(1) BRB-Array Tools (version 4.6.0; http://linus.nci.nih.gov/BRB-ArrayTools.html) were used to analyze the DEGs (12); (2) the Cytoscape software (version 3.7.1; http://www.cytoscape.org) was applied to construct FI network and the network-based functional modules; (3) the FunRich (version 3.1.3; http://www.funrich.org) was used to construct biological pathways (13); (4) the Omicshare online database (https://www.Omicshare.com/), a commercial database based on the Database for Annotation, Visualization, and Integrated Discovery (DAVID), was used for the visual analysis of Gene Ontology (GO), and Kyoto Encyclopedia of Genes and Genomes (KEGG) was used for enrichment analysis (14); (5) the MultiExperiment Viewer (MeV) application (version 4.9.0; http://mev.tm4.org/) was used to generate the heat map; and (6) some bioinformatics online tools such as the Oncomine database (https://www.oncomine.org/), the Human Protein Atlas (https://www.proteinatlas.org/), the interactive web application Gene Expression Profiling Interactive Analysis (GEPIA) (https://www.gepia.cancer-pku.cn), and the TIMER database (https://cistrome.shinyapps.io/timer/) were used. A detailed flowchart of identifying the targeted genes associated with pancreatic cancer is shown in Figure 1.

Between 2004 and 2008, a total of 90 human pancreatic cancer specimens were collected from the Taizhou People's Hospital of China for tissue microarray (TMA) detection, 60 of which were paired nonmalignant tissue. All the specimens were stored at the Biobank Center of National Engineering Center for Biochips at Alenabio and Shanghai Outdo Biotech Company, China. All the patients received no preoperative anticancer treatment or postoperative adjuvant chemo-radio-therapy. Histological differentiation grade and disease stages were classified according to the American Joint Committee on Cancer (AJCC) TNM Classification (7th Edition). The follow-up time was recorded from the date of the surgery to the date of death or last visit to the clinic. The use of pancreatic cancer samples and clinical data were reviewed and approved by the Ethical Committees of the National Engineering Center for Biochips at Shanghai along with the Taizhou People's Hospital. Informed consent was written in accordance with the Declaration of Helsinki from all participants.

To investigate biological characteristics of genes differently expressed between pancreatic cancer and normal pancreatic tissues, five data sets were retrieved from National Center for Biotechnology Information (NCBI)-GEO (http://www.ncbi.nlm.nih.gov/geo/) for the analysis, including GSE101448 (18 with pancreatic tumor and 13 non-tumor pancreatic tissue samples), GSE62165 (118 pancreatic tumor samples and 13 control samples), GSE62452 (69 pancreatic tumor and adjacent non-tumor tissues), GSE41372 (matched pancreatic ductal adenocarcinoma samples from 15 patients), and GSE28735 (45 matching pairs of pancreatic tumor and adjacent non-tumor tissues). NCBI-GEO is a public repository containing an initial database of 1,325 biologically defined gene profiles and was developed to help in the analysis and interpretation of the long lists of genes produced from microarray-based experiments. The details of the series data are listed in Table 1.

As microarray batch effects appear to be the major contributors to the differential expression, SangerBox online tool (http://sangerbox.com/) was used to eliminate them before taken into account (15). Later, data were imported into the BRB-Array Tools to identify the DEGs between pancreatic tumor and normal pancreatic tissues. The Affymetrix platform (Santa Clara, California, USA) was applied for data set and gene annotation. The probes were filtered when some gene expression data were missing. The Bioinformatics and Evolutionary Genomics web tool (http://bioinformatics.psb.ugent.be/webtools/Venn/) was depicted graphically in the form of Venn diagram to detect the common DEGs among the five data sets.

The Reactome pathway database (version 7.2.0; http://www.reactome.org) was applied to analyze the pathway enrichment for the sets of genes, investigate the functional relationships of the genes, and visualize the outcomes with diagrams that can be manually generated (16). Over 60% of human proteins were covered to help construct the Reactome FI network. The pairwise interaction information was uploaded into Cytoscape software (http://www.cytoscape.org) to establish the FI network (17). A false discovery rate (FDR) <0.05 was accepted as the cutoff criterion.

