Introduction

With 496,000 newly diagnosed cases globally and 466,000 related deaths in 20201, pancreatic cancer is categorized among the malignancies with the poorest prognostic outcome2. According to the cancer statistics of the International Agency for Research on Cancer (IARC) GLOBOCAN, the incidence rate of pancreatic cancer has been rising in recent decades, and it accounts for 4.9% and 4.5% of worldwide cancer incidence and related deaths, respectively1. Pancreatic ductal adenocarcinoma (PDAC), the most common subtype of pancreatic cancer, accounts for over 90% of the cases3. Despite being the 10th most prevalent cancer, PDAC is the seventh most common cause of cancer-related deaths worldwide due to its poor prognosis4. Although the 5-year survival rate of pancreatic cancer differs regionally, it is < 10% due to a lack of clear clinical manifestations until advanced stages5. The primary reasons for the low survival rate of pancreatic patients are that the disease remains asymptomatic until advanced stages due to the anatomical position of the pancreas in the retroperitoneum and the lack of valuable biomarkers for early stages can be considered as other reasons6,7. Clinical biomarkers play a pivotal role in diagnosing and managing various cancers, including pancreatic cancer. CA-19-9 is one such biomarker commonly used for pancreatic cancer. CA-19-9 is a carbohydrate antigen that can be detected in the blood of some pancreatic cancer patients. Elevated levels of CA-19-9 may indicate the presence of pancreatic cancer, but it is important to note that this biomarker is not specific to pancreatic cancer. Other conditions, such as liver disease, bile duct obstruction, and certain gastrointestinal tumors, can also cause increased CA-19-9 levels8. Although the primary aetiology of pancreatic cancer has not been identified, some genes have been previously shown to be associated with the various cancer subtypes, treatment responses, or poor prognosis in pancreatic cancer9,10,11. Many cancers cannot be effectively treated in the advanced stages of disease, therefore developing novel biomarkers for the early stage is a potential approach for diagnosis, prognosis, and treatment of pancreatic cancer12. Currently, the K-RAS gene is known to be one essential gene playing a crucial role in pancreatic cancer, with a prevalence of more than 85%. Furthermore, P53 and P16, as tumor suppressor genes, are inactivated in approximately 95% of pancreatic cancer patients6. To activate or inactivate proto-oncogenes and other related genes like those functioning as tumor suppressors, including HER2, MYB, AKT2, BRCA2, FHIT, CDKN2A, PALB2, STK11, and PRSS1 are involved. Furthermore, the analysis of mutations in BRCA1/2, MMR(mismatch repairing system), and NTRK1–3 fusions was performed for pancreatic cancer patients receiving the treatment of pembrolizumab, entrectinib, and larotrectinib13. Advanced technologies such as bioinformatics and artificial intelligence are developed to provide cancer research opportunities14,15. Machine learning is a part of artificial intelligence that can improve the accuracy of cancer diagnosis, prediction, and prognosis by employing various statistical techniques16,17,18,19,20.

MiRNAs are non-coding RNAs with a length of 19–24 nucleotides that regulate gene expression of more than 30% of human genes following transcription. They pair to their target's untranslated 3′(3′-UTR) region of mRNAs, resulting in inhibition or degradation of the mRNAs21. Up or down-regulation or misplacement of miRNAs may play crucial roles in cancer development, tumor cell proliferation, migration, invasion, and chemical resistance22,23,24,25,26. These modifications and abnormalities in the miRNA transcription levels have previously been reported in several human malignancies27,28,29. It is hypothesized that genes and miRNAs might be evaluated as biomarkers to initiate better diagnostic or predictive approaches for pancreatic cancer. Previous studies have targeted KRAS and other genes in pancreatic cancer. Some miRNAs, including hsa-miR-217, hsa-miR-96, miR-216a, and miR-148a/b, are reported downregulated, and some, such as the miR-221, miR-210, miR-155, and miR-21 upregulated in pancreatic cancer30,31,32,33,34.

