Integration of metabolites from meta-analysis with transcriptome reveals enhanced SPHK1 in PDAC with a background of pancreatitis

A systematic search of PubMed, Web of Science and Scopus, was carried out to identify biomarker studies that reported metabolites in CP and PDAC (Fig. 1). The following terms were applied to search criteria: “pancreas or pancreatic” and “cancer or carcinoma or chronic pancreatitis” and “metabolite profiling or metabolomics or metabolome or metabolite analysis or metabolic characterization” and “serum or plasma or biomarkers or markers”. The literature search was conducted from March 2005 till December 2021. The references that were found relevant from letters, thesis and posters presented during conferences were reviewed for further analysis.

Title and year of publications, authors and abstracts were extracted to EndNote X7(Thomson Reuters, NY, USA). Following redundancy check, an initial screening process was employed to screen the titles and abstracts of the eligible articles. Articles found relevant were further selected for reading. Authors have been advised to avoid bias and a consensus was made in article selection. Only studies that examined and validated using metabolomic platforms in CP and PDAC especially in serum or plasma-based marker discovery were eligible. Studies included the following criteria: (1) Biomarker in circulation (serum / plasma); (2) metabolomics platform (GC/MS, LC/MS, NMR); (3) biomarker selection was selected based on statistical significance and further evaluation by statistical learning for discrimination of PDAC from healthy control or CP; (4) accuracy, sensitivity, specificity and area under the curve (AUC-ROC, etc.) were deemed as suitable prediction metrics. Exclusion of studies was followed if (1) no suitable control groups were used; (2) the studies were conducted on animal tissues, cell lines and urine samples; (3) there was duplication of data and (4) they are letters, reviews, theses, and conference proceedings.

The data were abstracted by at least two independent researchers. A data extraction form was made in accordance with earlier systematic review guidelines [9]. For preliminary analysis, information pertaining to the study including first author, year of publication, patient/control characteristics (age, gender, clinical stage) were extracted. The nature of biological specimen, platforms employed for analysis, proper targeted / untargeted approach, sample collection, storage, sample preparation, pretreatment description, analytical standards, validation and metabolite identification methods were described. The list of biomarker panels were extracted. For each biomarker study comprising controls, CP and PDAC, fold changes values were noted or the biomarker levels were grouped as elevated or low respectively. Total patients for each study along with the biomarker panel, area under curve, sensitivity and specificity values were noted. Following extraction of biomarker data, a third independent researcher reviewed abstracted data. Data were discussed and sorted in case of discrepancy.

The study outcome measures are

Data pertaining to 25 biomarker studies data were identified and entered into MS-excel. Meta-analysis was performed using MedCalc software for all publications which reported AUC, specificity, sensitivity by deriving standard error. We computed the standard error from samples size(n) for each study from AUC and then grouped the studies as all metabolites, lipids, carbohydrates and amino acids. Using a two tailed critical value of normal distribution, and keeping sample size between two groups we computed AUC, the standard error and 95% confidence interval. AUC and standard error were used to generate the forest plots for all metabolite markers, carbohydrate, amino acid and lipid metabolites. Assessment of statistical heterogeneity by forest plots and publication bias were by funnel plot asymmetry were performed using Egger’s intercept. The compound name standardization was followed using Human Metabolome Database version 4.0 [14]. The pathway analysis was conducted using metaboAnalyst (Ver5) [15]. The Kyoto Encyclopedia of Genes and Genomes (KEGG) and the small molecule pathway (SMP) were used to identify pathways of significance to the study [16]. The analysis algorithm was a hypergeometric test for overrepresentation analysis, and the relative betweenness centrality was selected for pathway topology analysis. Pathways that had an adjusted P-value (false discovery rate, FDR) of less than 0.05 were considered to be significantly enriched. Subgroup analysis was screened by metaboAnalyst [17] and overall mapping pathways were derived for metabolic intermediates and networks of significance to differentiate healthy controls from PDAC, healthy controls/CP from PDAC and CP from PDAC.

For integrating the genes of altered metabolites identified from meta-analysis, transcriptome profiles were generated in pancreatic tissues with CP, PDAC-CP, PDAC and healthy tissue. Paraffin embedded tissue blocks (FFPE) were retrieved from the archived samples. A total of 12 tissues were included of which 3 were normal pancreatic tissues (from trauma cases), 3 with CP, 4 with PDAC-CP and 2 with PDAC. All the participants provided written informed consent and the study protocol conformed to the ethical guidelines of the 1975 Declaration of Helsinki. The study was approved by the Institutional Review Board (Protocol #AIG/AHF-IRB:02/31/20) of Asian Healthcare Foundation, a research wing of AIG hospitals. Transcriptome datasets are available online, (Vishnubhotla, Ravikanth (2021), “Transcriptome Dataset”, Mendeley Data, V1, doi: https://​doi.​org/​10.​17632/​2jjbpvfm72.​1). CP was diagnosed with the clinical symptoms: chronic abdominal pain, pancreatic calcifications, abdominal pain and morphological characteristics from duct pancreatograms and exocrine pancreatic insufficiency. PDAC was diagnosed by histopathologic examination. Patients who developed PDAC with long standing CP were considered as PDAC-CP (Additional file 1: Table S2). RNA was isolated from the FFPE blocks using RNeasy FFPE kit (Qiagen) as per manufacturer’s instructions. cRNA was synthesized and hybridized to Affymetrix Human Transcriptome arrays 2.0 and scanned. Transcriptome data was analyzed using Affymetrix Transcriptome Analysis Console 3.0.0.466. Transcriptome data is represented as splicing index which is similar to the fold change (FC) that is often used on the gene level. The exon expression is first normalized to the expression of the corresponding gene before calculating the ratio between CP and PDAC-CP. Transcript level dysregulation were identified between the patient groups. GO enrichment analysis were performed using PANTHER and GO ontology database [18].

