = 0.86), from OC stage I–II with a sensitivity = 0.70 and specificity = 0.75 (AUC = 0.74) and from OC stage III–IV with a sensitivity = 0.94 and specificity = 0.96 (AUC = 0.97) (Table 4). At a specificity of 0.96, as a cutoff value for the biomarker panel to be useful in population screening, the sensitivity to distinguish controls from OC stage I–IV was 0.66, from OC stage I–II it was 0.43 and from OC stage III–IV it was 0.92 (Table 4).

Since plasma samples from benign tumors and OC stage I–IV had been analyzed previously for CA125 we compared our CA125 values from PEA with those from the Architect CA125-II assay (Abbott Diagnostics, Abbott Park, IL, USA). There was a strong correlation between the two measurements (R² = 0.87, p value < 6.72e−128), except for the highest values which were outside the linear part of the range for PEA (Fig. 4, Additional file 2: Table 2).

For EC the 16-protein panel distinguished benign tumors from EC with a sensitivity = 0.77 and specificity = 0.79 (AUC = 0.83). OC stage I–IV could be distinguished from EC with a sensitivity of 0.84 and specificity of 0.88 (AUC = 0.89). Finally, the EC patients could be distinguished from population controls with a sensitivity = 0.69 and specificity = 0.78 (AUC = 0.76) (Table 4).

Replication step

Two focused multiplex PEA assays were developed for the 27 proteins identified in the discovery step, according to the protocol described by the manufacturer [15]. There was excellent correlation between the protein estimates from the multiplex assays used in the discovery step and the focused PEA assays used in the replication step (Fig. 5). Despite absolute differences in NPX-values (as expected since the dynamic range was shifted for some assays in the custom panel) the rank of different proteins was maintained (Fig. 5). Average intra- and interassay CV for 17 of the proteins included in the discovery and replication assays have previously been shown to be similar to the Olink’s commercial 92-plex screening panels [15]. Thus, the results of the PEA analyses used in the discovery and the replication steps were considered comparable.

The focused PEA panels were then used to study an independent set of plasma samples from women with either of the two cancers, benign tumors and population controls (Table 1). For OC, the association was replicated for 13 of the 15 proteins and only IL-10 and CDH3 showed no association with OC. The protein panel distinguished benign tumors from OC stage I–IV with a sensitivity = 0.83 and specificity = 0.87 (AUC = 0.89), from OC stage I–II with a sensitivity = 0.73 and specificity = 0.74 (AUC = 0.76) and from OC stage III–IV with a sensitivity = 0.88 and specificity = 0.92 (AUC = 0.92) (Table 4, Fig. 6). The protein panel selected for OC had a significantly higher AUC than using only CA125 and HE4 for distinguishing benign tumors from OC (stage I–IV, p = 1.41e−13; stage 3–4, p = 9.56e−22). At a sensitivity of 0.96, the specificity to distinguish between benign ovarian tumors and OC stage I–IV was 0.48, stage I–II it was 0.19 and stage III–IV it was 0.58 (Table 4).

The selected protein panel distinguished population controls from OC stage I–IV with a sensitivity = 0.77 and specificity = 0.85 (AUC = 0.86), from OC stage I–II with a sensitivity = 0.62 and specificity = 0.68 (AUC = 0.63), and from OC stage III–IV with a sensitivity = 0.85 and specificity = 0.92 (AUC = 0.89) (Table 4, Fig. 6). At a minimum specificity of 0.95, the sensitivity to distinguish population controls from OC stage I–IV was 0.62, from OC stage I–II it was 0.24, and from OC stage III–IC it was 0.78 (Table 4).

In a recent study, candidate biomarkers for OC were identified using the Olink Multiplex ONC I v2 panel [5]. That panel is discontinued by the manufacturer and replaced by the ONC II panel used here. However, some of the proteins overlap with the panels used here, and we therefore compared the AUC values for these proteins in the two studies (Fig. 7a). There was good correlation in the comparison of benign tumors to OC stage III–IV, but the AUC values were generally higher in the study by Boylan et al. [5]. By contrast, no correlation was found in the AUC values when comparing benign tumors and OC stage I–II (Fig. 7b).

