The tissue samples were collected under a protocol approved by the Institutional Review Board at Duke University Medical Center. 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. The tissue samples were collected from surgical resections of pancreatic ductal adenocarcinomas. Tissue from tumor resections were formalin fixed and embedded into paraffin blocks according to standard procedures. The tissue sections used for this study were from extra material not needed for patient evaluation. The status, procedures, and outcomes of the patients were recorded for at least 3 years.

We performed immunofluorescence on 5 μm thick sections cut from formalin-fixed, paraffin-embedded blocks. Paraffin was removed from 5 µm thick FFPE sections using CitriSolv Hybrid (Decon Labs, King of Prussia, PA) containing d-limonene and isopropyl cumene, and the tissue was rehydrated through an ethanol gradient of 100%, 95%, and 70% followed by washing with PBS. Following rehydration, antigen retrieval was achieved through incubating slides in citrate buffer at 100°C for 20 minutes. Slides were blocked in phosphate-buffered saline with 0.05% Tween-20 (PBST0.05) and 3% bovine serum albumin (BSA) for 1 hour at RT. Primary antibodies against CA19-9 (clone 9L426, US Biological Life Sciences) and TRA-1-60 (Novus Biologicals) were labeled for immunofluorescent staining with Sulfo-Cyanine5 NHS ester and Sulfo-Cyanine3 NHS ester respectively. After dialysis to remove unreacted conjugate, the antibodies were 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-containing solutions were decanted and the slides were washed twice in PBST0.05 and once in 1X PBS, each time for 3 minutes. The slides were dried via blotting and incubated with DAPI at 10 µg/mL in 1X PBS for 15 minutes at RT. Two five-minute washes were performed in 1X PBS, and then slides were cover-slipped and scanned using a fluorescent microscope (AxioScan.Z1, Zeiss, Oberkochen, Germany). The microscope collected 3 images at each field-of-view, each image corresponding to the emission maxima of Hoechst 33258, Cy3, and Cy5. We next quenched the fluorescence using 6% H₂O₂ in 250 mM sodium bicarbonate (pH 9.5-10) and performed another round of immunofluorescence using two different antibodies. The subsequent incubations and scanning steps were as described above.

Prior to the second round of detection with the TRA-1-60 antibody, we treated the slides with sialidase to remove terminal sialic acids. The slides were incubated with a 1:200 dilution (from a 50,000 U/mL stock) of α2-3,6,8 Neuraminidase in 5 mM CaCl₂, 50mM pH 5.5 sodium acetate overnight at 37°C. The subsequent incubations and scanning steps were as described above. The hematoxylin and eosin (H&E) staining followed a standard protocol.

All image data were quantified using SignalFinder. For each image, SignalFinder creates a map of the locations of pixels containing signal and computes various values for the output report, such as the percentage of tissue-containing pixels that have signal. This analysis required major computational resources, as each high-resolution image of the ~1.5 x 2.5 cm tissue comprises about 9 billion pixels at ~3000 pixels/mm resolution and ~500 GB file size. A 1-square-inch image at such resolution would scale to about 4000 square feet if changed to 72 pixels per inch as used for display.

We developed the SignalFinder software using MATLAB, Java, and C++. We used Microsoft Excel and MATLAB for analyzing numerical output; GraphPad Prism, the R language, and Microsoft PowerPoint for the preparation of graphs; and Canvas X for the preparation of figures

We analyzed whole-block tumor sections from 22 patients who underwent resection and adjuvant chemotherapy as curative treatment for PDAC ( Table 1 ). We immunostained each specimen for the CA19-9 and STRA biomarkers and detected the biomarkers using multicolor fluorescence imaging ( Figure 1A ). We then analyzed the fluorescence images by automated signal identification (24) to produce maps of the signal pixels ( Figure 1B ), which we overlaid onto the brightfield images of the stained slides ( Figure 1B ). This system enables a visualization of the locations of biomarker staining and of the histomorphologies of the cells producing the biomarkers ( Figure 1B ). But more importantly for the investigations of biomarker features associated with outcomes, the method provided automated, objective signal identification and quantification for use in subsequent biomarker analyses. This foundation allowed us to ask whether we could achieve a fully data-guided method of ranking the likelihood of recurrence, that is, a method of classifying the tumors that does not involve user selections of locations or settings.

The amounts of each biomarker—as determined by the number of signal pixels identified by SignalFinder normalized to the pixels in the tissue—were highly variable among the 22 subjects and between the biomarkers ( Figure 1C ). The subjects had widely varying outcomes, with some exhibiting no evidence of disease after several years and others succumbing to recurrent disease after less than one year ( Table 1 and Supplementary Table 1 ). A group between these extremes had extended survival but struggled with recurrence and continued progression of the disease. The outcomes did not associate with any demographic factors or type of systemic therapy ( Table 1 ). Neither the individual biomarkers ( Figure 1D ) nor the combination of the two (not shown) was associated with outcome. An alternative way to quantify the signal using average intensities above a threshold gave different quantifications but also produced no associations with outcome ( Supplementary Figure 1 ).