The Cancer Genome Atlas (TCGA) is a project that maps out the genome variation of human cancerous cells by RNA sequencing and using a non-malignant cell as a reference. These maps have identified many core genetic pathways activated in various cancers35,36. Therefore, in the current study, we performed gene expression proofing of pancreatic cancer using the TCGA database and Machine learning to identify differential expression genes (DEGs) and differentially expressed miRNAs (DEmiRNA). Survival was assessed using Kaplan–Meier analysis to predict prognostic biomarkers and the risk model. Additionally, the protein–protein interaction (PPI), the molecular pathways, the co-expression of DEGs, and the correlation between candidate genes and pancreatic cancer with clinical data were evaluated. Furthermore, the diagnostic markers were detected based on machine learning technology (Fig. 1A).

Figure 1
figure 1

(A) The overall workflow, (B, C) miRNAs and important genes identified by machine learning in PDAC at different stages. (D, E) The association of clinical variables with cancer that were not significant.

Full size image

Material and method

Data collection

The TCGA database (http://tcga-data.nci.nih.gov/tcga/) was utilized to extract pancreatic cancer gene and miRNA data from 183 and 193 samples including healthy and tumor samples. RNA gene expression, microRNA, and clinical data were downloaded.

Data preprocessing and the identification of DEGs (differential expression genes)

In the pre-processing step, gene expression data were filtered to eliminate the gene and miRNA with zero expression and duplicates. Then, data was normalized with Limma and DESEQ2 packages in R 4.0.3 software. Filtering and normalization are the most important step in data analysis performed before machine learning. Genes and miRNAs were adjusted between pancreatic cancer samples and healthy tissue samples based on the particular criteria, including P < 0.05 and − 1.5 <|Log2FC (fold change) |< 1.5, to evaluate the upregulated and downregulated genes of the data integrity and subsequent analysis. The heatmap was created by “cluster, dendextend, circlize, RcolorBrewer, ComplexHeatmap, d3heatmap, gplots, pheatmap, and gplots” packages in R software version 4.3.1.

Identifying predictive markers

Machine learning methods can be used to analyze the data collected from various biological data, such as genomics, transcriptomics, and metabolic data27. Our study used machine learning algorithms, including Random Forest (RF), Max Voting, Adaboost, Gradient Boosting machines(GBM), and Extreme Gradient Boosting (XGB), for the analysis of DEGs and identifying novel biomarkers.

Machine learning by stage

Ensemble learning

This method performs better than using a simple algorithm alone because it employs many algorithms to provide poor predictive outcomes in accordance with the features taken from various estimations of data and integrates the results using various voting methods37.

Random Forest (RF): A technique that involves a set of decision trees that naturally incorporate feature selection and interactions into the learning process and report their average as an acceptable label. This algorithm is also the most popular machine learning method.

Max voting

One of the well-known methods in decision-making is max voting. This process is done independently, and as the best class vote is estimated, the outcome with the highest vote is carried out38.

Adaboost

One of the most efficient recognition algorithms in machine learning is Adaptive Boosting, aka. Adaboost. This algorithm makes a pile of weak learners by keeping a set of weights over training data and modifying them after each weak cycle adaptively to make more precise and strong learners out of a collection of weak learners. This recognition algorithm is used for ensemble learning as it has outstanding classification performance that is beneficial in estimating fruit biochemical parameters, image recognition, and complex change prediction modeling39.

Gradient boosting machines(GBM): As decision trees develop, a group forms gradient boosting machines using the information previously generated by growing trees. This way, each decision tree stems from an original training set focused on the parts where earlier model iterations deliver poor prediction40.

Extreme Gradient Boosting (XGB): Extreme Gradient Boosting (XGB) is considered one of the applications of gradient boosted decision trees. To have optimized memory usage and get the most out of hardware computing power, we can use XGBoost. It also reduces the processing time with enhanced performance compared to other machine learning algorithms and deep learning models41.

Performance of machine learning methods

In both true positive and true negative machine learning, accuracy is a measure of an algorithm's effectiveness and performance. F1score is a measure mostly used in unbalanced data to evaluate the algorithm's performance in a false positive and false negative. Auc_curve is a measure to evaluate the correct performance of the algorithm in classifying each class. The confusion matrix is a table that identifies four types of classifications (TN, TP, FN, FP) and shows the algorithm's overall performance. R2 is mainly used in regression algorithms to evaluate the performance of machine learning methods.