To construct the FI network-based module analysis, the Microarray Data Analysis bioinformatics tool installed in the ReactomeFIViz was applied (18). After uploading the text file containing gene expression data, correlations between genes in the FI-network were analyzed. The Markov Cluster algorithm (MCL) was used to achieve the functional modules with a selected module size. In order to supervise the size of the modules, we inputted 5.0 for the inflation coefficient (19). MCL modules in size of three or more were applied, and the average Pearson's correlation coefficient was ≥0.25. Six different colors were used to label the nodes in six different network modules. Module size ≥7 and fluorodeoxyglucose (FDG) <0.05 were used as cutoffs for the KEGG pathway and GO term enrichment analysis of the modules in a separate manner.

The Human Protein Atlas (http://www.proteinatlas.org) presented an expression map of human protein across 20 cancer types and 44 normal human tissues. It measured the RNA level as well as used antibody profiling to precisely localize the corresponding proteins (20). With the Human Protein Atlas (http://www.proteinatlas.org), we detected the IFITM1 protein expression in both pancreatic cancer and normal pancreatic tissues. Oncomine microarray database (http://www.oncomine.org) is a cancer microarray database as well as an online data-mining platform which contains 65 gene expression data sets comprising nearly 48 million gene expression measurements from over 4,700 microarray experiments (21). With the Oncomine online tool (http://www.oncomine.org), we explore the expression of IFITM1 in pancreatic cancer and normal pancreatic tissues across eight different data sets.

Gene Expression Profiling Interactive Analysis (http://gepia.cancer-pku.cn/) is an online server, containing more than 8,000 normal and 9,000 tumor samples, providing interactive and customizable functions based on the Cancer Genome Atlas (TCGA) and Genotype-Tissue Expression (GTEx) database (22). Boxplot with jitter was applied to compare the expression level of IFITM1 mRNA in pancreatic cancer and paired normal pancreatic tissues. To validate the expression of IFITM1 correlating with survivals and pathological stages of patients with pancreatic cancer, the database was searched to obtain Kaplan–Meier survival curves and violin plots. The requested mRNA expressions above or below the median classified the patients into both high and low expression groups. The log-rank p-value and hazard ratio (HR) with 95% CIs were calculated and displayed on the website. A value of p < 0.05 was considered statistically significant.

To analyze the IFITM1 expression by immunohistochemical (IHC) staining experiment, 90 patients with pancreatic cancer were included in the TMA analysis. Sample slides were dried for adherence and deparaffinized in xylene for decreasing ethanol concentrations. The tissue sections were marked and incubated with rabbit anti-IFITM1 diluted 1:1,500 (ab224063, Abcam, Cambridge, UK) at 4°C overnight. The percentage of the IHC staining and the staining intensity were assessed and recorded by two pathologists from the department of pathology, who were both blinded to the features of patients. The H-scores ranged from 0 to 100, where 100 represented 100% of regions with strong intensity. The H-scores were dichotomized by the cutoff and the median value for the statistical analysis.

Three human pancreatic cancer cell lines, PanC-1, BxPc-3, and SW1990, and one normal pancreatic cell line, HPDE6-C7, were used in the study. The panC-1 and BxPc-3 cells were purchased from the Cell Bank of Shanghai Institute of Biochemistry and Cell Biology (Shanghai, China), and the others were purchased from the American Type Culture Collection (ATCC, Manassas, VA, USA). The panC-1 and HPDE6-C7 cells were cultured in Dulbecco's modified Eagle's medium (DMEM) (Gibco, Brooklyn, NY, USA), and the BxPc-3 and SW1990 cells were cultured in Roswell Park Memorial Institute Medium (RPMI) 1640 (Gibco, Brooklyn, NY, USA) with both cultures supplemented with 10% fetal bovine serum (FBS; Gibco, Brooklyn, NY, USA) and 100 U penicillin/streptomycin (Gibco, Brooklyn, NY, USA) in a humid atmosphere containing 5% CO₂ at 37°C. Short-tandem repeat (STR) profiles were authenticated to ensure the quality of the cell lines. All of the cells were passaged by standard cell culture techniques.

Two lentivirus-derived short hairpin RNAs (shRNAs) against IFITM1 were designed by Biolink Biotechnology Co. Ltd (Beijing, China) to silence the IFITM1 mRNA expression (the targeted sequences as follows: IFITM1 shRNA-1, 5′- CCTCATGACCATTGGATTCAT-3′; IFITM1 shRNA-2, 5′- CCTGTTCAACACCCTCTTCTT-3′. The procedure of transducing pancreatic cancer lines with shRNA lentiviral transduction was described in detail previously (23).