To gain insights into the sources of variability between the specimens, we examined the images of the whole-block sections ( Figure 2 ). At the microscopic level, we observed extreme diversity within and between tumors in the biomarker locations and amounts. The biomarker staining occurred primarily in epithelial and glandular areas—a potentially useful trait for assessing adenocarcinoma—but it also occurred in varying degrees among both benign and cancerous glands and some non-epithelial area. For example, specimen 17-213 showed CA19-9 staining in ampullary glands and low-grade PanIN and STRA staining in poorly differentiated carcinoma and focally in PanIN, but in another section (16–570), both STRA and CA19-9 were together in varying relative amounts in cancer glands. STRA staining appeared in some small-intestinal mucosa that was at the edge of the section. STRA also appeared alone or together with CA19-9 in cancer glands in a specimen (18–371). We also observed occasional CA19-9 staining of macrophages (16-570a) and red blood cells (18-371a) and STRA staining of stroma (17-213a). Such diversity appeared in the remainder of the specimens ( Supplementary Figure 2 ). The above observations indicated that simple quantifications of signal amount or intensity across entire sections would include large areas with low or zero biomarker production, as well as signals from both benign and cancerous glands.

To quantify features that potentially are more specific to tumor assessment, we tallied the signal only where it is part of a cluster of biomarker production ( Figure 3A ). Starting with the pixel maps of signal from the SignalFinder output, we produced a sum of signal abundance within a pixel-by-pixel moving box ( Figure 3B ). The number of regions with spatial concentrations above various thresholds provided the amount of clustering for each biomarker ( Figure 3B ).

We first asked whether this mode of quantification provides distinct information, or alternative whether it simply reflects random clustering that correlates with total signal amount. A view of cluster amounts with respect to signal amounts or intensities showed only weak correlation ( Figure 3C ). In some cases, the signals were highly clustered, such as the CA19-9 data in patient 15-658, and in other cases, the signals were more evenly distributed, as with the CA19-9 data in patient 19-296 ( Figure 3D ). This finding indicates that the level of clustering is distinct from total biomarker amount and that specimens have diversity between them in the amount of biomarker clustering.

The quantifications of clustered signal, however, did not associate with recurrence to a greater degree than signal amount ( Figure 3E ). A survey of the histological features of regions with high or low spatial concentrations gave insights into the tumor characteristics identified by this method of quantification ( Figure 3F ). The areas with higher spatial concentrations aligned with greater cellularity, and the areas with very low spatial concentrations were frequently free of epithelia, but areas with both high and low spatial concentration contained various levels of cancer and benign cells. These observations indicated that spatial concentration specifically quantifies the glandular-epithelial biomarker production but does not, by itself, distinguish types of histomorphology.

Given the varied relationship between the glycans ( Figure 2 ), we next asked whether the combined CA19-9 and STRA provide unique information about the biomarker clusters. We used the maps of the individual biomarker clusters to define the overlap between the glycan-defined clusters and to identify the regions that were exclusively STRA, exclusively CA19-9, or both ( Figure 4A , referred to as STRA-only, CA199-only, and dual). We then separately quantified the amount of each type of cluster to examine whether the overlap was simply a random event correlated with total signal amounts. Some of the tumors just had the STRA-only or the CA199-only clusters ( Figure 4B and Supplementary Figure 3 ), and the amounts of clusters containing both markers were not correlated with the sums of the signal amounts ( Figure 4B ). For example, subject 15-658 had a low amount of the biomarkers but proportionally high clusters; subject 17-213 largely had the STRA-only clusters, and 18-137 had high amounts and clusters ( Figure 4B ). Maps of the three cluster types confirm such various levels of heterogeneity and showed that the different types of clusters can either appear near each other or occupy separate regions ( Figure 4D ).

Given that the dual clusters were not resulting just from a mass-driven overlap of unrelated signals, we concluded that the combined glycans provide distinct information about the biomarker clusters. Furthermore, the specimens vary greatly in the types of clusters they exhibit. Specimens exist that are either relatively homogeneous in the types of clusters they contain or are heterogeneous with various relative proportions of the STRA-only, CA19-9-only, or dual clusters ( Figure 4C ).

The above findings suggested that a combined evaluation of the distinct cluster types may provide value for tumor assessment. An alignment of the individual biomarker amounts and clusters by subject showed the diversity between subjects ( Figure 5A ), both for subjects with and without recurrence of disease after 3 years. It also showed the lack of correspondence between the two methods of quantification, as with subject 19-296, who had a high amount but low clusters of CA19-9. None of the individual biomarker clusters was significantly associated with outcome ( Figure 5B ), prompting us to ask whether certain combinations of clusters have stronger associations with outcome than others. To categorize the subjects based on combined biomarker amounts or clusters, we dichotomized the presence or absence of each biomarker using cutoffs determined from receiver-operator-characteristic analysis ( Supplementary Table 2 ). The resulting values illustrated the prominence of only one cluster type in certain tumors and two or more in others ( Figure 5C ). Furthermore, it was evident that the co-occurrence of multiple cluster types was a dominant feature of the recurrent specimens. Accordingly, the occurrence of more than one type of cluster was significantly greater in the cases than in controls (11/13 cases vs. 2/8 controls, χ² = 7.5, p = 0.006). The improved classification accuracy resulted from the identification of heterogeneity in biomarker clusters in cases where biomarker amounts were not heterogeneous. For example, subject 18-371 had high biomarker amounts only of CA19-9 but high clusters of all three types ( Figure 5C ). An equivalent approach applied to the marker amounts resulted in no significant difference (5/13 cases vs. 2/8 controls, χ² = 2.5, p = NS) ( Figure 5D ).

