Diagnostic Performance of a Tumor Marker Gene Test to Personalize Serum CA19–9 Reference Ranges

Pancreatic cancer survival is slowly improving; 5-year survival is now ∼12% (1). Evidence from the Cancer of the Pancreas Screening (CAPS) study (2, 3) and other surveillance programs (4) indicates pancreatic imaging surveillance can improve early detection, especially the detection of Stage I pancreatic ductal adenocarcinoma (PDAC; ref. 2), and can reduce pancreatic cancer mortality. Nationally, the outcome of patients diagnosed and treated with Stage I PDAC is excellent (over 80% 5-year survival; ref. 5). However, some patients diagnosed with pancreatic cancer while under pancreatic surveillance will have higher-stage disease, and survival with PDAC depends heavily on the stage at diagnosis (6). The addition of blood-based biomarkers to pancreatic imaging surveillance could improve early detection. CA19–9 is the best available blood-based biomarker of pancreatic cancer and is used for monitoring disease status, but when studied as a diagnostic test achieves only ∼50% to 60% diagnostic sensitivity among patients with resectable-stage PDAC, at a specificity too low for diagnostic use (7).

The synthesis of CA19–9 and related metabolites involves multiple fucosyltransferases (FUT; ref. 8), and prior studies have reported that variants encoding the enzymes FUT2 and FUT3 influence CA19–9 levels (9–12). A functional FUT3 enzyme is needed to synthesize CA19–9, and subjects who lack functional FUT2 (known also as “non-secretors” for their lack of secretion of blood group antigens) have higher levels of CA19–9 than those with an intact FUT2 (12, 13). Control subjects who have only one functional FUT3 allele have lower levels of CA19–9 than those with two functional alleles; thus, individuals can be classified into one of four functional FUT groups that have different CA19–9 reference ranges (14). Assigning individuals their FUT group and CA19–9 reference range improved diagnostic performance (14) compared with the usual approach of applying the same standard cutoff for everyone.

We developed a “CA19–9 tumor marker gene test” in the Johns Hopkins Molecular Diagnostics Lab as a companion test to serum CA19–9 (see Supplementary Fig. S1). This tumor marker gene test was applied to control subjects enrolled in the CAPS study to establish clinical laboratory reference ranges for CA19–9 for each functional FUT group. We then applied this test to patients diagnosed with resectable-stage pancreatic cancer.

This study included 716 subjects, including 348 high-risk individuals (HRI) and other controls enrolled in the Cancer of the Pancreas Screening-5 (CAPS5) study (NCT02000089) without evidence of pancreatic cancer at least 1 year after their blood-draw, and 368 patients diagnosed with PDAC that had blood samples collected prior to pancreatic resection at the Johns Hopkins Hospital, including a testing set of 234 cases from the Abe and colleagues study that had available biospecimen (14), and 134 Hopkins PDAC cases diagnosed between 2005 and 2019 that formed the validation set. Patients with PDAC who had undergone neoadjuvant chemotherapy were excluded. Controls were excluded if they had previous surgery for pancreatic cancer or high-grade dysplasia, or had worrisome imaging findings, or if they had other diseases or abnormalities known to be associated with an elevated CA19–9 [such as pancreatitis, an abnormally dilated main pancreatic duct or common bile duct, or large (>2 cm) or innumerable hepatic or renal cysts; ref. 7]. The CAPS controls selected for this study had undergone prior FUT genotyping in the Goggins lab and were selected to have sufficient numbers for each of the main FUT groups (approximately 100 subjects for each functional group, with fewer needed for the FUT3-null group as its reference range was expected to be close to 0). The final diagnosis of PDAC was made by surgical pathology using World Health Organization criteria (15). Tumor stage was defined by the American Joint Committee on Cancer (AJCC) 8th edition. All authors had access to study data and reviewed and approved the final manuscript. This study was performed in accordance with the Declaration of Helsinki, approved by the Johns Hopkins Institutional Review Board, and written informed consent was provided from all enrolled patients.

