Detection of Chemotherapy-resistant Pancreatic Cancer Using a Glycan Biomarker, sTRA

Systemic therapy is considered necessary for all patients with pancreatic ductal adenocarcinoma (PDAC), even those with localized disease, because most patients already have occult metastases at the time of diagnosis (1). Chemotherapy particularly benefits patients who have surgical resection of the tumor. In a seminal study that established adjuvant chemotherapy (chemotherapy applied after surgery) as standard of care for PDAC, the median disease-free survival after surgery improved to 13.4 months with gemcitabine from 6.9 months with observation (2). Further improvements in chemotherapy were demonstrated using the stronger FOLFIRINOX regimen (3), or gemcitabine in combination with capecitabine (4) or nab-paclitaxel (5). Systemic therapy is increasingly applied prior to surgery, called neoadjuvant therapy, to increase the percentage of patients who receive chemotherapy (6), because some patients have a delay or reduction in their adjuvant chemotherapy as a consequence of surgery. Neoadjuvant therapy could have the additional advantages of identifying patients with rapid progression who would not benefit from surgical intervention, treating occult metastases earlier, and downsizing the tumors to increase the chance for a margin-free resection (6).

While the combination of surgery plus systemic therapy results in significant benefit relative to surgery alone, a subset of PDACs is highly resistant to systemic therapy. Nearly 40% of patients receiving surgery plus gemcitabine monotherapy experience relapse within 1 year of surgery (2). Even in the subset of fit patients who are candidates for more aggressive chemotherapy regimens, more than 25% relapse within 1 year (3). Currently, identifying this chemotherapy-resistant cohort prior to treatment remains a challenge, because conventional imaging, liquid biopsy, and molecular biomarkers are lacking.

The gene expression subtypes defined in prior research (7–9) potentially provide some guidance to this problem. The consensus subtypes have been termed as classical, basal (also referred to as quasi-mesenchymal), and exocrine, terms chosen to reflect the normal cell types that most closely correspond to the cancer cells. In retrospective evaluations of outcomes following curative resection, tumors with transcriptome profiles matching with the classical subtype had longer survival than the others (7–9). Likewise, among patients with metastatic PDAC, the classical subtype was associated with longer survival in retrospective analyses (10, 11). On the other hand, patients with the classical subtype demonstrated no benefit from adjuvant chemotherapy (9, 12), in contrast to patients with the basal type, and cell lines of the classical subtype were more resistant to chemotherapy than those of the basal type (8). Therefore, the predictive role of molecular subtyping in PDAC treatment remains to be established.

We recently identified a new biomarker of PDAC, which is a cell surface and secreted glycan, called sialylated tumor-related antigen (sTRA; refs. 13–15). Nothing was known about the relationship of this biomarker to subtype of cancer or drug resistance, but some clues were evident. A different subset of PDACs express sTRA, and the ones that do have outwardly different characteristics. The tumors primarily expressing sTRA tended to be sparse, poorly differentiated, or highly vacuolated, while those expressing mainly CA19-9 were part of well-differentiated or moderately differentiated secretory glands (14). These facts suggested that the two groups represent distinct subtypes of tumors having differing biology and clinical behavior. Of particular interest was the possibility that the drug-resistant group observed in clinical care corresponds to a glycan-defined subtype.

The PaTu-8988S and PaTu8988T cell lines were obtained from Creative Bioarray, and Colo357, L3.3, and L3.6PL cell lines were kindly provided by Dr. Isaiah J. Fidler (University of Texas, MD Anderson Cancer Center, Houston, TX). The remaining cell lines were obtained from the ATCC. All cell lines were cultured in RPMI1640 supplemented with 5% FBS, 2 mmol/L l-glutamine, and 100 IU/mL penicillin/streptomycin. The cells were grown at 37°C in a humidified atmosphere supplemented with 5% (v/v) CO₂. All cell lines were within 10 passages of collection when used in the described experiments. The authenticities of the cell lines were confirmed by comparing their RNA expression profiles with those of previously authenticated cell lines and/or by short tandem repeat profiling (ATCC). Cross-contamination between cell lines was excluded by Infinium QC Array (see Supplementary Materials and Methods). All cell lines were Mycoplasma free, as determined by RNA sequencing (RNA-seq) profiling and DAPI staining.

