Research ArticleCancer

Detection of early pancreatic ductal adenocarcinoma with thrombospondin-2 and CA19-9 blood markers

See allHide authors and affiliations

Science Translational Medicine  12 Jul 2017:
Vol. 9, Issue 398, eaah5583
DOI: 10.1126/scitranslmed.aah5583

Getting a head start on pancreatic cancer

Pancreatic ductal adenocarcinoma (PDAC) has a dismal prognosis due to a lack of diagnostics for detecting early-stage disease. Kim et al. genetically reprogrammed late-stage human PDAC cells to a stem cell–like state, enabling the reprogrammed cells to recapitulate human PDAC progression and revealing secreted candidate markers of early-stage disease. The protein thrombospondin-2 (THBS2) was screened against 746 cancer and control human plasma samples in a multiphase study. The authors report that, THBS2, in combination with the marker CA19-9, boosts detection of the early stages of PDAC in high-risk human populations.

Abstract

Markers are needed to facilitate early detection of pancreatic ductal adenocarcinoma (PDAC), which is often diagnosed too late for effective therapy. Starting with a PDAC cell reprogramming model that recapitulated the progression of human PDAC, we identified secreted proteins and tested a subset as potential markers of PDAC. We optimized an enzyme-linked immunosorbent assay (ELISA) using plasma samples from patients with various stages of PDAC, from individuals with benign pancreatic disease, and from healthy controls. A phase 1 discovery study (n = 20), a phase 2a validation study (n = 189), and a second phase 2b validation study (n = 537) revealed that concentrations of plasma thrombospondin-2 (THBS2) discriminated among all stages of PDAC consistently. The receiver operating characteristic (ROC) c-statistic was 0.76 in the phase 1 study, 0.84 in the phase 2a study, and 0.87 in the phase 2b study. The plasma concentration of THBS2 was able to discriminate resectable stage I cancer as readily as stage III/IV PDAC tumors. THBS2 plasma concentrations combined with those for CA19-9, a previously identified PDAC marker, yielded a c-statistic of 0.96 in the phase 2a study and 0.97 in the phase 2b study. THBS2 data improved the ability of CA19-9 to distinguish PDAC from pancreatitis. With a specificity of 98%, the combination of THBS2 and CA19-9 yielded a sensitivity of 87% for PDAC in the phase 2b study. A THBS2 and CA19-9 blood marker panel measured with a conventional ELISA may improve the detection of patients at high risk for PDAC.

INTRODUCTION

Pancreatic ductal adenocarcinoma (PDAC) is projected to become the second leading cause of cancer death in the United States by 2020 (1). Most of the PDAC patients are diagnosed at an advanced stage of disease, and their tumors are not surgically resectable, contributing to an overall 5-year survival rate of 7% (2). The lack of early diagnostics has made it challenging to develop therapeutics to slow or reverse PDAC (3). The CA19-9 serum marker is used to assess disease progression in PDAC patients (4, 5) but is not recommended for general screening (5, 6) because it is elevated in nonmalignant pancreatic conditions, such as chronic pancreatitis (7), and can produce false negatives in individuals who do not express Lewis blood group antigens (8). Other secreted markers have been reported for PDAC (912) including blood or urine proteins (1315), exosomes (11), microRNAs (16), and epigenetic marks in circulating nucleosomes (17). However, challenges include lack of translation to the clinic, small sample sizes precluding statistical robustness, lack of blinded design, or inappropriate construction of data sets from development to validation (1519). Most biomarkers were discovered in advanced PDAC or cell lines that are not representative of earlier stages, when detection would be most relevant, although recent candidates have been tested or discovered in prediagnostic samples of PDAC (2022). When agnostic biomarker panels are assessed in validation samples, the need to aggregate samples from multiple sources hampers achieving statistical power (23).

We reasoned that proteins released from progressing precursor lesions, such as pancreatic intraepithelial neoplasia (PanIN) stages 2 and 3 (PanIN2 and PanIN3) (24), might provide an effective opportunity for discovering diagnostic markers for PDAC. We previously reprogrammed recurrent, advanced human PDAC cells into an induced pluripotent stem cell (iPSC)–like cell line (25). This iPSC-like cell line (designated as 10-22) can be propagated indefinitely yet preferentially generates PanIN2 and PanIN3 ductal lesions after growing for 3 months as teratomas in immunodeficient mice. The lesions progress to invasive PDAC by 6 to 9 months. Proteomic analysis of conditioned media from 10-22 cell–derived precursor PanINs cultured as organoids, compared to control media from 10-22 cells grown under pluripotency conditions, revealed 107 secreted human proteins specific to the PanIN2/3 organoids (25). Of these, 43 proteins fell into interconnected transforming growth factor–β (TGFβ) and integrin networks for PDAC progression (26, 27), and 25 proteins were within a network for the transcription factor hepatocyte nuclear factor 4α (HNF4α), which also showed an increase in expression (25). Here, we report an analysis of proteins secreted or released from the 10-22 cell–derived PanIN organoids using a phased cancer marker development design that incorporated criteria for prospective specimen collection and retrospective blinded evaluation (PRoBE) (28, 29).

RESULTS

Discovery of candidate marker proteins for PDAC

Of the 107 proteins secreted and released selectively by human PanIN organoids (25), we focused on 53 proteins with a low abundance (≤2 nmol) in the healthy human plasma proteome and RNA sequencing (RNA-seq) databases (table S1) (3032). Enzyme-linked immunosorbent assays (ELISAs) from validated sources (33) were not available for most of these rarely expressed proteins. Of the proteins for which reliable ELISA kits were available and were not implicated as markers in other diseases, we focused on matrix metalloproteinase 2 (MMP2), MMP10, and thrombospondin-2 (THBS2) because they occur in integrated networks for TGFβ and integrin signaling, which drive PDAC progression (25). We investigated these three candidate markers in a screen of human plasma samples. All procedures were performed using a recommended biomarker phased design following the PRoBE criteria (28, 29). De-identified human plasma samples were provided by the Mayo Clinic pancreas research biospecimen repository. We then performed ELISA analyses blinded to disease status and then returned coded data to the Mayo Clinic team for statistical analysis and interpretation.