We further tested this relationship using Cox Proportional Hazards models for continuous analysis of survival with respect to spatial cluster abundance ( Figure 5E ). Because p values are unstable for survival analysis using small sample sizes, we utilized explained randomness in proportional hazard models according to the method of Xu and O’Quigley ( ρ X O Q 2 ) (25)​. The explained randomness of survival probability was higher using the combination of the cluster types than using the individual clusters ( ρ X O Q 2   0.52 compared to 0.10, 0.04, and 0.38 for STRA, CA19-9 and Dual, respectively) or using the combination of the biomarker amounts ( ρ X O Q 2   0.16) or the average intensities of 90^th percentile ( ρ X O Q 2   0.10). These quantifications – the amounts and the clusters – are from the same SignalFinder output. Therefore, the increased association with recurrence suggests that biomarker clusters are more informative for outcomes than biomarker amounts. Further, it suggests that heterogeneity in biomarker clusters is more informative than individual biomarker clusters.

We tested the reproducibility of these results for 5 new tissue specimens that were cut from the same blocks as used in the above analysis but separated in the block by 10-20 mm. The quantifications of clusters correlated positively between the first and second runs, and after applying thresholds to classify each case, the classifications matched exactly ( Supplementary Figure 4 ). We concluded that for these sections, the types of biomarker clusters are consistent between different sections of a tumor block within moderate variation. We concluded also that the analytical variability of the assay is low, although in this assay we cannot disentangle variability in the tissue from variability in the assay.

We also investigated the associations between heterogeneity and other tumor characteristics. The intensities of the biomarker signals ( Supplementary Figure 1 ) were positively but non-significantly correlated with cluster amounts (not shown), suggesting a higher per-cell production of the biomarkers in areas where cells are clustered, although intensity by itself is not informative of outcome. The total cellularity of the tumors, referring to amount of tumor filled by cells instead of extracellular matrix, varies greatly between PDACs. Using a quantification of the DAPI nuclear staining as a measure of cellularity, cellularity was higher, but not significantly, in the heterogeneous tumors. This result suggests that tumors with clonal heterogeneity proliferate and spread more than others, and it also could provide a basis for a previous observation that high cellularity in PDAC is moderately associated with short survival (26). Thus, heterogeneity is an independent factor of clinical significance that is moderately associated with signal intensity and total cellularity.

The association of cluster heterogeneity with recurrence suggests that the co-occurrence of divergent clusters identifies a subset of cancer cells. We therefore investigated whether areas with biomarker cluster heterogeneity are more associated with the presence of cancer cells than areas without cluster heterogeneity. We performed intra-tumoral comparisons of histomorphology in 4 specimens with heterogeneity in clusters but not in amounts and in 1 specimen with heterogeneity in both clusters and amounts ( Figure 6 ). For each, we identified regions containing more than one cluster type and regions containing just one cluster type within regions-of-interest (~3 mm) three times the size of the region used in calculating spatial concentrations (~1 mm, Figure 3B ). We then identified by surgical pathology review the regions containing cancer glands and compared the overlap with the clusters. At the macroscopic level, the areas with heterogeneous clusters corresponded well with the locations identified by surgical pathology as containing cancer cells ( Figure 6A ). In three subjects (16-250, 15-658, and 18-371), the identifications were equivalent, but in the others, cluster heterogeneity occurred in some regions not identified by surgical pathology. In all, cluster heterogeneity identified areas containing cancer cells in 28 out of 32 (88%) instances. The matched regions-of-interest taken from areas containing only 1 cluster type overlapped with cancer-cell-containing regions in 2 of 17 (11.7%) instances. This result indicates that heterogeneity in biomarker clusters preferentially identifies regions containing cancer cells, in contrast with the individual biomarkers.

Given that cluster heterogeneity associated with recurrence, we further examined the histological differences between the recurrent and non-recurrent specimens. By surgical pathology, no systematic differences were evident between the patients with recurrence and without recurrence, nor between the patients with cluster heterogeneity and without cluster heterogeneity ( Table 2 ). A feature present in 5 of the recurrent cases and 0 of the non-recurrent cases was positive surgical margins (typically an sign of poor prognosis in any cancer). Of the 8 cases with recurrence and negative margins, 7 were positive cluster heterogeneity. We concluded therefore that heterogeneity in biomarker clusters provides information about cancer-containing regions that is not discernable from histopathological review, in particular as a quantifiable feature of tumors with high likelihood of recurrence.