Flash-frozen tissue specimens were obtained from pancreatic resection specimens after pathologist review from five patients with stage IIB PDAC, five with IPMN, and five normal/non-neoplastic pancreatic tissue samples. Some of the normal/non-neoplastic tissues were derived from non-neoplastic regions adjacent to pancreatic tumors. This strategy was intentionally chosen to allow comparison of the metabolic changes in the cancer specimens with normal and non-specific tissue and metabolic changes that could arise in tumor adjacent tissue. Another reason for selecting adjacent normal tissue is that using both cancer and non-cancer tissues from the same patients allows for better control over tissue processing conditions as well as systemic patient factors. PDAC and IPMN cases were selected if patients had not undergone prior treatment, and if sufficient frozen tissue, including normal adjacent tissue was available. Flash-freezing was performed after pathologist review and sampling of the resection specimen for diagnostic purposes, after which the frozen tissues were placed in standard cryovials and stored in −70 °C freezers. Although the exact timing was not recorded for these cases, pancreatic resections typically require approximately 45 minutes to process, with a few additional minutes to take additional specimen for tissue banking. Normal and neoplastic tissues were snap frozen at the same time. The amount of tissue stored for each case was similar, but tissues were not completely uniform in thickness. All the tissues had been frozen for several years before being thawed for cryosectioning for this study. The standard SOP for the biobanking applied to all samples. A standard SOP was followed for all the cryosectioning. The collection and analysis of these tissues was performed in accordance with recognized ethical guidelines and written informed consent, under a protocol approved by the Johns Hopkins Medicine Institutional Review Board that is guided by the ethical principles in the Belmont Report.

Frozen pancreatic tissue samples were mounted with Epredia M-1 Embedding Matrix and cryosectioned at 10 μm thickness using a Leica CM1860-UV cryostat (Leica Biosystems, Wetzlar, Germany, RRID:SCR_025772). The sections were thaw-mounted onto indium tin oxide (ITO)-coated glass slides, Part no. CG-40IN-S115 (Delta Technologies LTD, Loveland, CO) which were washed with hexanes (10 minutes sonication) and ethanol (10 minutes sonication). Two adjacent sections were cut from each tissue sample: one for the MALDI-MSI experiment and another for H&E staining. Each ITO slide accommodated 7 to 8 tissue sections arranged in a randomized order. Prior to matrix application, the ITO slides were transferred to a vacuum desiccator for equilibration to room temperature and dried for at least 20 minutes. The dried slides were sprayed using an HTX M5 pneumatic sprayer (HTX Imaging, Chapel Hill, NC, RRID:SCR_023733), with 10 mg/mL 1,5-diaminonaphthalene (1,5-DAN) in 50% acetonitrile (ACN) with 0.2% trifluoroacetic acid (TFA). Four passes were used, with 0.1 mL/min flow rate, 1200 mm/min nozzle velocity, 2.5 mm track spacing, criss cross (CC) pattern, 2 L/min N2 flow at 10 psi, with a nozzle height of 40 mm, and no dry time between passes.

MALDI-MSI experiments were carried out using a RapifleX MALDI TOF/TOF instrument (Bruker Daltonics, Billerica, MA, RRID:SCR_026525) in both positive (m/z 30-1,000 Da) and negative (m/z 30-1,000 Da) ion modes on the same tissue sections by offsetting the laser position by 100 μm within the same raster. flexControl (v 4.1) was used to control the instrument and optimize the acquisition parameters, while flexImaging (v 5.1) was used to set up the MALDI imaging experiments and select the regions of interest for MALDI imaging. Prior to data acquisition, height adjustment (target profile generation), laser focus tuning, and external mass calibration with red phosphorus were conducted to achieve a mass error <5 PPM. A 100 μm imaging laser using the M5 Smartbeam option with 200 shots per pixel at 10 kHz was selected to generate 200 μm sized pixels, and the laser fluence was optimized for each imaging experiment. All imaging data were loaded into SCiLS Lab (v2023b Pro, Bruker Daltonics, RRID:SCR_014426) using the default settings with the parameters from FlexImaging. For all pixels defined within each tissue sample, peak intensities were averaged and normalized by total ion current. After imaging experiments were conducted, the slides were vacuum-sealed and stored at −80 °C until further use.