The chemotherapeutic reagents, cisplatin, etoposide, gemcitabine, and 5-fluorouracil (5-FU), were obtained from Sigma. Irinotecan, oxaliplatin, and paclitaxel were obtained from Cayman Chemical. All drugs were dissolved in DMSO or dimethylformamide. For the preparation of FOLFIRINOX, 5-FU was first prepared in dimethylformamide, and then leucovorin, irinotecan, and oxaliplatin were each added in a 1:5 ratio by weight to 5-FU. The concentration of FOLFIRINOX was calculated on the basis of the 5-FU concentration. Cells were seeded into 96-well plates at 2 × 10³ cells per well and cultured for 3 days before treatment with a drug or drug mixture at six different concentrations each. After 3 days, cell viability was estimated using CellTiter-Glo (Promega). The IC₅₀ values were calculated using GraphPad Prism 6 (GraphPad Software) with 5-parameter, variable-slope fits.

All immunofluorescent and chemical stains were performed using 5-μm-thick sections cut from formalin-fixed, paraffin-embedded blocks. Paraffin was removed from the sections by using CitriSolv Hybrid (04355121, Thermo Fisher Scientific), and tissue was rehydrated through an ethanol gradient. Following rehydration, antigen retrieval was achieved through incubating slides in citrate buffer at 100°C for 20 minutes. Slides were blocked in PBS with 0.05% Tween-20 (PBST0.05) and 3% BSA for 1 hour at room temperature. Primary antibodies (Supplementary Table S4) were labeled for immunofluorescence staining with either Sulfo-Cyanine5 NHS Ester (13320, Lumiprobe) or Sulfo-Cyanine3 NHS Ester (11320, Lumiprobe), so that two primary antibodies could be used simultaneously in each round of staining. Dialysis was performed following labeling to remove unreacted conjugate, and the primary antibodies were then diluted into the same solution of PBST0.05 with 3% BSA to a final concentration of 10 μg/mL. Slides were incubated overnight with this solution at 4°C in a humidified chamber.

The following day, the antibody solution was decanted, and the slides were washed twice in PBST0.05 and once in 1 × PBS, each time for 3 minutes. The slides were dried via blotting, and then incubated with DAPI at 10 μg/mL in 1 × PBS for 15 minutes at room temperature. Two 5-minute washes were performed in 1 × PBS, and then slides were cover-slipped and scanned using a fluorescence microscope. All slides were scanned for fluorescence using either Vectra (PerkinElmer) for the tissue microarrays (TMA) or the Axio Scan.Z1 (Zeiss) for the organoid sections. Each system collected data for each field of view at three different emission spectra. All image data were quantified using the SignalFinder-IF software (16).

Following scanning, slides were stored in a humidified chamber. Coverslips were removed for the subsequent rounds of staining by submerging the slide in deionized water at 37°C until the coverslip floated free (between 30 and 60 minutes). Fluorescence was quenched between rounds by incubating the slides with 6% H₂O₂ in 250 mmol/L sodium bicarbonate (pH 9.5–10) twice for 20 minutes at room temperature. Subsequent incubations and scanning steps were repeated as described above with different primary antibodies.

To perform sialidase treatment, slides were incubated with a 1:200 dilution (from a 50,000 U/mL stock) of α2-3,6,8 neuraminidase (P0720L, New England Biolabs) in 1 × reaction buffer (5 mmol/L CaCl₂ and 50 mmol/L sodium acetate, pH 5.5) overnight at 37°C. Slides were washed as described above, followed by subsequent antibody detections. Hematoxylin and eosin staining was performed following a standard protocol.

The immunoassays were based on the method described previously (15). The capture antibodies were CA19-9, anti-MUC5AC, and anti-MUC16, and the biotinylated primary antibodies were CA19-9 or TRA-1–60 (details in Supplementary Table S4). The secondary detection agent was Cy5-conjugated Streptavidin (Roche Applied Science). We diluted the samples of human plasma (8- and 32-fold) and cell line or organoid media (2- and 8-fold) into a buffer (1 × PBS with 0.1% Tween-20, 0.1% Brij-35, species-specific blocking antibodies, and protease inhibitor) and incubated each sample on an antibody array overnight at 4°C.

