Improving a Plasma Biomarker Panel for Early Detection of Pancreatic Ductal Adenocarcinoma with Aminopeptidase N (ANPEP) and Polymeric immunoglobin receptor (PIGR)

Abstract

Purpose:

Pancreatic ductal adenocarcinoma (PDAC) is typically detected too late for useful therapeutic interventions; hence we sought blood biomarkers to detect early-stage disease.

Experimental Design:

Using mass spectrometry and enzyme-linked immunosorbent assays on plasma pools from the University of Pennsylvania (Penn) and the Mayo Clinic (Mayo), we identified aminopeptidase N (ANPEP) and polymeric immunoglobin receptor (PIGR) as increased in early stage (Stage I/II) PDAC plasma compared to controls. We tested ANPEP and PIGR, along with prior data for Thrombospondin 2 (THBS2) and CA19-9, in retrospective Phase 2 studies using separate cohorts of PDAC plasmas at different stages vs. healthy or nonmalignant disease controls from Penn (n=135) and Mayo (n=537).

Results:

Comparing healthy controls to Stage I/II PDAC, we obtained area under the Receiver Operating Characteristic curves (AUCs) of 0.78 (0.68–0.86)/0.80 (0.74–0.85) (ANPEP) and 0.81 (0.70–0.88)/0.86 (0.82–0.90) (PIGR) for the Penn/Mayo Phase 2 studies, respectively. In multivariable models, CA19-9/THBS2/ANPEP, CA19-9/THBS2/PIGR and CA19-9/THBS2/ANPEP/PIGR elicited AUCs of 0.94–0.96 for Penn and 0.97 for Mayo. Notably, the four-marker panel elicited AUCs of 0.87 for the Mayo Stage I/II vs. disease controls and 0.91 for Stage I-IV vs. disease controls. At a specificity of 95%, a plasma biomarker panel composed of CA19-9 (≥35U/mL), THBS2 (≥42ng/mL), ANPEP (≥2995ng/mL), and PIGR (≥1800ng/mL) yielded a sensitivity of 91.9% for PDAC Stages I-IV and 87.5% for PDAC Stage I/II.

Conclusions:

Adding ANPEP and PIGR to a plasma biomarker panel of CA19-9 and THBS2 enhances the detection of early stage PDAC when comparing cancer vs. healthy or nonmalignant disease controls. Given the concordance of our data in two retrospective Phase 2 studies, assessments in pre-diagnostic cases are warranted.

Introduction

Due to most cases of pancreatic ductal adenocarcinoma (PDAC) being detected at late stages, beyond potential for effective treatment, PDAC is associated with a high mortality rate, low survival, and poor prognosis. Accounting for 92% of all pancreatic cancers, PDAC originates from the ductal cells of the exocrine pancreas^1,2. More than 80% of patients are ineligible for resection at the time of diagnosis because the cancer has either locally advanced to involve surrounding vascular structures or metastasized. The overall 5-year relative survival rate of less than 13% improves to 44% if PDAC is diagnosed in the early localized stages, but this only occurs in about 14% of all cases^3,4. Hence there is intense research to find markers to enable early-stage detection, which is the subject of the present study.

Given the low incidence of PDAC in the general population, routine screening is only indicated for patients at high risk for disease⁵. Stage I pancreatic cancers exhibit a higher survival rate of over 80% according to the NCI SEER database, if accurately diagnosed and treated⁶. Carbohydrate antigen 19–9 (CA19-9) is widely used to monitor PDAC treatment response in patients with an established diagnosis^7,8 but falls short as a standalone screening tool due to upregulation in benign disease conditions such as pancreatitis and bile duct obstruction^9,10, low sensitivity and specificity for early-stage PDAC, and because CA19-9 levels can be affected by a patient’s genetics^11,12.

Despite these limitations, prior studies have indicated that CA 19–9 can serve as an anchor biomarker for inclusion in early-detection biomarker panels with addition of other protein and genomic biomarkers to enhance sensitivity of early-stage detection¹³. Moreover, a mutant fucosyltransferase (FUT) enzyme genotype test for FUT2 and FUT3 can distinguish individuals unable to express or secrete the Lewis blood group antigens independent of PDAC, enabling personalized reference ranges to interpret CA19-9 levels^11,14. Using a retrospective Phase 2 study design as recommended by Pepe et al.^15–17, we previously showed that elevated plasma Thrombospondin-2 (THBS2) levels could be complementary to false-negative CA19-9 patients, i.e., PDAC patients with non-elevated CA19-9¹⁸; also, that elevated THBS2 levels predict poor prognosis for late-stage patients¹⁹. Various other studies have comparably assessed the utility of THBS2^20–23. Yet in Phase 3 retrospective studies across a population, not longitudinally, CA19-9 and THBS2 were only weakly elevated within a year of a PDAC diagnosis²⁴. In the present study, we used Phase 1 and Phase 2 studies to seek new plasma markers to improve upon the CA19-9/THBS2 panel.

Materials and Methods

Phase 1: Discovery of candidate biomarkers

To ensure the biomarkers selected were robust enough to overcome individual variation, we created plasma pools. Plasma collected from Mayo Clinic and the University of Pennsylvania were selected for pool composition based on diagnosis (control, chronic pancreatitis, or PDAC by stage) and no statistically significant difference was found for age, sex, or smoking status. Following statistical analysis of these variables, plasma were pooled (n = 10 cases per pool; Supplemental Table 1) to create representative samples for healthy control, chronic pancreatitis, early stage PDAC (stage I/II), mid-stage PDAC (Stage III), and late PDAC (Stage IV) from two collection sites: Mayo Clinic (prefix: M.) and University of Pennsylvania (prefix: P.). In addition, 6 cases identified as early stage PDAC (Stage I/II) with low CA19-9 (<35U/mL) from the University of Pennsylvania were pooled and referred to as “early-low” plasma pool to provide representation of individuals with PDAC Stage I/II that may be Lewis antigen null. Nine pooled samples (M./P.Pools) identified as M.HC, M.CP, M.EAR, M.LAT, P.HC, P.CP, P.EAR, P.EAR.LOW, and P.LAT were submitted to mass spectrometry analysis at two institutions. Following high abundant protein depletion and further purification, Washington University of St. Louis (WUSTL) (MTAC@MGI) performed LFQ-mass spectrometry and the Wistar Proteomics and Metabolomics Facility performed TMT-labeled mass spectrometry on technical replicates. M.MID and P.MID (Stage III) pools were not sent for mass spectrometry analysis due to our focus on early stage PDAC, but were utilized for subsequent testing of commercial ELISA kits (Supplemental Figure 1). Biomarker candidates were selected based on 1) a significant difference in expression identified by the mass spectrometry analysis in cancer vs. control plasma (see below), 2) literature search on expression and disease implications, and 3) the ability to test plasma using commercially available materials.

LFQ Mass Spectrometry - WUSTL

Label Free Quantification (LFQ) Mass Spectrometry analyses were performed by the Mass Spectrometry Technology Access Center at McDonnell Genome Institute (MTAC@MGI) at WUSTL (RRID:SCR_027393). Following protein concentration measurement (BCA), 600 ug of total protein from each pooled sample was subjected to abundant protein depletion using High Select^™ Top14 Abundant Protein Depletion Midi Spin Columns (#A36371). Depleted samples were purified by TCA precipitation, then reduced, alkylated, and digested with trypsin according to the facility’s optimized protocol. Digested peptides were desalted on C18 spin columns. To load the same amount of each peptide sample into the mass spectrometry, peptide concentration was measured by the colorimetric peptide assay. All samples were analyzed in technical duplicates randomized across the mass spectrometry run order. We searched the data against a Human database using the MaxQuant (RRID:SCR_014485) search engine and then performed label-free quantification (LFQ) based on the MS1 peptide intensity. Protein was filtered for >1 unique peptide, which yielded a total of 402 proteins, and the obtained LFQ intensities were Log2 transformed. Data were grouped into 9 groups, corresponding to the 9 plasma pools analyzed, and filtered to retain proteins with 2 LFQ values from technical duplicates in at least one group, resulting in 391 proteins. Finally, we carried out a t-test using a permutation-based FDR method²⁵ with a correction threshold of 0.05 for multiple hypotheses to determine significant differences for 13 comparison sets: P.EAR vs. P.HC, P.EAR vs. P.CP, M.EAR vs. M.HC, M.EAR vs. M.CP, P.EAR.LOW vs. P.HC, P.EAR.LOW vs. P.CP, P.EAR.LOW vs. P.EAR, P.LAT vs. P.HC, P.LAT vs. P.CP, P.LAT vs. P.EAR, M.LAT vs. M.HC, M.LAT vs. M.CP, and M.LAT vs. M.EAR. In each of these 13 comparisons, statistically significant proteins were identified using Perseus (RRID:SCR_015753) software²⁵ v2.0.7.0 resulting in a total of 285 unique proteins. The detailed data are available upon request from the corresponding author.