Investigation of the correlations of Clinical/Demographic with cancer

To explore the relationships between variables, R 4.1.3 was used to create a cancer correlation matrix to investigate the association between clinical data, including age, tumor size, lymph node involvement, distant metastasis, and stage. A correlation matrix visualizes connections by showing the coefficient of correlation between variables. The correlation coefficient is evaluated on a scale of − 1 to 1. While a negative correlation shows the variables moving in opposite directions, a positive correlation indicates that the variables are moving in the same direction. The cut-off for statistical significance was set ata p < 0.05.

Functional enrichment analysis of the genes and miRNAs

Functional analysis of Gene Ontology (GO) and Reactom, Do, GSEA pathways signaling pathways was performed. In these two analyzes, the categories that include biological process (BP), cellular components (CC), and molecular function (MF) are used. Results with enrichment score > 1, FDR < 0.25, and adjust p < 0.05 were determined as statistically significant results.

PPI network construction

The STRING v11.5 database (http://string-db.org/) was obtained to evaluate the interactions between the target genes of the selected miRNAs. The highest confidence score was set at 70.7 and was considered significant. Proteins were selected based on their interaction with other proteins. Cytoscape software was utilized to view the protein–protein interaction networks (PPIs). Selected miRNAs with several connections to other target genes propose their essential part in PPI.

Identifying prognostic markers

Kaplan‐Meier survival curves and Cox proportional hazard ratio (HR) were plotted for top-selected genes and miRNAs using SPSS version 20 and 95% CI . All the data were analyzed under screening criteria, including the cut-off threshold of HR > 1 and P < 0.05. The candidate genes and miRNAs presented as “prognostic genes”.

Identifying diagnostic markers

Diagnosing PDAC before the tumor spreads provides the best chance for treatment and survival. Here, we assessed the candidate genes, miRNAs, and every combination discovered through survival analysis and machine learning algorithms. In order to evaluate the diagnostic potency and create diagnostic models, a generalized linear model, and combined receiver operating characteristic (ROC) curve analysis were used. Additional diagnostic parameters such as sensitivity, specificity, cut-off value, positive predictive value, negative predictive value, and area under the ROC curve were assessed to evaluate the discrimination of individual or combined biomarkers. The entire procedure was applied using R 4.1.3’s combioROC package.

Quantitative real-time PCR

RNA was isolated from twenty-one Formalin-Fixed Paraffin-Embedded (FFPE) tissue samples using a Parstous kit (Parstous, Tehran, Iran). The extraction quality was evaluated on 1.5% agarose gel, and the quantity was assessed by a Nanodrop 2000 spectrophotometer (BioTek, USA EPOCH). cDNA was synthesized according to the manufacturer's instructions (Parstous, Tehran, Iran). Quantitative real-time PCR was performed using specific primers (Macrogene Co., Seoul, South Korea) and SYBR green master mix (Parstous Co. Tehran, Iran) using an ABI-PRISM StepOneTM instrument (Foster City, CA)18. To identify tissue-specific housekeeping genes for gene expression analysis and to avoid single control normalization error, accurate normalization of qRT-PCR data based on the geometric means of multiple internal control genes was performed. The housekeeping gene which was used as an internal control was GAPDH.

Statistical analysis

The RNA-Seq data analysis, including quality control, preprocessing, and identifying differential expression genes, was performed by R software version 4.3.1. The data were compared by paired t-test and were expressed as mean ± standard deviation (SD). A p-value < 0.05 was considered statistically significant.

Ethics approval and consent to participate

The data was downloaded from TCGA portal (https://tcga-data.nci.nih.gov/). TCGA generates over 2.5 petabytes of genomic, epigenomic, transcriptomic, and proteomic data. The data will remain publicly available for anyone in the research community to all procedures consisting of Ethical issues followed by the TCGA committee. This article does not contain any studies with animals performed by any of the authors. This study was approved by the Ethical Committee of Mashhad University of Medical Sciences (IR.MUMS.MEDICAL.REC.1401.430).