The tumor marker gene test was validated in the Johns Hopkins Molecular Diagnostics Laboratory. DNA samples were amplified using an Ampliseq custom primer pool panel sequenced on the Ion Torrent ion semiconductor sequencing platform similar to that used by Abe and colleagues (14). FUT variants were also sequenced using Sanger sequencing as an orthogonal method. Fifteen DNA samples were chosen from samples previously genotyped by Abe and colleagues (14) that contained one or more functionally impairing variants in either FUT2 or FUT3, including the common null variant in the FUT2 gene [rs601338, p.Trp154Ter, 461G > A (G428A in prior reference transcripts), (NM_000511.6)]; the weak FUT2 rs1047781 p.Ile140Phe, A418T variant (previously A385T) common in East Asians, which has five-fold less FUT2 enzyme activity (16); or one or more of seven FUT3 (NM_00149.3) variants (rs28362459, rs812936, rs778986, rs143012663, rs28362463, rs3745635, rs3894326), known to impair function alone, or in combination (17). One of these FUT3 null SNPs, (rs28362463; c.484G>A), was not in the Abe and colleagues study (14); it is rare in European populations, but relatively common in individuals with African heritage (∼12% minor allele frequency). Sanger sequencing was also used to characterize the FUT2 null variant, rs1047781. Prior studies have characterized multiple FUT3 haplotype groups (17), and phase analysis of 1000 Genomes data (18) identified 9 alleles including the wild-type (WT) allele. The population percentage of the variant haplotypes that affect function in the 1000 Genomes database for FUT3 and FUT2 is shown in Supplementary Table S1. An additional allele (FUT3^508A) was found in our patient population but not called by phase analysis of 1000 Genomes. The FUT3 variant rs143012663 (c.445C > A) was not found in the 1000 Genomes data or in our study population. One allele (containing the rs28362459 minor allele without any other variants) has been described as hypofunctional (17). We classified this allele as functional. The possible FUT3 genotypes associated with these haplotypes are shown in Supplementary Table S2. Individuals were assigned to one of four functional groups based on the predicted function of each FUT3 allele and their FUT2 functional status: (i) FUT3 null (homozygous or compound heterozygous for nonfunctional FUT3 alleles); (ii) FUT3 heterozygous (one FUT3 null allele) with intact FUT2; (iii) FUT3 WT (no nonfunctional FUT3 variants), with intact FUT2; or (iv) FUT2 null, with intact FUT3 (individuals homozygous null or compound heterozygous for the weak/null FUT2 variants (rs601338 and/or rs1047781) with at least one FUT3 functional allele. For simplicity and based on the diagnostic cutoffs for each group, we renamed the groups, FUT3 null, FUT-low, FUT-intermediate and FUT-high. (In our prior paper, we referred to these groups as Group A, Group B-het, Group B-wild and Group C.)

All case and control DNA samples were sequenced and interpreted blinded to CA19–9 levels or to any prior genotype data. Libraries were sequenced using ion semiconductor sequencing technology on an Ion Torrent S5XL. Sequencing results were mapped to the reference human genome UCSC version hg19 (NCBI build GRCh37). BAM files were reviewed in IGV and genotypes interpreted and recalled using the IonTorrent Variant Caller software (ITVC, tvc5.10–10). Each individual was then assigned a FUT3 allelic group, and to one of the four FUT functional groups (Supplementary Fig. S1).

All serum samples were measured in the Johns Hopkins Clinical Chemistry Reference Laboratory using the Tosoh Bioscience ST-AIA-PACK CA 19–9 immunoassay method on the AIA 900 analyzer. The reference range of <36 U/mL was determined using healthy individuals.

Summary statistics were used to describe the demographic, clinical and pathology information on all controls and PDAC cases. The PDAC cases were divided into a testing and validation set. Violin plots were used to present the distribution of CA19–9 in the controls based on the different FUT2/3 classes. The Fisher exact test and ANOVA were used to describe differences between FUT2/3 classes among the controls as appropriate. P values <0.05 were considered statistically significant. Estimated upper limits of the reference range for CA19–9 were calculated as 99th percentiles using nonparametric percentile of untransformed values using all data, including values <1 U/mL (considered as zero) with associated 90% confidence intervals (CI) calculated using the binomial distribution and a bootstrap approach. Receiver operating characteristic (ROC) curve analysis was used to evaluate discriminative ability of CA19–9 alone or CA19–9 with tumor marker gene test to distinguish PDAC cases from controls. Corresponding predictive probabilities were calculated from logistic regression models with CA19–9 as a continuous variable used alone or with FUT groups and terms for their interaction. GraphPad Prism version 9.5.0 for Windows (GraphPad Software, San Diego, California) and R version 4.2.1 were used for data analysis (RRID:SCR_001905).

The data generated in this study, including raw used to generate results, are available upon reasonable request from the corresponding author.

To validate the accuracy of the Johns Hopkins Molecular Diagnostics lab's CA19–9 tumor marker gene test at detecting variants in FUT2 and FUT3, a reference set of 15 DNA samples was analyzed that contained a representative set of FUT variants previously sequenced in the Goggins lab (14). Genotype interpretation was performed by J.R. Eshleman, who was blinded as to each sample's genotypes. The reproducibility of the assay was established by performing the next-generation sequencing (NGS) assay in triplicate on multiple days. The genotype calls were completely concordant across replicates, sequencing methods (all variants were confirmed by Sanger sequencing), and with the prior genotype calls on these samples by Abe and colleagues (14).