Using the provided output from Perseus, we sought to identify proteins that were upregulated when comparing cancer and control plasma. We looked at each comparison set and identified the total number of significant proteins followed by the number of proteins that exhibited an increase or decrease in cancer vs. control plasma. As seen in Supplemental Table 2.1, comparisons within the Penn plasma pools yielded fewer significant protein hits than the Mayo plasma pools and therefore we focused our analysis on the data from cancer vs. control Mayo plasma for this data set. When comparing M.EAR vs. M.HC and M.CP, 230 and 136 statistically significant proteins were identified, respectively. Further analysis showed that 106 proteins were identified as statistically significant in both M.EAR vs. M.HC and M.CP data. Of these 106 proteins, 36 were increased in early stage PDAC plasma when compared to both healthy control and chronic pancreatitis pools. As seen in Supplemental Table 2.2, both PIGR and ANPEP were significantly increased in Mayo early-stage PDAC pools when compared to both healthy control and chronic pancreatitis pools. PIGR and ANPEP were in the top 20 significantly increased proteins when comparing M.EAR vs. M.HC. For M.EAR vs. M.CP, PIGR and ANPEP were in the top 10 with several common inflammatory genes such as SAA1, SAA2, and CRP outranking one or both (data not shown).

TMT-labeled MS - Wistar

Nine pooled plasma samples (60 ul each) were provided to Wistar Institute’s Proteomics and Metabolomics Facility (RRID:SCR_010211). Each sample was subjected to protein depletion using High Select^™ Top14 Abundant Protein Depletion Midi Spin Columns (#A36371). Duplicate depletion was performed for each sample. Depleted samples were concentrated using a 10K MWCO concentrator and protein concentration was determined using BCA assay. Twenty five ug of each depleted sample was reduced with TCEP, alkylated with iodoacetamide, digested in-gel with trypsin and cleaned up using MiniSpin C18 columns. Each sample was labeled with a specific TMTpro 18-plex reagent. A pilot 1:1 mix was made from all labeled samples, cleaned up with a C18 microspin column and analyzed by LC-MS/MS on a Thermo Q Exactive HF mass spectrometer (RRID:SCR_020545). Sums of each TMT reporter ion intensity were determined (indicative of sample protein amount) and sample volumes were adjusted to achieve a final 1:1 mix. The final 1:1 mixed sample was subjected to high pH fractionation into 18 fractions (5%, 10%, 12%, 14%, 16%, 18%, 19%, 20%, 22%, 23%, 24%, 25%, 26%, 28%, 30%, 40%, 50% and 80% acetonitrile). The 5% fraction was not analyzed as it contained unlabeled TMTpro reagents. The 50% and 80% fractions were combined for analysis.

Each fraction was analyzed by LC-MS/MS on the Thermo Q Exactive HF mass spectrometer using a 180 min LC method. A total of 16 runs were performed. MS/MS data were analyzed using Thermo Proteome Discoverer (RRID:SCR_014477) v2.5. Spectra were searched against the UniProt (RRID:SCR_002380) human database (07/21/2022) and a common contaminant database using Sequest HT. Percolator target False Discovery Rate was set at 0.01 (Strict) and 0.05 (Relaxed). 766 proteins were identified with high Protein FDR Confidence (Experimental q-value<0.01), and 196 proteins with medium Protein FDR Confidence (Exp. q-value between 0.01 and 0.05). Using the same 13 comparisons previously described in WUSTL’s LFQ-labeled MS, abundance ratios were calculated for each comparison. Proteins that had a fold-change greater than or equal to 1.3 with a q-value <0.05 were considered statistically significant in each of the 13 comparisons. As seen in Supplemental Table 3.1, both Penn and Mayo comparisons yielded statistically significant results therefore both plasma sets were used for biomarker candidate selection. The detailed data are available upon request from the corresponding author.

We performed several downstream analyses on this data. First, we compared early stage PDAC pools to their respective healthy control or chronic pancreatitis pool. A total of 43 proteins were present in both M.EAR vs. M.HC and M.CP data with 17 shared proteins found to be increased (>1.3 fold). Among these 17, PIGR is within the top 10 highest abundance ratios for comparisons of both early PDAC vs. healthy control and chronic pancreatitis comparisons for Mayo pools. A total of 42 proteins were identified as statistically significant in both P.EAR vs. P.HC and P.CP data with 9 shared proteins, found to be increased (>1.3 fold). PIGR is found in both data sets and is within the top 5 for statistically significantly increased proteins when comparing P.EAR vs. P.HC with an abundance ratio of ~1.65. Finally, we compared the statistically significant proteins increased in both M.EAR vs. M.HC/M.CP and P.EAR vs. P.HC/P.CP. We found that there are 8 proteins overlapping between Penn and Mayo early stage vs. healthy control plasma. In addition, there are 5 proteins significantly increased in both Penn and Mayo early stage vs. chronic pancreatitis plasma. As seen in Supplemental Table 3.2, PIGR was significantly more abundant (>1.3-fold change) in all early-stage comparisons except for early PDAC-low CA19-9 comparisons which exhibit a modest increase. In addition, ANPEP was found to be significantly more abundant (>1.3-fold change) only in M.EAR vs. M.HC with modest increases in other comparisons and a slight decrease in P.EAR.LOW vs. P.CP (Supplemental Table 3.2). Like WUSTL’s LFQ-MS data, both Penn and Mayo cancer vs. control or chronic pancreatitis data had several inflammatory markers such as SAA1, SAA2, and CRP appear as statistically significant with high abundance ratios (data not shown), but they were not chosen for further study because they are elevated in many contexts. PIGR and ANPEP have forms naturally found in plasma and were less studied.

Phase 2: Validation of selected markers in human plasma

ELISA kits for human Polymeric immunoglobulin receptor (abcam Cat. #ab282302), human Aminopeptidase N/CD13 (R&D Systems Cat. #DY3815), human Thrombospondin-2 (R&D Systems Cat. #DTSP20), and human CA19-9 (Invitrogen^™ Cat. #EHCA199) were used as described by the manufacturers’ instructions. Each plasma sample was diluted in the manufacturer-provided buffer using a starting dilution factor of 1:500 (PIGR), 1:400 (ANPEP), 1:10 (THBS2), or 1:5 (CA19-9 – Penn Phase 2 only). ANPEP plates were prepared following manufacturer instructions using DuoSet ELISA Ancillary Reagent Kit 2 (R&D Systems Cat. #DY008B) by adding 100 uL of diluted capture antibody to each well and incubating at room temperature overnight then blocked with 300 uL of Protein-Free T20 blocking buffer (Thermo Scientific^™ Cat# 37573), before continuing the protocol provided by manufacturer. ANPEP samples were diluted in 1% BSA in PBS using R&D Systems^® Reagent Diluent Concentrate 2 (Catalog # DY995) from DuoSet ELISA Ancillary Reagent Kit 2. Plates for all assays were read at 450 nm using Molecular Devices SpectraMax 340PC384 Microplate Reader (RRID:SCR_020303) and SoftMax Pro (RRID:SCR_014240) version 5.3. Marker concentrations were determined from standard curves of manufacturer-provided positive control proteins with a four-parameter logistic nonlinear regression model using GraphPad Prism (RRID:SCR_002798) version 9.5.1.

