THBS2/CA19-9 Detecting Pancreatic Ductal Adenocarcinoma at Diagnosis Underperforms in Prediagnostic Detection: Implications for Biomarker Advancement

Pancreatic ductal adenocarcinoma (PDAC) is among the most lethal of cancers because it is usually diagnosed at an advanced stage and the tumors are often challenging to surgically resect. With a 5-year survival rate of less than 10% (1), PDAC may become the second leading cause of cancer death in the United States by the year 2030 (2). The late stage of initial PDAC presentation for the majority of patients has engendered a search for diagnostics to detect PDAC or its precursors early enough to be actionable (3). Because PDAC has an annual incidence of only 13 per 100,000 men and women (4), a diagnostic test must have an exceptionally high ability to detect disease (sensitivity ≥60%) and a comparable ability to truly distinguish individuals without PDAC (specificity ≥99%) to be useful, i.e., ROC area under the ROC curve (AUC) ≥0.8 (3). Nonetheless, there is a nearly 2% lifetime risk of PDAC for men and women. This risk is doubled or tripled for individuals who smoke, or who are first-degree relatives of PDAC patients, or individuals over age 55 with recent onset diabetes (6–9), and is highest among those with germline mutations in BRCA1, BRCA2, ATM, CDKN2A, TP53, MLH1, or PALB2 (6, 10). Thus, at-risk populations will be the initial beneficiaries of a PDAC diagnostic. In this study, we tested a biomarker panel that detects PDAC at the time of clinical diagnosis (5) for its performance in blood samples collected prior to a PDAC diagnosis. Our results have implications for likely heterogeneity in progression to PDAC and indicate that better strategies may be needed to develop early-detection biomarkers.

The most common precursor lesion for PDAC is pancreatic intraepithelial neoplasia (PanIN), which has been designated stages PanIN1, PanIN2, and PanIN3 based upon advancing histopathology (6). Although PanIN1 and PanIN2 are considered low-grade lesions with low risk of progression, PanIN3 is considered a high-grade lesion that can lead to PDAC (7). As well, a subset of intraductal papillary mucinous neoplasms (IPMN) increase risk for PDAC (8). Genomic analysis of PDAC cell populations arising from PanIN and IPMNs reveals common driver mutations in KRAS, CDKN2A, TP53, and SMAD4, among other genes (9–16). The sporadic accumulation of such mutations in PDAC tumor samples has suggested that precursor lesions could take many years to progress to cancer (17, 18).

On the other hand, recent studies indicate that PDAC may progress far more rapidly than previously appreciated. Patients with resectable stage I disease are, on average, only 1.3 years younger than patients with inoperable stage IV disease (19). Venous invasion of PDAC cells and access directly into the liver can explain frequent liver metastases (20–22). The challenge in discerning lesions that are benign from those that will progress to cancer can lead to overtreatment of patients being screened for precursors of disease (3, 23). Nonetheless, increased surveillance with endoscopic ultrasound can reveal actionable lesions and increase patient survival (24, 25). Only a fraction of high-risk patients undergo invasive screening; thus, there is an intense effort to develop blood biomarkers to detect PDAC (3, 26).

We previously established a system to detect candidate protein biomarkers that are secreted or released from human PanIN2/3 lesions harboring common PDAC genetic mutations and that will progress to an invasive phenotype. Briefly, we reprogrammed late-stage, recurrent human PDAC cells harboring mutations in KRAS, CDKN2A, TP53, and SMAD4 to an induced pluripotent stem cell–like state, and redifferentiated the cells to early-stage lesions in nonimmune mice that eventually progress to an invasive state (27). Mass spectroscopy analysis of secreted and released proteins from early-stage lesion explants revealed thrombospondin 2 (THBS2) as a candidate biomarker (27). We used a phased study using prospective-specimen collection, retrospective-blinded-evaluation (PRoBE) design (28, 29) that included a phase I plasma discovery phase and two phase II validation studies (5). THBS2 levels were elevated in plasma samples from patients with stage I/II PDAC, with AUCs above 0.8 in the phase II studies. CA19-9, a well-established PDAC biomarker, but lacking sufficient specificity and sensitivity to be clinically useful for early detection (30–34), exhibited AUCs comparable to THBS2 in our phase IIa/IIb studies (5). Yet THBS2 and CA19-9 were partially complementary in detecting PDAC, such that their use as a biomarker panel yielded AUCs for stage I/II PDAC above 0.9 in the phase studies (5). The 99th percentile value of THBS2 (42 U/mL) had a specificity of 98% and a sensitivity of 87%, which yielded positive and negative predictive values that appeared to be in a clinically useful range for at-risk individuals (5, 26, 35, 36), assuming a second stage of follow-up screening to eliminate false positives. In the present study, we performed a prospective, phase III study, adhering to PRoBE criteria (28, 29), to assess the THBS2/CA19-9 panel in blood samples taken prior to a clinical diagnosis of PDAC.