Results

Data description and Identification of differentially expressed genes (DEGs) and differentially expressed miRNAs (DEmiRNA)

The clinical features of the patients are shown in Supplement Table 1. TCGA database containing 193 patients was used as a source to download our required data. Then the data were filtered and finally normalized with the DEseq package. Genes compliant with criteria 1 | LogFC |> and p-value < 0.05 were selected. Using five different machine learning methods, including SVM, DTS, RF, LR, and KNN, some key genes were nominated and analyzed by five various criteria: Accuracy, f1score, r2score, auc_curve, and Confusion matrix. During each step, the best classification algorithm was introduced.

Identifying predictive markers for genes and miRNAs

As shown in Fig. 1B,C, three genes (ABCA12, B3GNT3, and BMF) and eight miRNA (hsa.miR.577, hsa.miR.503, hsa.miR.3613, hsa.miR.19a, hsa.miR.19b.2, hsa.miR.365a, hsa.miR.365b, and hsa.miR.4668) were found to be dysregulated in four different stages of pancreatic cancer.

Investigation of the correlations of Clinical/Demographic with cancer

No association was found between DEG and clinical data for the patients from whom the RNA samples were obtained; only age was significantly associated with prior malignancy. The correlation is considered low when less than 0.3, moderate between 0.3 and 0.6, and strong when more than 0.6 (Fig. 1D,E). The heat map depicted for visualizing DEGs and DEmiRNA across the samples based on the specific criteria (Fig. 2A–F).

Figure 2
figure 2

(A–C) The heatmaps of dysregulated genes, (D–F) The heatmaps of dysregulated miRNAs (created by “cluster, dendextend, circlize, RcolorBrewer, ComplexHeatmap, d3heatmap, gplots, pheatmap, and gplots” packages in R software version 4.3.1.).

Full size image

Functional enrichment analysis of the RNAs and miRNAs

A list of genes was generated, then the gene enrichment to determine the functionally related genes involved in different pathways was calculated, and the expression of other genes was adjusted by R software. Finally, the key genes were enriched to study the Reactom, Do, Go, GSEA pathways. In stage 1, the highest number of genes in the biological process (BP) portion is involved in regulating leukocyte activation and cell activation. As in cellular component (CC), most genes play a role in the receptor complex and side of the membrane pathways. Moreover, during stage 1 in the Molecular Function (MF), the highest number of genes are involved in the pathways of NAD + nucleosidase activity and hydrolyzing N-glycosyl compounds hydrolase activity (Fig. 3). In stage 2, most genes in the BP were involved in the positive regulation of cell death. In CC, the two pathways of hydrolyzing N-glycosyl mitochondrial membrane and mitochondrial envelope are affected by most of the genes, and in MF, the highest number of genes are involved in oxidoreductase activity (Fig. 4). During stage 3, in the BP portion, the highest number of genes are in the immune response-regulating signaling pathway; in the CC stage, the highest number of genes in the side of the membrane pathway, and in the MF section, the genes are equally involved in 3 molecular pathways including, protein and molecular sequestering activity, and NAD(P) H oxidase H2O2-forming activity (Fig. 5). Stage 4 in the BP involved three inflammatory response pathways, cell activation and activation regulation of leukocytes. Also, in the CC, the highest number of genes are involved in the two pathways of hydrolyzing N-glycosyl compounds, hydrolase activity, and NAD + nucleosidase activity (Fig. 6).

Figure 3
figure 3

(A) DO functional annotation in stage 1 of PDAC. (B) GSEA functional annotation in stage 1 of PDAC. (C) GO functional annotation in stage 1 of PDAC. (D) Reactome functional pathways in stage 1 of PDAC. The P-value is less than 0.05 and is shown by the color.

Full size image
Figure 4
figure 4

(A) DO functional annotation in stage 2 of PDAC. (B) GSEA functional annotation in stage 2 of PDAC. (C) GO functional annotation in stage 2 of PDAC. (D) Reactome functional pathways in stage 2 of PDAC. The P-value is less than 0.05 and is shown by the color.