There were 348 individuals who served as controls. Their characteristics are summarized in Table 1. There were no significant differences in the demographics between the FUT groups. DNA samples from controls were sequenced using the tumor marker genotype test in the Johns Hopkins Molecular Diagnostics lab and their genotypes interpreted independently by J.R. Eshleman, M. Dbouk, and Y. Ando, with the interpretation completely concordant. They were also completely concordant with prior sequencing results performed in the Goggins research lab.

The diagnostic cutoffs for the FUT3-null, FUT-low, FUT-intermediate and FUT-high groups for serum CA19–9 at the 99th percentile were <3, 34.9, 41.8, and 89.2 U/mL, respectively (see Table 2). Figure 1 shows violin plots of the distribution of CA19–9 levels for the overall control group as well as for each of the different FUT groups.

We did not detect significant differences in the level of CA19–9 between other potential genotype subgroups (such as between FUT2 WT versus FUT2 functionally heterozygous controls that were FUT3 WT, or between FUT3 WT versus FUT3 heterozygous controls that were FUT2 null), either in our prior study (14) or in this study (Supplementary Table S3). There was also no significant difference in CA19–9 levels within each FUT group stratified by pancreatic cyst status (present/absent; data not shown). One third of the controls in this study had pancreatic cysts (Table 1). Pancreatic cysts in the controls were representative of what we find in patients who meet criteria for pancreas surveillance (19); almost all were subcentimeter in diameter, without concerning features (note as part of this study design, we excluded subjects with pancreatic cysts with worrisome imaging features).

A summary of the characteristics of the PDAC cases is described in Table 3. The demographics of the controls were somewhat younger than the PDAC cases, reflecting the average age of patients who undergo surveillance (mean age 59.5 ± 11 years), with no significant difference between the groups with respect to gender or ethnicity (Supplementary Table S4). There was also no significant difference in the distribution of individual FUT variants in cases versus controls (Supplementary Table S5). The FUT2/FUT3 variant haplotypes of the cases and controls are listed in Supplementary Table S6.

We did not observe any significant difference in CA19–9 level according to age within any of the FUT groups (data not shown). Assigning each case and control to their appropriate CA19–9 reference range with the tumor marker gene test yielded a diagnostic sensitivity at 99th percentile cutoff of 66.7% (elevated in 157 of 234 PDAC cases in the testing set), with only 4 false-positives among the 348 controls, corresponding to a diagnostic specificity of 98.9% (Table 4; Supplementary Fig. S2). In contrast, using the uniform CA19–9 cutoff (<36 U/mL) for all the testing set PDAC cases yielded a sensitivity of 68%, but a specificity of 91.4%. Among Stage I cases diagnostic sensitivity using tumor marker gene test–determined cutoffs was 51.1% (Table 4). In the validation set of PDAC cases (Table 3), the diagnostic sensitivity of CA19–9 using tumor marker gene test–determined cutoffs was 65.7% (elevated in 88 of 134 cases), at 98.9% specificity (Table 4), compared with 67.2% sensitivity at 91.4% specificity using the uniform CA19–9 (36 U/mL) cutoff. The diagnostic sensitivity for Stage I PDAC cases in the validation set using the tumor marker gene test assigned reference range cutoffs was 41.9%.

Overall, the tumor marker gene test reclassified the CA19–9 result in 7.3% of the PDAC cases compared with CA19–9 alone (10 from negative-to-positive, 17 positive-to-negative). Among controls, 8.6% of subjects were reclassified, most reclassified from false-positive to true-negative (2 from negative-to-positive, 28 positive-to-negative; Supplementary Table S7).

Among all resectable-stage PDAC cases, the diagnostic sensitivity of the CA19–9 tumor marker gene test was 66.4%, and among Stage I PDAC cases, it was 46.4%.

We also performed an ROC analysis to compare the ability of CA19–9 alone or with the tumor marker gene test to discriminate between PDAC cases and controls without using the above cutoffs (Fig. 2). For the combined cohort of the two sets of PDACs and the controls, the AUC for CA19–9 alone was 0.84 (95% CI, 0.81–0.87), while adding the tumor marker gene test to the model increased the AUC to 0.92 (95% CI, 0.90–0.94), a difference of 0.08 (95% CI, 0.05–0.11; P < 0.001). For Stage I PDAC cases, there was a similar AUC difference of 0.08 (95% CI, 0.04–0.13; P < 0.001); 0.77 (0.71–0.84) for CA19–9 alone versus 0.86 (0.81–0.91) when the tumor marker gene test is added to the model.

