Data-driven analysis of a validated risk score for ovarian cancer identifies clinically distinct patterns during follow-up and treatment.

Background: Ovarian cancer is the eighth most common cancer among women and due to late detection prognosis is poor with an overall 5-year survival of 30-50%. Novel biomarkers are needed to reduce diagnostic surgery and enable detection of early-stage cancer by population screening. We have previously developed a risk score based on an 11-biomarker plasma protein assay to distinguish benign tumors (cysts) from malignant ovarian cancer in women with adnexal ovarian mass.
Methods: Protein concentrations of 11 proteins were characterized in plasma from 1120 clinical samples with a custom version of the proximity extension assay. The performance of the assay was evaluated in terms of prediction accuracy based on receiver operating characteristics (ROC) and multiple hypothesis adjusted Fisher's Exact tests on achieved sensitivity and specificity.
Results: The assay's performance is validated in two independent clinical cohorts with a sensitivity of 0.83/0.91 and specificity of 0.88/0.92. We also show that the risk score follows the clinical development and is reduced upon treatment, and increased with relapse and cancer progression. Data-driven modeling of the risk score patterns during a 2-year follow-up after diagnosis identifies four separate risk score trajectories linked to clinical development and survival. A Cox proportional hazard regression analysis of 5-year survival shows that at time of diagnosis the risk score is the second-strongest predictive variable for survival after tumor stage, whereas MUCIN-16 (CA-125) alone is not significantly predictive.
Conclusion: The robust performance of the biomarker assay across clinical cohorts and the correlation with clinical development indicates its usefulness both in the diagnostic work-up of women with adnexal ovarian mass and for predicting their clinical course.

Introduction

Ovarian cancer is currently the eighth most common cancer among women across the world, with over 300,000 cases and 200,000 deaths per year, and an estimated global incidence of 6.6 per 100,000 women per year[1]. Detection of the cancer is usually late with less than one-third of cases discovered in stage I or II, resulting in poor prognosis with an overall 5-year survival rate of only 30–50%[2]. The overall 5-year survival rate varies greatly depending on tumor stage at diagnosis, and it is close to 90% when the tumor is detected in stage I, but only 20% for stage IV[2]. The precursor states of ovarian cancers have proven difficult to identify. Precise knowledge of the etiology of the cancer could help determine an optimized screening interval in relation to cancer development. However, it has been suggested that serous tubal intraepithelial carcinomas (STIC), the presumed precursor to ovarian high-grade serous carcinomas, develop slowly with up to two decades from the first occurrence of genetic predisposing mutations[3]. Recent molecular evidence from patient material suggests that the developing of ovarian cancer from STIC can occur in a much shorter time, across an estimated timespan of 6–7 years[4,5]. Additional estimates based on tumor size and growth[6] indicate that ovarian cancer can spend over 4 years in situ, or as stage I and II, before progressing to stages III and IV. Today discovery is mainly symptom-driven and women who experience pelvic symptoms are typically examined with transvaginal ultrasound (TVU) or computer tomography, and when these indicate an adnexal ovarian mass, surgery provides the final diagnosis. However, a majority of patients undergoing surgery actually have benign cysts, and more effective and targeted preoperative tools to predict malignancy would reduce unnecessary operations and minimize potential complications and induced premature menopause.

Available biomarkers for ovarian cancer such as MUCIN-16 (CA-125) or WAP Four-Disulfide Core Domain 2 (WFDC2 or HE4) are used as a complement to imaging examinations. MUCIN-16 was introduced as a biomarker for ovarian cancer in 1983[7] and is currently the most important single biomarker for diagnosis and management of ovarian cancer[8]. However, MUCIN-16 alone has low sensitivity for early-stage cancer partly due to the large proportion of false positives linked to relatively benign gynecological conditions such as endometriosis, infections, or pregnancies[8]. Combinations of CA-125 with other biomarkers, including WFDC2, such as in the ROMA Score (Ovarian Malignancy Risk Algorithm), can achieve a sensitivity of up to 75% at a specificity of 90–95%[9,10]. But again, the low sensitivity for detection of early-stage ovarian cancer (stages I and II), and the resulting high cost and risk of over-treatment, prohibits population screening using these biomarkers. Previous studies predicting the risk of malignancy in adnexal ovarian mass using only TVU[11] report sensitivities ranging from 99.7 to 89.0%, with specificities of 33.7 to 84.7%; thus TVU in the hands of specialists can out-perform molecular tests[12]. However, these highly specialized units are scarce, whereas a molecular test could be objectively performed without the need for highly trained experts.

