Prognostic assessment capability of a five-gene signature in pancreatic cancer: a machine learning based-study

The gene expression data were obtained from four public datasets, including TCGA, Pancreatic Cancer (PAAD) (n = 182), GTEx (n = 167), GSE62452 (n = 130) and GSE28735 (n = 90). PAAD was downloaded from UCSC Xena (http://​xena.​ucsc.​edu/​), the expression data was normalized to log2(FPKM + 1). GTEx pancreatic cancer expression data was downloaded from UCSC Xena (http://​xena.​ucsc.​edu/​) and was normalized to log2(FPKM + 0.001). GSE62452 and GSE28735 were downloaded from GEO database (https://​www.​ncbi.​nlm.​nih.​gov/​geo/​), the expression data was normalized by RMA. The expression levels of TCGA-PAAD and GTEx were rescaled using FPKM as the unit of measure for subsequent analyses. Differentially expressed genes (DEGs) was identified using Limma package (version 3.51.8) in R. The cut-off value was set to |logFC|> 1 and p value < 0.05.

The univariate Cox regression, LASSO Cox regression models were performed with survival (version 3.1.1) and glmnet (version 4.1.4) R packages. The data screening process was conducted on TCGA-PAAD dataset.

The TCGA-PAAD gene expression values were then transformed to FPKM. FPKM were then transformed to TPM. After removing the mRNAs with low expression levels, the TCGA expression level were closer to gene chips. ComBat algorithm which is included in the sva R package (version 3.44.0) was used to remove batch effects.

All selected prognosis-related genes were permuted and combined, and used multivariate Cox regression analysis to model, respectively. Calculated AUC for each model separately. AUC value was used as the classification basis and classification was conducted with model-based hierarchical agglomerative clustering which is based on Gaussian finite mixture model. This process used the mclust R package (version 5.4.9) which is a contributed R package for model-based clustering, classification, and density estimation based on finite normal mixture modelling. The associations between prognosis-related gene expression levels and survival information were estimated by the Kaplan–Meier method. The cut-off value was calculated with Kaplan–Meier method.

All the statistical analysis was performed in R software (version 4.2.0). In the survival analysis, Cox proportional hazards regression and Kaplan–Meier analysis were used. We conducted paired t tests on paired samples. All statistical tests with a p-value of less than 0.05 were considered significant.

This study was conducted according to the flowchart in Fig. 1. TCGA-PAAD, GTEx, GSE62452, and GSE28735 were used to identify DEGs. Because of the lack of normal samples in TCGA-PAAD data, we combined TCGA-PAAD and GTEx-PAAD for analysis. The cut-off value is |logFC|> 1 and p-value < 0.05. As shown in Fig. 1A, there are 178 tumor samples and 4 normal samples in TCGA-PAAD dataset. 167 normal samples in GTEx-PAAD dataset, 45 tumor samples and 45 normal samples in GSE28735, 69 tumor samples and 61 normal samples in GSE62452. The TCGA-PAAD and GTEx-PAAD are RNA-seq data, the expression data in them were normalized by FPKM. GSE28735 and GSE62452 share the same platform which ID is GPL6244. The expression data in the array chips was normalized into RAM. Limma R package was used for differential expression analysis. After differential expression analysis, 208 mRNAs were selected by taking the intersection of these datasets (Fig. 1A).

These results suggested that 208 mRNAs have the same expression trends in different datasets (TCGA-PAAD, GTEx-PAAD, GSE62452, and GSE28735). Thus, these initially screened genes might be a promising parameter in patients with pancreatic cancer.

Firstly, univariate Cox analysis was performed by using TCGA-PAAD expression and clinical data. After excluding genes with P values greater than 0.05, 100 prognostic related genes were screened. To further conduct variable selection and regularization, lasso-penalized Cox analysis was used (Supplementary Fig. 2). The complexity adjustment in the lasso regression algorithm refers to controlling the complexity of the model through a series of parameters to avoid overfitting. As shown in Fig. 1A and Table 1, total 8 candidate genes were selected, which were prognosis related genes. Among these genes, ankyrin repeat domain 22 (ANKRD22), aryl hydrocarbon receptor nuclear translocator like 2 (ARNTL2), desmoglein 3 (DSG3), integrin subunit beta 6 (ITGB6), keratin 7 (KRT7), MET proto-oncogene receptor tyrosine kinase (MET) are up-regulated. The remaining two genes, serine protease 3 (PRSS3) and thyrotropin releasing hormone degrading enzyme (TRHDE), are downregulated (Fig. 1A). The expression trends of these 8 genes were consistent in the 3 datasets.

