Development of a Multiprotein Classifier for the Detection of Early Stage Ovarian Cancer

Abstract

Background: Individual serum biomarkers are neither adequately sensitive nor specific for use in screening the general population for ovarian cancer. The purpose of this study was to develop a multiprotein classifier to detect the early stages of ovarian cancer, when it is most treatable.

Methods: The Olink Proseek Multiplex Oncology II panel was used to simultaneously quantify the expression levels of 92 cancer-related proteins in sera.

Results: In the discovery phase, we generated a multiprotein classifier that included CA125, HE4, ITGAV, and SEZ6L, based on an analysis of sera from 116 women with early stage ovarian cancer and 336 age-matched healthy women. CA125 alone achieved a sensitivity of 87.9% at a specificity of 95%, while the multiprotein classifier resulted in an increased sensitivity of 91.4%, while holding the specificity fixed at 95%. The performance of the multiprotein classifier was validated in a second cohort comprised of 192 women with early stage ovarian cancer and 467 age-matched healthy women. The sensitivity at 95% specificity increased from 74.5% (CA125 alone) to 79.2% with the multiprotein classifier. In addition, the multiprotein classifier had a sensitivity of 95.1% at 98% specificity for late stage ovarian cancer samples and correctly classified 80.5% of the benign samples using the 98% specificity cutpoint.

Conclusions: The inclusion of the proteins HE4, ITGAV, and SEZ6L improved the sensitivity and specificity of CA125 alone for the detection of early stages of ovarian cancer in serum samples. Furthermore, we identified several proteins that may be novel biomarkers of early stage ovarian cancer.

Keywords: early detection; ovarian cancer; protein biomarkers.

Figure 1

Defining proteins that may be sensitive to preanalytical variation. (a) The standardized mean differences (SMDs, i.e., t-statistics) in protein levels between those with and without cancer were compared for Texas (TX) vs. Minnesota (MN). Blue points are proteins that were significantly differentially expressed between institutions (e.g., overexpressed in MN cancers vs. underexpressed in TX cancers). (b) The same plot as in (a), except only the proteins tested to determine their preanalytical variation in Shen et al. [13] are plotted. (c) Shen et al.’s measurement of protein instability (sum of ∆NPX values; higher values indicate more instability) was compared to the p-values for differential effects by institution, where small p-values are evidence of instability. The differential effect by institution p-values for all differentially expressed proteins are in Table S1, along with the mean differences by institution.

Figure 2

Hierarchical clustering of 452 serum samples from Cohort #1 based on 67 proteins (excluding FOLR3). Blue indicates proteins with high levels of expression, while yellow indicates proteins with low levels of expression. Three major clusters were identified. The three color bands at the bottom identify the samples. Band #1, sample type and institution: Light red, 70 early stage ovarian cancer (TX); dark red, 46 early stage ovarian cancer (MN); light green, 275 healthy (TX); dark green, 61 healthy (MN). Band #2, overall sample type: 116 cancer (red) and 336 healthy (green). Band #3, ovarian cancer subtypes: 53 serous (red), 21 endometrioid (blue), 14 clear cell (yellow), 15 mucinous (green), and 13 mixed (brown).

Figure 3

Development of a multiprotein classifier from samples in Cohort #1. (a) Proseek Oncology II NPX values were plotted with the median and 25th and 75th percentiles for the healthy patients (blue) and the 5 subtypes of early stage ovarian cancer (pink) for the four proteins included in the multiprotein classifier. (b) ROC curves for the multiprotein classifier and each of the four individual proteins included in the multiprotein classifier.

Figure 4

Validation of the multiprotein classifier for use with Cohort #2, a second cohort of early stage ovarian cancer samples. (a) Proseek NPX values of the proteins included in the early stage multiprotein classifier between healthy controls and early stage ovarian cancer samples from Cohort #2. The median and 25th/75th percentiles are shown in all plots. (b) ROC curves for the early stage multiprotein classifier applied to the Cohort #2 samples and individual proteins included in the early stage multiprotein classifier.

Figure 5

Validation of the multiprotein classifier with serum samples from women with benign ovarian conditions. (a) Comparison of the NPX values for the proteins included in the multiprotein classifier between healthy controls, benign samples, and the early stage ovarian cancer samples from both cohorts of samples. (b) Predicted cancer risk scores stratified by true cancer status for the early stage ovarian cancer, benign, and healthy control samples from Cohort #1 and Cohort #2. The median and 25th/75th percentiles are shown in all plots.

Figure 6

Validation of the multiprotein classifier with serum samples from women with late stage ovarian cancer. (a) Predicted cancer risk scores with median and 25th and 75th percentiles stratified by true cancer status for late stage ovarian cancer samples (n = 61 cancer samples; n = 88 healthy controls) from Skubitz et al. [12]. (b) ROC curve for the early stage multiprotein classifier applied to the late stage samples.