N6-methyladenosine methylation modification patterns reveal immune profiling in pancreatic adenocarcinoma

Gene expression annotation and clinical data were downloaded from the Cancer Genome Atlas database (TCGA, https://​cancergenome.​nih.​gov/​). Totally 176 PAAD samples from TCGA-PAAD were used for further analysis. TCGA RNA sequencing information (FPKM format) of gene expression was obtained from the Genomic Data Commons (GDC, https://​portal.​gdc.​cancer.​gov/​) and transformed into transcripts per kilo base million value. To remove the batch effects from non-biological technical biases, the “sav” R package with the ComBat algorithm was performed. The genomic mutation profiles, including simple nucleotide variation and copy number variation (CNV) of TCGA-PAAD cohort, were curated from the GDC. The CNV landscape of 23 m6A regulators in human chromosomes was plotted using “Rcircos” R package. Non-synonymous mutation (including frameshift, inflame, missense, nonsense, and splice site mutations) counts were recognized as TMB. The expression levels of the m6A regulators from the TCGA-PAAD project are listed in Additional file 7: Table S1.

According to the existing research on m6A modification, 23 known m6A methylation modification regulators have been utilized to uncover m6A methylation modification patterns [23‐26]. These m6A regulators comprise eight “writers” (KIAA1429, CBLL1, METTL3, METTL14, RBM15, WTAP, RBM15B, and ZC3H13), two “erasers” (FTO and ALKBH5), and 13 “readers” (FMR1, ELAVL1, HNRNPC, HNRNPA2B1, IGF2BP1, IGF2BP2, IGF2BP3, LRPPRC, YTHDC1, YTHDC2, YTHDF1, YTHDF2, and YTHDF3). An unsupervised clustering algorithm was used to determine different m6A modification patterns based on 23 m6A regulators to partition the PAAD subjects. R package “ConsensuClusterPlus” was employed to implement these analyses, and repetitions of 1000 times were performed for guaranteeing the stability of clustering.

GSVA [26] with the 'GSVA' R package was used to explore the variation in biological characterizations of distinct m6A modification patterns. The acknowledged biological processes were derived from the gene sets of “h.all.v7.4.symbols.gmt [Curated]’ (downloaded from the Molecular Signatures Database). The Entrez ID for each gene was obtained using the R package “org.Hs.eg.db”. The GO annotation for m6A phenotype-related genes was analyzed and visualized with R packages “clusterProfiler,” “enrichplot” and “ggplot2”.

Single-sample gene set enrichment analysis (ssGSEA) was used to estimate the relative level of immune infiltration. The relative abundance of each leukocyte subset was represented by an enrichment score in ssGSEA and normalized to a unity distribution from 0 to 1. The biosimilarity of infiltrating immune cells was estimated using multidimensional scaling and a Gaussian fitting model. The deconvolution approach, CIBERSORT (http://​cibersort.​stanford.​edu/​), was employed to quantify the components of 22 distinct immune cell types based on gene expression profiling. To uncover the potential role of m6A score in immune infiltration, seven algorithms (XCELL, TIMER, QUANTISEQ, MCPcounter, EPIC, CIBERSORT, and CIBERSORT-ABS) were used. Spearman correlation analysis was performed to investigate the correlation between the risk score and immune infiltration.

TMB was defined as the number of somatic, coding, base replacement, and insert-deletion mutations per mega base of the genome examined, using non-synonymous and code-shifting indels under a 5% detection limit. TMB scores for each sample were calculated by dividing the number of somatic mutations by the total exon length. The “maftools” R package [27] was employed to compute the total number of somatic non-synonymous point mutations within individual subjects.

Based on the results of the consensus clustering algorithm, the samples were partitioned into three distinct m6A modification phenotypes. Next, m6A modification-correlated DEGs were identified among these m6A modification patterns by employing R package 'limma'. DEGs with an adjusted P-value < 0.001 were considered significant and were employed for further analysis.

