Analysis of dynamic molecular networks for pancreatic ductal adenocarcinoma progression

Human PDAC tissues with four different stages and non-tumorous tissues were obtained from five patients who underwent surgical resection at Zhejiang Cancer Hospital. All specimens had a pathological diagnosis at the time of assessment. Studies were approved by the Ethics Committee of Zhejiang Cancer Hospital.

Microarray dataset GSE62165 containing staging pancreatic ductal adenocarcinoma (PDAC) and non-tumoral pancreatic tissue was obtained from Gene Expression Omnibus (GEO, http://​www.​ncbi.​nlm.​nih.​gov/​geo/​) database in the National Center for Biotechnology Information (NCBI), which was deposited by Janky and colleague [7]. The GSE62165 dataset included 118 surgically resected PDAC varying from stage I to stage IV and 13 control samples. The detail PDAC patient cohort consisted of 8 stage I samples, 30 stage IIa samples, 62 stage IIb samples, 5 stage III samples and 13 stage IV samples. Additionally, the gene expression profile was detected basing on HG-U219 (Affymetrix Human Genome U219 Array) platform.

The CEL file data of GSE62165 was read using the affy package in the R language software (version 3.2.3, https://​www.​r-project.​org/​). Background correction, normalization, expression calculation of the original array data and log2 transformation were processed by robust multichip average (RMA) algorithm. Empirical Bayes method was used to identify significant DEGs between different stages of PDAC and normal pancreatic samples basing on the limma package in R. P values were adjusted for multiple testing depending on the Benjamini–Hochberg False Discovery Rate (FDR) method. The strict thresholds for identifying DEGs were set as FDR < 0.01 and |log2 fold change (FC)| ≥ 2.

The software program Short Time-series Expression Miner (STEM) is designed for clustering, comparing, and visualizing gene expression data from short time series microarray experiments [8]. To search the differences of gene expression between PDAC stages, the STEM tool was applied to profile candidate genes simultaneously from normal status to stage IV. Briefly, the union of DEGs from five stages was input and normalized by STEM. The maximum number of model profiles was set as 100, and FDR was selected as correction method. The rest parameters were remained unchanged. Colored clusters indicated statistically significant number of genes assigned.

Gene ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) were applied for the functional annotation and pathway enrichment analysis of DEGs through using the Database for Annotation Visualization and Integrated Discovery (DAVID; https://​david.​ncifcrf.​gov/​) [9‐11]. Adjusted P values were calculated by Benjamini–Hochberg FDR method and the thresholds were set as FDR < 0.05 to indicate a statistically significant difference.

The online database Search Tool for the Retrieval of Interacting Genes (STRING, http://​string-db.​org/​) is commonly used to identify the interactions between known proteins and predict proteins and to construct a PPI network [12]. The DEGs were input into STRING to construct PPI network and further visualized by Cytoscape (version 3.6.1) software. The Cytoscape plugin DyNetViewer was applied to construct the stage-course dynamic protein interaction network [13]. The time course protein interaction networks (TC-PIN) algorithm was chosen to constructing sub-networks and the threshold was set as 2. For dynamic node centrality analysis, three typical centrality measures including degree centrality (DC), betweenness centrality (BC), local average connectivity-based method (LAC) were used. The top 20 DC, BC, and LAC of genes that uniquely elevated at each stage were firstly identified, respectively. Then, the common genes among the top 20 DC, BC, and LAC lists at each stage were considered as key nodes for dynamic network development. For dynamic module analysis, clustering algorithm molecular complex detection (MCODE) was conducted for analyzing clusters of dynamic networks. The degree cutoff was set as 5 and K-core was set as 2.

To identify the stage-specific activated kinase during PDAC progression, total protein kinase list was acquired from the human kinome database (kinase.com) [14] and was intersected with DEGs. Then, the dynamic DC, BC and LAC of intersection kinases were retrieved from the results of dynamic node centrality analysis described above. Furthermore, the substrates of kinases were searched from PhosphoSitePlus (https://​www.​phosphosite.​org) online tool. The sequence characteristics of kinase substrates were depicted by sequence logo to indicate the phosphosite markers.

The online database Kaplan–Meier plotter (www.​kmplot.​com) is capable to retrieve gene expression data and clinical information from GEO, European Genome-phenome Archive (EGA) and The Cancer Genome Atlas (TCGA) [15]. To evaluate the prognostic value of candidate genes, the patient samples were split into two cohorts according to the best cutoff of gene expression computed by Kaplan–Meier plotter. To analyze the relationship between particular pathways and overall survival, the mean expression of pathway associated signature genes and best cutoff were calculated by Kaplan–Meier plotter. Besides, the log rank P value and hazard ratio (HR) with 95% confidence intervals were also computed.