For sTRA detection, an extra step of enzyme treatment, before antibody detection, was needed. After sample incubation, we prepared α2-3 neuraminidase (P0728L, New England Biolabs) at a concentration of 250 U/mL in the supplied reaction buffer and incubated on arrays overnight at 37°C. We incubated each array with a biotinylated antibody (3 μg/mL in 1 × PBS with 0.1% Tween-20 and 0.1% BSA) and subsequently with Cy5-conjugated Streptavidin (43-4316, Invitrogen; 2 μg/mL in the same buffer as the primary antibody). The slides were scanned for fluorescence at 635 nm using a Microarray Scanner (Innopsys InnoScan 1100 AL). The quantification of fluorescence was performed using SignalFinder-MA (16). All plasma and media samples were repeated in at least three independent experiments. The CA19-9 values were obtained through the clinical laboratory services at the Medical College of Wisconsin (MCW, Milwaukee, WI) for the discovery set, and they were determined using a custom immunoassay previously validated in the Haab laboratory (15) for the test set.

Differences between marker-positive and marker-negative cell lines in IC₅₀ values and percent viability were tested using the Mann–Whitney U test. Differential expression in the RNA-seq data was tested using empirical Bayes quasi-likelihood F-tests, and the P values were adjusted using the Benjamini–Hochberg method. Differences in overall survival (OS) between patient groups in the survival analyses were evaluated with the log-rank test. Differences between patient groups in proportions of patients with long or short OS were analyzed with the Fisher exact test and the Breslow–Day test for homogeneity of ORs. Differences in sensitivity and the average of sensitivity and specificity was analyzed with the Wald test based on bootstrap SE estimate. P values of less than 0.05 were considered significant.

The plasma samples were collected under protocols approved by the Institutional Review Boards at the University of Pittsburgh Medical Center (Pittsburgh, PA) and the MCW (Milwaukee, WI). The donors consisted of patients diagnosed with pancreatic cancer who were scheduled to undergo neoadjuvant therapy. The plasma collections (EDTA plasma) took place prior to any surgical, diagnostic, or medical procedures, and were performed according to the standard operating procedure from the Early Detection Research Network. The samples were frozen at −70°C or colder within 4 hours of time of collection. Aliquots were shipped on dry ice and thawed no more than three times prior to analysis.

The tissue samples for TMAs were collected under approved protocols at the Medical University of South Carolina (Charleston, SC), University of Pittsburgh Medical Center (Pittsburgh, PA), and Memorial Sloan Kettering Cancer Center (New York, NY). All subjects provided written informed consent, and all methods were performed in accordance with an assurance filed with and approved by the U.S. Department of Health and Human Services and in accordance with the guidelines contained in the Belmont Report.

We initially tested for differences among PDACs classified by glycan expression using a panel of 27 cell lines. We classified each cell line on the basis of the sTRA glycan or the CA19-9 glycan (Fig. 1A). Both glycans are capped with sialic acid on type-1 N-acetyl-lactosamine (LacNAc), the disaccharide of galactose links β1,3 to N-acetyl-glucosamine (GlcNAc), and the CA19-9 epitope has a fucose attached to the GlcNAc, which is necessary for its recognition by selectin receptors. Type-1 LacNAc, as recognized by the TRA-1-60 antibody (17), is a marker for induced pluripotent stem cells, but the sialylated version has not been well studied because of lack of an effective antibody. We indirectly detected the sialylated structure using sialidase to uncover the TRA-1-60 epitope (Fig. 1A).

The cell surface expression of sTRA and CA19-9 was variable among the cell lines, with some primarily expressing only one glycan and others expressing both or neither (Fig. 1B and C). Organoid models of pancreatic cancer (18, 19) likewise showed variable expression of one, both, or neither of the glycans (Fig. 1D), with slightly different proportions among them (Fig. 1E).