ELISA confirmation of candidate markers in pooled plasma

The two mass spectrometry (MS) data sets yielded many provisional biomarker candidates that distinguished cancer from control plasma. Among our most promising new candidates, ANPEP and PIGR had statistically significant different abundances in both the LFQ and TMT-labeled MS. Aliquots of pooled plasma samples not sent to MS were used to confirm new candidates ANPEP and PIGR using commercial ELISA kits. As expected, both ANPEP and PIGR were markedly increased in cancer plasma pools when compared to both healthy control and chronic pancreatitis pools (Supplemental Figure 1). Pooled plasma samples were tested on a minimum of two commercial ELISA kits and the protocol was optimized before testing was moved to Penn Phase 2. Other statistically enriched markers did not yield promising results in initial Phase 2 tests with the Penn plasmas and were not studied further.

Penn Phase 2 validation of candidate markers for PDAC

De-identified samples were provided by the Carpenter laboratory at the University of Pennsylvania. After assay completion, the data and de-identification key were provided to an independent biostatistician at the University of Pennsylvania to maintain assay blinding. After obtaining written informed consent, plasma samples from 135 patients (Supplemental Table 4.1) were collected when they presented to the Hospital of the University of Pennsylvania (Penn) endoscopy clinic for an upper or lower endoscopy (controls) or for evaluation and treatment of PDAC, in accordance with recognized ethical guidelines under approved IRB protocol #822028, NCT02471170. All plasma samples were collected prior to endoscopy and neoadjuvant treatment.

Univariate and multivariable logistic regression models including 2, 3 and 4 biomarkers (i.e., combinations of CA19-9, THBS2, ANPEP, and PIGR) were developed for each binary outcome. The associated ROC curves were estimated. The area under the receiver operating curve (AUC) and its Wald 95% confidence interval (CI) were computed for each model. SAS (RRID:SCR_008567) Version 9.4 was used for statistical analyses as well as graphs.

Mayo Clinic Phase 2 validation of candidate markers for PDAC

We performed a second phase 2 validation study with an increased sample size from the Mayo Clinic. As we reported previously with the same phase 2 plasma set (Kim et al.)¹⁸, 537 patients were recruited at the Mayo Clinic including 197 PDAC cases, 140 healthy controls, 115 patients with IPMNs in the absence of PDAC, 30 patients with pNET, and 55 patients with a self-reported history of chronic pancreatitis (Supplemental Table 5). The plasma samples were collected with written informed consent from participants approved by the Mayo Clinic Institutional Review Board (ARB#s 1197–00, 354–06, and 356–06). The sample sets were constructed to ensure at least 80% power at the 0.01 level to distinguish an AUC of 0.95 from a null AUC of 0.88 (based on CA19-9 preliminary data) in cases versus healthy controls. All plasma samples were collected prior to diagnostic workup and neoadjuvant treatment. The remainder of sample aliquots used for testing THBS2¹⁸ were used to test PIGR and ANPEP. We assessed whether the proposed biomarkers could discriminate between cancer cases (n = 197) and healthy controls (n = 140) or other pancreatic conditions such as IPMNs in the absence of PDAC (n = 115), pNET (n = 30), and pancreatitis (n = 55) with an area under the Receiver Operating Characteristic (ROC) curve (AUC) analysis.

Univariate and multivariable logistic regression models were pre-planned and estimated in pancreatic cancer cases and healthy controls, to consider each candidate marker (i.e. CA19-9, THBS2, ANPEP, and PIGR) alone and in combinations of 2, 3 or 4 biomarkers. No model selection was performed, interactions were not considered, and no class weights were used. The response variable was coded as 1 to indicate the presence of, and as 0 for absence of, cancer. Candidate markers (including CA19-9) were entered as continuous variables. Given the goal of early detection with the new candidate markers rather than an epidemiological assessment, no covariates were included in the regression models. Coefficients were held fixed for application in other control groups. The associated ROC curve was plotted and the area under the receiver operating curve (AUC) was estimated for each model. A bootstrap percentile confidence interval (CI) approach was used to estimate a 95% CI for the AUC. This approach used random sampling with replacement of 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. Diagnostic thresholds are not proposed given the research grade assays and samples collected at time of diagnostic workup, rather than in asymptomatic patients. However, provisional biomarker levels corresponding to fixed levels of specificity were determined based upon distributions observed in the control samples, sensitivity, and a 95% CI were then estimated via bootstrap as described above. SAS (RRID:SCR_008567) Version 9.4 for Linux was used for modeling and figure generation.

Data Availability Statement

As noted in detail above, raw data for this study were generated at proteomics facilities at the Wistar Institute, Philadelphia, and at Washington University at Saint Louis. Readers can request the raw data from the corresponding author.

Results

Phase 1: Discovery of candidate biomarkers ANPEP and PIGR in independent early-stage PDAC plasma sources

We developed a strategy with marker discovery and initial validation using two different sources of pooled plasma from early stage PDAC patients and various controls, using two different approaches for mass spectrometry analysis on each source of plasma, and with different sources of pooled plasmas of patients at different stages of PDAC for initial candidate Phase 1 validation (Figure 1A). We assumed that the best biomarker candidates would be robust to the different methods and sources of test materials, which serve as as internal validations for our study. For discovery phase 1 (Pepe et al.’s defined Phase 1 study²⁶), we defined early-stage plasma pools as Stage I-IIA but include stage IIB in subsequent phase 2 early stage (stage I/II) classification and analyses. For continuity, early-stage will be referred to as stage I/II in the text.

Figure 1.

(A) Flowchart of experimental design for discovery phase 1 and validation phase 2. (B and C) Jitter plots of biomarker concentrations in plasma samples from Healthy Controls, Disease Controls, PDAC Stage I/II, and PDAC Stages I-IV. (B) Jitter plots of CA19-9 (U/mL), THBS2 (ng/mL), ANPEP (ng/mL), and PIGR (ng/mL) concentration in Penn Phase 2 plasma for healthy control (n = 47), disease controls (n = 29), Stage I/II PDAC (n = 35), and All Stages (I-IV) PDAC (n = 59). (C) Jitter plots of CA19-9 (U/mL), THBS2 (ng/mL), ANPEP (ng/mL), and PIGR (ng/mL) concentration in Mayo Phase 2 plasma for healthy control (n = 140), Stage I/II PDAC (n = 88), disease controls (n = 200) and All Stages (I-IV) PDAC (n = 197). PIGR concentrations maximized the assay at 6000 ng/ml in the Mayo samples.

Various proteins were enriched in stage I/II plasmas vs. controls from both Mayo and Penn in Wistar MS data. Detailed in the Methods, since the WUSTL MS data did not reveal significant protein expression in comparisons using Penn pools, we focused on proteins that were significant in Mayo pools for the LFQ-labeled MS data. Of note, aminopeptidase-N (ANPEP; APN; CD13) and polymeric immunoglobulin receptor (PIGR) were significantly increased in Mayo pools from both MS methods whereas in the Penn pools there was a borderline significant increase in ANPEP in Wistar’s data and no significant results from WUSTL’s data (see Methods). Despite the discrepancy, we continued with the two candidates selected because of their robust performance in both MS methods for the Mayo pools. For each marker, at least two commercial ELISA kits were tested on the plasma pools and the ELISA protocols were optimized.

Supplemental Figure 1 shows the ELISA data for the top two proteins, ANPEP and PIGR, both being elevated in the pooled early stage I/II plasmas compared to healthy control and chronic pancreatitis pools. As seen in data from The Cancer Genome Atlas (TCGA) (RRID:SCR_003193) in Supplemental Figure 2A, in pancreatic cancer tumor samples (n = 72) the ANPEP gene exhibits the 7^th highest median mRNA expression of all cancer types tested, while the PIGR gene exhibits the 4^th highest median mRNA expression of all cancer types tested (Supplemental Figure 2B).