These results confirmed that 8 candidate genes were associated with the prognosis of pancreatic cancer patients. Further analysis of 8 candidate genes might lead to better tools for assessing prognosis.

Due to various factors such as experimental personnel, technology, environment, time point, chip processing, etc., the chip expression matrix contains differences in nonbiological factors. Especially when two or more datasets are integrated, even on the same platform, the integration is even more magnified. After adjusting the magnitude of TCGA-PAAD, we used the combat R function to remove the batch effects of these datasets (TCGA, GTEx, GSE62452, GSE28735). As shown in Fig. 2, after de-batching effects, the expression data of the 3 datasets were merged. The expression heatmaps and boxplots generated from the datasets after the batch effect has been removed are shown in Fig. 3.

These results suggested that 8 candidate genes had the same expression level in the 3 datasets.

To further fit the model and predict patient prognosis more accurately, we conducted a classification based on Gaussian finite mixture model. After permutation and combination of 8 genes, a total of 255 multivariate Cox regression models were generated. All these models were evaluated for prediction accuracy using AUC. As shown in Fig. 4, 255 models were divided into 4 clusters, and cluster 4 had the highest AUCs. Model 168 that contained in cluster 4 had the highest AUC value of 0.755 (Supplementary table 1). The 5 genes contained in model 168 are ANKRD22, ARNTL2, DSG3, KRT7, and PRSS3. The results of multivariate regression analysis are shown in Table 2. Multivariate regression analysis showed that ANKRD22 (HR: 1.22, 95% CI 0.96–1.55; P = 0.098), ARNTL2 (HR: 1.71, 95% CI 1.19–2.46; P = 0.003), DSG3 (HR: 1.15, 95% CI 1.01–1.31; P = 0.029), KRT7 (HR: 1.19, 95% CI 0.97–1.46; P = 0.0093), PRSS3 (HR: 1.23, 95% CI 1.04–1.44; P = 0.013) were prognostic factors in pancreatic cancer patients. Survival analysis was performed on this 5-mRNA signature using TCGA training dataset and GSE validation datasets. Kaplan–Meier curves are shown in Fig. 5. According to the multivariate cox regression analysis results, the calculation formula of risk score is [0.19956 * Exp(ANKRD22)] + [0.538767 * Exp(ARNTL2)] + [0.141913 * Exp(DSG3)] + [0.17315 * Exp(KRT7)] + [0.205059 * Exp(PRSS3)]. The best cut-off value was calculated based on Kaplan–Meier analysis. The calculating result of cut-off value is shown in Supplementary Fig. 1A. By dividing the risk score according to its cut-off value (cut-off = 1.6), 176 patients were stratified into high-risk (n = 61) and low-risk (n = 115) groups (Fig. 6A). Kaplan–Meier curves showed that high-risk patients had worse survival outcome (P < 0.0001) (Fig. 6B). Additional ROC curves revealed that 5-mRNA signature model had the best AUC value (0.755), compared to ANKRD22 (0.655), ARNTL2 (0.688), DSG3 (0.671), KRT7 (0.696), PRSS3 (0.652) (Fig. 6C).

As noted above, these results suggested that 5-mRNA signature model was able to effectively predict the survival of pancreatic patients in the training dataset (TCGA-PAAD). The 5-mRNA signature model was better than the single gene prediction model.

Two GEO datasets, GSE62452 and GSE28735, were used for external validation. The sample size of a single dataset is limited. To improve the prediction accuracy, the two datasets were combined after removing batch effects. Risk scores were calculated with the same formula for each patient. The best cut-off value was calculated based on Kaplan–Meier analysis (supplementary Fig. 1B). Patients were divided into high-risk (n = 78) and low-risk (n = 29) groups according to the optimal cut-off value (Fig. 7A). Kaplan–Meier analysis with the 5-mRNA signature was used to compare the survival outcomes of patients in high-risk and low-risk groups. As shown in Fig. 7B, we confirmed that a higher risk score was associated with shorter survival time. We further analyzed the AUC value in the validation dataset, which was 0.91, significantly higher than AUC in the training dataset (0.755) (Fig. 7C).

In summary, the 5-mRNA signature's accuracy in predicting the prognosis of pancreatic cancer patients was validated, and the model performed better in the validation dataset.