Using the cell lines and organoids, we then asked whether similar differences exist in gene transcription programs. A total of 267 genes were differentially expressed between the sTRA-expressing cell lines (not including the three cell lines also expressing CA19-9) and all others (Fig. 2A; Supplementary Table S1). No individual genes were differentially expressed between the CA19-9 and sTRA groups at P < 0.05 after multiple hypothesis correction, possibly due to the lower number of CA19-9–positive lines. The sTRA-associated genes had ontologies that were enriched in developmental, drug metabolism, and glycan biosynthesis pathways (Fig. 2A). The developmental gene, BMP4, was a strong individual marker of sTRA cells, as was CYP3A5 (Fig. 2A), a gene previously identified as a marker of PDACs identified as classical and exocrine (20). In addition, nine of 14 sTRA-expressing cell lines were identified as classical, on the basis of the gene classifier called PDAssigner (8), compared with two of six CA19-9–expressing and 0 of 10 glycan-negative cell lines (Fig. 2A), suggesting that sTRA is more likely to recognize the classical subtype. Gene set enrichment analysis showed that sets defining stem-like differentiation, stem-like metabolism, and the classical subtype were enriched in the sTRA-expressing cells (Fig. 2B; Supplementary Table S1). Among individual genes that have been proposed as markers of classical, the expressions of GATA6 and CYP3A5 were higher in the sTRA cells (Fig. 2C). KRT81 and HNF1A showed weak associations with sTRA.

The epithelial/mesenchymal state of cancer cells has been widely explored as an indicator of their origin, invasiveness, or overall tumor-forming aggressiveness. All six of the CA19-9–expressing cell lines were epithelial, as determined by the gene expression of Zeb-1 and E-cadherin (Fig. 2C) and by morphology (Supplementary Fig. S1), but the sTRA-expressing cells and those expressing neither glycan were of various types (Fig. 2C). All the organoid models had epithelial morphologies, but three of the models had mesenchymal characteristics by gene expression (not shown). Two of these produced sTRA exclusively and the third produced neither glycan. The data from both model systems suggest that some sTRA-expressing cancer cells have the potential for mesenchymal-like differentiation, in contrast to CA19-9–expressing cells.

The type of KRAS mutation in cancer potentially can drive differences in phenotype (21). The less common Q61 alteration appeared exclusively in the cell lines and organoids that expressed only sTRA (Fig. 2D). The G12V mutation was present in three of 11 cell lines and one of three organoids that expressed only sTRA, in comparison with 0 of three cell lines and one of eight organoids that expressed only CA19-9. While these observations are based on relatively small sample sizes, they suggest that the Q61H/R mutation fosters cancers that express sTRA in the absence of CA19-9.

We determined the resistance of the 27 cell lines to eight chemotherapeutics that are either first-line or alternative treatments against pancreatic cancer. Our hypothesis was that sTRA-positive cancers are more treatment resistant than those that are sTRA negative. In a single-dose study, the sTRA-expressing cell lines were more resistant than the sTRA-negative cell lines in each case (Fig. 3A). In dose–response analyses to obtain the IC₅₀ concentrations (Fig. 3B and C; Supplementary Fig. S2), the sTRA cells were significantly more resistant than the non-sTRA cells (P < 0.05, Mann–Whitney U Test) to six of the eight drugs. For gemcitabine, the resistance was higher in the sTRA group, but with less statistical significance (P = 0.07, Mann–Whitney U Test); for oxaliplatin, resistance was similar between the groups. As a negative control, we performed the same analysis for CA19-9, because it is a glycan that is not a biomarker for resistance. CA19-9 did not define a resistant group (Supplementary Fig. S2), indicating that we were not nonspecifically detecting a general elevation in glycans.

We asked whether the high resistance corresponded to traits that have been associated with resistance. The sTRA-positive cell lines had higher levels of drug metabolizing enzymes from the cytochrome P450 family (P < 0.05, Mann–Whitney U test), and they had trends toward higher levels of the stem marker, ALDH1A3, and longer doubling times (Fig. 3D). Thus, some sTRA cell lines have resistance traits, but the mechanism of resistance may differ between cell lines.

We further tested the above relationships using sets of isogenic cell lines, where all cell lines in a set were from the same individual. We repeatedly cultured the L3.3 cell line in sublethal concentrations of each of four drugs, followed by recovery and outgrowth of the surviving cells, and found that sTRA expression was increased in several of the sublines, both in the percentage of stained cells and in total staining intensity. The sublines with higher sTRA coincided with significantly increased resistance to the drugs (Supplementary Fig. S3).