Phase 2: Validation of biomarker candidates in two independent cohorts

Sample composition for validation cohorts

The first patient cohort for our Phase 2 studies, provided by the Carpenter laboratory at UPenn, consisted of a collection of 135 patient plasma samples from the Hospital of the University of Pennsylvania (Penn) endoscopy clinic. At the time of enrollment, 59 patients presented with PDAC ranging from stages I-IV, 47 were healthy controls (HCs) with no known cancers at the time, and 29 were disease controls (DCs) (Supplemental Table 4.1). The disease controls included samples with chronic pancreatitis, pancreatic cysts, pancreatic intraepithelial neoplasia (PanINs), and intraductal papillary mucinous neoplasms (IPMNs); one was diagnosed with hepatocellular carcinoma 92 days after a blood draw and one was diagnosed with papillary thyroid carcinoma 283 days aftera blood draw. A full disease control diagnosis list, abstracted from electronic health record, can be found in Supplemental Table 4.2. In these Penn Phase 2 plasma samples, using 5–15 ul for each, we tested for ANPEP, PIGR, THBS2, and used a research grade test for CA19-9. The second patient cohort consisted of Phase 2 plasma sample collection from the Mayo Clinic (n = 537) used in our prior study¹⁸, including 197 PDAC cases, 140 healthy controls, 115 patients with IPMNs in the absence of PDAC, 30 patients with pancreatic neuroendocrine tumors (pNETs), and 55 patients with a self-reported history of chronic pancreatitis (Supplemental Table 5). For the purposes of this study, we compared ANPEP and PIGR data in this sample set to our previously published THBS2 and CA19-9 data¹⁸, the latter with a clinical grade test. At all times the experimental operator at Penn remained blinded to sample identities, and biostatisticians at Penn and Mayo assessed marker performance and discrimination from controls. Here we present data using healthy patient controls first, then with nonmalignant disease controls, and finally with combined healthy and nonmalignant disease controls.

Performance of ANPEP, PIGR, THBS2 and CA19-9 in different Phase 2 univariate studies

We first assessed whether our proposed biomarkers could discriminate between all cancer cases and healthy controls (Table 1). The estimated area under the Receiver Operating Characteristic (ROC) curves (AUC) and 95% confidence intervals are shown in the Table. Figure 1B, C shows relevant jitter plots of the primary data and the ROC curves are displayed in Figure 2. CA19-9 performed similarly in Penn and Mayo samples, with samples for all stages PDAC (I-IV) vs. HC, yielding an AUC of 0.91 (Penn; 95% CI: 0.83–0.95) (Figure 2A) and 0.93 (Mayo; 95% CI: 0.90–0.95) (Figure 2B). For the other markers, when comparing all stages PDAC (I-IV) vs. HC, the following AUCs were achieved for Penn/Mayo: THBS2 = 0.85 (95% CI: 0.77–0.91)/0.88 (95% CI:0.85–0.90), ANPEP = 0.72 (95% CI: 0.63–0.81)/0.78 (95% CI: 0.74–0.81), and PIGR = 0.74 (95% CI: 0.64–0.82)/0.85 (95% CI: 0.81–0.88) (Figure 2A–B, Table 1). Focusing on early stage PDAC (Stage I/II) vs. HC, while no single marker was able to outperform CA19-9 alone [Penn AUC = 0.90(95% CI: 0.82–0.96) and Mayo AUC = 0.90(95% CI: 0.86–0.94)], the AUCs were THBS2 = 0.86(95% CI: 0.76–0.92)/0.89(95% CI: 0.85–0.92), ANPEP = 0.78(95% CI: 0.68–0.86)/0.80(95% CI:0.74–0.85), and PIGR = 0.81(95% CI: 0.70–0.88)/0.86(95% CI: 0.82–0.90) (all Penn/Mayo) (Figure 2C–D, Table 1). We note that THBS2, ANPEP, and PIGR all performed better in the larger Mayo Phase 2 cohort (n = 537) than the Penn Phase 2 cohort (n = 135).

Table 1:

Univariate Biomarker Analysis of Penn Phase 2 and Mayo Phase 2 – Healthy Control Reference

Penn Phase 2 Data
PDAC versus healthy controls	CA 19–9				THBS2			ANPEP			PIGR
	n	AUC	95% CI		AUC	95% CI		AUC	95% CI		AUC	95% CI
All stages	59/47	0.91	0.83	0.95	0.85	0.77	0.91	0.72	0.63	0.81	0.74	0.64	0.82
Stage I/II	35/47	0.90	0.82	0.96	0.86	0.76	0.92	0.78	0.68	0.86	0.81	0.70	0.88
Stage I	21/47	0.88	0.78	0.95	0.83	0.71	0.91	0.76	0.65	0.86	0.76	0.65	0.86
Stage II	14/47	0.94	0.84	0.98	0.90	0.80	0.96	0.81	0.68	0.89	0.87	0.76	0.94
Stage III/IV	24/47	0.92	0.83	0.97	0.84	0.74	0.92	0.64	0.53	0.76	0.63	0.51	0.75
Mayo Phase 2 Data
PDAC versus healthy controls	CA 19–9				THBS2			ANPEP			PIGR
	n	AUC	95% CI		AUC	95% CI		AUC	95% CI		AUC	95% CI
All stages	197/140	0.93	0.90	0.95	0.88	0.85	0.90	0.78	0.74	0.81	0.85	0.81	0.88
Stage I/II	88/140	0.90	0.86	0.94	0.89	0.85	0.92	0.80	0.74	0.85	0.86	0.82	0.90
Stage I	10/140	0.93	0.86	1.00	0.90	0.83	0.97	0.83	0.73	0.92	0.93	0.88	0.99
Stage II	78/140	0.89	0.85	0.94	0.88	0.84	0.92	0.79	0.73	0.84	0.85	0.80	0.90
Stage III/IV	109/140	0.95	0.92	0.97	0.87	0.83	0.90	0.76	0.72	0.81	0.84	0.80	0.88

Figure 2.

Univariate modeling for All Stage (I-IV) and Early Stage (I/II) vs. Controls in two phase 2 studies. (A and B) ROC curves for univariate CA19-9, THBS2, ANPEP, and PIGR in All Stages PDAC (I-IV) vs. Healthy Control for Penn Phase 2 (A) and Mayo Phase 2 (B). (C and D) ROC curves for univariate CA19-9, THBS2, ANPEP, and PIGR in Early Stage (I/II) PDAC vs. Healthy Control for Penn Phase 2 (C) and Mayo Phase 2 (D).

Performance of ANPEP, PIGR, and THBS2 in two-variable models with addition of CA19-9

Using two-variable models of ANPEP, PIGR, or THBS2 with the addition of CA19-9, the estimated AUCs for PDAC vs. healthy controls were an improvement over any single marker alone (Figure 3). When comparing all stages PDAC (I-IV) vs. HC, the following AUCs were seen for Penn/Mayo: THBS2 = 0.93(95% CI: 0.88–0.98)/0.98(95% CI:0.97–0.99), ANPEP = 0.92(95% CI: 0.86–0.97)/0.94(95% CI:0.92–0.96), and PIGR = 0.91(95% CI: 0.86–0.96)/0.96(95% CI: 0.95–0.98) compared to 0.91(95% CI: 0.83–0.95)/0.93(95% CI: 0.90–0.95) for CA19-9 alone (Figure 3A, B, Table 2). Focusing on early stage PDAC (stage I/II), two-variable AUCs for Penn/Mayo were: THBS2 = 0.94(95% CI: 0.88–1.00)/0.97(95% CI: 0.95–0.98), ANPEP = 0.89(95% CI: 0.81–0.97)/0.92(95% CI: 0.88–0.95), and PIGR = 0.93(95% CI: 0.87–0.98)/0.94(95% CI: 0.92–0.97) compared to 0.90(95% CI: 0.82–0.96)/0.90(95% CI: 0.86–0.94) for CA19-9 alone (Figure 3C–D, Table 2).

Figure 3.

Two-variable modeling for Early Stage and All Stage PDAC vs. Healthy Control in two Phase 2 studies. (A and B) ROC curves for two-variable CA19-9 + THBS2, CA19-9 + ANPEP, and CA19-9 + PIGR in All Stages PDAC (I-IV) vs. Healthy Control for Penn Phase 2 (A) and Mayo Phase 2 (B). (C and D) ROC curves for two-variable CA19-9 + THBS2, CA19-9 + ANPEP, and CA19-9 + PIGR in Stage I/II PDAC vs. Healthy Control for Penn Phase 2 (C) and Mayo Phase 2 (D).