An m6A scoring system was developed using the principal component analysis (PCA) to evaluate m6A modification patterns in each sample. The patients were clustered into different subtypes using an unsupervised clustering algorithm to analyze overlapping DEGs identified from previous distinct m6A clusters. Prognostic analysis was performed for each gene using univariate Cox regression analysis. Genes with prognostic significance were further analyzed. Then, the expression profiles of the final identified genes were extracted, and the PCA was used to develop m6A relevant gene signatures. Principal components 1 and 2 were curated to serve as the m6A signature (m6sig) score.

The superiority of this algorithm is that it focuses on the score of the set with the largest block of well-correlated (or inverse-correlated) genes in the set, while down-weighting contributions from genes that do not track with other set members. A formula similar to that used in previous studies was adopted to quantify the m6Ascore [28, 29]. The formula is as follows.where i is the expression of m6A phenotype-related genes.

We further evaluated the TNFRSF21 expression in PAAD and normal pancreatic cell lines. HPNE and PANC-1 cells were cultured in DMEM (Biological Industries) containing 10% fetal bovine serum (Biological Industries), whereas BxPC-3 cells were cultured in RPMI-1640 (Biological Industries) supplemented with 10% fetal bovine serum. For the assessment of TNFRSF21 in cells, we extracted the total RNA using an RNA-Quick purification kit according to the manufacturer’s instructions (Vazyme, RN001). The PrimeScriptTM RT Reagent Kit with gDNA Eraser (Takara, RR047A) was used to reverse transcribe the total RNA into cDNA. Quantitative real-time polymerase chain reaction (qPCR) was conducted in an Applied Biosystems 7500 Fast Real-Time PCR System using TB Green Premix EX Taq™ II (Takara, RB820A). The expressions of TNFRSF21 were analyzed using the 2-ΔΔC method, and β-actin was used as an endogenous control. The primer sequences used for PCR are listed in Additional file 7: Table S7.

To estimate the sensitivity of chemotherapy, the R package pRRophetic was used to estimate the half-maximal inhibitory concentration (IC50) of PAAD samples in different groups. By constructing a ridge regression model based on the Genomics of Drug Sensitivity in Cancer (www.​cancerrxgene.​org/​) cell line expression spectrum and TCGA gene expression profiles, the pRRophetic package could estimate the IC50 of chemotherapeutic drugs.

In this study, the roles of 23 m6A modification regulators (“readers”: IGF2BP1, IGF2BP2, IGF2BP3, YTHDC1, YTHDC2, YTHDF1, YTHDF2, YTHDF3 ELAVL1, FMR1, HNRNPA2B1, HNRNPC, and LRPPRC; “erasers”: ALKBH5 and FTO; and “writers”: CBLL1, KIAA1429, METTL3, METTL14, RBM15, RBM15B, WTAP, and ZC3H13) were investigated in pancreatic cancer. GO annotation analyses of 23 m6A regulators was performed, and significant enrichment of the biological pathways was observed (Fig. 1A). We discovered that 8/178 (4.49%) samples possessed somatic mutations of m6A regulators through the delineation of mutation profiles of 23 m6A regulators in TCGA-PAAD samples. These mutations include: missense mutations, multi-hit and deep deletions (Fig. 1B). We identified mutated genes in PAAD with ranked percentages, including METTL3 (1%), WTAP (1%), RBM15 (1%), METTL14 (1%), ZC3H13 (1%), YTHDC1(1%), YTHDC2 (1%), YTHDF1 (1%), YTHDF3(1%), FMR1 (1%), and ALKBH5(1%). The prevalence of CNV mutations in the 23 m6A regulators was further analyzed, revealing that VIRMA, IGFBP2, ALKBH5, FMR1, and HNRNPA2B1 experienced widespread CNV amplification, whereas HNRNPC, YTHDF2, METTL3, RBM15B, and YTHDC1/2/3 possessed general CNV deletions (Fig. 1C). The CNV mutation locations of the 23 m6A regulators in the chromosomes are shown in Fig. 1D. Spearman’s correlation was analyzed to explain the mutual connection among the 23 m6A regulators (Fig. 1E). The results revealed that the “readers” IGF2BP1 and IGF2BP2, experienced a prominent negative relationship with other m6A regulators, but the other 21 m6A regulators had positive correlations with each other. Subsequently, we explored the difference in expression between tumor tissues and adjacent tissues of PAAD and observed that WTAP was notably downregulated in tumor tissues, whereas RBM15, IGFBP1/3, and YTHDF1 seemed upregulated in tumor samples (Fig. 1F). Kaplan–Meier survival analysis of the low and high expression of these m6A regulators indicated that 13/23 m6A regulators were significantly associated with survival. The overexpression of ALKBH5, FTO, IGFBP2, METTL3, and METTL16 exhibited a prominent advantage in survival, whereas the low expression of FMR1, HNRNPC, IGFBP3, LRPPRC, RMB15, VIRMA, YTHDF3, and ZC3H13 was associated with a better prognosis (Additional file 1: Fig. S1). The above research supports the remarkable distinctions and intrinsic relations in transcriptomic and genomic maps of m6A modification regulators between tumor and normal tissues of PAAD. Thus, the variation in the expression and genetic alterations of m6A regulators might provide novel insights into the occurrence and development of PAAD.