Next, we tested whether the sTRA levels in primary tumors are associated with resistance to systemic chemotherapy. We determined sTRA levels in two ways, by a gene expression classifier and by immunofluorescence. To develop a gene expression classifier, we identified the significantly upregulated or downregulated genes (P < 0.02, Bayes quasi-likelihood F-test after multiple testing correction) associated with sTRA expression in the panel of 27 cell lines (Supplementary Table S1) and used the algorithm from PDAssigner (8) to assign classes. We used this algorithm because of its previous robust performance and its simplicity for adoption with new gene sets. We applied the classifier to 150 cases of PDAC from The Cancer Genome Atlas (TCGA; ref. 22) and to 180 cases from the International Cancer Genome Consortium (ICGC; ref. 23) that had survival information. In both cohorts, distinct groups of patients showed overall differences in expression between genes associated with high sTRA and those associated with low sTRA (Fig. 4A). Tests of group differences in central tendency using the classifier genes showed significance (P = 0.001, Adonis test, Vegan R package). This finding confirmed the consistency between the cell lines and both cohorts in the differential expression of the gene groups, and it supports the idea that the classifier identifies true subtypes, rather than random variation in expression patterns.

We assigned the patients to an sTRA or non-sTRA group on the basis of the median of the calculated score of the classifier. No difference in OS was evident between the sTRA and non-sTRA groups of patients, but among the patients assigned to the non-sTRA group, those receiving adjuvant therapy had significantly longer OS than those who did not (P < 0.001, log-rank test; Fig. 4B). Among the patients assigned to the sTRA group, no difference was observed. Both TCGA and ICGC datasets showed this relationship. An analysis of progression-free survival instead of OS showed the same results, but on a shorter timescale (Supplementary Fig. S4), indicating that OS results were not biased by differences in treatments applied after progression. Other classifiers for the classical subtype gave similar results (Fig. 4C), but not as consistently between datasets as the sTRA classifier.

In a parallel approach, we asked whether the directly measured amount of sTRA in primary tumors associated with a lack of response to adjuvant therapy. This experiment used TMAs that included tumor tissue collected at the Memorial Sloan Kettering Cancer Center (New York, NY) from patients who either did or did not receive adjuvant therapy and who had long (>3 years) or short (<1 year) OS following surgery (ref. 24; Supplementary Table S2). The adjuvant chemotherapy consisted mainly of gemcitabine or fluorouracil, based on the institutional treatment paradigm at the time.

We measured sTRA and CA19-9 in the TMAs using multimarker immunofluorescence (14, 25) in conjunction with previously developed software that enables unbiased, automated quantification of multimarker immunofluorescence data (refs. 14, 16; Supplementary Fig. S5A; Supplementary Table S2). The sTRA and CA19-9 levels showed little correlation with each other (Supplementary Fig. S5B), but the sTRA levels were higher in the short OS group (P = 0.0077, Wilcoxon rank-sum test; Supplementary Fig. S5C). Among the patients receiving adjuvant chemotherapy, those with high sTRA showed a significantly lower proportion of long OS than those with low sTRA (P = 0.003, Fisher exact test; Supplementary Fig. S5D). Among the patients with high sTRA, those receiving adjuvant chemotherapy had a significantly lower proportion of long OS than those not receiving neoadjuvant chemotherapy (P = 0.01, Fisher exact test). No other comparison showed a significant difference. The Breslow–Day test for homogeneity of ORs (for association between survival and therapy) across biomarker-defined subgroups was highly significant for sTRA (P = 0.006), but was not significant for CA19-9 (P = 0.18). These results indicate a differential effect of adjuvant therapy on patient survival between the sTRA-high and sTRA-low groups.

Another pair of TMAs provided complementary information, in that the tumors had been exposed to neoadjuvant therapy. The TMAs included tumor tissue from PDAC resections (14) performed at the University of Pittsburgh Medical Center (Pittsburgh, PA). We found that the tumors that were dominated by cells producing only sTRA or only CA19-9 were in the short OS group, with few exceptions, while tumors without such clonal outgrowth were evenly distributed in the short OS and long OS groups (Supplementary Fig. S6; Supplementary Table S2). This observation suggests that the persistence of sTRA-dominant cells following neoadjuvant chemotherapy portends poor outcome.