Cells in logarithmic growth phase were seeded in 96-well Falcon Petri dishes (3,000 cells/well) the day before the assays were performed. The cell viability was assessed using a Cell Counting Kit (Dojindo, Kumamoto, Japan) based on sensitive colorimetric assays with the absorbance measured at 450 nm (OD 450 nm), according to the instructions of the manufacturer. The results were confirmed with three independent experiments.

To extract total RNA, cells were lysed with the TRIzol Reagent (Invitrogen Corporation, Carlsbad, CA, USA) and reversed into complementary DNA (cDNA) with the QuantiTect Reverse Transcription Kit (TaKaRa, Dalian, China). The sequences of the designed primers used for the real-time PCR (qRT-PCR) experiments were as follows: IFITM1 forward, 5′-AGCCAGAAGATGCACAAGGA-3′ and reverse, 5′-GATCACGGTGGACCTTGGAA-3′; GAPDH forward, 5′- GAAGGTCGGTGTGAACGGATTTG-3′ and reverse, 5′- CATGTAGACCATGTAGTTGAGGTCA-3′ (Sangon, Shanghai, China).

Radioimmunoprecipitation assay (RIPA; Invitrogen Corporation, Carlsbad, CA, USA) was added to protease, and phosphatase inhibitors were used to gain total protein samples. The proteins were separated by the sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE) gel (Bio-Rad, Hercules, CA, USA) and transferred to the polyvinylidene fluoride (PVDF) membrane (Sigma Chemical Corporation, St. Louis, MO, USA). The membranes were blocked for 1 h and then incubated overnight with primary antibodies against IFITM1 and β-actin (Abcam, Cambridge, UK), followed by the horseradish peroxidase (HRP)-linked secondary antibodies. SuperSignal West Femto Maximum Sensitivity Substrate (Thermo Fisher Scientific, Waltham, MA, USA) was used to detect the probed proteins.

The cells were seeded in 48-well Falcon (Thermo Fisher Scientific Inc. Waltham, Massachusetts, USA) Petri dishes. Once 80–90% confluence was reached, a 200-μl sterile plastic pipette tip was used to make a single wound by gently scratching in every well. Cell debris was removed using phosphate-buffered saline and refreshed with serum-free medium immediately. Cells that extended the borders of the wound were photographed and quantified in three randomly selected regions per well.

The cells in the logarithmic growth phase were seeded in 6-well Falcon Petri dishes (500 cells/well) and incubated for 7–14 days. The medium was refreshed every 3 days. The colonies were fixed with 70% ethanol and stained with crystal violet (0.5% w/v). The cell colonies containing at least 50 cells were photographed and quantified. The results were presented as mean ± SD for three independent experiments.

Statistical analyses were performed using SPSS 23.0 (SPSS Inc., Chicago, IL, USA) and Prism 8 (GraphPad Software, San Diego, CA, USA). Correlations between protein expression and clinicopathological features in patients with pancreatic cancer were analyzed by the χ²-test. Overall survival (OS) was evaluated using the Kaplan–Meier method and the log-rank test. A Cox proportional hazards regression model was used to assess the univariate and multivariate analyses. The Enter method was used to select the independent variables in the multivariate analysis.

Five data sets from GEO containing gene expression profiles of pancreatic cancer as well as normal pancreas samples were downloaded for microarray analysis. The random variance mode method along with the paired t-test was applied for the differential expression calculation. Considering the criteria of |log2 (Fold change) |>1 adjusted p < 0.05, 483 DEGs were obtained (Figure 2A). Of those, 231 DEGs were identified upregulated and 252 downregulated (Supplementary Table 1).

Functional interaction data derived from the Human Protein Reference Database (http://www.hprd.org/) (24), BioGrid (https://thebiogrid.org/) (25), I2D (http://ophid.utoronto.ca/) (26), IntACT (https://www.ebi.ac.uk/intact/) (27), MINT (https://www.ebi.ac.uk/intact/) (28), the Database of Interacting Proteins (http://dip.doe-mbi.ucla.edu) (29) as well as multiple high-throughput assays (30). We established the FI network by mapping the pancreatic cancer-related DEGs to the FIs data. Of 245 isolated nodes, 227 in eight clusters were identified with an effective mean degree of 2.20 (Figure 2B). The connectivity of a node represents the number of its neighbors, and the neighborhood connectivity represents the average connectivity of all neighbors of the selected node. Under the criteria of neighborhood connectivity ≥20 and the shortest path length ≥3.5 for non-single modules, eight genes (EDIL3, RSAD2, AHR, GBP2, IFITM1, AOX1, KIF11, and CEP170) among all the nodes in the FI network were filtered out (Supplementary Table 2).