The PAAD dataset with the corresponding clinical data was included in the combined cohort. The m6A regulator network showed a comprehensive view of 23 m6A regulator connections and their corresponding prognostic value (Fig. 2A), from which we found that m6A regulators with the same biological function experienced significant correlation, which was similar to the cross-talk among “readers”, “writers”, and “erasers”. Therefore, the correlation among m6A regulators plays an important role in the development and prognosis of PAAD. The “ConsensusClusterPlus” R package was employed to stratify samples into two m6A modification patterns based on the expression of 23 m6A regulators (Additional file 2: Fig. S2A–D). Then, two different modification patterns, pattern A (127 samples) and pattern B (49 samples), were finally determined using unsupervised clustering. The patterns were designated as m6Acluster A and m6Acluster B. Kaplan–Meier survival analysis of the two m6A modification clusters indicated that m6Acluster A experienced a notable advantage in survival, while m6Acluster B had a poor clinical outcome (Fig. 2B). The heatmap shows the differential expression of 23 m6A regulators based on the classification of the m6A cluster, survival status, gender, tumor stage, and age (Fig. 2C). The two m6A modification patterns exhibited different expression levels of m6A regulators, and it seemed like almost all m6A regulators were upregulated in the m6Acluster B subtype (Fig. 2C).

GSVA against the hallmark gene set was applied to further investigate the biological behaviors of the two m6A modification patterns (Fig. 2D). GSVA presented that m6Acluster B possessed enrichment pathways related to stromal pathways and carcinogenic activation, such as TGF-β, Wnt-β-catenin, Notch, PI3K-AKT-mTOR signaling, and hedgehog signaling pathways.

Furthermore, m6Acluster B was also significantly enriched in both protein secretion and glycolysis. Subsequently, the relative subpopulations of 23 infiltrating immune cell abundances were confirmed by comparing the two m6A modification patterns using ssGSEA and CIBERSORT (Fig. 2E, F). Inflammatory-response cell subpopulations that participate in anti-tumor processes, such as naive B lymphocytes, CD8 + T lymphocytes, and plasma cells, were markedly up-regulated in the m6Acluster A. We discovered that M0 macrophages were mainly rich in m6Acluster B. Although there were no significant differences, M1 macrophages experienced a higher trend in the m6Acluster A, whereas M2 macrophages presented more fraction in the m6Acluster B. Taken together, these two m6A modification patterns were characterized by distinct profiles of infiltrating immune cells. A large quantity of tumor-infiltrating lymphocytes, such as CD8 + T lymphocytes and plasma cells, are markedly correlated with anti-tumor immune status [30, 31]. Thus, m6Acluster A could be considered an immune-inflamed phenotype with abundant activated lymphocytes. Due to the retention of stroma peripheral cancer nests, numerous infiltrating immune cells could not penetrate the tumor parenchyma in tumors of the immune-excluded phenotype [32]. We inferred that overexpression of stromal elements would weaken the efficacy of anti-tumor immunotherapy in the m6Acluster B. In general, m6Acluster A was characterized by abundant infiltration of immune cells and immune activation, identified as the immune-inflamed phenotype. However, m6Acluster B was characterized by infiltrating innate immune cells, and it exhibited abundant pathways markedly related to stromal pathways and carcinogenic activation. Therefore, m6Acluster B is considered an immune-excluded phenotype.