We have previously developed a risk score for separating benign from malignant tumors based on analysis of eleven (11) plasma proteins (MUCIN-16, SPINT1, TACSTD2, CLEC6A, ICOSLG, MSMB, PROK1, CDH3, WFDC2, KRT19, and FR-alpha) plus age[13]. Previously, we used one discovery cohort and two independent validation cohorts to select proteins and evaluated the performance of the models using relative protein concentrations as reported by the proximity extension assay (PEA)[14]. Based on separate validation cohorts the risk score model was finalized with fixed coefficients based on measurements in absolute concentrations[13]. In the previously studied cohort, we achieved a sensitivity of 0.85 and a specificity of 0.93 in separating ovarian cancer tumors in stages I–IV from benign tumors.

In the present study, we validate the performance of the multiplex protein assay in two independent Swedish patient cohorts at time of diagnosis. We also analyze serially collected samples from one of these cohorts to study the development of the risk score during treatment and follow-up of ovarian, endometrial and cervical cancer. The risk score is also analyzed in samples collected from healthy women in a third, cross-sectional Swedish cohort. We show that the performance of multiplex protein assay is robust and that the risk score pattern after diagnosis follows common clinical responses during treatment and relapse/progression and may therefore also be useful in monitoring clinical developments during follow-up.

Methods

Clinical cohorts

The samples were from three separate cohorts: the Biomovca cohort[15], the UCAN biobank[16], and the Northern Swedish Population Health Study (NSPHS)[17]. NSPHS is a population-based health study[17] from which 87 healthy age-matched controls were selected. These samples were collected in 2006. The Biomovca is a clinical multicenter prospective cohort with samples from five secondary care centers and one tertiary care center in the region of Western Sweden[15] and contains samples collected at time of diagnosis. These samples were collected between 2013 and 2016. A detailed clinical description of the characteristics of this cohort has been published before[15]. In the current study, a total of 610 Biomovca samples (Table 1) were analyzed. The UCAN (Uppsala Cancer Cohort) is a prospective cohort collected in the Uppsala-Örebro region consisting of women who were treated at the Akademiska Sjukhuset, Uppsala, Sweden. The UCAN cohort includes samples collected from women at time of diagnosis, as well as serial samples from the same women collected during follow-up and treatment (Table 1). These samples were collected in 2012–2018. Inclusion criteria on diagnoses were epithelial ovarian cancer, fallopian tube cancer, and peritoneal cancer. At diagnosis, the samples from UCAN were characterized as high-grade serous carcinomas (HGSC, 54%), low-grade serous carcinomas (LGSC, 11%), endometroid carcinomas (12%), clear cell carcinomas (5%), combined clear cell and endometroid carcinomas (1%), mucinous carcinomas (3%), non-epithelial ovarian cancer (6%) and carcinosarcomas (5%). Two (2) percent of the samples had no histologic annotation available. The UCAN samples were assigned any of four groups depending on the clinical timepoint. Samples denoted “Primary” were collected when the tumor had been diagnosed but prior to commencing treatment. The category “Treatment ongoing” included samples collected from the commencing of surgery or chemotherapy until the end of the course of treatment. A common treatment span is typically six months, but this could be shorter or longer for individual cases. The category “Response to treatment” included samples collected during follow-up after completed treatment, and comprised women with partial or complete remission, and consequently decreased tumor burden. The fourth category, “Relapse/progression” included samples collected either when cancer was recurring after an initial positive response to treatment or when cancer was progressing. These two types were combined into one group, since they represent increasing tumor burden. Clinical follow-up was limited to 5 years after initial diagnosis. A proportion (N = 48, 40%) of the UCAN biobank samples collected at time of diagnosis were used in the 2nd replication cohort in Enroth et al.[13] (Table 1). These overlapping samples were not used to select the proteins in the model nor to establish the model-coefficients or cut-off for malignancy. The overlapping samples have not been previously analyzed using the quantitative proximity extension assay (PEA) used here.

Table 1

Characteristics of clinical samples.