For the phase III analysis, we included incident pancreatic cancer cases whose clinical coding were assessed and found to be histologically PDAC (ICD-9 codes 8500-3 and 8140), and controls matched for age, sex, race, and follow-up duration. The samples consisted of 179 cases and 475 controls from the Prostate, Lung, Colorectal, and Ovarian (PLCO) cohort, and 37 cases and 140 controls from the Mayo Clinic cohort. All subjects had THBS2 results available, and CA19-9 results were available for all but 2 control samples (1 each from PLCO and Mayo).

The PLCO Etiology and Early Markers Studies access committee approved this project (EEMS-2016-0041) in February 2017. The PLCO cohort is a resource available to the scientific community for secondary studies and has been well described https://cdas.cancer.gov/plco/. Briefly, the PLCO cohort was generated from a randomized, controlled trial of screening tests for prostate, lung, colorectal and ovarian cancers. Approximately 155,000 participants were enrolled between November 1993 and July 2001. Data were collected on cancer diagnoses and deaths from all causes that occurred through December 31, 2009. Median follow-up time was 12.4 years. For the phase III study, consented participants with no prior history of pancreatic cancer who provided biospecimens were considered. For each of the 238 participants for whom it was determined through study follow-up to have developed pancreatic cancer within 36 months of blood collection, two subjects (controls), who did not develop pancreatic cancer through a comparable follow-up duration, were matched on gender, age at study entry (<60, 60–64, 65–69, 70+ years), race, and date of blood draw (±365 days). THBS2 (CA19-9) assay results were unavailable for 1 case and 2 controls.

The study was reviewed and approved by the Mayo Clinic Institutional Review Board (IRB). Mayo participants in this study provided informed written consent to research under IRB protocols #354-06 and #356-06. The Mayo Clinic IRB is in compliance with the requirements of FDA regulations 21 CFR Parts 50 and 56 and HHS regulations 45 CFR 46, which are guided by the Belmont Report. In addition, the Mayo Clinic IRB complies with ICH guideline on Good Clinical Practices where they are consistent with FDA and HHS regulations. The Mayo Clinic phase III cohort samples consisted of 70 incident PDAC cases and 140 controls. We used two sources of participants who had not been diagnosed with PDAC, recruited from 2001 to 2017: (i) a prospectively collected research registry of at-risk individuals (family members of pancreatic cancer families; IRB 355-06) and healthy controls ascertained through primary care clinics (356-06), and (2) the Mayo Clinic Biobank (https://www.mayo.edu/research/centers-programs/mayo-clinic-biobank/overview), which consists of 50,000 participants recruited from primary care clinics. From these sources, consented participants who had provided a biospecimen at least 1 month prior to an incident clinical diagnosis of PDAC were identified. For each of these participants, two controls without cancer were matched on age and gender from their respective cohort sources. One incident cancer case was found not to have sufficient sample available and was excluded. CA19-9 assay results were unavailable for 1 control.

Among the Mayo Clinic incident cases, 17 were recruited into the Biospecimen Resource for Pancreas Research (IRB 354-06; previously described in ref. 37). Briefly, this is a clinic-based registry that uses prospective ultra-rapid case ascertainment to recruit patients with PDAC and subsequently, their relatives, with an estimated 80% participation rate. Enrolled participants provided a blood sample at time of PDAC diagnosis, thus enabling analysis of the paired samples for the substudy reported here.

Procedures followed a biomarker-phased design following with PRoBE criteria (28, 29). Deidentified human plasma samples from the Mayo Clinic pancreas research biospecimen repository or from the PLCO (received at the Mayo Clinic) were shipped to the Zaret laboratory, which performed ELISA analyses blinded to disease status, and then returned coded data to the Mayo Clinic team for statistical analysis and interpretation. An aliquot of serum was assayed for CA19-9 at the Mayo Clinic Immunochemical Core Laboratory as recommended by the ELISA kit manufacturer (Cobas/Roche). Demographic and clinical characteristics in each group are shown in Table 1.