The Limma package was used to determine 169 DEGs, which were regarded as the pivotal distinct index of two m6A modification patterns, to further explore the latent alterations of m6A-associated transcriptional expression between two m6A modification phenotypes (Additional file 3: Fig. S3A). Figure 3A, B shows the biological processes in the GO enrichment analysis of DEGs. GO enrichment pathway analysis revealed that DEGs were mainly enriched in endoplasmic reticulum to golgi vesicle-mediated transport, vesicle budding from membrane and establishment of vesicle budding in biological processes; golgi-associated vesicle, coated vesicle and collagen-containing extracellular matrix in cellular components; cell adhesion molecule binding, cadherin binding and extracellular matrix structural constituent in molecular function. Additionally, GO enrichment pathway analysis indicated that the DEGs were closely related to the PI3K-Akt signaling pathway, focal adhesion, and human papillomavirus infection.

The 169 most representative m6A phenotype-related genes were analyzed by unsupervised clustering analyses, and the samples were divided into different transcriptomic phenotypes, from which we could better investigate the underlying molecular mechanisms (Additional file 4: Fig. S4A–D). These stratifications classified the samples into two m6A gene signature subgroups with different clinical variables, which were named m6A gene clusters A and B.

Kaplan–Meier survival analysis of two m6A gene signature subgroups indicated that gene cluster A, which contained 117 patients, exhibited an absolute advantage in overall survival. In contrast, 59 patients in gene cluster B had poorer prognosis (Fig. 3C). A heatmap was used to visualize the distinct clinicopathologic features and transcriptomic variation between the two m6A gene signature subgroups (Fig. 3D).

GSVA against the hallmark gene set was performed to further explore the biological behaviors between the two m6A gene subgroups (Fig. 3E). The results of GSVA revealed that m6A gene cluster B presented enrichment pathways prominently associated with almost all biological processes, including mitotic spindle, protein secretion, glycolysis, and apoptosis, as well as stromal pathways and carcinogenic activation, such as the TGF-β, Wnt-β-catenin, Notch, p53 pathway, and PI3K-AKT-mTOR signaling pathways. To explore the latent impact of m6A modification patterns on TIME, CIBERSORT and ssGSEA algorithms were further performed to assess the relative subpopulations of immune infiltration. The results showed that m6A gene cluster A was significantly enriched with naive B cells, plasma cells, and activated CD8 + T cells, which were abundant immune cells (including humoral immunity and cellular immunity), while m6A gene cluster B was remarkably rich in infiltration of M0 and M2 macrophages (Fig. 3F, G).

All of these analyses implied that m6A methylation modification had an indispensable impact on the regulation of shaping TIME landscapes.

We revealed the potential roles of m6A modification or gene patterns in prognostic prediction and information on TIME. Nevertheless, the above studies only focused on the sample population of the database and might not be precisely applied to individuals. Therefore, on account of the complexity and heterogeneity of m6A modification in tumors, we constructed a scoring system, named as the m6Ascore, which is based on two m6A phenotype-associated signature genes, to explore the m6A modification pattern of individual PAAD patients.