Phase 1 validation of candidate markers for PDAC

We examined whether MMP2, MMP10, or THBS2 could discriminate between cancer cases (n = 10) and controls (n = 10) with an area under the receiver operating characteristic (ROC) curve (AUC) analysis of the sensitivity and specificity of the markers. All cancer cases for the phase 1 study were selected to have CA19-9 concentrations equal to or above 55 U/ml. MMP2 was unable to discriminate effectively between cancer cases and controls, and MMP10 signals were undetectable in all plasma samples (Fig. 1A). By contrast, THBS2 exhibited a c-statistic of 0.76 considering all cases versus controls (n = 10) and 0.886 when considering resectable and locally advanced PDAC (n = 7). Given that human THBS2 has 80% amino acid sequence homology with THBS1, we demonstrated the specificity of each of the reagents in the THBS2 ELISA (figs. S1 and S2).

Fig. 1. Phase 1 validation studies and THBS2 expression in PDAC and other human tumors.

(A) AUC analysis of blinded ELISA data for the proteins MMP2, MMP10, and THBS2 in plasma samples from 10 patients with PDAC at various stages of disease compared to 10 healthy controls. CI, confidence interval. (B) Box plots of THBS2 mRNA expression measured in various human tumors (sample sizes in parentheses) assessed by RNA-seq. Tumors are sorted in order of decreasing median expression of THBS2 mRNA. Of the pancreatic cancer samples from the TCGA database (n = 179), we analyzed only PDAC (n = 134). All expression values are log2(RSEM values = 1)–transformed.

After the phase 1 validation analysis, we performed a mass spectrometry study of the pooled cancer plasma samples (n = 10) and the pooled control plasma samples (n = 10) after the plasma samples were individually depleted of the 14 most abundant plasma proteins (for example, serum albumin). At 5% false discovery rate (FDR), four unique peptides for THBS2 were identified, of which two were from sequences specific to THBS2 and the other two were from sequences that are conserved between THBS1 and THBS2. One of two peptides specific to THBS2 was present at a threefold greater concentration in the cancer plasma pool compared to the control plasma pool, and another peptide specific to THBS2 was detected only in the cancer pool and not in the control pool (table S2, A and B). At 1% FDR, a THBS2-specific peptide was detectable only in the cancer pool (table S2C). Computational analysis of RNA expression data deposited in The Cancer Genome Atlas (TCGA) database (https://cancergenome.nih.gov/) showed that, of all cancers tested, PDAC (n = 134) was second only to mesothelioma for expression of THBS2 mRNA (Fig. 1B; medians are denoted by vertical black bars within the red boxes). Taking together the phase 1 validation study by ELISA, mass spectrometry data, and TCGA RNA-seq data, we concluded that THBS2 merited further study.

Phase 2a validation of THBS2 and CA19-9 markers for PDAC

Further validation of THBS2 was performed on human plasma samples in a phase 2a study (Table 1) that contained CA19-9–negative and CA19-9–positive cases. The median ELISA value for THBS2 at all PDAC stages (n = 81) in the phase 2a group, 29.7 ng/ml, was 12.2 ng/ml higher than that observed in controls (n = 80) (Fig. 2A), consistent with the mass spectrometry data for the phase 1 study. THBS2 exhibited a c-statistic of 0.842 for all PDAC samples compared to controls (n = 161; Table 2, all PDAC stages). In the same sample set, CA19-9 had a comparable c-statistic of 0.846 for all PDAC samples compared to controls (Table 2, all PDAC stages).

Table 1. Demographic and clinical characteristics of patients.

Continuous variables (age, body mass index, and CA19-9 concentration) are presented as mean (SD); categorical variables (male gender, personal history of diabetes, and stage of disease) are presented as frequency (%). IPMN, intraductal papillary mucinous neoplasm; PNET, pancreatic neuroendocrine tumor.

View this table:
Fig. 2. THBS2 and CA19-9 concentrations in plasma samples from patients with PDAC versus healthy controls.

(A and B) Scatter plots of THBS2 concentrations in plasma samples from patients at all stages of PDAC versus controls for the phase 2a (A) and phase 2b (B) validation studies. (C and D) ROC curves for THBS2, CA19-9, and THBS2 + CA19-9 concentrations in plasma samples from patients with all stages of PDAC versus healthy controls for phase 2a (PDAC, n = 81; controls, n = 80) (C) and phase 2b (PDAC, n = 197; controls, n = 140) (D) studies. (E and F) Scatter plots showing THBS2 and CA19-9 concentrations in plasma samples from patients with all stages of PDAC versus healthy controls for phase 2a (E) and phase 2b (F) studies.

Table 2. AUC calculations of ELISA results for PDAC at different stages, bootstrapped (1000 repetitions) at 95% CIs.
View this table:

The data for the THBS2 analyses were reproducible using three different lot numbers for ELISA kits testing the same subset of phase 2a human plasma samples over a 2-year period, with an average 10% coefficient of variation (CV) across the samples (fig. S3). The samples included four plasma samples that were refrozen and thawed twice and three plasma samples that were refrozen and thawed three times. We concluded that the THBS2 assay was robust from the point of view of differences in plasma sample handling and assessment.

To determine whether CA19-9 and THBS2 together could constitute a more discriminatory panel than either marker alone, we performed logistic regression to estimate the combined probability of their discriminatory ability. A combination of CA19-9 and THBS2 for all cases versus controls with the phase 2a data yielded a c-statistic of 0.956 (95% CI, 0.93 to 0.98) (Fig. 2C, all PDAC stages, and Table 2), indicating the utility of the two-marker panel.

Phase 2b validation of THBS2 and CA19-9 markers for PDAC

We performed an independent phase 2b validation study (see Table 1 for specimens) with an increased sample size. We accomplished temporal validation (18) because the phase 2b analysis was conducted more than 1 year after the phase 2a study. The distribution of THBS2 values across the phase 2a and 2b studies is shown in fig. S4, and the range and median values of THBS2 and CA19-9 are shown in table S3.

