The role of m6A-related genes in the prognosis and immune microenvironment of pancreatic adenocarcinoma

Comparison of the expression levels of m6A-related genes between PAAD and normal pancreatic tissues

Genes closely associated with the installation, recognition and removal of m6A marks were identified in the literature and analyzed in the present study. As the main cohort, this study incorporated transcriptome data of PAAD and normal pancreatic tissue samples originally available in the TCGA-PAAD and GTEx projects and further analyzed the data via Gene Expression Profiling Interactive Analysis (GEPIA) (http://gepia.cancer-pku.cn) (Li et al., 2017; Tang et al., 2017), which is an online platform, to visualize the TCGA-PAAD and GTEx data. The validation cohort consisted of the following two parts: 1. transcriptome data of PAAD and matched normal tissues collected from the GEO database for validation at the transcriptome level and 2. immunohistochemical data collected from the Human Protein Atlas (https://www.proteinatlas.org/) for validation at the protein level.

An ANOVA was used to compare the expression levels of m6A-related genes in GEPIA with those in the TCGA and GTEx data. GEO2R, which is a built-in tool in the GEO database, was used to validate the differential expression of the m6A-related genes based on the GEO datasets.

Construction of a prognostic gene model

The clinical data of the main cohort included the survival time, survival status, TNM classification, stage, age and sex and were downloaded from TCGA-PAAD. First, we applied a univariate Cox regression analysis to identify the m6A-related genes significantly associated with the prognosis of PAAD (P < 0.05). Then, a lasso regression was conducted to calculate the risk coefficient of each gene after the removal of some genes with a risk of overfitting according to the partial likelihood deviance and lambda value (the lambda value is determined by the smallest likelihood deviance; the coefficient-lambda curve shows the genes that are eligible when the lambda value is determined). The lasso risk was calculated using the following formula: lasso risk = ${\sum }_{i=1}^{n}Coef\times xi$. Finally, the remaining genes were utilized to construct a predictive model of the prognosis of PAAD. The samples with a top 50% risk value were regarded as “high risk”, while the samples with a bottom “50%” risk value were considered “low risk”. A ROC curve was generated to assess the predictive value of the constructed model. Univariate and multivariate Cox analyses were performed to identify the independent prognostic factors of PAAD (P < 0.05). A validation cohort was established from the GEO dataset and used to confirm the accuracy of the prognostic model based on the TCGA cohort. We adjusted the expression level of each m6A-related gene in the GEO dataset due to the use of different sequencing platforms to evaluate the accuracy of our prognosis model, which ensured optimized comparability between the validation cohort and TCGA cohort. First, we standardized each gene’s expression level using the following formula: $x_{std}=\frac{x_i-\overline{x}}{s}$, $\overline{x}$ = $\frac{1}{n}{\sum }_{i=1}^n{x}_i$, s = $\sqrt{\frac{1}{n-1}{\sum }_{i=1}^n{\left(x_i-\overline{x}\right)}^2}$; Then, we adjusted each X_std to match the training data of TCGA using the following formula: $x_{adj}=x_{std}\times s_{train}+{\overline{x}}_{train}$. A survival analysis was performed using the R package “survival”. The “glmnet” package was used to perform a Cox proportional hazards regression analysis with least absolute shrinkage (glmnet, version 2.0-18). Univariate and multivariate Cox regression analyses were performed to provide evidence that the prognostic gene model was independent of other clinicopathological factors. A lasso regression was conducted using the packages “glmnet” and “survival”. The package “survivalROC” was used to depict the ROC curve. The risk score plot was generated by the package “pheatmap”. P < 0.05 was considered statistically significant.

Exploration of potential m6A modification targets

Initially, we acquired the top 100 genes that harbored the most m6A-modified sites from among 23,391 genes in RMBase (Xuan et al., 2018) and then plotted the survival curve to evaluate the role of these genes in PAAD prognosis (Zhu et al., 2019). Additionally, we screened the top 20 genes that were closely associated with PAAD survival and then determined the number of m6A-modified sites in their transcriptional products. The mutation information of the m6A-related genes was obtained and visualized in cBioPortal (http://www.cbioportal.org/) (Cerami et al., 2012; Gao et al., 2013). The survival analysis compared the overall survival time and progression-free time between patients with mutated m6A-related genes and those without.