GEPIA is a newly developed interactive web server for estimating the RNA sequencing expression data from the TCGA and Genotype-Tissue Expression (GTEx) dataset projects [16]. Candidate genes were queried by GEPIA to explore their expression levels at different stages using TCGA-pancreatic adenocarcinoma data.

Paraffin-embedded sections were deparaffinized and rehydrated firstly. Then, 1 mM EDTA (pH 8.0) was applied to retrieve the antigen. Endogenous peroxide activity was block by 0.3% hydrogen peroxide. Before incubating with primary antibody, 5% goat serum in TBS was used to avoid non-specific binding. The primary antibodies contained rabbit anti-MFL1IP (1:100, Proteintech, USA), rabbit anti-LAMA3 (1:200, Abbkine, China) and rabbit anti-LAMB3 (1:200, Abbkine, China). After incubating with biotinylated secondary antibodies, sections were developed using DAB (Beyotime, China) and counterstained with hematoxylin.

The microarray dataset GSE62165 was acquired from GEO. Differentially expressed genes (DEGs) (FDR < 0.01, |log2 FC| ≥ 2) of different PDAC stages were screened out basing on the R analysis, respectively. Compared with normal tissues, the total DEGs number for each stage were 994 (stage I), 967 (stage IIa), 965 (stage IIb), 1027 (stage III), 925 (stage IV). Then, the heatmaps of the top 10 up and down-regulated DEGs at each stage were hierarchically clustered and displayed, respectively (Fig. 1a–e). As the results showed in Fig. 1a–e, these top 20 DEGs could clearly distinguish each PDAC stage from normal pancreatic tissues.

To explore the differences of gene expression between PDAC stages, the STEM tool was applied to profile stage-course gene expression patterns. A total of 30 expression manners were enriched among 100 assumed expression models (Fig. 1f). Intriguingly, we found some clusters exhibited stage-specific expression trends throughout PDAC progression. Genes in the Model 89 (115 genes enriched) were specifically up-regulated at stage I. The Model 6 contained 166 genes and particularly elevated at stage IIa. The Model 3 (113 genes enriched) and Model 2 (45 genes enriched) were stage IIb- and stage III-specific gene clusters, respectively. Additionally, the Model 1 associated genes specifically increased at stage IV. These dynamic expression patterns suggested that distinct molecular signals were required for particular PDAC stages. Notably, we also found six genes that dramatically changed from normal status to stage IV (Fig. 1g–h). TNNT1, the troponin T type 1 encoding gene, was up-regulated increasingly from cancer initiation to advanced stage. Other five genes, including BSPRY, C8ORF47, FAM3B, HOOK1 and REG1A, were clustered in Model 16 and significantly decreased during cancer progression. These six genes may act as key regulators for driving PDAC progression via their special expression manners.

To further investigate the biological function of DEGs between different PDAC stages and normal tissues, gene ontology (GO) analysis was performed using the DAVID online analysis tool. Basing on the results of GO biological process (BP) enrichment (Fig. 2a–f), we found up-regulated DEGs of five stages were commonly enriched in six biological processes, including cell adhesion, extracellular matrix organization, immune response, collagen fibril organization, collagen catabolic process, and endodermal cell differentiation. Moreover, extracellular matrix disassembly was shared by stage I, IIa, IIb, and IV. Notably, wound healing was enriched at stage IIa, IIb, III, and IV. Leukocyte migration was found in stage I, IIa, and III. Besides, type I interferon signaling pathway was commonly enriched by stage I, IIb, IV. Angiogenesis was particularly found in advanced PDAC including stage III and IV. Interestingly, inflammatory response was only found at early stage (stage I).

Functional annotation of down-regulated DEGs indicated that 8 biological processes were commonly enriched at each PDAC stage (Fig. 2a–e, g), including lipid digestion, cellular response to zinc ion, lipid catabolic process, reactive oxygen species metabolic process, cobalamin metabolic process, proteolysis, cellular amino acid metabolic process, digestion. Transmembrane transport was shared by stage I and IIa, while negative regulation of growth was conspicuously found at stage IIb, III, and IV. Additionally, meiotic gene conversion, cellular amino acid biosynthetic process, amino acid transport, cellular response to cadmium ion, and metabolic process were specifically enriched at each stage, respectively.

Subsequently, KEGG pathway analysis demonstrated that the up-regulated DEGs of different PDAC stages were commonly enriched in five key pathways including pathways in cancer, small cell lung cancer, ECM-receptor interaction, amoebiasis, focal adhesion (Fig. 3a–f). Staphylococcus aureus infection was enriched at every stage except stage IV. Phagosome was shared by stage I, IIa, III, while rheumatoid arthritis was only found at stage I and IIa. Protein digestion and absorption occurred at all stages except stage IIa. PI3K-Akt signaling pathway was enriched at stage I, III, and IV. Three pathways including p53 signaling pathway, platelet activation, and cell cycle were specifically enriched at stage IV.