Of the 16 proteins selected for EC, an association was replicated for nine (PRSS8, MK, WFDC2 (HE4), ADM, MMP-7, ST2, VEGF-A, IL-6, HGF), while seven proteins showed no association (IL-10, IDUA, FABP-4, PCSK9, CTSZ, CCL16, Dkk-4). The protein panel distinguished between benign tumors and EC with a sensitivity = 0.70 and a specificity = 0.67 (AUC = 0.72). Population controls could be distinguished from EC with a sensitivity = 0.64 and specificity = 0.72 (AUC = 0.71).

Of the 13 proteins that differed significantly between OC stage I–IV and EC in the discovery data, six (hK11, MUCIN-16, FR-alpha, NTRK3, EN-RAGE, HGF) showed a significant difference in the replication step (Tables 2, 3). OC stage I–IV could be distinguished from EC with a sensitivity = 0.73 and a specificity = 0.77 (AUC = 0.78).

Proteins identified

Among the proteins we identified as biomarker candidates for OC some have been discussed previously [5], such as Mucin-16 (CA125) and Major epididymis-specific protein E4 (HE4), Midkine (MK) [16, 17], Kallikrein-11 (hK11) [18], Folate receptor alpha (FR) [19, 20] and Prostasin (PRSS8) [21, 22]. Boylan et al. [5] also using PEA technology further reported an association of OC with Interleukin-6 (IL-6) [23–25], Kallikrein-6 (KLK6) [26], Furin (FUR) [27], Chemokine (C-X-C motif) ligand 13 (CXCL13) and Tumor necrosis factor ligand superfamily member 14 (TNFSF14), but none of these proteins were among our top candidates. A number of the proteins we selected for the replication step were not studied by Boylan et al. [5] such as Interleukin-8 (IL-8), Nectin-4 (PVRL4), Interleukin-17C (IL17C), Poly (ADP-ribose) polymerase-1 (PARP-1), Superoxide dismutase (Mn), mitochondrial (SOD2) and Protein S100-A12 (EN-RAGE). Several of these have previously been noted in connection to OC. The level of Interleukin-8 has been proposed as a diagnostic and prognostic biomarker for OC [28–30]. Nectin-4 is overexpressed in epithelial cancers including OC and has been proposed as a therapeutic target [31]. Interleukin-17C has been shown to be tumor-promoting in OC cell models [32]. Poly (ADP-ribose) polymerase 1 is overexpressed in OC and may enhance angiogenesis by upregulating Vascular endothelial growth factor A (VEGF-A) [33]. Genetic variation in Neurotrophic tyrosine receptor kinase 3 (NTRK3) has been associated with prognosis of OC and suggested to predict platinum resistance in OC patients [34]. Superoxide dismutase (Mn), mitochondrial, is highly expressed in OC and has been shown to increase tumor development and metastatic spread [35].

Among the nine proteins (Prostasin (PRSS8), Midkine (Mk), Major epididymis-specific protein E4 (HE4), Adenomedulin (ADM), Matrilysin (MMP-7), Interleukin-1 receptor-like 1 (ST2), Vascular endothelial growth factor A (VEGF-A), Interleukin-6 (IL-6), Hepatocyte growth factor receptor (HGF)) associated with EC in the replication step, several have been discussed in relation to EC. Midkine has been proposed as a serum biomarker for high-risk EC patients, and preoperative serum levels have been shown to correlate with lymph node metastasis [36]. Adrenomedullin expression is upregulated in post-menopausal endometria, and is increased during progression from benign endometrium to type-1 adenocarcinoma [37, 38]. Fatty acid-binding protein (FABP4) has been proposed as a diagnostic biomarker for EC [39, 40] and showed a significant association in our discovery step analysis, but not in the replication step.