The above findings presented the possibility that high plasma sTRA identifies PDACs that do not benefit from systemic chemotherapy. We investigated this possibility using plasma samples from patients scheduled to receive neoadjuvant therapy at the MCW (Milwaukee, WI). The patients received treatments as determined by stage of disease, resectable cancer receiving chemoradiation, and borderline resectable cancer receiving chemotherapy followed by chemoradiation. The chemotherapy consisted mainly of FOLFIRINOX or gemcitabine/nab-paclitaxel, as determined by age and performance status. No association with outcome was detected for any particular regimen. Many of the patients received additional adjuvant chemotherapy after surgery, but the use of additional treatment was not associated with outcome.

We measured the plasma levels of sTRA glycan using sandwich immunoassays, in which we detected sTRA on the proteins captured by one of three different capture antibodies (Fig. 5A). The use of this assay as a surrogate for tumor sTRA was supported by separate analyses. First, the agreement of the cell surface expression of the glycan and the amount in the conditioned media was high for both the cell lines and organoids (Supplementary Fig. S7). Furthermore, in a previous study, we found that the peripheral blood glycans correlated with tumor glycans for cell line xenograft mouse models, patient-derived xenograft mouse models, and human patients with PDAC (14).

We asked whether any of the biomarkers could serve as an indicator of short time-to-progression (TTP), with disease progression diagnosed on the basis of CT scans at 3–4 months intervals for the first 2 years and at 6 months intervals thereafter. We dichotomized the patients using a cutoff of 18 months from the time of diagnosis (Supplementary Table S3), based on the approximate rate of 50% recurrence within 1 year after the completion of treatments and on the clinical observation that such recurrence strongly suggests treatment resistance. We did not dichotomize by pathologic response (based on surgical pathology from resection), as it is a subjective assessment that does not correlate with TTP or OS, or by radiographic response prior to surgery, because it is weakly associated with TTP and OS.

In a discovery set, two of the sTRA immunoassays were significantly higher in subjects with short TTP than in subjects with long TTP (Fig. 5B). Patients elevated in two or more of the sTRA assays (using thresholds optimized for each marker; Fig. 5B) were especially likely to have short TTP, as 16 of 17 had short TTP (94% positive predictive value; Fig. 5C; Table 1). To minimize the effect of overfitting on the estimate of panel performance, we further assessed the panel performance using 5-fold cross-validation with 200-fold bootstrapping (resamplings of the cohort). The improvements in cross-validated sensitivity and the average of sensitivity and specificity was statistically significant (P < 0.0001, Wald test based on bootstrap SE estimate; Table 1).

We then applied the panel to a blinded test set. We applied the thresholds that were derived from the discovery set to the test set and made case/control calls on the blinded samples. The calls were sent to the collaborators who collected the samples, and upon comparison with the true outcomes data, the result was 96% specificity (27/28 with long TTP were high in 1 or less) and 56% sensitivity (15/27 with short TTP were high in 2 or more; Table 1). The improvements in sensitivity and the average of sensitivity and specificity was statistically significant (P < 0.0001, Wald test based on 1,000-fold bootstrap SE estimate; Table 1).

Adjustments to the individual marker thresholds (naïve performance) gave 94% positive predictive value (17/18 with two or more marker elevations had short TTP), 96% specificity (27/28 with long TTP), and 63% sensitivity (17/27 with short TTP; Fig. 5D; Table 1). The three individual sTRA assays showed strong associations with short TTP (P = 0.008 to 0.00008, Mann–Whitney U test), as did CA19-9 (P = 0.007; Supplementary Fig. S8). The sTRA panel also identified differences in TTP in Kaplan–Meier analysis (Fig. 5E). Kaplan–Meier analysis is appropriate here because the cohort was a random selection of the patients seen in the clinic. In both sets, patients positive in the panel had shorter TTP than the rest of the patients. In the test set, the difference was highly significant (P < 0.0001, log-rank test) for sTRA and moderately associated (P = 0.04) for CA19-9.

To further test these associations, we performed additional analysis on data from a previous study of these markers (15) that used plasma collected at the University of Pittsburgh Medical Center (Pittsburgh, PA). We obtained outcomes for a subset of patients who were scheduled to receive neoadjuvant therapy (data in Supplementary Table S3). The study was not designed for this question, but nevertheless could provide insights. In two separate cohorts, two of the sTRA assays trended with short OS (P = 0.05 and 0.06; Supplementary Fig. S8). CA19-9 showed no such trend in either cohort. These findings substantiate the use of a blood test for sTRA to identify a subtype of pancreatic cancers that is resistant to chemotherapy.