We find that the diagnostic performance improves significantly when individuals are assigned the CA19–9 reference range that corresponds to their functional FUT group. The FUT group cutoffs include those both below and above the uniform CA19–9 cutoff (36 U/mL), so the FUT gene test identifies one of the main causes of both false-negative and false-positive CA19–9 tests. Assigning diagnostic cutoffs by FUT group yields much significantly higher diagnostic sensitivity at the same ∼99% specificity compared with the usual one size fits all approach to diagnostic cutoffs. Biomarkers are often assigned a uniformly high specificity cutoff, but the reduced diagnostic sensitivity that results arises particularly among patients with smaller, earlier-stage cancers that generate lower biomarker levels, the ones that when treated yield the most survival benefit (20, 21). On the other hand, the false-positives that result when the lower uniform cutoff is used for CA19–9 (our clinical lab cutoff is 36 U/mL) are mainly those that are FUT2-null; the upper limit of reference range for CA19–9 for FUT2-null subjects was 89.2 U/mL. Inevitably, although overall diagnostic performance was improved, the trade-offs between diagnostic sensitivity and specificity also resulted in more false-negatives in cases, mainly in the FUT-high group.

The FUT2 and FUT3 variants included in the CA19–9 tumor marker gene test are inactivating, common, and have large effects on CA19–9 levels. There is considerable interest in identifying and evaluating variants that influence expression of biomarkers and other proteins through genome-wide association studies (GWAS; refs. 10, 22) and other approaches. Many of the identified variants to date have only small biomarker effect sizes, though further study is needed. For pancreatic cancer, variants that associate with CA125, CEA, and CPA1 levels have been evaluated and have been shown to add to diagnostic accuracy, though their effects are more modest (14, 23).

In this study, we set diagnostic cutoffs to achieve 99% specificity so as to limit the number of false-positive tests. False-positives predominate when testing for a low prevalence disease like pancreatic cancer. The reference ranges for CA19–9 in this study were generated from subjects enrolled in the CAPS study, consisting mostly of patients undergoing pancreatic surveillance for their familial/genetic risk, who would be one of the main target populations of such a test. Even in the context of pancreatic surveillance, limiting false-positives is important as they create diagnostic confusion, worry, and unnecessary and potentially burdensome further testing. CA19–9 is not currently recommended as a stand-alone test for patients undergoing pancreatic surveillance, though it is often measured in patients with worrisome pancreatic imaging features as a possible indicator of invasive cancer (24, 25). CA19–9 has also been evaluated in initial studies as part of a multi-marker signature test (26). For the typical patient enrolled for pancreas surveillance such as CAPS with a relatively high lifetime pancreatic cancer risk (∼5% to 20%, depending on family history extent and germline mutation and other factors) the pretest probability of having pancreatic cancer detected at any given time is low. A typical CAPS patient might have a pretest probability of having a pancreas-imaging test-detected pancreatic cancer of only 0.25% (2). In this scenario, a patient with a positive CA19–9 with the tumor marker gene test (sensitivity 66%, specificity 99%) would have a negative predictive value of 99.8% and a positive predictive value for PDAC of ∼20%, with ∼4 of 5 positive CA19–9 tests being false-positive. This example highlights the importance of maintaining high diagnostic specificity cutoffs (∼99%) even in a high-risk cohort without any concerning findings. The consequences of false-positives tend to become more significant as downstream tests are performed, each with their own false-positive rate. It should be noted that the overall diagnostic yield of an elevated CA19–9 test in a CAPS cohort could be modestly higher than it is for PDAC alone as CA19–9 elevations also arise from other cancers (27), as well as from other diseases/abnormalities that can cause an elevated CA19–9 (28, 29).

With the CA19–9 tumor marker gene test available for clinical investigation, the next step is a study evaluating its diagnostic performance as an early detection test, and ultimately its diagnostic yield when added to existing pancreatic surveillance imaging tests. Patients enrolled into pancreas surveillance programs need long-term surveillance; for patients who might undergo regular CA19–9 tests over many years, adding a one-time tumor marker gene test would involve little additional cost. Until its diagnostic performance is better understood, the CA19–9 tumor marker gene test is best applied to patients who are undergoing surveillance pancreatic imaging, as abdominal imaging can identify many of the sources of benign elevations of CA19–9 (such as large hepatic, renal, or splenic cysts, biliary or pancreatic duct dilation associated with inflammation, etc.). Some of these diseases that cause elevated CA19–9 are uncommon in the CAPS surveillance setting, and they often cause symptoms (e.g., cholangitis, acute pancreatitis; ref. 30), and require diagnostic evaluation anyway. The clinical scenarios surrounding these diseases are very different to the scenario of an asymptomatic HRI undergoing tumor marker surveillance who presents with a positive tumor marker test, though it is important to evaluate the impact of false-positive tests in this setting in future studies