Cohort	Diagnose	Typea	Nb	Aged	BMId	OC Stagec				MUC16d	WFDC2d
						I	II	III	IV	(U/ml)	(pg/ml)
NSPHS
	Healthy	n.a.	87	63.7 (10.9)	28.3 (5.1)					n.r.	n.r.
Biomovca
	Benign	Prim	443 (5)	50.7 (15.5)	25.9 (4.7)					54.7 (163.9)	64.6 (43.4)
	Borderl.	Prim	31 (1)	55.6 (14.6)	25.6 (4.1)	29	1	1		195.5 (459.7)	79.8 (36.6)
	Ovarian	Prim	136 (1)	62.4 (12.0)	27.9 (14.6)	37	16	68	15	972.1 (1477.3)	528.8 (512.5)
UCAN
	Benign	Prim	55 (3)	57.5 (14.4)	27.0 (5.2)					69.0	n.r
	Ovarian	Prim	66 (2)	61.5 (13.3)	25.7 (7.3)	15 (8)	7 (4)	22 (11)	22 (12)	1216.9	n.r.
	Ovarian	Resp	66 (1)	62.4 (10.5)	25.4 (7.5)	18 (2)	4	26	18	17.6	n.r.
	Ovarian	Rela	33 (1)	70.4 (11.6)	27.2 (8.6)	1	1	16 (2)	15	259.0	n.r.
	Ovarian	Treat	108	60.1 (12.3)	25.4 (12.2)	17 (2)	9 (1)	45 (2)	37 (3)	348.9	n.r
	Cervix	Prim	9	54.2 (9.6)	n.r.					n.r.	n.r.
	Cervix	Treat	28	53.3 (10.7)	n.r.					n.r.	n.r
	Cervix	After	12	57.6 (12.8)	n.r.					n.r.	n.r
	Endo	Prim	19 (1)	73.2 (11.1)	n.r.					n.r.	n.r.
	Endo	Treat	8 (1)	71.7 (10.7)	n.r.					n.r.	n.r
	Endo	After	18	71.2 (11.3)	n.r.					n.r.	n.r.
	Endo	Rela	1	83.5	n.r.					n.r.	n.r
			1120 (16)

n.a. not applicable, n.r. not recorded.

aSee “Methods” for full description of these categories. Prim—primary, Resp—responding to treatment, Rela—progress/relapse, Treat—ongoing treatment, After—after treatment was completed.

bNumber in parenthesis indicate samples that were excluded upon quality control (“Methods”).

cNumber in parenthesis indicate samples that were included in Enroth et al. as 2nd replication cohort.

dIndicated number is group mean (sd).

To study the cancer specificity of the risk score we also included samples from the UCAN biobank collected at time of diagnosis and at one follow-up occasion from 25 women (2 samples from each woman) diagnosed with invasive cervical cancer and 25 women (2 samples from each woman) diagnosed with endometrial cancer (Table 1). The 25 cases with cervical cancer included the following diagnoses: squamous cell carcinomas (60.0%), adenocarcinomas (20.0%), adenosquamous carcinomas (12.0%), glassy cell carcinoma (4.0%), and leiomyosarcoma (4.0%). The endometrial cancers had the following diagnoses: endometrial carcinoma (79.2%), carcinosarcoma (4.2%), clear cell carcinoma (4.2%), endometrial stromal sarcoma (4.2%), mixed tumor, corpus (4.2%) and serous carcinoma (4.2%). In total, 423 samples from the UCAN cohort were analyzed.

The studies and use of the samples have been approved by the appropriate local ethics committees; Biomovca (Gothenburg University, Ref 139-13), U-CAN (Regionala Etikprövningsnämnden, Uppsala, Dnr: 2016/145), and NSPHS (Regionala Etikprövningsnämnden, Uppsala, Dnr. 2005:325 with approval of an extended project period on 2016-03-19). Written informed consent was obtained from all participating individuals.