The down-regulated DEGs of different PDAC stages were commonly enriched in five key pathways including metabolic pathways, glycine, serine and threonine metabolism, pancreatic secretion, protein digestion and absorption, fat digestion and absorption (Fig. 3a–e, g). Three pathways including mineral absorption, bile secretion, proximal tubule bicarbonate reclamation were found in all PDAC stages except stage IV. Biosynthesis of amino acids occurred at early stage (stage I, IIa, IIb). Maturity onset diabetes of the young was shared by stage I, IIb, III, IV. Metabolism of xenobiotics by cytochrome P450 was found in advanced PDAC at stage III and IV. For stage IV, three pathways were particularly enriched, including chemical carcinogenesis, drug metabolism—cytochrome P450, and retinol metabolism.

To analyze the prognostic relevance of five key pathways shared by all PDAC stages, mean expression value of pathway associated genes and best cutoff were calculated by Kaplan–Meier plotter. As shown in Fig. 4a–e, pathways in cancer (P = 0.0047), small cell lung cancer (P = 0.0024), ECM-receptor interaction (P = 0.039) and focal adhesion (P = 0.048) were negatively associated with overall survival (OS). Moreover, the intersection of five pathways associated genes indicated that LAMB3, LAMA3, COL4A1, LAMC2 and FN1 were commonly enriched in these pathways (Fig. 4f). Kaplan–Meier survival analysis suggested that these five genes were unfavorable factors of prognosis (Fig. 4g–k).

The PPI networks of the DEGs between PDAC and normal pancreatic tissues at different stages were constructed by the online database STRING. The typical PPI network of stage IV was shown in Fig. 5, and the network consisted of 868 nodes interacting via 4717 edges. Expression level of up-regulated DEGs and down-regulated DEGs in the PPI network was shown in red and blue.

DyNetViewer is a novel Cytoscape app for constructing, analyzing, and visualizing dynamic molecular interaction networks. To analyze the dynamic cluster attribution for PDAC progression, the networks from five stages were input into DyNetViewer and calculated by MCODE algorithm over time. The clusters at each stage which exhibited highly interconnected regions were considered as key regulators for promoting networks development. A total of 45 featured molecular modules were identified for all PDAC stages (Additional file 1: Figures S1–S5). At stage I, 11 clusters that contributed to network were specifically screened out, among which the cluster 11 also contributed to the network of stage IIa and IIb (Fig. 6a). These 11 gene clusters were enriched in pathways of influenza A, cell cycle, rheumatoid arthritis, Staphylococcus aureus infection, amoebiasis, protein digestion and absorption, ECM-receptor interaction and platelet activation (Fig. 6f). Furthermore, we found that key genes including IL1A, CXCL8, ICAM1, ITGB2, ITGAM, COL1A1 and COL1A2 participated in as much as four pathways at stage I, respectively (Fig. 6f). Then, we found that ten clusters were particularly involved in the network of stage IIa, and the cluster 13 was also identified at stage IIb (Fig. 6b). Pathway enrichment showed that these clusters were mainly enriched in cell cycle, chemokine signaling pathway, influenza A, vascular smooth muscle contraction, focal adhesion, leishmaniasis, Staphylococcus aureus infection, protein digestion and absorption, ECM-receptor interaction, complement and coagulation cascades, basal cell carcinoma and platelet activation (Fig. 6g). CXCL8, ADCY7, ITGAM, ITGB2, ITGB1, IL1A, ICAM1, ITGA2, THBS2, SDC1, COL3A1, COL1A2 and COL1A1 were commonly shared by four or more pathways, respectively (Fig. 6g). In the network of stage IIb, ten clusters were found acting as key modules for the protein interaction network (Fig. 6c). These clusters were associated with cell cycle, vascular smooth muscle contraction, influenza A, Staphylococcus aureus infection, focal adhesion, complement and coagulation cascades, protein digestion and absorption and ECM-receptor interaction (Fig. 6h). ADCY7, ITGAM, ITGB2, MYL9 were shared by four or more pathways, respectively (Fig. 6h). For stage III, we found 9 unique clusters were essential for its network (Fig. 6d). These gene clusters mainly involved in cell cycle, vascular smooth muscle contraction, arrhythmogenic right ventricular cardiomyopathy, amoebiasis, complement and coagulation cascades, protein digestion and absorption and ECM-receptor interaction (Fig. 6i). Two key genes including CXCL8 and TGFB3 were shared by as much as four pathways, respectively (Fig. 6i). Moreover, a total of 8 clusters specifically participated in the network of stage IV (Fig. 6e). These clusters mainly enriched in cell cycle, vascular smooth muscle contraction, dilated cardiomyopathy, IL-17 signaling pathway, protein digestion and absorption and ECM-receptor interaction (Fig. 6j). No genes were found shared by more than two pathways at stage IV (Fig. 6j).