Investigation of the role of m6A-related genes in the immune microenvironment of pancreatic cancer

According to the similarities in the gene expression levels, the PAAD samples in the TCGA dataset were clustered into different groups using “ConsensusClusterPlus” (50 iterations, resample rate of 80%, and Pearson correlation; http://www.bioconductor.org/). We performed a PCA via an R package in R v3.6.1 to study the gene expression patterns in different PAAD groups. A single sample gene-set enrichment analysis (ssGSEA) was conducted based on the expression level of 29 immunity-associated signatures using the R package “GSEAbase”. The activity of immune signatures and the immune score were compared between two subgroups using a t-test. The tumor purity of the two subgroups was compared using a Wilcox test. The package “Estimate” was used to calculate the immune score, stroma score and tumor purity of each tumor sample. The relationship between the mutation state or expression level of m6A-related genes and infiltration of immune cells was established and visualized using Tumor Immune Estimation Resource (https://cistrome.shinyapps.io/timer/).

We selected the gene that showed the highest hazard ratio with PAAD prognosis and analyzed its role in the alteration of immune signatures. Quartiles were used to assess the expression level of this gene. The samples with the top 25% expression level of this gene are regarded as “high expression”, while the samples with the bottom 25% expression level of this gene are classified as “low expression”.

Differential expression of m6A-related genes between tumor and normal samples

The results of this study are summarized in a flow chart (Fig. S1). We selected 21 genes (METTL3, METTL14, METTL16, WTAP, KIAA1429, EIF3, IGF2BP1-3, RBM15, RBM15B, ZC3H13, YTHDC1, YTHDC2, YTHDF1, YTHDF2, YTHDF3, HNRNPC, HNRNPA2B1, FTO and ALKBH) based on the existing literature (Batista, 2017; Chai et al., 2019; Dai et al., 2018). The translated products of these genes are essential for the installation, removal and recognition of m6A marks (Table 1). We identified 14 differential m6A-related genes in the TCGA-GTEx cohort. Specifically, five m6A writers (METTL14, METTL16, KIAA1429, RBM15 and ZC3H13) were upregulated in PAAD, and seven m6A readers (YTHDF1, YTHDF2, YTHDF3, IGF2BP2, IGF2BP3, HNRNPA2B1 and HNRNPC) were highly expressed in PAAD. Similarly, two major RNA demethylases, i.e., FTO and ALKBH5, were overexpressed in the tumor samples. In addition, four Gene Expression Omnibus (GEO) datasets (GSE15471, GSE28735, GSE62452 and GSE11838) were used as validation cohorts (Table S1). Genes with the same expression pattern in the TCGA-GTEx cohort and more than three of the four GEO datasets are defined as “validated genes”. The results revealed that HNRNPC, IGF2BP2 and YTHDF1 are highly expressed in PAAD with powerful evidence, while METTL3 and YTHDC2 are less likely to have differential expression (Fig. 1A).

In addition, we used immunohistochemical data from the Human Protein Atlas to evaluate the expression of the m6A-related genes at the protein level (Table S1), although data for some m6A-related genes were unavailable on the website. The results showed that the protein expression of RBM15 is upregulated in pancreatic cancer, which is similar to the results of the TCGA-GTEx cohort and two GEO datasets.

Pearson coefficients were used to quantify the magnitude of the co-expression of the m6A-related genes. Ten gene pairs were found to be highly co-expressed with a Pearson coefficient greater than 0.6 (Fig. 1B). Among these pairs, the YTHDC1-YTHDC2 pair had the strongest correlation magnitude (0.74); however, YTHDF1 and IGF2BP1-3 exhibit a lower correlation with the other m6A-related genes.

Prognostic model based on three m6A-related genes can predict the survival of patients with PAAD with intermediate accuracy

Three m6A-related genes significantly associated with the prognosis of PAAD were screened through a univariate Cox analysis (Fig. 2A). To avoid the overfitting phenomenon in the subsequent model construction, a least absolute shrinkage and selection operator (LASSO) regression was used to detect whether dimensionality reduction is possible by eliminating redundant genes. According to the partial likelihood deviance and lambda values, all three genes are suggested to be included in the model construction (Figs. S2A to S2C). The lasso regression provided a risk score to each sample based on the expression level of the three included genes. The samples were divided into two groups in a dichotomous fashion according to the ranking of the risk score (Fig. 2B). Overall, the group of patients with the low lasso risk score showed a longer survival time than the group of patients with a high lasso risk score (P < 0.005) (Fig. 2C). Then, a receiver operating characteristic (ROC) curve was generated to assess the predictive accuracy of this model (Fig. 2D). The lasso risk score was obtained as follows: lasso risk score = (0.2623*expression level of RBM15) + (0.0168*expression level of HNRNPC) + (0.0367*expression level of IGF2BP2). The model exhibited intermediate accuracy with an area under the ROC curve (AUC) of 0.700. One GEO dataset (GSE71729) comprising 125 pancreatic ductal adenocarcinoma (PDAC) samples was used as a validation cohort to verify the accuracy of this prognostic model. Fifty-five samples were considered high-risk, and the other 70 samples were considered low-risk based on the coefficient calculated by the TCGA-based model (Fig. S3A). Similarly, the samples with low risk scores showed longer overall survival (OS) times than those with high risk scores (P < 0.05), supporting the accuracy and practicability of this model (Fig. S3B).