MS/MS fragmentation analysis was utilized to confirm the preliminarily identified m/z features. For comparison with pure standard compounds, N-acetylglutamic acid, histidine, malate, creatine, N-methylnicotinamide, citrate, xanthine, taurine, ascorbate were purchased from Sigma-Aldrich (Burlington, MA). The isolation width for the parent ion was set to 1.00 Da width or narrower to fragment isolated single m/z peaks. On-tissue MS/MS spectra were acquired in tissue regions with high detected intensity of the parent ion peak, selected from the respective MALDI ion images displayed in SCiLS Lab. MS/MS spectra from collision induced dissociation (CID) with nitrogen were acquired directly from the pancreatic tissue samples using a timsTOF fleX MALDI-2 instrument (Bruker Daltonics, Billerica, MA, RRID:SCR_023615) in both positive and negative ion modes for the identification of metabolites. Spectra were collected using timsControl (v 5.1.1). Prior to data acquisition, height adjustment (target profile generation), laser focus tuning, and external mass calibration with red phosphorus were conducted to achieve a mass error <1 PPM. A 50 μm custom laser using the single M2 Smartbeam option with 1000 shots per pixel at 1 kHz was selected for profiling, and the laser fluence, collision energy (30-50 eV), and isolation width were optimized for each m/z. On-tissue MS/MS spectra were acquired in tissue regions with high detected intensity of the parent ion peak, selected from the respective SCiLS ion images. Spectra were saved and analyzed in DataAnalysis (v 6.1) and ChemDraw (v 20.0.0.41, RRID:SCR_016768) to confirm diagnostic peak fragmentation and assign an identity. Briefly, bonds with lower bond dissociation energy were targeted for breakage in ChemDraw to computationally determine fragment m/z values. Such m/z values were compared to the MS/MS spectra. If sufficient diagnostic fragments existed in the spectra, the identification was considered confirmed. For metabolites where MS/MS spectra were acquired from both tissue samples and reference standards, dot product similarity scores were calculated to assess spectral matching, with higher scores indicating greater confidence in compound identification. The dot product was calculated using the formula: DotProduct=∑(I1×I2)√∑(I12)×√∑(I22) where I1 and I2 are the normalized intensities of matching fragment ions in the sample and reference spectra, respectively. The dot product was calculated for all fragment ions exceeding a normalized intensity threshold of 2% in both tissue and standard spectra.

A standard H&E protocol was employed for staining tissue sections. Sections were sequentially immersed in 95% ethanol, 70% ethanol, and distilled water for 30 seconds each, followed by staining in hematoxylin for 3 minutes. After rinsing in distilled water, sections were dehydrated in 70% ethanol and 95% ethanol for 30 seconds each, then stained in eosin for 1 minute. The sections were subsequently dehydrated in 95% ethanol and 100% ethanol for 30 seconds each and cleared in xylene for 2 minutes. Finally, cover slips were applied using xylene based mounting media (Epredia #4112) and allowed to dry. High-resolution digital images of the H&E stained tissue sections were obtained using a Hamamatsu NanoZoomer S210 Digital slide scanner (RRID:SCR_023760).

Total RNA was extracted from frozen tissue using a Qiagen kit (Qiagen #79656) and RNeasy mini column kit (Qiagen #74106) following the manufacturer's protocol. cDNA was synthesized using an iScript cDNA synthesis kit (Bio-Rad #1708891). qRT-PCR was performed from cDNA using IQ SYBR Green supermix (Bio-Rad #1725121) and gene specific primers in the PCR machine (Bio-Rad, RRID:SCR_026760). The primer sequence for the variant 1 human SLC6A6 gene (NM_003043) was used. 5'-aminolevulinate synthase 1 (ALAS1, NM_000688.6) was used as a housekeeping gene. Fold change was determined using 2^−ΔΔCT.

Statistical analyses were performed using GraphPad Prism (version 6.0, GraphPad Software, RRID:SCR_002798) and MetaboAnalyst 6.0 (www.metaboanalyst.ca, RRID:SCR_015539). The normalized peak intensities were used as an input parameter to carry out Partial Least Squares Discriminant Analysis (PLS-DA) in the MetaboAnalyst web tool. PLS-DA score plots were generated to visualize group separations, with ellipses representing 95% confidence intervals for each group. Heatmaps were generated based on the Euclidean distance measure and the Ward clustering algorithm. Correlation heatmaps were constructed using pairwise Spearman rank correlation analysis between samples, allowing visualization of metabolomic profile similarity and clustering behavior across tissue groups. To determine the statistical significance of the quantified data, an unpaired two-tailed Student’s t-test was performed. p-values < 0.05 were considered significant unless otherwise stated. To evaluate significance, values from all the pixels were consolidated into a single mean value for each sample. Receiver Operating Characteristic (ROC) curve analysis was performed for the significantly dysregulated identified compounds to evaluate their diagnostic performance in distinguishing PDAC from non-malignant pancreatic tissues. Metabolites with Area Under the Curve (AUC) > 0.85 were considered highly discriminative. Bar plots were generated for all the significant compounds (p-value < 0.05) to show differences across the pancreatic tissue samples. qRT-PCR gene expression data were analyzed using one-way ANOVA followed by Šídák’s multiple comparisons test to assess statistical significance across sample groups.