The THBS2 ELISA kit from Bio-Techne (R&D Systems DTSP20) was used as per the recommendations of the manufacturer except that to minimize lot-to-lot variation in the kit's diluent buffer RD1-75, we developed and used our own diluent buffer consisting of 0.2 mg/mL porcine intestinal mucosa heparin (Sigma #H3149) in protein-free phosphate-buffered saline with Tween 20 (Thermo Fisher #37573). Manual THBS2 ELISAs were performed on plasma samples with the new diluent buffer as otherwise described previously (5) and automated ELISAs were performed on a Dynex DS2 device in a preclinical biomarker laboratory, working with the Dynex to ensure reproducible and accurate performance (see Supplementary Fig. 1). CA19-9 levels were detected with a clinical ECLIA assay (Roche) at the Mayo Clinic.

The primary analysis evaluated the combination of CA19-9 and THBS2 in the combined PLCO and Mayo Clinic cohorts as an early-detection marker of pancreatic cancer. Due to use of specimens from time of diagnosis for prior studies and a change of assay, previous studies were not used to establish a fixed algorithm for combining these markers. Thus, a multivariable time-to-event Cox proportional hazards model assuming incident cases over a range of follow-up times was fit to assess the time-varying performance of CA19-9 and THBS2 together (38). CA19-9 was dichotomized as 0 = normal (<55 U/mL) or 1 = elevated (≥55 U/mL), and THBS2 as a continuous variable. The AUC was used as a basis for assessing marker performance. Due to the weak performance, a cross-validated estimate was not generated. Secondary analyses were performed within PLCO and Mayo Clinic cohorts. Additional exploratory analyses were performed via logistic regression, together with a bootstrap 95% percentile confidence interval (CI) approach. This approach resampled the data set 1,000 times, estimating the logistic regression models each time to calculate area under the ROC curve (AUC) on each bootstrapped data set to approximate the sampling distribution of the AUC. The 2.5th and 97.5th percentiles from this distribution of AUC values were then used as estimates of lower and upper bounds for the 95% CI for the AUC. Subsets of pancreas cancer patients with samples drawn within 3 years (2 years and 1 year) prior to diagnosis are presented. The THBS2 control distribution 99th percentile value from our prior publication (42 U/mL) is indicated on graphs.

Given the increased sample size of the phase III study, we migrated our previously developed manual ELISA for THBS2 onto an automated ELISA platform and developed a customized diluent buffer, which ensured assay stability across different lots of the procured THBS2 ELISA kit (see Materials and Methods). Using the custom diluent buffer and comparing the original manual ELISA versus the new automated ELISA for THBS2, the automated ELISA exhibited much lower variation among the replicates (Supplementary Fig. S1A) with comparable results (Supplementary Fig. S1B). We therefore performed testing herein with the automated assay.

PLCO samples included consented participants with no prior history of pancreatic cancer who provided biospecimens. For each of the 238 participants who was determined through study follow-up to have developed pancreatic cancer, two controls (subjects who did not develop pancreatic cancer through a comparable follow-up duration) were matched by gender, age at study entry (<60, 60–64, 65–69, 70+ years), race, and date of blood draw (±365 days). THBS2 or CA19-9 results were unavailable for 1 or 2 controls, respectively.

Mayo samples included consented participants with incident PDAC who provided biospecimens at least 1 month prior to a clinical diagnosis were identified. For each of these 70 participants, two controls (participants without cancer) were matched on age and gender. One incident PDAC case ended up having insufficient sample available and was excluded. CA19-9 assay results were unavailable for 1 control.

A subset of 615 healthy controls and 216 (PLCO = 179, Mayo = 37) cases who had PDAC (ICD-O codes 8500-3 and 8140), based upon histologic data provided, were considered in the final analysis. All subjects had THBS2 results available, and CA19-9 results were available for all but 3 control samples.

We evaluated THBS2 in plasma and CA19-9 in serum in 238 prediagnostic case samples and 475 controls from the PLCO, and 69 prediagnostic case samples and 140 controls from the Mayo Clinic. Similar to our approach in the phase I and IIa/IIb studies (5), we used all 615 healthy controls and 216 PDAC samples (Table 1) to directly assess performance of these validated assays in the phase III samples in Table 2, in contrast to other studies that split samples into training and test sets (39). Participants who were incident cases, developing PDAC, were matched to controls who did not develop PDAC across the 15-year time periods in each study (Table 1). CA19-9 data had been obtained from clinical assays of matched serum or plasma samples.