Preoperative diagnostic

Our protein panel has a sensitivity = 0.83 and specificity = 0.87 to distinguish between benign tumors and OC stage I–IV. The specificity of our protein panel is somewhat lower than the 0.91 of Overa^®, while the sensitivity is higher than the 0.69 reported for Overa^® [4]. Focusing on OC stage III–IV we have a sensitivity = 0.88 and specificity = 0.92, which is significantly higher than using only CA125 and HE4.

A test for triaging women with adnexal ovarian mass should have better performance than TVU. A recent study showed that TVU in the hands of specially trained gynecologic sonographers can achieve an AUC = 0.92 [41]. However, the performance of ordinary gynecologists is generally lower. For instance, among women with a TVU indication of adnexal ovarian mass that are diagnosed by surgical sampling, 58% have been reported to have benign tumors, 30% have OC stage I–IV and the remaining 15% borderline tumors [42, 43]. Among the 30% of women with OC, 15% have OC stage III–IV. Based on these estimates, clinical diagnosis by surgical sampling has a specificity for OC stage I–IV of 0.30 and for OC III–IV of 0.15. At a minimum sensitivity of 0.96, used as a threshold for a preoperative diagnostic test, our protein panel can distinguish between benign tumors and OC stage III–IV with a specificity of 0.58. This indicates that the number of women with OC stage III–IV among those stratified for surgical sampling based on TVU could be increased from 15% when using TVU to 58% by using the protein panel. For OC stage I–IV, the protein panel show a specificity of 0.48, while the specificity of TVU is 0.30. This correspond to a 50% increase of the specificity. The combined use of both TVU and the biomarker test is likely to give even higher specificity.

Boylan et al. [5] used the ONC Iv2 panel to search for candidate biomarkers for OC. Several of the proteins found to be associated with OC in their study were also on our list of candidates, such as Major epididymis-specific protein E4, Midkine, Kallikrein-11, Folate receptor alpha. Interleukin-6, and Prostasin. Their list of top proteins had consistently higher AUC values than our estimates (Fig. 7). This may reflect differences in design between the studies. Boylan et al. [5] used a single set of cases and controls for identification of proteins, while we used a two-step analysis with two independent set of clinical materials, and in the replication step we also included OC patients from two different hospitals. Our study thus includes several factors that could introduce variation, such as multiple patient cohorts from both the same and different hospitals, different PEA analysis rounds and using both 92-plex and focused PEA panels, and finally the use of a replication step to verify initial findings. Together these factors are likely to reduce the overall performance characteristics of the assay, while at the same time result in more realistic performance indicators.

We also identified a set of protein biomarkers that can be used to distinguish between benign tumors and EC. Biomarkers have previously been described for EC, and good diagnostic accuracy has been reported for e.g. Major epididymis-specific protein E4 (HE4), Growth/differentiation factor 15 (GDF-15), C-Jun-amino-terminal kinase-interacting protein 4 (JIP-4), JNK-interacting protein 4 (SPAG9), Chitinase-3-like protein 1 (YKL-40), Interleukin-31 (IL-31) and Interleukin-33 (IL-33) [11]. Further studies are needed to compare the performance of these candidates with the ones identified in the present study.

Population screening

Our results indicate that the focused protein panel at a specificity of 0.96 has a high sensitivity to distinguish population controls from women with OC stage I–IV and stage III–IV, while lower for stage I–II. To determine the potential of using the protein panel in population screening, further studies are needed of based on samples collected at distinct time-points prior to diagnosis. A recent evaluation of four markers (CA125, HE4, CA72.4, and CA15.3) for OC showed that the performance declined with increasing time between sample collection and time of diagnosis [44]. Serial sampling could enable the use of individual baseline values, and testing women at 3 months rather than 6-12-month intervals using ROCA has been shown to result in better sensitivity and high specificity for detection of early-stage disease [45].

In summary, we have used a two-step strategy to identify plasma proteins that can be used to distinguish benign tumors from OC or EC and for differential diagnostic procedures of women with suspicious pelvic mass. The biomarker panels could be useful in population screening and to identify women in need of TVU examination at a specialized gynecologic ultrasound unit.