A total of 15 pancreatic tissue samples were analyzed, comprising 5 PDAC, 5 IPMN, and 5 normal/non-neoplastic samples. Clinicopathological details of the patient samples used in this study are summarized in Supplementary Table S1. Spatial mapping of metabolites and lipids in PDAC was achieved using MALDI-MSI using the workflow presented in Figure 1A that outlines tissue sample preparation, acquisition of MALDI-MSI data, and subsequent spatial data analyses.

Representative examples of MALDI spectra obtained from PDAC, IPMN, and normal/non-neoplastic tissue samples are presented in Figure 1B. These examples demonstrate the metabolic differences across the three groups. A data matrix consisting of 970 features in positive ion mode and 950 features in negative ion mode was generated from the analysis. 135 significantly dysregulated m/z features (p < 0.05), consisting of 42 features detected in negative ion mode and 93 detected in positive ion mode were identified.

Significantly dysregulated m/z features were initially identified using the Human Metabolome Database (HMDB, RRID:SCR_007712) and Lipid Maps (RRID:SCR_003817), providing preliminary annotations of metabolites and lipids. To confirm these preliminary annotations, MS/MS analysis was conducted, to validate the identification of these analytes. All identifications were made from on-tissue MS/MS fragmentation analysis and are based on diagnostic ions detected in this analysis which are shown in the Supporting Information in Supplementary Figure S1. For metabolites, we additionally performed comparative analysis of MS/MS fragmentation from standard compounds that were compared to on-tissue MS/MS fragmentation data (Supplementary Figure S1). For metabolites where MS/MS spectra were acquired from both tissue samples and reference standards, dot product similarity scores were calculated to assess spectral matching, with higher scores indicating greater confidence in compound identification. Identifications were considered confirmed when all fragment ions from standard compounds were detected in on-tissue MS/MS fragmentation and the dot product similarity between the on-tissue and standard spectra, calculated without the precursor (parent) ion, was ≥ 0.8. If the dot product value without the precursor ion was <0.8, even when diagnostic fragments were present, the compound was classified as a tentative identification. All structural identifications were carried out manually, by matching detected on-tissue MS/MS fragments with possible fragmentations occurring in chemical structures under investigation. Lipids were classified as confirmed when the precursor m/z matched the theoretical mass within ±0.02 Da and the on-tissue MS/MS fragmentation pattern showed the expected head-group and fatty acyl chain–derived fragment ions. When the precursor m/z deviated from the theoretical mass by more than ±0.02 Da, but the MS/MS spectrum still showed the characteristic head-group and acyl-chain fragment ions expected for the candidate lipid and no alternative lipid species could account for the observed fragments, the identification was reported as tentative. Representative MS/MS spectra for selected compounds, demonstrating their characteristic fragmentation patterns, are presented in Figure 2. Among the validated compounds, 9 were identified in negative ion mode and 28 in positive ion mode. The MS/MS analysis, conducted to validate the identification of all identified analytes in this study, is presented in Supplementary Figure S1, showcasing the fragmentation patterns that were critical for confirming their molecular structures and ensuring accurate annotation of compounds. These validated metabolites included key intermediates involved in lipid metabolism, energy pathways, and redox regulation.

To evaluate the adequacy of the sample size, we performed a post hoc power analysis across the 135 significantly altered m/z features. Power was calculated using the observed means, standard deviations, and pooled standard deviations (with n = 5 per group and α = 0.05). The results showed that power values ranged from 0.53 to 1.00, with a mean power of 0.73 and median power of 0.70. Notably, 67 of the 135 features (~50%) achieved power ≥ 0.70, suggesting that our study was sufficiently powered to detect moderate-to-large biological effects despite the small group size.