Risk score based on three m6A-related genes is a reliable indicator of a shorter OS

The age and lasso risk score were significantly associated with a worse prognosis of PAAD in both the univariate and multivariate Cox regression analyses, suggesting that these factors are independent prognostic factors. Interestingly, the grade, stage, T stage, N stage and sex were unrelated to the survival of patients with PAAD in this study (Figs. 2E to 2F).

Potential m6A modification targets and summary of the mutation information of the m6A-related genes

A survival analysis was performed with the top 100 genes that harbored the most m6A-modified loci, which ranged from 116 to 518 in number (Table S3). Sixteen of these genes were found to be strongly associated with the prognosis of PAAD (P < 0.05) (Fig. 4). Additionally, we determined the number of m6A-modified sites in the top 20 genes that were closely associated with PAAD survival (Table S4) and found that 17 of the 20 genes harbored between five and 38 m6A-modified loci.

Four cohorts (ICGC, QCMG, TCGA and UTSW) were included in this study to analyze the mutation information of the m6A-related genes. Overall, these genes were mutated in 118 samples (118/848) in which deletion mutations occurred mainly in METTL14, WTAP, YTHDC1 and YTHDF2, whereas amplification mutations occurred mostly in KIAA1429 and YTHDF1 (Fig. S5). The patients with samples that did not exhibit mutations in m6A-related genes showed longer OS and disease/progression-free survival times than those with genetic alterations by a Kaplan–Meier analysis (Figs. S6A to S6B). Among the genes with mutations crucial for PAAD tumorigenesis, ZNF814 was frequently concurrently mutated with the m6A-related genes (Fig. S6C). We further compared the copy number alteration frequency of four driver genes between the samples with and those without mutations of the m6A-related genes. The results showed that the copy number alteration frequency of TP53 was significantly increased in the samples with mutations in the m6A-related genes, suggesting that TP53 may be involved in m6A regulation in PAAD (Figs. 6C to 6D).

Subgroup of PAAD with a specific expression pattern of m6A-related genes exhibits different immune signatures

A clustering analysis based on the expression levels of the m6A-related genes in the PAAD samples was performed, and an increasing trend of the cumulative distribution function (CDF) value with respect to the consensus index was considered indicative of an appropriate classification (Figs. S7A to S7B). Finally, two subgroups with similar sample sizes were clearly classified (Fig. 3A). We further confirmed the independence of the two subgroups by a principal component analysis (PCA), which further supported the clustering results (Fig. 3B). We also evaluated the difference between the two subgroups in terms of the stromal score, immune score, immune signatures and tumor purity (Fig. 3C). The results demonstrated that group 1 had higher stromal and immune scores but a lower tumor purity (Fig. 3D). Moreover, group 1 exhibited a better cytotoxic response and more check point signatures, HLA, macrophages, mast cells, neutrophils, etc., suggesting that the difference in the expression pattern of m6A-related genes leads to an altered immune microenvironment (Fig. 3E).

Expression level and mutation state of m6A-related genes in the immune microenvironment of pancreatic cancer

First, we correlated the expression of the eight most studied m6A-related genes with immune cells in pancreatic cancer (Fig. 4). All eight genes were positively associated with the number of CD8+ T cells after adjusting for the tumor purity, suggesting that m6A modification potentially regulates CD8+T cell aggregation. However, the presence of arm-level deletion or gain mutations was negatively correlated with the CD8+T cell number (Fig. 5). Specifically, both the arm-level gain and deletion of ALKBH5 decreased the infiltration of CD8+T cells (P < 0.05 and P < 0.01, respectively). RBM15 was the gene with the highest hazard ratio to PDAC survival; hence, we further analyzed the correlation between RBM15 and 29 immune signatures. The results showed that a higher expression of RBM15 is positively associated with APC co-stimulation, inflammation promotion, MHC class I, Th2 cells and Treg cells (Fig. 6).