The workflow of our study is shown in online supplemental figure S1. RNA sequencing data (count values) for gene expression analysis, genetic mutations (VarScan), and clinical data were downloaded from the Genomic Data Commons (https://portal.gdc.cancer.gov/).27 The count values were transformed into transcripts per kilobase million (TPM) values (the gene lengths used for the above transformation were measured as total non-overlapping exon length) and the Ensembl gene IDs of the RNA-seq data were converted to gene symbols by referring to the annotation file (https://www.gencodegenes.org/human/release_22.html). The copy number variation (CNV) data were downloaded from the Broad GDAC Firehose (https://gdac.broadinstitute.org/). The normalized data from another ccRCC cohort (91 cases) were downloaded from the International Cancer Genome Consortium (ICGC, https://dcc.icgc.org/).

Gene expression levels were quantified using the metric log2 (TPM +1), then used to identify m⁶A modification patterns based on the expression of 24 m⁶A regulators by model-based clustering analysis implemented in the R package ‘mclust’.28 The optimal number of clusters was determined based on the Bayesian information criterion.

CIBERSORT—a deconvolution algorithm reported by Newman et al29 and verified by fluorescence-activated cell sorting—was used to quantify the 22 infiltrated immune cells according to normalized gene expression profiles. The 22 immune cells included memory B cells, naïve B cells, plasma cells, resting/activated DCs, resting/activated natural killer (NK) cells, resting/activated mast cells, eosinophils, neutrophils, monocytes, M0–M2 macrophages, and seven T-cell types (CD8+ T cells, regulatory T cells (Tregs), resting/activated memory CD4+ T cells, follicular helper T cells, naïve CD4+ T cells and gammadelta T cells (γδ T cells)) For each sample, the sum of all estimated values for the proportion of immune cells was equal to 1. CIBERSORT results were obtained from the following website: https://gdc.cancer.gov/about-data/publications/panimmune (online supplemental table S1).30 The relative abundance of Th1/Th2/Th17 cell infiltration in the ccRCC TIME was quantified by single-sample gene-set enrichment analysis. The gene sets for marking the Th1/Th2/Th17 cell types were obtained from a study published by Thorsson et al.30 The prognostic value of infiltrated immune cells was assessed by univariate Cox regression analysis.

Values of key immune characteristics (including leukocyte fraction, Th1/Th2/Th17 cells, single nucleotide variant neoantigens, indel neoantigens, proliferation, aneuploidy score, intratumor heterogeneity (ITH), B-cell receptor (BCR) evenness, T-cell receptor (TCR) evenness and cancer testis antigens (CTA) score) and measures of DNA damage (including CNV burden (number of segments and fraction of genome alterations, respectively), loss of heterozygosity (LOH; number of segments with LOH events, and fraction of bases with LOH events, respectively), homologous recombination deficiency, and mutation load (non-silent mutation)) were obtained from the following website: https://gdc.cancer.gov/about-data/publications/panimmune (online supplemental table S1).

A list of 78 immunomodulators (IMs) was obtained (online supplemental table S2),30 three of which (HLA-DRB3, HLA-DRB4 and KIR2DL2) had no corresponding mRNA expression and were excluded from subsequent analysis. The median expression levels of the samples were used to represent the expression of each ccRCC subtype. In order to examine the differences in IM expression among different subtypes, we carried out the Kruskal-Wallis test on the gene expression levels for each ccRCC subtype. CNVs for each IM gene were obtained from the following website: https://gdc.cancer.gov/about-data/publications/panimmune.30 We calculated the difference between the observed and expected amplification frequencies (deletions) for each IM gene in each ccRCC subtype, where the expected frequency is the overall amplification frequency (deletions) of all ccRCC cases.

Gene set variation analysis (GSVA)—a non-parametric and unsupervised method commonly used for estimating pathway variations in the samples of expression datasets—was performed to explore the differences in biological processes among m⁶A modification patterns.31 The ‘c2.cp.kegg.v6.2.symbols’ gene sets for GSVA were downloaded from the Molecular Signatures Database (MSigDB). A p<0.05 was considered statistically significant.

To identify genes related to m⁶A modification patterns, we classified patients into m6Aclusters based on the expression of 24 m⁶A regulators. Differentially expressed genes (DEGs) among these clusters were determined using the R package ‘DESeq2’, which was applied using the raw count values of RNA sequencing data. Genes with adjusted p<0.01 and at least two-fold changes in expression were identified as DEGs.

We applied a methodology to quantify the m⁶A modification pattern (m6Ascore) of individual ccRCC patients. The m6Ascore was established as follows. First, we extracted the overlapping DEGs among m6Aclusters and classified the ccRCC patients into several groups using model-based clustering to analyze overlapping DEGs. Univariate Cox regression analysis was performed to evaluate the prognosis of each overlapping DEG. Genes with a significant prognosis (p<0.05) were extracted for further analysis. Next, principal component analysis (PCA) was performed to establish the m⁶A gene signature. We selected both principal components 1 and 2 as signature scores. Finally, the m6Ascore was defined using a method similar to Genomic Grade Index26 32 33:

m6Ascore=∑(PC1i+PC2i)

where i is the expression of overlapping genes with a significant prognosis of DEGs among m6Aclusters.