Individual changes in patient attributes are summarized in an alluvial diagram (Fig. 4A). We analyzed the correlation between the m6Ascore with the well-established signatures. A heatmap of the correlation matrix shows the biological function of the m6Ascore. The m6Ascore was significantly positively correlated with most biological processes, such as the TGF-β, Wnt-β-catenin, PI3K-AKT-mTOR, and Notch signaling pathways (Fig. 4B). Moreover, the patients were classified into subgroups of high- or low-m6Ascore to reveal the prognostic value of the m6Ascore system in predicting overall survival time. The low-m6Ascore was associated with a better prognosis in the combined cohort (Fig. 4C).

Univariate (hazard ratio (HR) 1.057(1.026–1.090), P < 0.001, Fig. 4D) and multivariate (HR 1.061(1.027–1.097), P < 0.001, Fig. 4E) Cox regression analyses were performed to explore the role of the m6Ascore in forecasting the prognosis of PAAD patients, which demonstrated that the m6Ascore was an independent and powerful prognostic biomarker for predicting prognosis. These analyses strongly implied that the m6Ascore system could reflect the features of m6A modification patterns and predict prognosis in PAAD. Moreover, we explored the difference of m6Ascore in m6A methylation modification patterns and m6A gene signature subgroups. Notably, patients in m6Acluster B and gene cluster B had higher m6Ascore (Fig. 4F, G).

The potential role of m6Ascore in clinicopathological variables was further investigated. The subgroup with ≤ 65 had obviously higher m6Ascore (P < 0.05, Fig. 4H). The higher m6Ascores were concentrated in grade 3–4 (P < 0.05, Fig. 4I) and T3-4 (P < 0.05, Fig. 4J), while lower m6Ascores were accumulated in the subgroup of > 65, grade 1–2 and T1-2. However, the m6Ascore was not significant among the subgroups of the clinical stage (P > 0.05, Additional file 3: Fig. S3B).

Based on different clinical features, we divided PAAD patients into distinct subgroups and further explored whether the risk score could identify differences in the prognosis of PAAD patients using stratification analysis. When samples were divided according to age, the m6Ascore was still predictive of prognosis, and a lower m6Ascore was significantly correlated with a better outcome (Additional file 5: Fig. S5A, B). Similarly, m6Ascore acted as an excellent prognostic marker for patients based on sex (Additional file 5: Fig. S5C, D), early or late grade cancer (Additional file 5: Fig. S5E, F), and stage I-II cancer (Additional file 5: Fig. S5G, H). These results indicate that the m6Ascore system is closely related to clinical features. Therefore, the m6Ascore system could serve as a useful prognostic predictor independent of other clinical traits.

Based on previous research, we revealed that m6A modification patterns and immune cell infiltration are closely related. Next, we explored the potential role of the m6Ascore in terms of the complexity and diversity of the TIME context. The results illustrated that the m6Ascore was negatively correlated with the subpopulations of B lymphocytes and CD8 + T lymphocytes, whereas it was positively correlated with myeloid dendritic cells, M0 macrophages, M2 macrophages, and cancer-associated fibroblasts (Fig. 5A). The results of Spearman’s correlation analysis are shown in detail in Additional file 7: Table S6. The results indicated that the m6Ascore had an indispensable impact on the immunological cross-talk of the TIME.

Immunotherapy using anti-PD-1/CTLA-4 drugs has achieved encouraging success in clinical anti-cancer treatment regimens.. The immunotherapeutic hub targets, such as CD274, PDCD1LG2, CTLA‐4, etc. [33‐35], are strongly associated with the clinical outcome of immunotherapy. Correlation of the m6Ascore with immunotherapy targets was further analyzed. It was observed that the m6Ascore was positively and significantly correlated with CD274 (r = 0.42; P = 4.4e − 09) and PDCD1LG2 (r = 0.2; P = 0.0091; Fig. 5B, C). The difference in immunotherapy response was further characterized by more immunotherapeutic genes, from which most targets were significantly upregulated in subjects with higher m6Ascores (Fig. 5D). The results strongly supported that m6Ascore was latently related to immunotherapies, and it may further predict prognosis accordingly.