Heat-map analysis was conducted to visualize the individual abundance profiles of the dysregulated m/z features across the PDAC, IPMN, and normal/non-neoplastic pancreas tissue samples, as shown in Figures 3A and B. The heat map uses a color-coded representation, where color temperature indicates metabolite abundance with a warmer color for higher abundance, and a colder color for lower abundance. A clear difference was evident in the abundance levels of compounds among the three groups, with several metabolites showing marked dysregulation in PDAC tissue compared to both IPMN and normal/non-neoplastic tissue, identifying the significant metabolic reprogramming that occurred in PDAC.

To characterize the metabolomic variations in normal/non-neoplastic, IPMN, and PDAC tissue, partial least squares discriminant analysis (PLS-DA) was performed. The PLS-DA results, presented in Figures 3C and D, revealed distinct clustering patterns that identified clear separation between PDAC and normal/non-neoplastic tissue and the overlap of IPMN tissue with normal/non-neoplastic and PDAC tissue in both positive (Figure 3C) and negative (Figure 3D) ion modes. The overlap of IPMN with normal/non-neoplastic and PDAC tissue reflects the transitional metabolic phenotype of IPMN, bridging the characteristics of benign and malignant states. Each dot in the score plots corresponds to an individual tissue sample, representing its unique metabolic profile. In the positive ion mode, principal component 1 (PC1) accounted for 45% of the variance, and principal component 2 (PC2) accounted for 25% of the variance. Similarly, in the negative ion mode, PC1 and PC2 accounted for 49% and 23% of the variance, respectively.

To further examine the relationships among tissue samples based on their metabolic/lipid profiles, Spearman rank correlation analysis was performed using the normalized intensities of all significantly altered metabolites. The resulting heatmap illustrated the degree of similarity between individual samples and revealed distinct clustering patterns corresponding to tissue classification (Supplementary Figure S2). PDAC samples exhibited strong intra-group correlation and formed a distinct cluster, clearly separated from both IPMN and normal/non-neoplastic tissues, reinforcing their distinct metabolic signature.

ROC curve analysis was conducted to determine the diagnostic performance of the significantly dysregulated metabolites in PDAC. Nearly all of these compounds achieved AUC values exceeding 0.85, demonstrating strong diagnostic accuracy in distinguishing PDAC from non-malignant pancreatic tissues (Supplementary Table S2).

H&E staining combined with MALDI-MSI provided a framework to correlate histopathological features with metabolic heterogeneity in the tissues. The histopathological analysis of pancreatic tissue sections, visualized through H&E staining, demonstrated marked differences in tissue architecture between normal/non-neoplastic, IPMN, and PDAC samples as shown in the representative example for taurine in Figure 4. Normal/non-neoplastic tissue exhibited well-preserved acinar structures, while PDAC tissue displayed disrupted architecture with dense desmoplastic stroma, characteristic of the PDAC TME. Elevated taurine levels localized within regions of dense desmoplastic stroma and invasive tumor clusters in PDAC tissue. Some normal/non-neoplastic pancreatic tissues analyzed in this study were obtained from regions adjacent to pancreatic cancer. One challenge in comparing tumor and normal/non-neoplastic pancreatic tissues is that while normal/non-neoplastic pancreata are predominantly composed of acinar tissue, PDAC exhibits a dense tumor stroma microenvironment that includes inflammatory cells, fibroblasts, and other components. Even with no tumor involvement, the histology of the pancreas is often not entirely normal, especially in older individuals where features such as PanIN, mild atrophy, fatty replacement, and small areas of acinar-to-ductal metaplasia (ADM) may be observed. Histological images for all normal/non-neoplastic pancreatic samples are shown in Supplementary Figure S3.

Spatial maps of key metabolites together with the corresponding H&E staining are presented in Figure 5 for all three groups of tissues. Spatial metabolic heterogeneity was evident in normal/non-neoplastic tissue, IPMN and PDAC as seen in the corresponding metabolic maps of taurine, acetylcarnitine, creatine, malate, ascorbic acid, and lipidomic maps of phosphatidylcholine (PC) 34:3, PC 32:0, and m/z 835.64 (tentatively identified as sphingomyelin SM 42:2;O2). The spatial heterogeneity in normal/non-neoplastic tissue was structured and closely followed tissue morphology in the H&E sections. Unstructured metabolic heterogeneity became evident in IPMN, and was most evident in PDAC. To quantitatively support these visual observations, we performed Haralick feature analysis on the ion images and determined values for the feature ‘homogeneity,’ which provides an inverse index of heterogeneity [19]. A decrease in homogeneity values from normal/non-neoplastic to PDAC (Supplementary Figure S4) quantitatively confirmed the increased heterogeneity observed visually. One advantage of MSI is the ability to investigate the relationship between the spatial distributions of metabolites in the multiplexed data. The images in Figure 5 clearly demonstrate the spatial heterogeneity of each metabolite, and the lack of overlap of this spatial heterogeneity for the different metabolites and lipids that occupied distinct spaces in the tissues.