Protein measurements

A custom 11-plex proximity extension assay (PEA)[14] with a read-out in absolute concentration was used, as in Enroth et al.[13]. Description of the process for combining protein assays into a custom multiplex reaction and the technology used to achieve a readout in absolute concentrations have been described in a white paper[18]. In brief, standard curves with known concentrations are run together with the clinical samples and the standard PEA output is then transformed to absolute concentrations. All the samples were analyzed at the same time and randomized with respect to cohort and diagnosis across plates. All protein measurements were carried out at the Olink Proteomics AB service laboratory in Uppsala, Sweden. Protein concentrations were reported in pg/mL except for KRT19 and MUCIN-16 that were reported in mU/mL. Basic quality control of the data was carried out by Olink Proteomics AB. This procedure flags individual measurements above or below pre-defined limits of quantification. No measurements were reported to be below the limit of detection. A total of 73 measurements, corresponding to 0.6% of the total 12,320 measurements (11*1120 = 12,320), were reported to be as high or above the upper limit of detection and subsequently replaced with the respective upper limit. In the data, 1 of 1120 was replaced for FR-alpha, 7 of 1120 for KRT19, 6 of 1120 for CDH3, 40 of 1120 for MUCIN-16, and 19 of 1120 for SPINT1. Here, 16 (1.4%) of the 1120 samples had one or more of the 11 analytes flagged in the quality control carried out by Olink Proteomics AB and these were removed from further analyses. Seven of these 16 samples (1 malignant, 1 borderline, and 5 benign) were from the Biomovca cohort, and 9 samples (4 malignant, 3 benign, and 2 endometrial cancers) were from the UCAN biobank. No normalization of the protein concentrations was applied.

Risk score calculation

The raw protein concentrations and individual age were log2-transformed and truncated to the ranges observed in the development of the risk score models as described in Enroth et al.[13]. In total, 296 individual data-points out of 13,440 (11 protein and age in 1120 samples) were truncated for the risk score model. The truncated values were used to calculate the risk scores according to the models previously reported in Supplementary Data 4 of Enroth et al.[13].

Statistical analysis

All calculations were done using R[19] (version 4.0.3). All statistical tests for differences were computed using the Wilcoxon ranked sum test and two-sided unless specified otherwise. Calculation of receiver operating characteristics (ROC) and area under curve (AUC) were performed using the ‘pROC’[20] R-package. Evaluation of differences in AUC were done using the DeLong’s test as implemented in the ‘pROC’[20] R-package. Statistical differences in sensitivities and specificities were estimated using a Fisher’s exact test based on counts of false and true positives and negatives. The Cox proportional hazards analyses were conducted using the ‘survival’[21] R-package (version 3.2-11). The bee-swarm plots were produced using the ‘beeswarm’[22] R-package (version 0.4.0). All other figures were generated using custom scripts and basic R-functions.

Table 2

Comparisons of AUC for separating ovarian malignant and benign tumors between the developmental cohort[13] and the two validation cohorts for samples collected at the time of diagnosis.

Benign vs stage	Dev.a	Biomovca			UCAN			UCANb
	AUC	AUC	N	p-valc	AUC	N	p-valc	AUC	N	p-valc
I–IV	0.95	0.92	573	0.24	0.93	116	0.56	0.92	80	0.63
I–II	0.89	0.86	491	0.58	0.79	73	0.27	0.77	61	0.45
III–IV	0.98	0.96	520	0.27	0.99	89	0.18	0.99	71	0.33

aDevelopmental cohort from Enroth et al.[13].

bWith only the non-overlapping samples with Enroth et al.[13] (“Methods”).

cTwo-sided difference compared to the AUC achieved by the models in Enroth et al.[13].

Table 3

Comparisons of AUC, sensitivity, and specificity at time of diagnosis for the risk score and clinical MUCIN-16.

Benign vs stage	AUC			Sensitivity			Specificity
	Mucin-16	Risk score	p-vala	Mucin-16	Risk score	p-valb	Mucin-16	Risk score	p-valb
I–IV (N = 689)	0.89–0.94	0.89–0.95	0.41	0.88–0.96	0.81–0.90	0.037	0.64–0.73	0.85–0.91	4.4 × 10−14
I–II (N = 564)	0.77–0.88	0.78–0.9	0.40	0.72–0.90	0.61–0.82	0.17	0.64–0.73	0.85–0.91	4.4 × 10−14
III–IV (N = 615)	0.95–0.98	0.95–0.99	0.73	0.96–1.00	0.89–0.98	0.10	0.64–0.73	0.85–0.91	4.4 × 10−14

AUC, sensitivity, and specificity are given as 95% confidence intervals. The comparisons were made using both cohorts (Biomovca and UCAN) with the total number of samples as indicated in the first column.

aDeLong’s test.

bFishers’ exact test.