To gain insights into the role of the putative targets involved in various BPs and molecular functions (MFs), the preliminary GO function and KEGG pathway analysis were performed using the OmicShare online database. As expected, among the top 20 KEGG pathways, extracellular matrix (ECM)–receptor interaction, pancreatic secretion, and pathways in cancer had significant relations to tumorigenesis and progression of pancreatic cancer (Figure 2C; for details see Supplementary Table 3). Among the most highly enriched functions in the BP category, cell process, biological regulation, and regulation of BP were found associated with the tumorigenesis and progression of pancreatic cancer. Meanwhile, cell, cell part, organelle, and membrane were top enriched items in the cellular component (CC) category. In the MF category, the DEGs were mainly enriched in binding activities (Figure 2D; for details see Supplementary Tables 4–6).

The Markov Cluster algorithm (MCL) was applied for the network-based module-clustering algorithm, and the absolute value of the Pearson's correlation coefficient of the interacting genes was used to weigh each correlation edge. A total of 6 modules involving 63 genes were obtained, ranging from 7 to 19 genes per module (Table 2, Figure 3A). Later, GO function and pathway analyses were performed on the six network-based functional modules. Through enrichment analysis (Supplementary Table 7), we identified that Module 0 containing 19 genes was in relation to the classical cell–cell adhesion process. Module 1 was of the highest average correlation and was related to the classical cancer signaling pathway in the KEGG and IFN-gamma pathway in NCI-PID. Module 2 was found related to cell motility and proliferation in KEGG and Module 3 to cell migration and cytoskeleton. Fibro-relevant families in Reactome were enriched in Module 4. Meanwhile, the enrichment of GO term was also analyzed. The calculation results showed that the GO functions of six modules complied with the pathway annotations (Supplementary Tables 8–10).

Five genes (IFITM1, AHR, AOX1, GBP2, and RSAD2) common to both Module 1, with an average correlation equal to 1.00, and the set of previously selected eight genes were selected for the volcano plot analysis of the five GEO data sets (Figure 3B, IFITM1 marker in the red triangle). Through comprehensive literature searching, studies relative to AHR and pancreatic cancer have been taken, but not enough background data was found about AOX1, GBP2, and RSAD2 (31). Combining the results from the functional analysis of the FI network and literature mining, we were interested to see if IFITM1 could be a potential biomarker in pancreatic cancer.

First, we used the Oncomine database to explore the transcriptional level of IFITM1 in pancreatic cancer and normal pancreatic tissues (Figure 4D). Based on the data from the Oncomine database, the transcriptional level of IFITM1 was significantly elevated in pancreatic cancer tissue vs. normal pancreatic tissue. Same results can also be seen in colorectal, esophageal, gastric, liver, and many other cancer types. Eight studies reported from 2003 to 2009 were filtered out with the screening condition of “IFITM1; Cancer vs. Normal Analysis; Cancer Type: Pancreatic Cancer” and assessed with a heat map analysis (Figure 4A). Segara et al. (32) found that the level of IFITM1 in pancreatic cancer was significantly increased compared to adjacent normal pancreas samples, with a fold change of 5.521 and a p-value of 6.78E-7. The same results were also reported by Badea et al. (33) (fold change of 3.065 and p-value of 8.21E-9), Logsdon et al. (34) (fold change of 4.104 and p-value of 1.14E-4), Ishikawa et al. (35) (fold change of 2.350 and p-value 0.008), and Iacobuzio-Donahue et al. (36) (fold change of 2.855 and p-value of 0.005), respectively. While Buchholz et al. (37) and Grutzmann et al. (38) found no significant upregulation in pancreatic cancer with p-values of 0.412 and 0.231, respectively. Recently, Pei et al. (39) studied 36 samples of pancreatic carcinoma and 16 paired normal samples, and statistical significance was reported with a fold change of 2.304 and a p-value of 0.002. Meanwhile, we found the protein expression of IFITM1 with the help of the Human Protein Atlas, and the positively strong level was also found in pancreatic cancer specimens compared with normal tissues (Figures 4B,C). Furthermore, the GEPIA online tool (https://www.gepia.cancer-pku.cn) was also used to validate the differential mRNA expression analysis of IFITM1 in pancreatic cancer and normal tissues (Figure 4E).