The c-statistics for CA19-9 and THBS2 alone, 0.881 and 0.875, respectively, slightly improved with the larger sample size of the phase 2b study (n = 337), compared to the phase 2a study (n = 161), and the combination of the two markers yielded a c-statistic of 0.970 (95% CI, 0.96 to 0.98) (Table 2, all PDAC stages; Fig. 2, B and D). With regard to the distribution variability, the 75th percentile of the control values fell below the 25th percentile of the case values. Furthermore, the 95th percentile of the controls fell below the median measure observed in the case samples. The fact that 50% of the case values exceeded 95% of the control values was likely driving the AUC we observed for THBS2 with regard to being able to discern between cases and controls. We compared individual and combined marker performance in the phase 2a and 2b studies at the stages of resectable PDAC (stages I and II) and locally advanced and metastatic PDAC (stages III and IV). Notably, the combination panel of CA19-9 and THBS2 performed well across all stages of PDAC (Table 2).

More detailed analysis of the distribution of ELISA signals provided insight into how the combination of CA19-9 and THBS2 performed. As observed in the scatter plots in Fig. 2 (E and F), various cases (red plus) had essentially zero CA19-9 signal (that is, along the bottom of the plot), consistent with the likelihood that they were from PDAC patients who were Lewis antigen–negative; however, many of these cases had elevated THBS2 concentrations. Similarly, several cases exhibited THBS2 concentrations that overlapped with the upper range of the group of controls, and these cases exhibited high CA19-9. Thus, the two markers appeared to be complementary in their ability to detect PDAC.

Although stages I, IIA, and IIB are classified as resectable tumors in the sixth edition of the American Joint Committee on Cancer Pancreatic Cancer Staging System (34), only stages I and IIA are considered to be early. Therefore, we directly compared the AUCs for combinations of stages: I + IIA + IIB + II (unspecified), I + IIA + II (unspecified), and I + IIA in our phase 2a and 2b studies. The AUC and 95% CI values were comparable for the two-marker combination across these three subsets, indicating that the exclusion of the questionable early-stage IIB samples had limited impact on marker performance (table S4).

We evaluated the relationship between THBS2 plasma values and age, sex, and the presence of diabetes mellitus in the cohort (table S5, A and B). We observed no apparent associations between these parameters for any of the diagnosis groups of PDAC adenocarcinoma stages I/II, adenocarcinoma stages III/IV, pancreatitis, intraductal papillary mucinous neoplasm, insulinoma (islet cell), and healthy controls. Given the overall lack of associations, we did not include any of these factors as adjustor variables in subsequent modeling analyses.

Establishing a provisional cutoff plasma concentration for THBS2

To determine a THBS2 plasma concentration to use as a cutoff point for discriminating healthy controls from PDAC cases in the clinic, we first considered the distribution of THBS2 values based on the 230 healthy controls from our combined phase 1, 2a, and 2b studies. From this distribution, we chose six cutoffs that represented a range of approximate false-positive rates (FPRs) from 0 to 5%. These cutoffs were then evaluated for their sensitivity in detecting PDAC in the phase 2a and 2b plasma samples. As seen in Table 3 for the phase 2b study, a concentration of THBS2 at or above 42 ng/ml detected about half of the PDAC cases (sensitivity) with 99% specificity. Combining the conventional CA19-9 cutoff of ≥55 U/ml and a cutoff of 42 ng/ml for THBS2 in the phase 2b samples, we observed 98% specificity and 87% sensitivity (Table 3).

Table 3. THBS2 concentration cutoff points based on percentiles of distribution in control plasma samples.
View this table:

Testing the THBS2/CA19-9 panel against other benign pancreatic conditions

When considering all PDAC cases versus chronic pancreatitis (phase 2a, n = 109; phase 2b, n = 252), c-statistics including CA19-9 increased from 0.774 or 0.816 (alone) to 0.842 or 0.867 (with THBS2) with the phase 2a or 2b data, respectively (Table 4 and Fig. 3A). The THBS2/CA19-9 panel was able to discriminate all PDAC cases tested (stages I to IV) versus intraductal papillary mucinous neoplasms (n = 312) with a c-statistic of 0.952 (Table 4 and Fig. 3B). Thus, the THBS2/CA19-9 panel was able to distinguish PDAC from intraductal papillary mucinous neoplasms and helped to distinguish PDAC from pancreatitis compared to CA19-9 alone.

Table 4. AUC calculations of ELISA results for patients with all-stage PDAC versus patients with benign pancreatic diseases, bootstrapped (1000 repetitions) at 95% CIs.
View this table:
Fig. 3. THBS2 and CA19-9 concentrations in plasma samples from all-stage PDAC cases versus benign pancreatic disease cases.

(A to D) ROC curves for THBS2, CA19-9, and THBS2 + CA19-9 concentrations in plasma samples from patients with PDAC in the phase 2b study (n = 197) versus pancreatitis (n = 55) (A), PDAC versus intraductal papillary mucinous neoplasm (IPMN) (n = 115) (B), PDAC versus PNET (n = 30) (C), and PNET (n = 30) versus healthy controls (n = 140) (D) are shown.

THBS2 lacked the ability to discriminate between all PDAC cases and PNETs and hindered, rather than enhanced, the c-statistic of CA19-9 (Table 4 and Fig. 3C). Considering a lack of markers available for PNETs and the poor performance of THBS2 in the discrimination of PDAC from PNETs, we examined whether THBS2 could discriminate PNETs (n = 30) from healthy normal controls (n = 149). CA19-9 alone did not discriminate PNET samples, as previously reported (35). However, THBS2 alone could discriminate PNETs from healthy normal controls with a c-statistic of 0.751 (Table 4 and Fig. 3D).