Full size image
Figure 5
figure 5

(A) DO functional annotation in stage 3 of PDAC. (B) GSEA functional annotation in stage 3 of PDAC. (C) GO functional annotation in stage 3 of PDAC. (D) Reactome functional pathways in stage 3 of PDAC. The P-value is less than 0.05 and is shown by the color.

Full size image
Figure 6
figure 6

(A) DO functional annotation in stage 4 of PDAC. (B) GSEA functional annotation in stage 4 of PDAC. (C) GO functional annotation in stage 4 of PDAC. (D) Reactome functional pathways in stage 4 of PDAC. The P-value is less than 0.05 and is shown by the color.

Full size image

PPI network construction

Figure 7 illustrates the interaction of DEGs checked and plotted using the STRING (interaction score: 0.4). In accordance with PPI network, the CD22 gene has the highest binding capacity, followed by CTSW and BTNL9 (Figs. 7C,D, 8B).

Figure 7
figure 7

(A, B, E, F, G, H, I) Kaplan–Meier plots of prognostic genes and miRNAs, (C, D) PPI network of novel genes from STRING.

Full size image
Figure 8
figure 8

(A) Kaplan–Meier plot of CTSW gene, (B) PPI network of CTSW gene from STRING, (C) The level of CTSW in PDAC tumor tissue, as detected by RT-PCR.

Full size image

Identifying prognostic markers for RNAs and miRNAs

Kaplan Meier analysis was applied to identify key prognostic signature genes in pancreatic cancer. The outcome revealed survival is associated with three genes, including BTNL9 (HR = 1.02), CD22 (HR = 1.7), and CTSW (HR = 2.03) and five miRNAs, including hsa.miR.539 (HR = 1.3), hsa.miR.412 (HR = 1.04), hsa.miR.153.1 (HR = 1.5), hsa.miR.642a (HR = 1.00), and hsa.miR.363 (HR = 1.5) in PDAC patients. All analyses were performed by R software (Figs. 7, 8A).

Identifying diagnostic markers for RNAs and miRNAs

For stages 1 and 2, GLM model analysis for HCK and SIGLEC7 combination in diagnostic biomarkers with coefficients of 1.2920 and − 0.5562 (AUC of 0.74, 95% CI with sensitivity of 0.85 and specificity of 0.66). For stage 3, the combination of B3GNT3, ABCA12, and ADAP2 with 0.8409 (AUC of 0.86, 95% CI with sensitivity of 0.8 and specificity of 1. In stage 4, our finding showed that the Coefficients of combination AIF1 and RASGRP3 were 4.233 and − 7.841 (AUC of 0.86, 95%CI with 0.8 sensitivity and one specificity). Furthermore, three miRNAs (Has.mir.194.2, hsa.mir.194.1, and hsa.mir.192) had the highest AUC value, sensitivity, and specificity and coefficients of 4.932, 5.531, and 3.584, respectively (Supplement Table 1).

Validation of CTSW in an additional cohort of PDAC

The clinical data are shown in Supplement Table 2; our population consisted of 52.4% males and 47.6% females. The mean age was 61.66 years and 52.4% underwent advanced stage. We further evaluated the value expression of CTSW in PDAC cases using RT-PCR. This data showed the significant downregulation of this gene in tumor tissue (P < 0.05) (Fig. 8C).

Discussion

To the best of our knowledge, this is the first study showing the potential of downregulation of hsa.miR.642a, hsa.mir.363, CD22, BTNL9, and CTSW and overexpression of hsa.miR.153.1, hsa.miR.539, hsa.miR.412 with shorter survival of patients with PDAC (Supplement Fig. 1) The result indicated the diagnostic value of the combination of AIF1 and RASGRP3 in an advanced stage with the Coefficients of combination AIF1 and RASGRP3 were 4.233 and -7.841 (AUC of 0.86, 95%CI with 0.8 sensitivity and one specificity). The result of the survival analysis showed that the CTSW gene is a novel prognostic marker. CTSW (Cathepsin W), also known as LYPN is a novel human cysteine proteinase member of the C1 peptidase family expressed in CD8 + T and NK cells and regulated by interleukin-2. This gene has a specific function in the cytotoxicity-mediated mechanism by NK cells and CD8 + T cells. Various T cell populations can act differently in regulating a tumor's degree, stage, and ability to invade endometrial cancer. CTSW is an immunomodulatory gene that functions similarly to the CTSF gene42. In research done by Song and colleagues, the expression of CTSF in non-small cell lung cancer was evaluated, and downregulated levels of CTST were observed in NSCLC samples despite normal tissues and good prognosis of NSCLC being correlated with high expression of CTSF. Besides using GeneMANIA, the gene–gene interaction network was established for CTSF and showed that CTSF had a similar function as CTSW genes43. A study on endometrial cancer reported that the CTSW gene had a positive correlation with tumor infiltration levels of B cells, CD8 + T cells, CD4 + T cells, macrophages, and dendritic cells42.