Immunohistochemical experiments were applied to validate IFITM1 expression in 90 patients with pancreatic cancer. Representative IHC-stained images showed that IFITM1 is mainly located in the cell membrane of the pancreatic cancer tissues. A significantly upregulated expression of IFITM1 was found in 76.7% of the patients (69/90), which was consistent with the results analyzed by the microarray data (p < 0.001, see Figures 5A,B).

The χ²-test was applied to assess the relationship between the expressions of IFITM1 and the clinical and pathologic features of patients with pancreatic cancer, including gender, age at diagnosis, tumor size, histologic differentiation, TNM stage, nodal status, and distant metastasis. Histologic differentiation (p = 0.032) and TNM stage (p = 0.044) were found to have close correlations with IFITM1 expression (Table 3). Higher expression of IFITM1 correlated with poor histologic differentiation and larger TNM stage. The significant correlation between IFITM1 and the TNM stage was also verified using the GEPIA (https://www.gepia.cancer-pku.cn; Figure 5C). Consistent with our investigation, IFITM1 might be an oncogene for pancreatic cancer that participates in tumorigenesis.

A total of 184 patients with pancreatic cancer in the TCGA database, with 5 of them having no documented T or N stage, leaving 179 patients for the confirmation of the relationship between IFITM1 and human pancreatic cancer. The Kaplan–Meier analysis and the log-rank test were used for analysis. As shown in Figure 6A, patients with a low level of IFITM1 had better disease-free survival than those with high IFITM1 (log-rank p = 0.036) as well as had a better OS (Figure 6B, log-rank p = 0.014).

For the survival analysis, the median IHC value was used as an objective cutoff to stratify 90 patients with pancreatic cancer with low and high expressions. We used the Kaplan–Meier analysis and the log-rank test to explore the correlation. The follow-up time was 81 months, and a high IFITM1 level was found significantly related to decreased OS (p = 0.032; Figure 6C). The median OS of the patients with pancreatic cancer with the low and high IFITM1 expression was 43.0 vs. 10.0 months, respectively.

As for the univariable Cox regression analysis, positive nodal status (p = 0.008), TNM stage (p = 0.012), histologic differentiation (p = 0.004), and IFITM1 expression (p = 0.040) were significantly associated with worse outcome (Table 4). In the multivariable analysis, IFITM1 expression (HR = 2.134; 95% CI 1.073, 4.245; p = 0.031) remained an independent prognostic marker for OS of patients with pancreatic cancer (Table 4).

IFN-induced transmembrane protein 1 was found involving in inflammatory processes and immune cells infiltration, which might further influence the clinical outcomes of patients with pancreatic cancer. We used the TIMER database (https://cistrome.shinyapps.io/timer/) to explore a comprehensive relation between IFITM1 levels and immune infiltrations. The positive correlations were assessed between IFITM1 and B cell (Cor = 0.171, p = 2.51E-2), CD4+ T cell (Cor = 0.307, p = 4.90E-5), macrophage (Cor = 0.204, p = 7.34E-3), neutrophil (Cor = 0.475, p = 5.20E-11), and dendritic cell (Cor = 0.421, p = 1.01E-8). IFITM1 level was in negative association with purity (Cor = −0.165, p = 3.1E-2). Only CD8+ T cell was found having no significant relation to IFITM1 expression (Figure 7).

Three types of pancreatic cancer cell lines and normal human pancreatic duct epithelial cell line HPDE6-C7 were used to examine the expressions of IFITM1. As shown in Figures 8B,C, higher mRNA and protein levels of IFITM1 were found in BxPc-3, PanC-1, and SW1990 compared to HPDE6-C7. Two shRNA targeting IFITM1 were transfected into PanC-1 and SW1990 to detect the IFITM1 silencing effects, and the efficiency was examined both by qRT-PCR and Western blot analysis (Figures 8D,E). We performed cell counting kit-8 assay to assess the cancer cell proliferation. The results showed that the downregulation of IFITM1 significantly decreased the proliferation of PanC-1 and SW1990 cell lines (Figure 8A). Moreover, the wound-healing assay indicated that IFITM1 silencing inhibited cell migration capability, and the colony-formation assay showed that it greatly reduced the cancer stem cell-like properties of pancreatic cancer (Figures 8F–I).