PDAC can result in obstructive jaundice that can confound plasma assays (20, 36). Of the 288 adenocarcinoma cases included in these studies, we retrieved clinical total serum bilirubin information for 279 cases (96.9%) (table S6A). Of the 279 samples with this information, 70 (25.1%) were inferred to have obstructive jaundice based on total bilirubin concentrations of ≥3.5 mg/dl. Slightly lower median CA19-9 concentrations (208.5 U/ml versus 220 U/ml) as well as elevated median THBS2 concentrations (56.4 ng/ml versus 33.0 ng/ml) were observed in PDAC subjects with jaundice when compared to those without jaundice, indicating that obstructive jaundice influenced both CA19-9 and THBS2 concentrations. Yet, 14 of 55 (25%) PDAC patients with normal CA19-9 and without jaundice had elevated THBS2 concentrations (≥42 ng/ml) (table S6B). Also, 8 of 13 (62%) patients with normal CA19-9 and with jaundice showed elevated THBS2 concentrations (≥42 ng/ml) (table S6B). Therefore, the THBS2 concentration in the plasma identified a subset of nonjaundiced adenocarcinoma cases with normal CA19-9 concentrations. Furthermore, stratifying the marker panel performance by overall PDAC or PDAC without jaundice, versus controls, in the phase 2a and 2b studies affected the AUCs by less than 0.01 (table S6C). Because of limited availability of benign biliary disease samples, we did not compare THBS2 and CA19-9 concentrations between benign biliary disease, nonjaundice PDAC, and jaundice PDAC.

Cross-validation of THBS2 measurements in different laboratories

After these analyses, an independent biomarker development laboratory at the University of Pennsylvania tested a subset of the phase 2b plasma samples for THBS2 concentrations. Thirty-eight samples were randomly selected to cover the entire range of THBS2 concentrations, focusing on those around the cutoff value. The samples were de-identified and provided without communication other than the manufacturer’s instructions for the ELISA and Materials and Methods. The ELISAs for THBS2 were performed more than a year after the original study and with reagents with different lot numbers. The THBS2 concentrations in the original and cross-validated assays were highly concordant and yielded Pearson and Spearman correlation coefficients of 0.95 and 0.968, respectively (fig. S5A).

We noticed that the THBS2 signals were slightly lower in the cross-validated data, including those for the normal human control plasma samples used on each plate. The original studies yielded an average value of 17 ng/ml for the normal control plasma samples, whereas the cross-validation study yielded a value of 13.25 ng/ml for the normal control plasma samples. The lower overall values of unknowns caused 4 of the 38 samples that were just over the cutoff of 42 ng/ml to fall below the cutoff (table S7A).

To accommodate operational differences, we created a scalar where our original cutoff of 42 ng/ml to detect cancer was divided by the original average value of 17 ng/ml (the normal plasma control value) to yield a scalar cutoff of 2.47. We therefore divided the THBS2 result for each unknown in the cross-validation study by the value (13.25 ng/ml) of the normal control plasma, where the cutoff value would be 2.47 times the control value. Scaling did not affect the correlation coefficient (fig. S5B). With the data scaled in this fashion, two samples that were below the cutoff of 42 ng/ml in the original samples were now above the cutoff in the cross-validation data (table S7B). Thus, although the scaling method improved the outcome of the cross-validation assay, careful calibration was needed to ensure consistency in the assay results over time and with different batches of reagent, once a cutoff for clinical practice was determined.

THBS2 expression in different stages of human PDAC

We sought to determine the cells expressing THBS2 in a total of 42 cases of human PDAC and 4 cases of incidental PanIN and intraductal papillary mucinous neoplasms by immunohistochemistry. All 42 cases of PDAC and all 4 cases of incidental PanIN/intraductal papillary mucinous neoplasms exhibited detectable THBS2 (Fig. 4, fig. S6, and table S8). Two different antibodies detected THBS2 in PanIN2 epithelia found incidentally in the PDAC tumor, but little THBS2 was found in PanIN1 epithelia (Fig. 4, A and B). Both antibodies also detected THBS2 in stage II and stage III PDAC. A 10-fold excess of peptide specific to the second antibody blocked staining by this antibody (Fig. 4, C to K). Epithelial cells, but not stromal cells, predominantly stained positive for THBS2 expression in PanIN/intraductal papillary mucinous neoplasm tissue (four of four) (Fig. 4, A and B, and fig. S6, B and C). In PDAC tumor tissue, 32 cases were labeled with THBS2 in epithelial cells and 21 cases were labeled in both epithelial and stromal cells, and in 8 cases of poorly differentiated PDAC tissue, the staining was mostly in stromal cells (table S8).

Fig. 4. Expression of THBS2 in human PanIN tissue and PDAC tumor tissue.

(A and B) Representative THBS2 immunohistochemistry analysis of incidental PanIN1/2 tissue derived from the head and neck of a pancreas from a patient with pancreatic periampullary cancer using two different antibodies (Ab). The arrows indicate that PanIN2 tissue stained positive for THBS2; dotted arrows indicate weak or negative staining of PanIN1 tissue. (C to K) THBS2 expression, designated by arrows, was also confirmed in stage II (C to E) and stage III (F to K) PDAC pancreatic cancer tissue arrays. Competitive assays were performed for antibody #2 by preincubating the antibody with a 10-fold excess of antigen peptide (E, H, and K) to confirm target specificity. Brown color indicates THBS2 staining, and blue color indicates hematoxylin nuclear staining. THBS2 was detected in the epithelial cells of noninvasive lesions (PanINs and intraductal papillary mucinous neoplasms) and poorly differentiated PDAC tissue as well as in fibroblasts in invasive PDAC tissue (see table S8 and fig. S6).

DISCUSSION

With a 5-year survival rate for patients with stage I PDAC being at least four times the overall survival rate for PDAC (34, 37), the THBS2/CA19-9 marker panel may help to detect early-stage tumors that are resectable and should improve the prognosis of PDAC. The performance of THBS2 in early-stage cancer may be a consequence of our discovery that it is secreted or released from human precursor PanIN organoids (25), reflecting the ability of the iPSC-like 10-22 PDAC cell line to recapitulate human pancreatic cancer progression. Although most patients with PDAC are diagnosed at advanced stages, the time from the occurrence of the initiating mutation to the birth of PDAC founder cells (38) can be a decade, suggesting that there may be a time period to identify progressing disease before PDAC can be clinically imaged. PanINs have been identified in the pancreas up to 10 years before the development of infiltrating PDAC (39), underscoring the importance of early diagnosis. However, we note that PanINs can also be observed in the absence of PDAC.