The PLCO phase III cohort samples consisted of 179 incident PDAC cases and 475 healthy controls, where two controls were each matched to a case on age, sex, and time on study in the PLCO Cancer Screening cohort. Scatterplot analysis showed that CA19-9 results on average appear elevated up to a few years prior to a PDAC diagnosis (Fig. 1A), with THBS2 less clearly so (Fig. 1B). For comparison, we plotted the biomarker levels in control samples taken at various time points prior to study exit, illustrating the occasional elevated signals among the controls (Fig. 1C and D). Although CA19-9 and THBS2 exhibited AUCs for incident pancreatic cancer of 0.557 and 0.547, respectively, across the entire PLCO data set and of 0.587 when combined (n = 179; Table 3A), CA19-9 exhibited AUCs of 0.649 and 0.697 at 3 years (n = 64) and 2 years (n = 44) prior to a PDAC diagnosis, respectively, with THBS2 AUCs of 0.536 and 0.568, respectively, and only elevating CA19-9 in combination to 0.713 at 2 years prior to a PDAC diagnosis (Table 3A).

The Mayo Clinic phase III cohort samples consisted of 37 incident PDAC cases and 140 controls, where 2 controls were matched to a case on age, sex, and time on study. These individuals had been recruited into the Mayo Clinic pancreatic cancer family studies or from primary care clinics from 2001 to 2017. An analysis of the Mayo Clinic phase III prediagnostic samples is shown in Fig. 1E and F (cases). Scatterplot analysis of CA19-9 and THBS2 levels showed sporadic positive values, i.e., above the 99th percentile value from phase IIa/IIb studies, and no marked trend toward elevated signals within 1 or 2 years prior to a PDAC diagnosis (Fig. 1E and F). These impressions were confirmed by a statistical analysis, which showed that CA19-9 and THBS2 exhibited AUCs of 0.537 and 0.552, respectively, across the entire data set (Table 3B, n = 37). The combination of the two markers only elevated the AUC to 0.585 (Table 3B). More specifically, for the samples taken 3 years prior to a PDAC diagnosis (n = 16), the markers yielded AUCs below 0.6 alone or in combination, while samples taken 2 years prior to a diagnosis (n = 10) yielded AUCs of 0.596 and 0.548 for CA19-9 and THBS2, respectively, and of 0.665 for the combination (Table 3B).

The combination of the dichotomized CA19-9 and THBS2 in the overall PLCO + Mayo Clinic cohort generated an AUC of 0.572 (0.53, 0.61; Table 3C). Although our work to date has used CA19-9 with a 55 U/mL case versus control cutoff, using CA19-9 levels as a continuous variable yielded an AUC of 0.591 alone and 0.595 with THBS2 (Table 3C). When assessing AUCs for the combined biomarker panel as they occur within 10 years prior to a diagnosis, a modest upward trend was observed starting about 4 years prior to PDAC diagnosis, but it did not reach a clinically useful level (Fig. 2A). More specifically, CA19-9 alone exhibited AUCs of 0.679 and 0.752 at 2 years (n = 54) and 1 year (n = 29), respectively, prior to a PDAC diagnosis; combining the data with those for THBS2 1 year prior did not increase the AUC (0.746; Table 3D).

For a prediagnostic study, it is helpful to demonstrate that a biomarker from a phase II study performs at the time of diagnosis in a phase III study, given the differences in sample collection, storage, and freeze/thaws over time. Unfortunately, matched diagnostic samples of incident PDAC patients in cohorts are rarely obtained or made accessible; such was the case for the PLCO cohort. In the Mayo Clinic biorepository, specimens from before and at the time of diagnosis were available for only 17 patients (Supplementary Table S1). Although THBS2 levels did increase from prediagnosis to time of diagnosis for all but 3 patients, there is no trend of the prediagnostic values increasing before diagnosis (Fig. 2B and C). We conclude that THBS2 exhibited an upward trend at PDAC diagnosis, but did not exhibit a trend of increased levels prior to the diagnosis in the samples studied.

We adhered to the recommended PRoBE criteria for three phases of biomarker validation (Fig. 3A and B; refs. 28, 29), following a biomarker discovery platform to detect proteins that are secreted or released from early-stage PanIN lesions that harbor common PDAC mutations and that will progress to invasive PDAC (27). We found that neither CA19-9 nor THBS2, nor the combination of the two, performed sufficiently in a phase III, prospective study of two different cohorts of samples. The results have implications for future study designs to discover biomarkers for relatively rare diseases, such as pancreatic cancer, and for screening for PDAC.

Strengths of our experimental design include access to plasma samples from the Mayo Clinic and the PLCO Cancer Screening Trial (40); samples were collected 1 to 15 years prior to a PDAC diagnosis; the availability of time-of-diagnosis samples of a subset of the incident Mayo Clinic cases; the development of an automated ELISA assay for THBS2 that reduces variability among replicates, and operator blinding during sample identification and analysis. Factors that could have affected our work were the advanced age of numerous frozen plasma specimens, multiple freeze–thaw cycles (41), and that only a small number of patients had paired prediagnostic and time-of-diagnosis samples, limiting the ability to validate biomarker stability over the time frame. Still, the fraction of THBS2-high patients at the time of PDAC diagnosis in the Mayo cohort was about the same as seen in our phase IIa/IIb studies (Fig. 2B and C). We suggest that pre- and time-of-diagnosis paired samples be used for future phase III studies to validate the utility of the biomarker with the particular sample set. We did not have access to data on the incident PDAC patients to explore biomarker levels with tumor characteristics.