The goal of pancreatic surveillance is to detect early-stage pancreatic cancer, preferably when it is a small asymptomatic Stage I PDAC, or when there is high-grade dysplasia without invasive cancer (31), as intervening at these points in the natural history offers the best opportunities for cure. CA19–9 has also been evaluated in cohort study biorepositories (32–35), where evidence reveals elevations beginning mainly in the year prior to diagnosis, consistent with preclinical cancer progression. Perhaps the best criterion of a blood test for early detection is its diagnostic performance for Stage I PDAC. In our study, the CA19–9 tumor marker gene test had an overall diagnostic sensitivity for Stage I PDAC of 46.4%, at 98.9% specificity. Pancreatic imaging with EUS and/or MRI/CT detects many Stage I PDACs when the CA19–9 level is below even the tumor marker gene test defined cutoffs, so until more sensitive blood-based tests are available, pancreatic imaging will need to be the primary modality of pancreatic surveillance for CAPS subjects. How best to incorporate blood-based testing for patients undergoing pancreatic imaging surveillance remains to be determined, but one option would be to time the test in between annual pancreatic imaging tests.

There is allelic variation across different ancestral populations in the FUT3 and FUT2 null variants (rs601338 and rs1047781; refs. 36, 37). Genetic variation across populations is likely one of many factors influencing the diagnostic performance of CA19–9 when using a uniform cutoff. The tumor marker gene test can be expected to limit this source of variation. Some FUT2/FUT3 null variants are limited to certain population groups, though the overall prevalence of FUT3 and FUT2 null subjects in different populations does not vary nearly as much as individual variants. Common variants are often under evolutionary pressure, and there is evidence that FUT2 null alleles are under selection (38).Applying a FUT-based tumor marker gene test to other populations should take into account local variant prevalence.

There are limitations to this study. Our study was retrospective in nature. It allowed us to select similar numbers of controls from each of the functional FUT groups for the purposes of generating a reference range from our CAPS study population. As a result, the number of subjects in each functional group somewhat differs from the general population (for example, FUT2-null subjects represented a somewhat higher percentage of our controls, 27%, compared with the ∼20% to 22% prevalence in the U.S. population).

In conclusion, CA19–9 achieves greater diagnostic performance when combined with a tumor marker gene test that assigns individuals to the appropriate reference range for their FUT genotype group. The CA19–9 tumor marker genotype test merits prospective evaluation in HRIs undergoing pancreatic surveillance.

J. He reports grants from AbbVie outside the submitted work. L. Sokoll reports grants from NIH during the conduct of the study. No disclosures were reported by the other authors.

M. Dbouk: Formal analysis, investigation, visualization, writing–original draft, writing–review and editing. T. Abe: Investigation, visualization, writing–review and editing. C. Koi: Investigation, visualization, methodology, writing–review and editing. Y. Ando: Investigation, visualization, writing–review and editing. H. Saba: Data curation, writing–review and editing. E. Abou Diwan: Data curation, writing–review and editing. A. Macgregor-Das: Investigation, visualization, writing–review and editing. A.L. Blackford: Formal analysis, writing–review and editing. E. Mocci: Software, formal analysis, writing–review and editing. K. Beierl: Formal analysis, investigation, writing–review and editing. A. Dbouk: Data curation, writing–review and editing. J. He: Resources, writing–review and editing. R. Burkhart: Resources, writing–review and editing. A.M. Lennon: Resources, writing–review and editing. L. Sokoll: Investigation, writing–review and editing. M.I. Canto: Resources, supervision, writing–review and editing. J.R. Eshleman: Formal analysis, investigation, methodology, writing–review and editing. M. Goggins: Conceptualization, resources, supervision, funding acquisition, writing–original draft, project administration, writing–review and editing.

This work was supported by NIH grants (U01210170, R01CA176828 CA62924), Susan Wojcicki and Dennis Troper, and by a Stand Up To Cancer–Lustgarten Foundation Pancreatic Cancer Interception Translational Cancer Research Grant (Grant number: SU2C-AACR-DT25–17). The indicated SU2C research grant is administered by the American Association for Cancer Research, the scientific partner of SU2C.

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.