Table 2:

Two-variable Biomarker Analysis of Penn Phase 2 and Mayo Phase 2 – Healthy Control Reference

Penn Phase 2 Data
PDAC versus healthy controls	CA 19–9 + THBS2				CA 19–9 + ANPEP			CA 19–9 + PIGR
	n	AUC	95% CI		AUC	95% CI		AUC	95% CI
All stages	59/47	0.93	0.88	0.98	0.92	0.86	0.97	0.91	0.86	0.96
Stage I/II	35/47	0.94	0.88	1.00	0.89	0.81	0.97	0.93	0.87	0.98
Stage I	21/47	0.91	0.82	1.00	0.81	0.67	0.95	0.91	0.83	0.98
Stage II	14/47	0.98	0.95	1.00	0.90	0.76	1.00	0.95	0.90	1.00
Stage III/IV	24/47	0.91	0.84	0.99	0.88	0.77	0.99	0.92	0.85	0.98
Mayo Phase 2 Data
PDAC versus healthy controls	CA 19–9 + THBS2				CA 19–9 + ANPEP			CA 19–9 + PIGR
	n	AUC	95% CI		AUC	95% CI		AUC	95% CI
All stages	197/140	0.98	0.97	0.99	0.94	0.92	0.96	0.96	0.95	0.98
Stage I/II	88/140	0.97	0.95	0.98	0.92	0.88	0.95	0.94	0.92	0.97
Stage I	10/140	0.97	0.92	1.00	0.93	0.86	1.00	0.95	0.90	1.00
Stage II	78/140	0.97	0.95	0.98	0.92	0.88	0.96	0.95	0.92	0.98
Stage III/IV	109/140	0.98	0.97	1.00	0.96	0.94	0.99	0.98	0.96	0.99

Multivariable analysis of ANPEP, PIGR, THBS2, and CA19-9

We created three 3-marker panels (2 markers and CA19-9) and a 4-marker panel (Figure 4). Multivariable analysis of all stage PDAC vs. healthy controls yielded AUCs ranging from 0.89–0.94 for Penn (Figure 4A, Table 3) and 0.96–0.98 for Mayo samples (Figure 4B, Table 3). Focusing on early-stage performance, the combination of CA19-9/THBS2/ANPEP outperformed the other 3-marker models with an AUC of 0.96 (95% CI: 0.91–1.00)/0.97 (95%CI: 0.95–0.98) in Penn/Mayo samples (Figure 4C, D). A panel of CA19-9/THBS2/PIGR also yielded an AUC of 0.94 (95% CI: 0.88–1.00)/0.97 (95% CI: 0.95–0.99) with the Penn/Mayo early-stage samples (Figure 4C, D, Table 3). The weakest performing panel in both cohorts was CA19-9/ANPEP/PIGR which had AUCs of 0.91 (Penn; 95% CI: 0.84–0.99) and 0.94 (Mayo; 95% CI: 0.92–0.97) in early stage vs. healthy controls. Notably, the strongest performing panel in both cohorts was the 4-marker panel of CA19-9/THBS2/ANPEP/PIGR, which yielded AUCs of 0.96(95% CI: 0.91–1.00)/0.97(95% CI: 0.95–0.99) in stage I/II vs. HC for Penn/Mayo (Figure 4C, D), outperforming CA19-9 alone which had an AUC of 0.90 in both Penn and Mayo comparisons. It was striking how well the univariate and multivariate AUCs agreed with the two sources of Phase 2 plasma cohorts.

Figure 4.

Multivariable modeling for Early Stage and All Stage PDAC vs. Controls in two Phase 2 studies. (A and B) ROC curves for multivariable CA19-9 + THBS2 + ANPEP, CA19-9 + THBS2 + PIGR, CA19-9 + ANPEP + PIGR, and CA19-9 + THBS2 + ANPEP + PIGR in All Stages PDAC (I-IV) vs. Healthy Control for Penn Phase 2 (A) and Mayo Phase 2 (B). (C and D) ROC curves for multivariable CA19-9 + THBS2 + ANPEP, CA19-9 + THBS2 + PIGR, CA19-9 + ANPEP +PIGR, and CA19-9 + THBS2 + ANPEP + PIGR in Stage I/II PDAC vs. Healthy Control for Penn Phase 2 (C) and Mayo Phase 2 (D). (E and F) ROC curves for multivariable CA19-9 + THBS2 + ANPEP, CA19-9 + THBS2 + PIGR, CA19-9 + ANPEP +PIGR, and CA19-9 + THBS2 + ANPEP + PIGR in Stage I/II PDAC vs. Disease Control (E) and All Stages PDAC (I-IV) vs. Disease Control (F) for Mayo Phase 2.

Table 3:

Multivariable Analysis of Penn Phase 2 and Mayo Phase 2 – Healthy Control Reference

Penn Phase 2 Data
PDAC versus healthy controls	CA19-9 + THBS2 + PIGR				CA19-9 + THBS2 + ANPEP			CA19-9 + ANPEP + PIGR			CA19-9 + THBS2 + ANPEP + PIGR
	n	AUC	95% CI		AUC	95% CI		AUC	95%CI		AUC	95% CI
All stages	59/47	0.93	0.88	0.98	0.94	0.89	0.99	0.89	0.83	0.96	0.94	0.89	0.99
Stage I/II	35/47	0.94	0.88	1.00	0.96	0.91	1.00	0.91	0.84	0.99	0.96	0.91	1.00
Stage I	21/47	0.91	0.82	1.00	0.94	0.85	1.00	0.90	0.81	0.98	0.94	0.85	1.00
Stage II	14/47	0.98	0.95	1.00	0.99	0.98	1.00	0.92	0.78	1.00	1*	--	--
Stage III/IV	24/47	0.93	0.86	0.99	0.91	0.81	1.00	0.87	0.76	0.98	0.91	0.81	1.00
Mayo Phase 2 Data
PDAC versus healthy controls	CA19-9 + THBS2 + PIGR				CA19-9 + THBS2 + ANPEP			CA19-9 + ANPEP + PIGR			CA19-9 + THBS2 + ANPEP + PIGR
	n	AUC	95% CI		AUC	95% CI		AUC	95% CI		AUC	95% CI
All stages	197/140	0.98	0.97	0.99	0.98	0.97	0.99	0.96	0.95	0.98	0.98	0.97	0.99
Stage I/II	88/140	0.97	0.95	0.99	0.97	0.95	0.98	0.94	0.92	0.97	0.97	0.95	0.99
Stage I	10/140	0.96	0.93	1.00	0.96	0.92	1.00	0.95	0.90	1.00	0.96	0.93	1.00
Stage II	78/140	0.97	0.95	0.99	0.97	0.95	0.99	0.95	0.92	0.98	0.97	0.95	0.99
Stage III/IV	115/140	0.98	0.97	1.00	0.98	0.97	1.00	0.98	0.96	0.99	0.98	0.97	1.00

^*The predicted probabilities from both models 1 and 2 predict which group (HC vs Stage II) each patient is in perfectly. HC patients have predicted probabilities between 0.0 and 0.045 and Stage II patients have predicted probabilities between 0.96 and 1.

Univariate, Two-Variable, and Multivariable analysis of ANPEP, PIGR, THBS2, and CA19-9 in Disease Controls

We next assessed whether our proposed biomarkers could discriminate between all cancer cases and disease controls in the Penn (Supplemental Tables 6.1–6.3) and Mayo Phase 2 (Table 4) cohorts. The Penn Phase 2 cohort had insufficient n for comparable analysis but all univariate, two-variable, and multivariable analysis is available in Supplemental Tables 6.1–6.3.