Interestingly, CA19-9 considered across the entire phase III study, i.e., Mayo and PLCO together, performed at the level of modest discrimination (AUC-0.75) in the subset of 29 cases 1 year prior to PDAC diagnosis (Table 3D). In the phase III study with the Mayo Clinic samples, THBS2, CA19-9, and the combination were not able to discriminate PDAC cases prior to a diagnosis. That THBS2 increased in the Mayo Clinic diagnostic samples (Fig. 2B and C), albeit with a small sample size, suggests that the negative results with THBS2 in the Mayo prediagnostic samples are valid. Similarly, THBS2 performed poorly with the PLCO prediagnostic samples and only raised the AUC with CA19-9 (0.697 alone) to 0.713, for 2 years prior to a diagnosis. The CA19-9 discrimination is similar to the AUC of 0.656 found by Nolan and colleagues with a different subset of a PLCO samples (42).

We suggest that the high dynamic range of CA19-9 signals in the PDAC population, compared with the modestly changing THBS2 (Fig. 3B, see phase IIa/IIb), makes CA19-9 a more sensitive indicator of emerging PDAC, though with fewer samples exhibiting elevated signals. Thus, it may be relevant to focus on new biomarkers that, like THBS2, are partially complementary in detecting PDAC, but have a wider dynamic range in affected individuals compared with controls.

The 17 matched pre- and peridiagnostic samples provided a qualitative sense of THBS2 biomarker dynamics longitudinally. Our observations suggest that PDAC progression can be heterogeneous. Indeed, the difference in biomarker concentration over time could be more informative for detecting disease than the absolute level over a general threshold, suggesting that future phase III studies would benefit from serial samples. In addition, our findings underscore the importance of discovering biomarkers in patient samples as close as possible to targeted clinical utility for screening goals.

The most recent recommendations for high-risk patient screening, based on the International Cancer of the Pancreas Screening (CAPS) consortium, is for annual surveillance in the absence of concerning lesions (24, 25). Based on the ability of the THBS2/CA19-9 panel to detect PDAC in early-stage I/II patients (5) but underperformance in the phase III prediagnostic studies described here, including plasma samples at all stages from one source (Mayo Clinic), it is possible that some tumors develop more rapidly than appreciated. If the progression of PDAC is more rapid than generally appreciated, more research is needed on frequency of screening in high-risk individuals.

W.R. Bamlet reports grants from NIH–Pancreas Cancer Detection Consortium U01 during the conduct of the study; in addition, Dr. Bamlet has a patent for Methods for Diagnosing Pancreatic Cancer pending. A.L. Oberg reports grants from the NCI during the conduct of the study. D. Herman reports grants and non-financial support from Roche Diagnostics outside the submitted work. J. Kim reports a patent for U.S. Provisional pending to 62/529/970. G.M. Petersen reports a patent for U.S. Provisional 62/529/970 entitled “Methods for Diagnosing Pancreatic Cancer” pending. K.S. Zaret reports a patent for 62/529/970 pending. No disclosures were reported by the other authors.

S. Udgata: Data curation, formal analysis, and investigation. N. Takenaka: Supervision and investigation. W.R. Bamlet: Data curation and formal analysis. A.L. Oberg: Data curation, formal analysis, and supervision. S.S. Yee: Formal analysis and supervision. E.L. Carpenter: Formal analysis and supervision. D. Herman: Formal analysis. J. Kim: Conceptualization. G.M. Petersen: Conceptualization, resources, supervision, writing–original draft, writing–review, and editing. K.S. Zaret: Conceptualization, resources, formal analysis, supervision, writing–original draft, project administration, writing–review, and editing.

The work was supported by NIH U01 CA210138 to G.M. Petersen and K.S. Zaret; by NIH U01CA208517 to G.M. Petersen; Pancreatic Cancer Action Network grant 17-80-ZARE, funding from the Abramson Cancer Center, Pancreatic Cancer Research Center, and the Institute for Regenerative Medicine at the University of Pennsylvania to K.S. Zaret; and Pancreatic Cancer Action Network grant 17-80-PETE and NIH P50CA102701 Mayo Clinic SPORE in Pancreatic Cancer to G.M. Petersen.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.