Genes corresponding to the altered metabolites differentiating CP and PDAC were extracted from relational database of metabolic pathways [19] that integrates KEGG reactome [20] and HMDB [14]. We retrieved dysregulated gene list from the transcriptome data of CP and PDAC-CP generated from our laboratory. The two gene sets were compared to identify common and unique genes between CP and PDAC-CP with a likely role in malignant transformation. Additionally, dysregulated gene list from PDAC-CP was also compared to that of PDAC. TSGene 2.0 [21] database was compared to retrieve tumour suppressor genes from our identified gene datasets.

To validate the findings of integrating meta-analysis of metabolome and in-house transcriptomic data, we performed immunohistochemistry on the paraffin embedded tissue sections of CP(n = 6), PDAC-CP (n = 5) and PDAC (n = 4) obtained from department of pathology archives. 4 um sections were stained with anti-human SPHK1-HRP antibody (mouse origin, Santacruz biotechnologies Inc., USA), CD31 antibody (mouse origin) using BioGenex detection system. Each tissue was observed using a light microscope at a total magnification of 40X and documented using a computer with cellSens imaging software and a camera that had been integrated with Olympus BX63 microscope. Photographs were taken randomly with a total of ten visual fields per one tissue for acini, blood vessel, duct, islet and standardized using a global white balance. The brown colour intensity was calculated using the plugin program in Image J, IHC profiler [22] which quantified the intensity of the images. The results of quantification were expressed as H-Score [23] based on the formula: (% low positive × 1) + (% positive × 2) + (% high positive × 3). Fisher’ exact test was used to compare SPHK1 expression among the two groups: CP, PDAC-CP and PDAC-CP, PDAC with an α of 0.05 as a cut-off to denote statistical significance.

Studies were included from Web of Science, Scopus, PubMed and were screened for title, abstract and duplications were deleted. Inclusion and exclusion criteria were followed as described earlier. Letters, reviews, theses, conference posters or proceedings were also excluded. 54 studies were found to be eligible for further full-text assessment, and only 25 studies could be included for systematic review (Fig. 1).

The characteristics of studies included are described in Additional file 2: Dataset S1. A total of 10,951 controls (either healthy controls or CP patients were considered as controls) and 2248 patients with PDAC, pancreatic cancer were included in the analysis (Additional file 2: Dataset S1). Healthy controls were used in 15 studies while 8 studies used either healthy controls or CP patients to compare to PDAC and 2 used CP as controls. Apart from one study [24] which included 240 PDAC and 7772 controls, the sample size of the remaining studies were relatively small, of which six studies out of 25 comprised more than 100 PDAC patients. Tumour stage was provided in 16 studies, out of which 6 studies included PDAC patients who were in resectable stages [25‐30].

The study design and metabolomic methods followed in the studies were tabulated (Additional file 2: Dataset S1). Untargeted metabolomics approach was followed by 17 studies (68%) while 2 studies used only lipidomics. Among studies that used targeted metabolomics, focus was on amino acids [24, 25, 31‐33] and fatty acids [34]. Most preferred platform was Mass spectrometry based assessment of novel biomarkers (20 studies), followed by NMR (5 studies) [31, 35‐38]. Eleven studies collected samples under fasting condition, and only two studies performed outlier detection, using Principal component analysis.

Studies included in the meta-analysis identified panels of lipids, carbohydrates and amino acids to differentiate CP from PDAC. The heterogeneity was more than 95% in all classes of biomarkers and all were statistically significant. Among the studies which reported AUC, heterogeneity was higher in amino acids followed by phospholipids and carbohydrates (Fig. 2). Heterogeneity was more than 90% when random effects meta-analysis was performed for studies reporting sensitivity and were statistically significant (SI Additional file 1: Fig. S1). A total of 19 metabolomic studies compared either healthy, CP and PDAC patients and reported AUC values. Standard error was calculated from AUC values for generating forest plots. For all the circulatory biomarkers detected, the AUC was observed to be 0.86 (95% confidence interval 0.81 to 0.91), with heterogeneity 97.92% using the random effects model (Fig. 2A). We further classified these biomarker studies to amino acid, carbohydrate and lipid metabolites. Compared to other metabolite biomarkers we observed that, lipid metabolites (Fig. 2B) differentiate PDAC from CP with higher AUC 0.93 (95% confidence interval 0.90 to 0.95 with heterogeneity 77%, P-value (< 0.0001) followed by amino acids with AUC 0.83(95% confidence interval 0.77 to 0.90 with heterogeneity 98.51%, P-value (< 0.0001) (Fig. 2D) and carbohydrates (Fig. 2C) with AUC 0.863 (95% confidence interval 0.82 to 0.91 with heterogeneity 79%, P-value (0.0002). Egger’s test and funnel plot (Additional file 3: Dataset S2; p value = 0.5) suggest no publication bias among the selected biomarker studies (Additional file 1: Fig. S1).