This research demonstrates the use of the sTRA glycan to identify the PDAC cases that are highly resistant to chemotherapy. The precise performance of the biomarker for treatment prediction will need to be determined and validated in larger, prospective studies, but the work here establishes the relationship and the validity of the finding in clinical samples, as well as model systems. The immediate implication of this result relates to the development of treatment plans for patients with resistant PDAC. For patients with resectable PDAC, potentially morbid operations could be avoided if rapid relapse following surgery could be predicted a priori. For patients with metastatic disease and patients undergoing neoadjuvant therapy, a practical biomarker could guide the choices and comparisons of the treatment options. For example, FOLFIRINOX is suggested to be slightly better than gemcitabine for the classical subtype (10, 26), possibly indicating a difference between the sTRA-positive and sTRA-negative types. Patient stratification could also improve the selection of patients that receive nab-paclitaxel (5). Furthermore, the sTRA assay could be used in subgroup analyses of the many human trials currently underway, given that many trials do not meet primary objectives, but show evidence of efficacy in subgroups. Trials could involve targeted therapies suggested from studies on the cell culture and organoid models that are available for sTRA-positive and sTRA-negative PDACs. Thus, the biomarker could have value for patient stratification using current options, as well as for research using model systems and in biomarker-guided drug trials.

A blood test has particular value in the clinical setting because physicians could stratify patients prior to any treatments, without a biopsy. Furthermore, a blood test would capture secretions from the whole tumor, rather than just the cells that are sampled by a biopsy, which may not reflect the heterogeneity of the tumor. Various blood tests have potential value for detecting or diagnosing PDAC, including mutated DNA in the circulation (27, 28), tumor exosomes, and metabolites (29–31), but they do not predict therapeutic responses. Highly elevated CA19-9 in the blood and the failure to drop to normal levels following neoadjuvant therapy or surgery (32, 33) are unfavorable prognostic factors (34), but its value is as a tumor volume indicator, rather than as a subtype indicator or a means to predict resistance to chemotherapy (32). The sTRA marker, in contrast, differentiates biological subtypes. As with any new test, blinded, prospective studies using a clinical assay will be required to fully assess the value of the sTRA test to patients and physicians.

The sTRA and CA19-9 tests could be used in combination, however, in the surveillance setting (15). For example, if a patient is high in either sTRA or CA19-9 (or both), the patient would be further evaluated to confirm or negate a diagnosis of cancer. In the event of a diagnosis of pancreatic cancer, the sTRA-elevated cancers would be predicted to be more resistant to chemotherapy than the sTRA-nonelevated cancers. The cutoff used for sTRA would likely be different in the surveillance setting. The previous study of sTRA in the surveillance setting (15) used cutoffs that captured about 65% of the patients, whereas cutoffs optimized for treatment response prediction captured a smaller percentage, suggesting that both resistant and sensitive PDACs would be detected by the combined sTRA and CA19-9 panel.

In the use of sTRA for treatment response prediction, a current limitation is that some of the resistant PDACs would be missed because they do not make the sTRA antigens. In our study of patients receiving neoadjuvant therapy (Fig. 5), 47% of the combined cohorts had short TTP, but 25% of the patients were high in plasma sTRA, working out to 53% of the subjects with short TTP. Consistent with this observation, one of the very resistant cell lines, AsPC-1, makes cell surface sTRA, but does not secrete it well (Supplementary Fig. S7), suggesting that a subset of the resistant tumors do not secrete the antigen in enough quantity to be detectable in blood. Nevertheless, the test picks up a substantial proportion of the resistant cancers and could represent a valuable step forward that can be built upon. Additional research could test for biomarkers to detect the remainder of the resistant PDACs.

This work extends previous findings relating to the prognostic and predictive value of the molecular subtypes. The classical subtype in previous research indicated a better prognosis than the basal subtype, but, on the other hand, it tended to benefit less from chemotherapy (9, 12). This result is also consistent with an in vitro study in cultured cells, which indicated that classical subtype cells were more resistant to gemcitabine (8). But the relative value for prognosis versus treatment prediction was not clear, and a recent study suggested that the classical subtype is more sensitive to chemotherapy than the basal subtype (11). The differences between the studies may result from several sources. The latter study included nonresectable, advanced PDAC (the COMPASS trial), while the former studies involved resectable PDAC. Advanced cancer could be less responsive in general than localized disease. The COMPASS study also used a different classifier that was more stringent than the original, and the study was not designed to distinguish prognostic from predictive value, because it did not include a nontreatment control group. Overall, the results indicate that native prognosis and sensitivity to chemotherapy are not necessarily linked, and that the classical subtype is suboptimal for decoupling these traits.