We propose that high specificity outweighs considerations of increased sensitivity because of heightened anxiety in patients over suspected pancreatic cancer plus the costs of subsequent diagnostic evaluation. We found that, with a THBS2 concentration cutoff of 42 ng/ml, THBS2 could discriminate PDAC patients from healthy primary care controls with a specificity of 99% (1% FPR) and a sensitivity of 52%. Impressively, combining CA19-9 (>55 U/ml) with THBS2 (>42 ng/ml) showed a specificity of 98% and a sensitivity of 87% in our larger phase 2b study. An important strength of this study was the ability to use large defined plasma samples obtained from a single institution that followed standardized processing protocols (23).

Decreased THBS1 concentrations by mass spectroscopy analysis have been reported in plasma samples from PDAC patients and in samples obtained before the cancer diagnosis (20, 40). Jenkinson et al. (20) found that reduced concentrations of THBS1 occurred in PDAC patients with diabetes, but not in PDAC patients without diabetes. In contrast, we observed that elevated concentrations of THBS2 were associated with PDAC, but we did not find an association of elevated THBS2 concentrations with diabetes mellitus, age, or sex (table S5, A and B). THBS1 and THBS2 share 80% of their protein sequence but have diverged in function and in their genetic regulation (41, 42). We showed that elevated THBS2 did not correspond to THBS1 in PDAC. First, the antibodies in the ELISA kit used for our study had negligible cross-reactivity with THBS1 (figs. S1 and S2). Second, we confirmed by mass spectrometry that the peptides specific to THBS2 were more abundant in cancer patient plasma samples than in plasma samples from normal healthy controls (table S2). Thus, elevated THBS2 concentrations in PDAC were independent of THBS1 concentrations reported in the literature.

THBS2 is a glycoprotein that may be an angiogenesis inhibitor, and mutation of the mouse TSP-2 gene increases susceptibility to cancer (43). We found that THBS2 antigen is expressed in normal pancreas cells, but the baseline concentration of THBS2 is very low in normal human plasma by both mass spectrometry and ELISA. We found that THBS2 antigen is robustly expressed in PDAC tumor tissue, perhaps concordant with the poor vascularization associated with PDAC, and is also increased in the plasma of PDAC patients. THBS2 is down-regulated in gastric cancer cells (44). Further work is needed to understand how the release of THBS2 into the plasma is increased in the patient cohorts studied here.

A group of scientists has initiated the STARD (Standards for Reporting of Diagnostic Accuracy) with guidelines to improve the reporting of diagnostic accuracy (45). It will be useful to follow these standard guidelines in the clinic by reporting imprecision as the CV (%CV) and precision as 95% CIs near clinical decision points obtained by repeating the test over several independent days. Also, to reduce even small differences in the assay occurring between different laboratories, presenting the likelihood ratio with 95% CIs along with specificity and sensitivity at several cutoff points is recommended. Our cross-validation study was an initial attempt to address these issues, and more work is needed for a determination of clinical decision points with confidence.

The combination of THBS2 and CA19-9 improved the discrimination of patients with PDAC from those with chronic pancreatitis. Longitudinal studies are needed to determine whether a subset of the pancreatitis patients who scored positive for THBS2 but were clinically assessed to be PDAC-negative harbored early-stage PDAC. Likewise, larger studies are needed to determine the effectiveness of THBS2 for diagnosing PNETs where CA19-9 is not applicable, and other cancers showing high THBS2 mRNA expression (Fig. 1B). Further research with larger numbers of pancreatic cancer cases without jaundice as well as patients without cancer but with jaundice will be necessary to quantify the value of THBS2 and CA19-9 for detecting nonjaundice pancreatic cancer.

There are limitations to our study. The prevalence of PDAC in different populations affects the positive predictive value (PPV) and negative predictive value (NPV) for determining the utility of a biomarker in a population. The PPV is the probability that subjects with a positive screening test have the disease, and the NPV is the probability that subjects with a negative screening test do not have the disease. Given the low prevalence of 4 to 12.4 cases of pancreatic cancer per 100,000 in the general population (https://seer.cancer.gov/statfacts/html/pancreas.html), our marker panel with a combined 98% specificity and 87% sensitivity would have a PPV of 0.002, with an NPV of 1.0 (2). Yet, when viewed in terms of the 1.5% lifetime risk of PDAC in the general population (2), the PPV becomes about 0.4 with an NPV of 0.99. For patients older than 55 years who are newly diagnosed with diabetes (46), with a prevalence of 1% in the general population for PDAC, the PPV is 0.31 and the NPV is 1.0. For first-degree relatives of PDAC patients and smokers in the general population, each group with a lifetime risk of 3.75% (47), the PPV is 0.63 and the NPV is 0.99. For carriers with relevant germline mutations (in aggregate, including BRCA1, BRCA2, CDKN2A, and PALB2), the lifetime risk is 40% (47), the PPV rises to 0.97, and the NPV is 0.92. On the basis of these considerations, we suggest that the THBS2/CA19-9 marker panel could serve as a low-cost, nonintervention screening tool in asymptomatic individuals who have a high risk of developing PDAC (3, 47, 48) and also in patients who are newly diagnosed with diabetes mellitus that developed as a result of pancreatic injury (49), but not in the general population.

Another limitation of our work is that our histological analysis of THBS2 expression at different stages of pancreatic cancer was limited by the portion of tissue available from each of the resections. Furthermore, it is unclear how expression of THBS2 in cells under normal or pathological conditions may relate to the extent to which the protein is secreted or released into the plasma and how stable it is in the plasma. Also, further work is needed to refine clinical decision points for high-risk individuals, to determine the panel’s utility for detecting earlier-stage progression to PDAC, and to determine the specificity for pancreatic cancer versus cancers of other tissue types and other disease states.