Table 4:

Univariate, Two-Variable, and Multivariable Analysis of Mayo Phase 2 Cohort – Disease Control and All Control (HC+DC) Reference

A.Univariate Analysis of Mayo Phase 2 Cohort – Disease Control Reference
PDAC versus benign pancreatic disease	CA 19–9				THBS2			ANPEP			PIGR
	n	AUC	95% CI		AUC	95%CI		AUC	95% CI		AUC	95% CI
PDAC (I-IV) vs. all controls	197/340	0.90	0.88	0.93	0.80	0.77	0.82	0.73	0.70	0.76	0.78	0.75	0.81
PDAC (I-II) vs. all controls	88/340	0.87	0.83	0.91	0.81	0.77	0.84	0.75	0.70	0.80	0.80	0.75	0.84
PDAC (I-IV) vs. disease controls	197/200	0.89	0.86	0.92	0.74	0.71	0.78	0.70	0.66	0.74	0.74	0.70	0.78
PDAC (I-II) vs. disease controls	88/200	0.85	0.81	0.89	0.75	0.70	0.80	0.72	0.66	0.76	0.75	0.71	0.79
PDAC vs. pancreatitis	197/55	0.87	0.83	0.91	0.73	0.67	0.79	0.63	0.57	0.68	0.61	0.55	0.68
PDAC vs. IPMN	197/115	0.91	0.88	0.94	0.78	0.75	0.82	0.75	0.70	0.79	0.80	0.76	0.84
PDAC vs. pNET	197/30	0.87	0.82	0.92	0.39	0.34	0.71	0.63	0.55	0.71	0.72	0.65	0.79
pNET vs. healthy controls	30/140	0.60	0.50	0.72	0.75	0.66	0.84	0.69	0.61	0.78	0.67	0.57	0.76
B.Two-variable Analysis of Mayo Phase 2 Cohort – Disease Control Reference
	CA19-9 + THBS2				CA19-9 + ANPEP			CA19-9 + PIGR			Intentionally Blank
PDAC versus benign pancreatic disease	n	AUC	95% CI		AUC	95% CI		AUC	95% CI
PDAC (I-IV) vs. all controls	197/340	0.94	0.92	0.96	0.92	0.90	0.94	0.93	0.91	0.95
PDAC (I-II) vs. all controls	88/340	0.92	0.89	0.94	0.89	0.86	0.92	0.90	0.87	0.93
PDAC (I-IV) vs. disease controls	197/200	0.91	0.89	0.94	0.91	0.88	0.93	0.91	0.89	0.93
PDAC (I-II) vs. disease controls	88/200	0.89	0.85	0.92	0.86	0.83	0.90	0.87	0.83	0.90
PDAC vs. pancreatitis	197/55	0.90	0.86	0.93	0.87	0.84	0.91	0.87	0.84	0.91
PDAC vs. IPMN	197/115	0.95	0.93	0.97	0.93	0.90	0.95	0.94	0.93	0.96
PDAC vs. pNET	197/30	0.87	0.81	0.92	0.87	0.82	0.92	0.88	0.83	0.93
pNET vs. healthy controls	30/140	0.75	0.66	0.85	0.62	0.52	0.74	0.67	0.57	0.77
C. Multivariable Analysis of Mayo Phase 2 Cohort – Disease Control Reference
PDAC versus benign pancreatic disease	CA19-9 + THBS2 + PIGR				CA19-9 + THBS2 + ANPEP			CA19-9 + ANPEP + PIGR			CA19-9 + THBS2 + ANPEP + PIGR
	n	AUC	95% CI		AUC	95% CI		AUC	95%CI		AUC	95% CI
PDAC (I-IV) vs. all controls	197/340	0.93	0.92	0.95	0.93	0.92	0.95	0.93	0.91	0.95	0.93	0.91	0.95
PDAC (I-II) vs. all controls	88/340	0.90	0.88	0.94	0.91	0.88	0.94	0.90	0.87	0.93	0.90	0.88	0.94
PDAC (I-IV) vs. disease controls	197/200	0.91	0.89	0.94	0.91	0.89	0.94	0.91	0.88	0.93	0.91	0.89	0.94
PDAC (I-II) vs. disease controls	88/200	0.87	0.84	0.91	0.88	0.84	0.92	0.87	0.83	0.90	0.87	0.84	0.91
PDAC vs. pancreatitis	197/55	0.90	0.86	0.94	0.90	0.85	0.93	0.88	0.84	0.92	0.90	0.86	0.94
PDAC vs. IPMN	197/115	0.95	0.94	0.97	0.95	0.94	0.97	0.94	0.92	0.96	0.95	0.94	0.97
PDAC vs. pNET	197/30	0.88	0.83	0.94	0.87	0.81	0.93	0.88	0.83	0.94	0.89	0.84	0.95
pNET vs. healthy controls	30/140	0.73	0.64	0.84	0.76	0.68	0.86	0.69	0.59	0.80	0.74	0.66	0.85

When comparing PDAC vs. disease controls (DC) in the Penn Phase 2 cohort, no single marker outperformed CA19-9 alone (AUC = 0.91; 95% CI: 0.83–0.96) (Supplemental Table 6.1). However, in the larger Mayo Phase 2 cohort, THBS2, ANPEP, and PIGR each outperform CA19-9’s ability to discriminate pNET from controls with AUCs of 0.67–0.75 exceeding CA19-9’s AUC of 0.60 (95% CI: 0.50–0.72; Table 4A).

For the Penn Phase 2, of the two-variable models discriminating PDAC from disease controls (DC), only CA19-9/ANPEP (AUC = 0.93; 95% CI: 0.86–0.99) showed improvement over CA19-9 alone (AUC = 0.91; 95% CI: 0.83–0.96) (Supplemental Table 6.2). For the Mayo two-variable models discriminating PDAC vs. IPMN, the addition of THBS2, ANPEP, or PIGR to CA19-9 increased the AUC from 0.91 (CA19-9 95% CI: 0.88–0.94) (Table 4A) to 0.95 (THBS2 95% CI: 0.93–0.97), 0.93 (ANPEP 95% CI: 0.90–0.95), or 0.94(PIGR 95% CI: 0.93–0.96), respectively (Table 4B). PDAC vs. pancreatitis showed minor improvement, raising the AUC of CA19-9 alone from 0.87 (95% CI: 0.83–0.91) (Table 4A) to 0.87–0.90 in two-variable models, the strongest model being with CA19-9/THBS2 (AUC = 0.90; 95% CI: 0.86–0.93) (Table 4B).

We applied the same multivariable models using disease controls instead of healthy controls for Penn (Supplemental Table 6.3) and Mayo (Table 4C). The panel that performed the best in the Penn cohort was CA19-9/ANPEP/PIGR with an AUC of 0.92 for all stages PDAC (95% CI: 0.85–0.98) and early stage PDAC (95% CI: 0.86–0.99) vs. disease controls (Supplemental Table 6.3). Multivariable models for Mayo’s cohort of PDAC vs. benign pancreatic disease controls showed only minor improvement over univariate CA19-9 and two-variable models. In PDAC vs. pancreatitis, there was no improvement in multivariable models compared to the two-variable model with CA19-9/THBS2 (AUC = 0.90; 95% CI = 0.88–0.93) (Table 4B). Instead, all the models except one achieved the same AUC of 0.90 (Table 4C), with the model of CA19-9/ANPEP/PIGR achieving a slightly lower AUC of 0.88 (95% CI= 0.84–0.92). In PDAC vs. IPMN, we saw a similar trend of comparable AUCs between two-variable and multivariable models with models CA19-9/THBS2/PIGR, CA19-9/THBS2/ANPEP, and CA19-9/THBS2/ANPEP/PIGR all having an AUC of 0.95 (Table 4C). Comparing PDAC vs. pNET, each multivariable model performed similar to the best two-variable model CA19-9/PIGR (AUC = 0.88; 95% CI: 0.83–0.93) (Table 4B), with AUCs ranging from 0.87–0.89 (Table 4C). In the case of pNET vs. controls, the best model composed of CA19-9/THBS2/ANPEP slightly increased the AUC to 0.76 (95% CI: 0.68–0.86) (Table 4C) while the other 3 multivariable models had AUCs ranging from 0.69–0.74 which are either comparable or dampen the performance of both THBS2 alone (AUC = 0.75; 95% CI: 0.66–0.84) (Table 4A) and the CA19-9/THBS2 two-variable model (AUC = 0.75; 95% CI: 0.66–0.85) (Table 4B).