BMF (Bcl-2 modifying factor) is a proapoptotic protein that belongs to the BCL-2 protein family. This gene has been identified in the BH3-only proteins subgroup and initiates the innate apoptotic pathway44. Consequently, BMF is linked with various cellular activities, including chemical sensitivity. For example, the YAP/TEAD/SLUG axis suppressed apoptosis by suppressing BMF transcription45. Badr et al. reported that upregulation of livin and downregulation of BMF and p53 expression are significantly correlated with more tumor aggressiveness (advanced TNM stage), making metastasis progress more rapidly and decreasing overall survival in colon cancer patients. Thus, we can apply these genes as crucial prognostic markers related to poor results46. Another research showed that STARD13 3′UTR could play as a ceRNA for BMF to enhance apoptosis and be used as a potential therapeutic target in breast cancer cells47. FERM is a superfamily of proteins, and one of its members is FERM domain-containing 4A (FRMD4A); these proteins are ubiquitous parts of the cytocortex and are involved in cell transport cell structure and signaling functions. Moreover, tumor progression and metastasis are the cellular events in which the proteins of the FERM family are involved. These proteins function as regulators or scaffolding units and are involved in many membrane-associated factors' functions48. In another study on tongue squamous cell carcinoma and squamous cell carcinoma, the expression of FMRD4A was increased, contrasting our findings.48,49. ArfGAP with dual PH domains 2 (ADAP2) belongs to the ArfGAP family of genes, which is the GTPase activating protein. This gene is expressed for ARF6, which acts as a scaffold in the innate and membrane immunosuppressive phosphate signaling pathways50. It is reported that the ADAP2 gene expression was decreased in primary lower-grade glioma51. Contrary to this, the expression of this gene was increased in radiation-resistant esophageal cancer cells52. Protein phosphatase 1 regulatory subunit 17(PPP1R17), also known as C7orf16, is a negative regulator that inhibits phosphatase activities of protein phosphatase 1 (PP1) and protein phosphatase 2A (PP2A) complexes which their substrates are the S6 ribosomal protein53. Contrary to our results, research in lung cancer adenocarcinoma has demonstrated that PPP1R17 can be used as biomarkers as it was specifically detected in stage III, which can help us detect cancer stage in tumor progression through cleft junction incompatibility, Wnt signaling, and GPCR signaling pathways54. Another study reported that PPP1R17 is a HAR-regulated gene that slows the progression of the neural precursor cell cycle while increasing cell cycle length, which is mainly observed in the neural growth of primates, especially humans.55. The CACNG3 gene encodes a transient AMPA regulatory protein (TARP) known as an auxiliary subunit of the calcium channel γ3. This gene is involved in the neurons formation, and has also been reported as a potential source of epilepsy56. In line with our results, several other studies have shown that the CACNG 3 gene in Gliomas has been predicted as an oncogene and significantly dysregulated in glioblastoma tissue compared to healthy controls57. Other studies have also reported dysregulation of CACNG3 gene in breast cancer58. In our study, hsa.miR.153.1, also known as MIRN153-1, was found to be a new microRNA that had not been used in any other diseases or cancers and had increased expression in pancreatic cancer.

In conclusion, we have identified some specific genes that are differentially expressed at different stages of pancreatic cancer. CTSW gene was reported as a novel prognostic biomarker and validated by Real-time PCR in pancreatic tumor tissue. Eventually, we highly recommend using machine learning to detect biomarkers in other cancers as well.