MATERIALS AND METHODS

Study design and populations

All procedures were performed using a recommended biomarker phased design following the PRoBE criteria (28, 29). De-identified human plasma samples from the Mayo Clinic pancreas research biospecimen repository were shipped to our laboratory, where ELISA analyses were performed blinded to disease status, and then coded data were returned to the Mayo Clinic team for statistical analysis and interpretation.

Collection of plasma samples was approved by the Mayo Clinic Institutional Review Board (IRB). After rapid case finding (50) and informed consent, participants with PDAC provided venous blood samples before initiation of cancer therapy. Samples were frozen at −80°C until used. Similarly, blood samples were obtained from the Mayo Clinic through primary care (healthy controls) and gastroenterology clinics (participants diagnosed with chronic pancreatitis, intraductal papillary mucinous neoplasms, and PNETs). An aliquot of the 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.

Exploratory set (phase 1). Plasma samples from 20 non-Hispanic Caucasian subjects recruited at the Mayo Clinic included 10 healthy primary care controls and 10 [6 early stage (I/II) and 4 late stage (III/IV)] patients with clinically or histologically proven PDAC. All cancer cases for phase 1 were selected to have CA19-9 concentrations above 55 U/ml.

Validation set (phase 2a). Plasma samples from 189 non-Hispanic Caucasian subjects recruited at the Mayo Clinic included 81 (58 early stage and 23 late stage) patients with clinically or histologically proven PDAC, 80 healthy primary care controls, and 28 patients with a personal history of chronic pancreatitis; patients with hereditary pancreatitis were excluded given their increased risk for PDAC. The controls were matched to the cases by age and sex. About 15% of the healthy controls self-reported a personal history of diabetes.

Validation set (phase 2b). Plasma samples collected from 537 non-Hispanic Caucasian subjects recruited at the Mayo Clinic included 197 (88 early stage and 109 late stage) patients with clinically or histologically proven PDAC, 140 healthy primary care controls, 115 patients with intraductal papillary mucinous neoplasm without PDAC, 30 patients with PNET, and 55 patients with a self-reported personal history of chronic pancreatitis; patients with hereditary pancreatitis were excluded. About 11% of the controls self-reported a personal history of diabetes.

Measurement of markers in human plasma

After the ELISAs for the phase 1 study were completed, the remaining PDAC (n = 10) and control samples (n = 10) were each separately depleted of abundant serum proteins by filtration and then high-performance liquid chromatography using a Seppro IgY14 LC10 column (Sigma-Aldrich). The resulting 10 samples of cancer plasmas, depleted of abundant proteins, were pooled separately from a pool of the controls, and the two pools were subjected to two-dimensional strong cation exchange (SCX) chromatography/tandem mass spectrometry analysis as previously described (51). In brief, an SCX tip column was made with a 200-μl tip packed with 20-μl PolySULFOETHYL resin (Nest Group). The SCX tip was prewashed with buffer B [500 mM KCl, 10 mM NaH2PO4, and 30% acetonitrile (pH 2.6)], followed by equilibration with buffer A [10 mM NaH2PO4 and 30% acetonitrile (pH 2.6)]. The lyophilized digested peptides (100 μg) were reconstituted in 50-μl buffer A. The reconstituted digested peptide solution was loaded into the SCX tip column twice, followed by washing with 50-μl buffer A. All flow-through fractions were combined (“flow-through”). The following 100-μl KCl concentration buffers, made by mixing the different proportions of buffer A and B, were used to successively wash the column: 30, 40, 50, 60, 70, 85, 100, 150, and 500 mM. A total of 10 fractions were dried and desalted using the homemade C18 Stage Tips. About 3 μg of the digested peptides was injected into a 75-μm inner diameter × 25-cm C18 column with a pulled tip. EASY-nLC 1000 was run at a flow rate of 300 nl/min for a 180-min gradient. Online nanospray was used to spray the separated peptides into an Orbitrap Fusion Tribrid mass spectrometer (Thermo Electron). The raw data were acquired with Xcalibur, and pFind2.8 software was used to search UniProt Human database. A 5% FDR for the protein spectrum measurement was used initially to filter the peptide search results. Table S2 shows the results for a total of four THBS2 peptides that were detected in the pooled cancer plasma samples. The pFind2.8 search engine revealed two unique peptides in each of the pooled plasmas; the cancer pool had a peptide specific to THBS2 (VCNSPEPQYGGK) and a peptide shared between THBS1 and THBS2 (NALWHTGNTPGQVR), and the normal pool had a peptide specific to THBS2 (TRNMSACWQDGR) and a peptide shared between THBS1 and THBS2 (FYVVMWK) (table S2A). We quantified THBS2 levels in each of the pooled plasmas by measuring the area under the peptide signals based on mass and retention time in original MS1 and MS2 windows. We then normalized the total spectral counts in each sample (table S2A).

To more stringently assess THBS2 peptide levels, we searched with 1 and 5% FDR settings against the UniProt Human database (89,796 entries in total) with the updated pFind3.0 search engine (52). Search parameters were set for a precursor mass tolerance of ±7 parts per million, fragment mass tolerance of ±0.4 Da, trypsin cleaving after lysine and arginine with up to two miscleavages, carbamidomethyl (C)/+57.021 as the fixed modification, and acetyl (protein N-term)/+42.011, deamidated (NQ)/+0.984, and oxidation (M)/+15.995 as the variable modifications. The target-decoy approach was used to filter the search results, in which the FDR was less than 1 or 5% at both the peptide and protein level. At both 1 and 5% FDR, the peptide specific to THBS2 (VCNSPEPQYGGK) and the peptide shared between THBS1 and THBS2 (NALWHTGNTPGQVR) were seen in the cancer pooled sample, consistent with the original pFind2.8 search (table S2, B and C). However, at either 1 or 5% FDR with the pFind3.0 search, no peptides specific to THBS2 sequence were identified, and only peptides shared between THBS1 and THBS2 were identified in the normal pooled sample (table S2, B and C). Therefore, more stringent analysis verified THBS2 sequence in cancer pooled sample.