MALDI-MSI ion maps of selected analytes, presented in Figure 6, provide a detailed visualization of their spatial distribution within the analyzed tissue samples. These maps are accompanied by the corresponding bar plots that summarize the relative abundances of these compounds across different tissue types. MALDI-MSI ion maps and bar plots for the most significantly dysregulated ions are provided in Supplementary Figures S5 and S6, respectively.

Untargeted metabolomic and lipidomic analyses identified significant changes in the metabolite and lipid profiles across normal/non-neoplastic, IPMN, and PDAC tissue in both positive and negative ion modes. Lipid dysregulation was predominantly observed in the positive ion mode, whereas MALDI-MSI conducted in the negative ion mode identified significant dysregulation of metabolites. Dot product similarity scores were calculated both with and without the precursor ion as presented in Supplementary Table S3. Dot product values ≥ 0.8 without the precursor ion were considered as reliably identified. These were taurine, ascorbic acid, creatine, xanthine and malate. The remaining m/z values, specified as tentative identifications, included citrate (m/z 191.02), N-methylnicotinamide (m/z 137.07), histidine (m/z 154.06), N-acetylaspartate (m/z 174.03), and N-acetylglutamate (m/z 188.06). In PDAC, taurine, ascorbic acid, and tentatively assigned histidine, N-acetylaspartic acid and N-acetylglutamic acid were significantly (p < 0.05) upregulated, while malate, creatine and PE 36:3 were significantly (p < 0.05) downregulated compared to normal/non-neoplastic tissue. Additionally, xanthine levels were significantly decreased in IPMN samples compared to normal/non-neoplastic pancreatic tissue.

In the positive ion mode, we observed dysregulation of the carnitine system and lipid species across normal/non-neoplastic, IPMN, and PDAC tissue samples. PDAC tissues exhibited the most pronounced changes, with significant alterations observed in several key lipid classes, including phosphatidylcholines (PCs), sphingomyelins (SMs), and lysophosphatidylcholines (LPCs).

Lipid species that were significantly dysregulated in our study included, PC 34:3, PC 38:3, SM 36:1;O2, LPC 16:0, PC 32:0, PC 28:3;O3, SM 34:1;O2, PC 32:1, PC 28:2;O3, SM 34:2;O2, PC 30:4, PC 34:2, PC 34:1, PC 34:0, PC 36:3, PC 36:1, PC 38:4, PC 36:3;O3, PC 40:6, PC 38:3, SM 42:2;O2, PC 40:5, PC 38:2, SM 42:1;O2 and PC 40:4. In addition to lipid alterations, several metabolites were also significantly dysregulated. The classification of these analytes as confirmed or tentative identifications is provided in Supplementary Tables S4a and S4b, respectively. Of the dysregulated lipids and metabolites, PC 34:3, PC 38:3, PC 32:0, PC 28:3;O3, SM 34:1;O2, PC 28:2;O3, SM 34:2;O2, PC 30:4, PC 34:1, PC 34:0, PC 36:3, PC 36:1, PC 38:4, PC 36:3;O3, PC 40:6, PC 38:3, SM 42:2;O2, PC 40:5, PC 38:2, SM 42:1;O2, PC 40:4 and tentatively assigned N-methylnicotinamide were significantly upregulated in PDAC compared to normal/non-neoplastic tissue. Compounds such as SM 36:1;O2, PC 32:1, PC 34:2, LPC 16:0, and creatine were downregulated in PDAC compared to normal/non-neoplastic tissue. Acetylcarnitine, butyrylcarnitine, and PC 34:3/PC 32:0 were upregulated significantly in PDAC compared to IPMN. Metabolites such as arginine and tryptophan were downregulated in IPMN compared to normal/non-neoplastic tissue. These findings highlight the lipidomic and metabolomic alterations that occur with progression from normal/non-neoplastic tissue to IPMN and to PDAC.

To further understand the mechanism underlying elevated taurine in PDAC, qRT-PCR of the taurine transporter SLC6A6 was performed. SLC6A6 expression significantly increased (p < 0.05) by 6.74-fold in PDAC and 2.62-fold in IPMN relative to normal/non-neoplastic tissue (Supplementary Figure S7).