To identify metabolic pathways differentiating PDAC from CP, we have categorized the studies into 3 groups; i) controls vs PDAC ii) controls/CP vs PDAC iii) CP vs PDAC. Glycerophospholipid metabolism with 45 metabolites (impact value of 0.22, FDR of 1) was the most prominent pathway in differentiating healthy controls from PDAC (Fig. 2E) Glutamate (impact value of 0.5, FDR of 1.36E-02) and arginine metabolism (impact value of 0.19, FDR of 3.45E-03) were the two prominent pathways derived from list of 13 biomarkers (Fig. 2F) differentiating healthy controls/CP from PDAC. Sphingomyelin pathway (impact value of 0.29, FDR of 0.45) and TCA cycle (impact value of 0.18, FDR of 0.06) were the two prominent pathways derived from a list of 30 biomarkers (Fig. 2G) (Additional file 4: Dataset S3) that differentiated CP from PDAC. We further grouped all the biomarkers from these studies and an overall metabolic flux among various intermediates was generated (Fig. 3A). A total of 517 genes corresponding to the 30 metabolites differentiating CP from PDAC were extracted from relational databases to integrate with transcriptome data sets (Additional file 5: Dataset S4).

Transcriptome datasets were generated in pancreatic tissues with CP, PDAC-CP and PDAC (Fig.4 and Additional file 1: Fig. S2). Comparison of transcriptome data between controls and CP identified 4902 genes to be dysregulated (2622 upregulated and 2280 downregulated). Likewise, comparison of data between controls and PDAC-CP identified a total of 6691 genes to be dysregulated (3571upregulated and 3120 downregulated). While comparison of controls and PDAC identified 7350 genes to be dysregulated (5421 upregulated and 1929 downregulated).

We first compared transcriptome data sets of CP vs control and PDAC-CP vs control. The dysregulated genes in CP vs control (4902 genes), PDAC-CP vs control (6691 genes) were then compared with 517 dysregulated genes obtained from altered metabolites separately. This resulted in identifying a total of 102 genes common between transcriptome and metabolome data sets (Fig. 3B). Distribution of 102 genes and types of metabolites are represented in (Fig. 3C) (Additional file 5: Dataset S4). Of the 102 genes, 28 in CP (23 upregulated and 5 downregulated) and 33 genes in PDAC-CP (1 upregulated and 32 downregulated; AKR1B1, − 2.76, Fig. 4D and KMT2C, − 2.67 Fig. 4E were tumour suppressors) were unique. The uniquely upregulated gene in PDAC-CP was SPHK1(Fig. 4A). Of the remaining 41 genes, three (PIGC, PKM and PPIB) were commonly upregulated with oncogenic potential (Fig. 4B) and 31 were commonly downregulated among CP and PDAC-CP (Fig. 4C), of which, 3 (AZGP1, EGLN1 and GNMT) were tumour suppressors (Fig. 3A).

Integrating 517 genes corresponding to altered metabolites and comparing with transcriptomes datasets of PDAC-CP and PDAC resulted in a total 127 genes that were dysregulated. There were 11 dysregulated genes common between PDAC-CP and PDAC; of which 3 were upregulated with oncogenic potential (PIGC, PKM and PPIB) and 8 were downregulated with no functional relevance. While 24 genes were down regulated (tumour suppressors AKR1B1 and KMT2C), SPHK1(2.11 fold) was the only upregulated gene in PDAC-CP. Splicing index of identified genes are listed in Additional file 6: Dataset S5.

We performed gene ontology (GO) enrichment analysis using PANTHER for the 102 dysregulated genes and found that majority of them were with hydrolase and transferase activity with predominant cytoplasmic localization affecting metabolic activity (Additional file 1: Fig. S4).

SPHK1 expression in acini and islets was comparatively low in PDAC-CP and PDAC in comparison to CP (H-score: acini- CP 262, PDAC-CP 138, PDAC 133; Islets-CP 250, PDAC-CP 150, PDAC 163). While the expression was marginally high in ducts (H-score: CP 175, PDAC-CP 193, PDAC 238), SPHK1 expression was significantly high in blood vessels in PDAC-CP (high positives 12/13, low positives 7/15) in comparison to CP (high positives 1/13, low positives 8/15) with a P-Value (0.015) (H-score: blood vessels- CP 189, PDAC-CP 242, PDAC 233), Fig. 5, Additional file 1: Table S5.