ELISA kits for human MMP2 (Millipore), human MMP10 (RayBiotech), and human Thrombospondin-2 Quantikine (DTSP20, R&D Systems) were used as described by the manufacturers’ instructions. Duplicate 5-μl plasma samples were diluted 10-fold with calibrator diluent RD5P buffer and all 50 μl used for THBS2. Marker concentrations were determined from standard curves of positive control proteins from the kits with a four-parameter logistic nonlinear regression model using SoftMax Pro Software (Molecular Devices). Normal pooled human plasma (IPLA-N, Innovative Research) was tested in duplicate on each ELISA plate. Across 15 independent ELISA plates, THBS2 in duplicate control samples of commercial normal pooled human plasma ranged between 15 and 21 ng/ml, with a CV of 13%. Also, the inclusion (or exclusion) of occasional plasma samples that were orange or reddish in color, indicating hemolysis, had a negligible impact on the data.

RNA-seq analysis from TCGA

THBS2 mRNA amounts were assessed in TCGA RNA-seq data sets (http://cancergenome.nih.gov/) using the cBioPortal for Cancer Genomics (53, 54). Data were downloaded from the University of California, Santa Cruz Xena data hub and sample IDs curated using the Broad Institute’s Genome Data Analysis Center Firehose. THBS2 mRNA values were estimated by the RNA-Seq by Expectation Maximization (RSEM) algorithm (55) and log2(RSEM + 1)–transformed for Fig. 1B, as parsed and plotted using scripts in Python, R.

Western blot and ELISA for validation of THBS2 ELISA kits for cross-reactivity with THBS1

The recombinant THBS proteins were obtained from R&D Systems and performed Western blot with polyclonal goat anti-THBS2 (detection antibody; working concentration, 0.15 nM) and monoclonal mouse anti-THBS2 (capture antibody; working concentration, 3 nM) to check the cross-reactivity. A detection antibody and a 100-fold molar excess of recombinant THBS1 or THBS2 proteins were incubated in 5% nonfat milk for 30 min at room temperature for competition assay. The incubated solution was centrifuged at 10,000 rpm for 15 min to remove any immunocomplexes before applying onto a polyvinylidene difluoride (PVDF) membrane; a total of two 10-ng proteins were transferred for detection antibodies. For competition assay of capture antibody, a 10-fold molar excess of recombinant THBS1 or THBS2 proteins was incubated with capture antibody in 5% bovine serum albumin for 30 min at room temperature. The incubated solution was centrifuged at 15,000 rpm for 15 min to remove any immunocomplexes before applying onto a PVDF membrane; a total of 10 50-ng proteins were transferred for detection antibodies. The presence of THBS2 in a gel was confirmed by silver staining or reprobing membranes with detection antibody in THBS2-competed membranes. To determine whether the presence of THBS1 interferes with the THBS2 ELISA, we spiked a recombinant THBS1 protein (200 ng/ml) into various concentrations of recombinant THBS2 proteins (0 to 20 ng/ml) or human plasma of wide range of THBS2 in THBS2 ELISA.

Immunostaining of THBS2 in human pancreatic cancer tissue

The pancreatic tumor tissue sections were obtained from US Biomax (catalog #PA1002), and each tissue spot was individually examined by their own pathologists certified according to World Health Organization published standardizations of diagnosis, classification, and pathological grade. Incidental PanIN1/2 tissue section was derived from the head and neck of the pancreas of a pancreatic periampullary cancer patient at the Fox Chase Cancer Center (FCCC) under IRB 09-801 to K.S.Z., and its histology was confirmed by pathologist J. Anderson at FCCC. These tissue blocks do not correspond to plasma samples where we measured plasma THBS2 concentrations. The paraffin-embedded tissues were antigen-retrieved by boiling in citric acid buffer (pH 6.0) after deparaffination. Next, the endogenous peroxidase activity in tissue slides was quenched in hydrogen peroxide solution for 15 min at room temperature. Tissues were blocked with avidin/biotin blocking (Vector Laboratories) for 15 min each, followed by nonprotein blocker (Thermo Fisher Scientific) for 30 min at room temperature. Primary antibodies were applied and incubated for 12 to 16 hours at 4°C. Two primary antibodies for THBS2 were used for our study: goat polyclonal THBS2 antibody (dilution 1:25; sc-7655, Santa Cruz Biotechnology) and rabbit polyclonal THBS2 antibody (dilution 1:100; TA590658, OriGene). It is not clear where TA590658 antibody recognizes and whether it detects secreted THBS2. Yet, sc-7655 antibody can detect both secreted and cytoplasmic THBS2 because it targets the epitopes of 15 to 20 amino acids in length that are located within the first 50 amino acids of the peptide sequence for THBS2, whose signal peptides are located between 1 and 18 amino acids. Only two amino acids of the epitope are overlapped to the signal peptides, and the remaining amino acids are overlapped over the main body of the peptides. A peptide was available for sc-7655 from Santa Cruz Biotechnology, thus the sc-7655 antibodies were incubated with the corresponding peptides in 10-fold excess for 30 min before being applied onto tissue section to confirm the specificity of signals. Also, no primary antibody controls for sc-7655 and TA590658 antibodies were used for negative controls. After washing twice, tissues were incubated with biotinylated anti-goat immunoglobulin G (IgG) or rabbit IgG (Vector Laboratories) at 37°C for 30 min. Tissue sections were conjugated with avidin–horseradish peroxidase by using VECTASTAIN Elite ABC kit (Vector Laboratories) at 37°C for 30 min, followed by developing with DAB (3,3′-diaminobenzidine) peroxidase substrate kit (Vector Laboratories) for peroxidase for 4 to 5 min. Developed tissue sections were stained with hematoxylin for nucleus, dehydrated, and mounted. We confirmed the THBS2 sequence of the peptide by mass spectrometry.