To evaluate performance against a combination of benign disease controls, we also compared all PDAC (Stage I-IV) and early stage PDAC (Stage I-II) vs. all Mayo Phase 2 disease controls (n = 200) for each univariate (Table 4A), two-variable (Table 4B), and multivariable model (Table 4C). No single marker outperformed CA19-9 in early stage (AUC = 0.85; 95% CI: 0.81–0.89) or all stage (AUC = 0.89; 95% CI: 0.86–0.92) comparisons (Table 4A). Two-variable models show improvement over CA19-9 alone with AUCs ranging from 0.86–0.89 for early stage vs. disease controls and 0.91 for all three two-variable models comparing all PDAC vs. disease controls (Table 4B). Additional analysis of multivariable models showed similar results with AUCs ranging from 0.87–0.88 for early stage vs. disease controls and 0.91 for all multivariable panels comparing all PDAC vs. DC (Table 4C). Interestingly, in early stage PDAC vs. disease controls, the two-variable model composed of CA19-9 and THBS2 is the best performing model with an AUC of 0.89 (95% CI: 0.85–0.92). The multivariable model composed of CA19-9 + THBS2 + ANPEP + PIGR yields an AUC of 0.87 (Stage I/II vs. DC; Figure 4E) and 0.91 (All PDAC vs. DC; Figure 4F).

Finally, Tables 4A–C compare PDAC I-IV and PDAC I-II with all controls, i.e., combined healthy and disease controls, in the larger Mayo Phase 2 cohort. While none of the univariate models outcompeted CA19-9 alone (AUC = 0.90; CI: 0.88–0.93 for PDAC I-IV vs. all controls and AUC = 0.87; CI: 0.83–0.91 for PDAC I-II vs. all controls) (Table 4A), CA19-9 was outcompeted by bivariate models CA19-9/THBS2 (AUC = 0.94; CI: 0.92–0.96 and AUC = 0.92; CI: 0.89–0.94, respectively) and CA19-9/PIGR (AUC 0.93; CI 0.91–0.95 and AUC 0.90; CI 0.87–0.93, respectively) (Table 4B), with AUCs of the multivariate models (Table 4C) comparable to the bivariate models.

Sensitivity at prespecified levels of specificity

To determine a provisional biomarker plasma concentration to use as a cutoff point for discriminating healthy controls from PDAC cases, we assessed the distribution of ANPEP, PIGR, and THBS2 values based on the 140 healthy controls in the Mayo Phase 2 validation study. Using approximate false-positive rates (FPRs) ranging from 0 to 5%, we chose six representative cutoffs to evaluate their sensitivity and specificity for distinguishing PDAC vs. healthy control. As seen in Supplemental Table 7 for the Mayo Phase 2 study 4-marker panel, combining cutoffs for CA19-9 of ≥35 U/mL and cutoff of approximately 1% FPR for each respective marker (THBS2 = 42 ng/mL; ANPEP = 2995 ng/mL; PIGR = 1800 ng/mL), yielded a 91.94% (95% CI: 89.07–95.20) sensitivity with a specificity of 95% (95% CI: 92.31–97.73) for all stage PDAC (I-IV) and 87.53% (95% CI: 82.46–93.28) for early stage PDAC I/II. As individual markers, PIGR is the most sensitive marker we tested after CA19-9. A cutoff of 35U/mL for CA19-9 results in a sensitivity of 82.71% (95% CI 78.63–86.53) and specificity of 96.46% (94.32–98.89) for all stage PDAC; a sensitivity of 76.24% (69.60–83.64) and specificity of 96.38% (94.25–98.82) results when comparing to early stage PDAC. For all stage PDAC, a concentration of 1800 ng/mL or greater for PIGR results in a 59.12% (95% CI: 54.03–64.31) sensitivity with 99.3% (95% CI: 98.75–100) specificity whereas applying the same FPR of 1% to ANPEP (2995 ng/mL) results in a 14.23% (95% CI: 10.36–17.53) sensitivity with 99.29% (95% CI 98.75–100) specificity. Applying the same concentration cutoffs and specificity of 99.29% to early-stage comparisons, PIGR results in a 59.05% (95% CI: 50.86–66.97) sensitivity and ANPEP results in 16.97% (95% CI: 10.91–22.86) sensitivity, displaying minor improvement in early stage over all stage sensitivity for ANPEP. As previously described, a cutoff of 42 ng/mL for THBS2 detects approximately 52% with a specificity of 99%¹⁸.

Assessing performance of two-variable models for new candidates ANPEP and PIGR, we note that a combination of CA19-9 and ANPEP minimally improves sensitivity for both early and all stage PDAC comparisons over CA19-9 alone whereas CA19-9 and PIGR two-variable models improve the sensitivity from 82.71% (CA19-9 alone) to 90.40% (95% CI 93.14–97.98) for all stages and from 76.24% sensitivity (CA19-9 alone) to 84.08% (78.22–90.82) for early stage PDAC. All sensitivity and specificity analysis at each provisional cutoff for both all stage and early stage PDAC can be found in Supplemental Table 7.

Discussion

Diagnosing PDAC in early, localized stages drastically improves 5-year survival rates from less than 13% to 44%, but this only occurs in about 14% of all cases^3,4. Using two different MS methods on plasma pools sourced from Penn and Mayo, we identified promising biomarkers that were elevated in early-stage PDAC vs. controls and advanced our top performers, ANPEP and PIGR, to test in two independent Phase 2 retrospective collections.

Our four-marker panel composed of CA19-9, THBS2, ANPEP, and PIGR yields an AUC of 0.96 in Penn Phase 2 (n = 135) and 0.97 in Mayo Phase 2 (n = 537) in early stage PDAC vs. healthy control comparisons (Figure 4C,D) and an AUC of 0.94 (Penn) and 0.98 (Mayo) in all PDAC (I-IV) vs. healthy control comparisons (Figure 4A, B). Additionally, we assessed performance of the four-marker panel in Mayo Phase 2 early stage PDAC vs. disease control (Figure 4E) and all PDAC (I-IV) vs. disease control (Figure 4F) which had AUCs of 0.87 (Stage I/II vs. DC) and 0.91 (Stage I-IV vs. DC). Attributed to the predictive power of THBS2, our three-marker panels performed similarly well with CA19-9/THBS2/ANPEP and CA19-9/THBS2/PIGR both yielding an AUC of 0.97 in Mayo Phase 2 when comparing early stage PDAC vs. healthy controls (Figure 4D). In Penn Phase 2, these three-marker panels with CA19-9/THBS2 + ANPEP or PIGR had AUCs of 0.96 (+ ANPEP) or 0.94 (+PIGR) for the same comparison (Figure 4C). Comparisons for all stage PDAC vs. healthy controls yield similar results (Figure 4A, B).

Considering patient care, our proposed method of detection uses blood-based screening, which is a relatively low-cost and minimally invasive procedure. To further assess marker performance, we calculated sensitivity and specificity for each new model (Supplemental Table 7) for Mayo Phase 2 plasma by defining biomarker concentration cutoff points based on percentiles of distribution (95–100%) in control plasma samples (n = 140). At a provisional specificity of 95%, a plasma biomarker panel composed on CA19-9 (≥35 U/mL), THBS2 (≥42ng/ml), ANPEP (≥2995ng/mL), and PIGR(≥1800ng/mL) yielded a sensitivity of 91.94% for all stages and 87.53% for early stage I/II PDAC detection in our larger Phase 2 study (Supplemental Table 7). The excellent discriminatory performance of this panel in the Phase 2 studies summarized here supports further exploration in Phase 3 studies, specifically to evaluate its ability to detect early stages of PDAC in high-risk individuals. Additionally, a plasma-based screening method could serve as a cost-saving measure to advise physicians which patients should prioritize diagnosis through imaging-based approach while also minimizing the burden on patients by using a less invasive procedure with a high level of sensitivity and specificity.