The sTRA subtype seems to distinguish the traits better: it was indicative of chemotherapy resistance, but not of a poor prognosis. It encompassed resistant cancers of both subtypes and was more consistent than the classical/basal system in identifying resistance in the primary tumors. A valid model is that the sTRA subtype more precisely identifies resistant tumors, but the classical subtype could be more effective for stratifying by native prognosis. Ultimately, the typing of cancers and prediction of drug responses could involve both glycans and other types of markers. Future studies should focus on clarifying additional markers that are suggestive of other subtypes. The genes HNF1A, CDH17, LGALS4, and CYP3A5 have been variously assigned as markers of nonbasal subtypes, including exocrine and classical, but without good agreement between studies (35). The elevation of these genes in most sTRA cancers could indicate that a third subtype is at least partially encompassed by sTRA.

Building on these findings, our next steps will involve the validation of clinical assays for prospective studies and the analyses of model systems to understand the susceptibilities of sTRA-positive cancers. On the basis of the gene expression results, a successful path may involve metabolic approaches (36). Alternatively, probing the sTRA-positive subtype for dependencies on particular nutrient sources may be feasible. These directions in research are made possible because we now have a practical assay to detect chemoresistant PDAC using either tissue or plasma. The use of such an assay in model systems, and then in clinical specimens to detect and follow the resistant subtype, should help both the development and the application of new treatments against PDAC.

L. Wisniewski reports grants from NCI during the conduct of the study. Y. Liu reports grants from NCI during the conduct of the study. D. Plenker reports other from German Research Foundation (DFG) and grants from Lustgarten Foundation during the conduct of the study. R. Drake reports grants from NIH during the conduct of the study. R.E. Brand reports grants from NCI during the conduct of the study, and grants from Immunovia and Freenome, Inc. outside the submitted work. D.A. Tuveson reports other from Leap Therapeutics, Surface Oncology, and Cygnal Therapeutics, grants from Mestag Therapeutics, ONO, and FibroGen, and personal fees from Merck outside the submitted work. B.B. Haab reports grants from NCI, Lustgarten Foundation, and German Research Foundation during the conduct of the study, as well as has a patent for 15/216,768 pending. No disclosures were reported by the other authors.

C. Gao: Conceptualization, investigation. L. Wisniewski: Investigation, methodology. Y. Liu: Investigation. B. Staal: Investigation. I. Beddows: Formal analysis. D. Plenker: Resources. M. Aldakkak: Resources. J. Hall: Software, investigation. D. Barnett: Formal analysis. M. Kheir Gouda: Resources. P. Allen: Resources. R. Drake: Resources. A. Zureikat: Resources. Y. Huang: Formal analysis. D. Evans: Resources. A. Singhi: Resources. R.E. Brand: Resources. D.A. Tuveson: Resources. S. Tsai: Resources. B.B. Haab: Conceptualization, resources, formal analysis, supervision, funding acquisition, writing-original draft.

We thank the core services at the Van Andel Institute for expert support of this research, particularly the Optical Imaging, Bioinformatics and Biostatistics, Genomics, and Pathology and Biorepository cores. We thank Zachary Klamer, MS, at the Van Andel Institute for support in the processing and analysis of the plasma biomarker data; Toshinori Hinoue, PhD, at the Van Andel Institute for advising on the analysis of TCGA and ICGC datasets, and Christine Decapite at the University of Pittsburgh Medical Center for assistance with data coordination. This work was supported by grant no., U01CA152653, Early Detection Research Network, NCI (to R.E. Brand and B.B. Haab); U01CA226158, Alliance of Glycobiologists for Cancer Detection, NCI (to B.B. Haab); D.A. Tuveson was supported by Distinguished Scholar and Director of the Lustgarten Foundation–Designated Laboratory of Pancreatic Cancer Research. D. Plenker was supported by the German Research Foundation (DFG, PL 894/1-1).

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