Statistical analysis

The primary comparison for this study was defined as PDAC cases (all stages) versus healthy controls. To explore any relationship between patient demographic information and THBS2 plasma concentrations, we calculated a Spearman correlation coefficient for continuous variables (age) and median expression concentrations for categorical variables (sex, male versus female; presence or absence of diabetes mellitus). On the basis of the data obtained from our phase 1 and 2 studies, we observed no apparent associations between age, sex, diabetes mellitus, jaundice, and THBS2. Given this lack of association, any concerns regarding the potential for confounding were mitigated, and these clinical factors were not included in subsequent multivariable modeling.

Univariate and multivariable logistic regression models were developed to consider each candidate marker (THBS2) alone and combined with CA19-9. The response variable was coded as 1 to indicate the presence of cancer (0 for controls). Candidate markers (THBS2) were entered as continuous variables. CA19-9 was dichotomized as 0 (normal) (<55 U/ml) or 1 (elevated) (≥55 U/ml). The AUC was calculated for each model considered. To assess whether the difference observed between AUCs from the CA19-9 and THBS2 and the CA19-9 alone models was statistically significantly different from 0, we considered a test statistic T (T = AUCCa199 − AUCCa199+THBS2)2/(s2Ca199 + s2Ca199+THBS2) (56), which looks at the difference in AUC between the two models divided by the sum of the variances from the two models. The fact that this test statistic followed a χ² distribution with 1 degree of freedom under the null hypothesis was used to calculate a resulting P value. A bootstrap percentile CI approach was used to estimate a 95% CI for the AUC. This approach resampled the data set (1000 times) and then ran the logistic regression models to calculate the 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.

A similar approach was considered for each of the subanalyses that stratified by stage (early stage and late stage) and other comparison groups (intraductal papillary mucinous neoplasms, chronic pancreatitis, or PNET). A κ statistic was calculated to assess agreement (below cutoff versus above cutoff) in the THBS2 assay results from each of the two independent laboratories in the cross-validation study. Analyses were performed using SAS 9.4 on Linux.

SUPPLEMENTARY MATERIALS

www.sciencetranslationalmedicine.org/cgi/content/full/9/398/eaah5583/DC1

Fig. S1. Validation of cross-reactivity of antibodies enclosed in THBS2 ELISA with THBS1 by Western blot.

Fig. S2. Validation of cross-reactivity and interference of THBS1 in THBS2 ELISA.

Fig. S3. Reproducibility of the ELISA for THBS2.

Fig. S4. Distribution of THBS2 values in phase 2 samples.

Fig. S5. Cross-validation tests, performed 1 year apart, of THBS2 concentrations in the same set of plasmas as determined in different laboratories.

Fig. S6. Representative immunohistochemistry images of THBS2 in human normal pancreas, pancreatitis, and PDAC tissue.

Table S1. List of 53 proteins secreted or released from 10-22 cell, PanIN-stage lesions that are at low abundance in healthy human plasma proteome and RNA-seq databases.

Table S2. Mass spectrometry assessment of THBS2 concentrations in phase 1 plasma samples.

Table S3. Range and median values of THBS2 and CA19-9 in this study.

Table S4. Impact of excluding stage IIB (and unspecified stage II) subjects.

Table S5A. THBS2 values by sex and diabetes mellitus status.

Table S5B. Spearman correlation analysis of age and THBS2 values.

Table S6A. Obstructive jaundice cases in the PDAC cohorts.

Table S6B. THBS2 and CA19-9 values and obstructive jaundice status.

Table S6C. AUC values for CA19-9, THBS2, and combined markers by jaundice status in phases 2a and 2b of PDAC cases versus controls.

Table S7A. Cross-tabulation of normal versus elevated THBS2 values, given a cutoff of 42 ng/ml, for the original cross-validation THBS2 assays (κ = 0.786).

Table S7B. Cross-tabulation of normal versus elevated scaled THBS2 values, given a cutoff of 2.47, for the original and cross-validation THBS2 assays (κ = 0.895).

Table S8. Summary of THBS2 immunohistochemistry in a total of 42 human PDAC and 4 cases of incidental PanIN and intraductal papillary mucinous neoplasm by immunohistochemistry.

REFERENCES AND NOTES

  1. Acknowledgments: We thank S. Vedula for preparing plasma samples for proteomics; H. Collins in the Penn Diabetes Center Biomarkers Core; Z.-F. Yuan for peptide searches using pFind3.0; D. Balli for PDAC data curation from the TCGA database; P. Kanuparthi for immunohistochemistry; R. Vonderheide, A. Rustgi, and J. Becker for comments on the manuscript; and E. Hulme for help with manuscript preparation. Funding: This work was supported by NIH grant no. R37GM36477; the Abramson Cancer Center Pancreatic Cancer Translational Center for Excellence; the Institute for Regenerative Medicine at the University of Pennsylvania; NIH grant nos. P30DK050306 and its cell culture core (to K.S.Z.), U01CA21038 (to G.M.P. and K.S.Z.), and P50CA102701 Mayo Clinic Specialized Programs of Research Excellence in Pancreatic Cancer (to G.M.P.); Department of Defense grant no. BC123187P1 (to B.A.G.); and NIH grant no. P30DK19525 supporting the Penn Diabetes Research Center Bioassay Core. Author contributions: J.K. and K.S.Z. curated biomarker candidates from a previous study. W.R.B. and G.M.P. selected the patient and control populations, analyzed the data, and prepared figures and tables. J.K. and G.D. analyzed TCGA data. J.K. performed the ELISAs blinded to sample identity. A.L.O., K.G.C., and S.C. advised on patient selection, statistics, and data analysis. J.K., X.-J.C., and B.A.G. obtained mass spectrometry data. W.R.B., G.D., X.-J.C., and G.M.P. obtained and statistically analyzed the mass spectrometry data. J.K. performed the immunohistochemistry. Competing interests: K.S.Z. has consulted for BetaLogics/J&J and RaNA Therapeutics. J.K. and K.S.Z. have a patent pending (application no. 61/837,358) for the dual marker panel entitled “Methods for Diagnosing Pancreatic Cancer.” The other authors declare that they have no competing interests.
View Abstract

Navigate This Article