The lifetime risk of pancreatic cancer in the general population is estimated to be approximately 1.7%. Patients with pancreatic cysts and pancreatitis are known to be at greater than average risk of pancreatic cancer. A meta-analysis of 7 studies noted a pooled relative risk of 13.3% for pancreatic cancer in patients with chronic pancreatitis; the standardized incidence rate is nearly three times higher for hereditary pancreatitis compared to other etiologies of chronic pancreatitis^27,28. For estimates of pancreatic cancer incidence in patients with pancreatic cysts, a meta-analysis published in 2017 indicated that the risk of progression to cancer within 10 years was 8–25%²⁹. Subsequently, microsimulation analysis has estimated the 20-year incidence of pancreatic cancer for an individual with pancreatic cysts at age 50 years to be lower at 1.8%–6.1%³⁰. More recently, emerging evidence from long-term follow-up studies indicates that this cancer risk is primarily driven by the small subset of pancreatic cysts with worrisome and high-risk imaging features whereas the risk of pancreatic cancer in patients with the more prevalent low-risk cysts is comparable to those without cysts^31,32. There are other patient populations at increased risk of pancreatic cancer such as those with familial and germline risk factors and patients with new-onset diabetes that were not included in the control group of this current study, but are of interest to the early detection biomarker research community and will need to be included as controls in future studies aimed at further validation of these results.

The Penn disease control cohort had a significant number of pancreatic cysts and IPMNs (Supplemental Table 4.2) that were not further investigated as a subgroup, due to small n and multiple benign diagnoses. To explore the relationship between patient demographic information and ANPEP or PIGR plasma concentrations, we calculated the median expression concentrations for categorical variables (sex, male versus female; presence or absence of diabetes mellitus) (Supplemental Tables 8.1 - 8.4) and a Spearman correlation coefficient for continuous variables (age) (Supplemental Table 8.5) for the larger Mayo Phase 2 cohort (n = 537). Potential relationships between patient demographic data and THBS2 were analyzed using the same methods in our previous publication¹⁸. Based on the data obtained from our Mayo phase 2 studies, we observed no apparent associations between age (Supplemental Table 8.5), sex (Supplemental Tables 8.1–8.2), diabetes mellitus (Supplemental Tables 8.3–8.4), and THBS2, ANPEP, or PIGR levels. This suggests that a single provisional cutoff may be used rather than age-, sex-, or diabetes mellitus-specific cutoffs pending further investigation in pre-diagnostic and high-risk individual studies. Given the limited clinical data available from electronic health records for our plasma pools and phase 2 cohorts, we plan an additional validation using Pancreatic Cancer Detection Consortium (PCDC) reference set samples. The PCDC samples will have more comprehensive clinical data available to advance the findings presented here and to further address the clinical spectrum bias.

ANPEP/APN/CD13 is a membrane-bound ectopeptidase that is also expressed in a soluble form³³ in blood plasma. It exists in various cells and tissues and is upregulated in tissue or serum from patients with non-small cell lung cancer, breast cancer, hepatocellular carcinoma, pancreatic cancer, and more^34–38. Gene expression profiling reveals that ANPEP is upregulated in a subtype of PDAC tumors alongside other genes that are involved in lipid and protein metabolism³⁹. Serum ANPEP has been described as a diagnostic and prognostic biomarker of pancreatic cancer when used in combination with CA19-9, with immunohistochemical analysis in several studies finding that about half of resected tumors stained positive for ANPEP, primarily in cell membranes and cytoplasm^38,40. Our study advances these findings by presenting ANPEP performance in a panel with other PDAC early-stage biomarkers.

PIGR is a transmembrane transporter of polymeric IgA and IgM through the intestinal epithelium and has been discovered in esophageal, gastric, lung, liver, colon, and pancreatic cancer^41–43. Gene expression profiling reveals that PIGR is upregulated in a subtype of PDAC tumors alongside other genes that are involved in carbohydrate metabolism³⁹. Additionally, immunohistochemical analysis on the chemoresistance of resected pancreatic cancer patient tumor tissues (n=77) showed that high expression of PIGR is significantly associated with poor prognosis⁴⁴. Serum PIGR was previously found to be significantly elevated in patients with pancreatic cancer and was used in a biomarker panel^45–47. Staining of 88 tissue microarray tumors showed that a normal pancreas does not express PIGR, but it is expressed in PDAC and increasingly so in premalignant PanIN stages⁴⁷.

Our Mayo Phase 2 data showed that PDAC vs. pancreatitis yields an AUC of 0.61 (95% CI: 0.55–0.68) whereas PDAC vs. other pancreatic conditions such as IPMNs yield an AUC of 0.80 (95% CI: 0.76–0.84) for PIGR alone (Table 4A). Our study provides more insight into the early-stage detection potential of PDAC with the use of plasma collected from two separate institutions and more detailed performance with the inclusion of CA19-9 in two-variable (Table 2) and multivariable (Table 3) analyses for both Penn and Mayo cohorts. Arumugam et al. investigated the expression and stromal activity of PIGR in PDAC using manipulated (siRNA and shRNA) cells lines for both 2D and 3D organotypic models and tissue microarrays to determine its utility as a biomarker⁴⁷. Their staining of 88 tissue microarray tumors collected from patients at the Royal London Hospital (Whitechapel, England) using duodenum as a positive control showed that a normal pancreas does not express PIGR, but it is expressed in PDAC and premalignant PanIN stages. Notably, their staining shows there is a trend of increasing expression of PIGR with PanIN progression stages found to be associated with a decreased expression of E-cadherin leading to the conclusion that PIGR may be involved in PDAC progression with a possible link to stromal activity.

Our new data confirms the performance of THBS2 in an independent Phase 2 cohort of PDAC cases and healthy controls, sourced from Penn, and builds on previous data¹⁸ by introducing new biomarkers ANPEP and PIGR tested on Phase 2 plasma sourced from Penn (Philadelphia, PA) and Mayo (Rochester, MN). Our data agrees with published data showing an elevated level of ANPEP⁴⁰ and PIGR⁴⁵ in serum samples from PDAC patients vs. controls, strengthening their discoveries with evidence of elevated levels in PDAC plasma from two institutions located in the United States of America. To our knowledge, this is the first time a biomarker panel composed of CA19-9, THBS2, ANPEP, and PIGR has been proposed for early detection of pancreatic ductal adenocarcinoma in retrospective Phase 2 plasma collections from two institutions. We believe that a blood-based assay presents an opportunity for screening methods that are low-cost, minimally invasive, and low-anxiety for patients while preserving the performance characteristics necessary for early-stage detection of pancreatic ductal adenocarcinoma. A panel comprised of CA19-9/THBS2/ANPEP/PIGR may be suitable for early detection of pancreatic ductal adenocarcinoma based on results showing a high sensitivity and specificity in the larger Mayo Phase 2 cohort, but would require pre-diagnostic cohorts for verification. At a 95% specificity, this panel yields a 91.94% sensitivity when for all stage PDAC (I-IV) and 87.53% for early stage (I/II). A notable limitation, our panel was tested on samples at the time of diagnosis and has not been tested on pre-diagnostic samples or high-risk surveillance cohorts which would provide a more accurate representation of performance in the clinic. We hope to test these markers on a prospective longitudinal cohort to fully evaluate biomarker panel performance as it relates to clinical application in the future.

Translational Relevance:

The overall 5-year relative survival rate of pancreatic ductal adenocarcinoma (PDAC) is less than 13%, hence there is intense research to find markers to enable early-stage detection. If accurately diagnosed and treated, stage I pancreatic cancers exhibit a survival rate of over 80%, according to the NCI SEER database. Carbohydrate antigen 19–9 (CA19-9) is widely used to monitor PDAC treatment response in patients with an established diagnosis, but falls short as a stand-alone screening tool due to upregulation in benign disease conditions such as pancreatitis and bile duct obstruction, limited sensitivity and specificity for early-stage PDAC, and CA19-9 levels being affected by a patient’s genetics for the relevant enzymes. By adding new markers ANPEP and PIGR to a biomarker panel of THBS2 with CA19-9, we have improved the ability to detect early stage PDAC in two retrospective Phase 2 studies, including when compared to disease controls.