To elucidated the key genes that participated in the dynamic networks, the node centrality at each PDAC stage were calculated. Three typical centrality measures including Betweenness Centrality (BC), Degree Centrality (DC), Local Average Connectivity-based method (LAC) were applied in our study. We screened out 19 key genes that potentially involved in the dynamic network development (Fig. 7a–c). Five genes including NDC80, KIF2C, KIF20A, OIP5, ZWINT were specifically found at stage I with highest DC, BC, and LAC. FGA, CRP and ITGB1 were featured nodes for stage II. WNT5A was essential for maintaining the network at stage IIa and IIb. F5 was uniquely found at stage IIb. Additionally, five genes including ITGA4, COL6A2, ITGA9, THBS1 and SERPINE1 were characteristic nodes at stage III. Four nodes including ITGB4, MLF1P, TRIM22 and CDC25B were potential key genes for promoting the advanced status of PDAC. Finally, we also sorted the top ten nodes ranked by standard deviation of DC, among which NCD80, KIF20A, ITGB1 and KIF2C also existed in the 19 key nodes mentioned above (Fig. 7d). The remaining genes including NCAPG, CENPE, KIAA0101, RACGAP1, ITGB5 and AURKA were also dynamically changed at particular stage (Fig. 7d).

According to the node centrality analysis above, we chose the featured nodes of stage IV to determine their clinical relevance. Kaplan–Meier survival analysis showed that high expression level of MLF1IP (also known as CENPU) and ITGB4 were significantly correlated with shorter overall survival, respectively (Fig. 8a). Furthermore, we also analyzed the expression manner of featured nodes (MLF1IP/CDC25B/ITGB4/TRIM22) during PDAC progression (Fig. 8b). The expression level of these four genes at stage IV showed no significant changes compared with other stages, which indicated that these nodes potentially promoted cancer progression by directly maintaining the molecular network without depending on their expression level (Fig. 8b).

Considerable studies have shown a causal role of protein kinase mutations or dysregulations in tumorigenesis and cancer progression. Cancer research has been trying to turn these molecules into valid drug candidates for emergence targeted therapies. Depending on the dynamic networks, the PDAC stage-specific activated protein kinases were also identified. As shown in Fig. 9a–f, a total of 15 kinases involved in the dynamic networks with distinct patterns. TTK, AURKA, BUB1, CDK1 and NEK2 were fundamental kinases with high node degree, which may be required for maintaining the molecular signals underlying tumor progression (Fig. 9a). Besides, we also found CHEK1, the checkpoint kinase 1 coding gene (degree = 32), specifically activated at stage IV (Fig. 9d), which may indicate a potential drug target for advanced PDAC. The kinase ABR was the featured node for stage III, though getting low node degree (Fig. 9d). Additionally, PDGFRB was particularly found at stage IIa, IIb and III in the dynamic network (Fig. 9d). Finally, we also predicted the phospho-targets of eight candidate kinases. The protein kinases AurA (encoded by AURKA), CDK1, Chk1 (encoded by CHEK1), NEK2 and TTK phosphorylated their substrates by targeting serine, threonine and tyrosine, while BUB1 and skMLCK (encoded by MYLK) targeted the serine and threonine for phosphorylation (Fig. 9g). Tyrosine was the only phosphorylation site of PDGFRB substrates (Fig. 9g). Moreover, the consensus sequences of kinase substrates were depicted by sequence logo to indicate the phosphosite markers (Fig. 9g).

To investigate the expression patterns of MLF1IP, LAMA3 and LAMB3 during PDAC progression, we performed the immunohistochemistry (IHC) analysis. As a featured node of stage IV, MLF1IP expressed weekly in normal pancreatic tissue (Fig. 10). With the progression of PDAC, MLF1IP dramatically increased in PDAC tissues and was cytoplasmic and nuclear localization. LAMA3 and LAMB3 were common genes shared by five fundamental pathways throughout different PDAC stages. The IHC analysis showed that LAMA3 and LAMB3 were positively stained in PDAC tissues with a progressive increase manner (Fig. 10). Interestingly, we found that LAMB3 positively expressed in normal pancreatic connective tissue. With the initiation of PDAC, LAMB3 mainly existed in stroma and cytoplasm of cancer cells. However, LAMB3 gradually translocated into the nucleus of cancer cells at advanced PDAC stages. This special expression manner may indicate the pathologic molecular basis of tumorigenesis and progression.