Somatic mutations are important factors in tumor development. We investigated the correlation between the m6Ascore and TMB, and we observed that the m6Ascore was positively and significantly correlated with TMB in PAAD (Fig. 6A). Furthermore, according to the distribution patterns of TMB (Additional file 3: Fig. S3C), there was no significant difference in the mutation frequency between the low- and high-m6Ascore subgroup. Moreover, the survival curve indicated that high TMB levels were associated with a poor overall survival rate (P < 0.05, Fig. 6B). We found no interaction between m6Ascore and TMB status in terms of prognostic predictive value according to the stratified survival curve. The m6Ascore subgroups presented prominent differences in prognosis in the high and low TMB value subgroups (P < 0.001; Fig. 6C). Subsequently, we explored the significantly mutated gene (SMG) in the two m6Ascore subgroups.

According to the SMG mutational landscapes, KRAS (100% vs. 72%) and TP53 (90% vs. 56%) presented higher somatic mutation rates in high m6Acore samples, and SMAD4 (24% vs. 20%) and TTN (15% vs. 10%) exhibited higher somatic mutation rates in low m6Ascore samples (Fig. 6D, E). These results confirmed the impact of m6Ascore classification on genomic variation, and the potentially complex interaction between individual somatic mutations and m6A modifications.

To better reveal the potential role of targets from differentially expressed genes between m6A gene clusters A and B (Fig. 3D), we analyzed TNFRSF21 in PAAD, which has not been previously reported. Using TCGA database, the expression level of TNFRSF21 in PAAD tumor tissues and normal pancreatic samples was analyzed. Furthermore, expression condition of TNFRSF21 was validated using the GEPIA website [36]. The results showed that TNFRSF21 was significantly upregulated in tumors compared with its normal counterpart (Fig. 7A). We further found that TNFRSF21 was prominently upregulated in PAAD cell lines, such as PANC-1 and BxPC-3, compared to the normal pancreatic cell line HPNE (Fig. 7B). To explore the effect of TNFRSF21 on survival and prognosis, survival curve analysis was performed for high- and low-expression of TNFRSF21 in PAAD patients. It was revealed that the lower TNFRSF21 expression presented a higher OS probability (Fig. 7C), longer disease-free survival time (Fig. 7D), and had a prominent advantage of survival. Regarding the relationship between TNFRSF21 and immune infiltration, the results indicated that TNFRSF21 was consistently correlated with B lymphocytes (r = 0.156; P = 4.15e − 02), CD8 + T lymphocytes (r = 0.224; P = 3.19e − 03), CD4 + T lymphocytes (r = − 0.106; P = 1.96e − 01), macrophages (r = 0.03; P = 6.96e − 01), neutrophils (r = 0.149; P = 5.13e − 02), and dendritic cells (r = 0.22; P = 3.80e − 03; Fig. 7E). Moreover, the TNFRSF21 mutation was correlated with the infiltration of B lymphocytes, CD4 + T lymphocytes, and neutrophils (Fig. 7F).

We performed GSEA for PAAD to further explore the molecular mechanisms of TNFRSF21 regulation in tumors. KEGG results indicated that spliceosome and steroid hormone biosynthesis were positively correlated with TNFRSF21, whereas neuroactive ligand receptor interaction and primary immunodeficiency were negatively associated with TNFRSF21 (Fig. 7G). The GO analysis showed the most significant terms that included detection of chemical stimulation, humoral immune response, regulation of trans-synaptic signaling, sensory perception of smell, and external side of the plasma membrane (Fig. 7H).

Considering the positive relationship between TNFRSF21 and the detection of chemical stimulation and humoral immune response, we evaluated its clinical therapeutic value with chemotherapy and immune checkpoint therapy in PAAD. We found that patients with low TNFRSF21 expression were more sensitive to bleomycin, whereas rapamycin, sunitinib, and vinblastine were more effective in patients with high TNFRSF21 expression (Fig. 7I–L). However, as shown in Additional file 6: Fig. S6A–D, there was no significant difference in the IPS–CTLA4 and PD1/PDL1/PDL2 blocker scores between the low- and high-expression levels of TNFRSF21 in samples.