Spearman’s correlation analysis was performed to investigate the correlation between m6Ascore and other relevant biological processes using the gene sets reported by Mariathasan et al (online supplemental table S3),18 including (1) antigen processing machinery (APM), (2) effector CD8 T-cell signature, (3) immune checkpoint, (4) nucleotide excision repair, (5) mismatch repair, (6) DNA replication, (7) DNA damage repair, (8) epithelial-mesenchymal transition markers, (9) Wnt targets, (10) pan-fibroblast transforming growth factor-β response signature, and (11) angiogenesis signature.

A systematic search for the genomic and transcriptomic datasets of ccRCC patients treated with anti-PD-1 therapy was performed. We ultimately included one immunotherapeutic cohort—advanced ccRCC with treatment of PD-1 blockade and mammalian target of rapamycin (mTOR) inhibition—obtaining the genomic, transcriptomic and clinical data (online supplemental table S4) from the online supplemental data appended to the published paper.34

Statistical significance for three or more groups was estimated using the Kruskal-Wallis test and association between categorical variables was explored using the χ² test. The correlation coefficient was calculated via Spearman’s correlation analysis. Continuous variables were dichotomized for patient survival using optimal cut-off values determined by ‘survminer’ R package. The Kaplan-Meier method was used to generate survival curves and the log-rank test was used to determine the statistical significance of differences. The independent prognostic factors, determined by multivariate Cox regression analysis, were visualized by ‘forestplot’ R package. The ‘oncoplot’ function of R package ‘maftools’ was used to depict the mutation landscape of The Cancer Genome Atlas Kidney Renal Clear Cell Carcinoma (TCGA-KIRC) cohort and immunotherapeutic cohort. The protein–protein interaction (PPI) networks among m⁶A regulators were identified based on the STRING interaction database35 and visualized by Cytoscape.36 All tests were two sided, and p<0.05 was regarded as significant. All analyses were performed with R software V.3.62 (http://www.R-project.org).

On reviewing the literature, we identified 24 genes that mainly regulate RNA methylation including 8 writers (RBM15/RBM15B, METTL14, METTL3, WTAP, CBLL1, VIRMA and ZC3H13), 2 erasers (FTO and ALKBH5) and 14 readers (FMR1, ELAVL1, HNRNPC, HNRNPA2B1, YTHDF1/2/3, YTHDC1/2, RBMX, IGF2BP1/2/3 and LRPPPRC). Somatic mutations and CNVs were integrated to explore the prevalence of m⁶A regulator variations in ccRCC. The overall average mutation frequency of m⁶A regulators was low, with only 27 of 336 samples having m⁶A regulator mutations (figure 1A). We then studied the CNV alteration frequency of the m⁶A regulators and demonstrated that CNV alterations were prevalent (figure 1B). The mRNA expression levels of m⁶A regulators in ccRCC and adjacent tissues were also explored, revealing that 21 out of 24 m⁶A regulators were differentially expressed (figure 1C). The above analyses showed that the genetic and expressional variations in m⁶A regulators were highly heterogeneous between ccRCC and adjacent tissues, suggesting a crucial role for the imbalance of m⁶A regulator expression in the development and progression of ccRCC. Moreover, the function of genes is not isolated, in that it has been shown that collaboration among m⁶A regulators exists in the context of cancer.37 38 Thus, the correlation of mRNA expression among m⁶A regulators was explored. We identified that writers, erasers, and readers had a high expression correlation (figure 1D, (online supplemental table S5) and interacted with each other frequently in PPI networks (online supplemental figure S2). Taken together, these results indicate crucial cross-talk roles among m⁶A regulators of RNA methylation in the formation of distinct m⁶A modification patterns.

Next, the clinical relevance of m⁶A regulators in ccRCC patients was explored. We found that many m⁶A regulators were related to prognosis in patients with ccRCC (figure 1D). Several m⁶A regulators (eg, IGF2BP1 and IGF2BP3) presented oncogenic characteristics, with higher expression levels of these genes related to poor prognosis in ccRCC patients. In contrast, we found that several m⁶A regulators (eg, LRPPRC and METTL14) presented characteristics of tumor suppressors, with higher expression levels of these genes correlated with favorable prognosis in ccRCC patients.

Model-based clustering was performed to classify ccRCC patients based on the expression of 24 m⁶A regulators. We ultimately uncovered three distinct methylation modification patterns (identified as m6Aclusters C1–C3), including 118 cases in m6Acluster-C1, 110 cases in m6Acluster-C2, and 285 cases in m6Acluster-C3 (figure 2A). The expression of m⁶A regulators with the greatest differences among subtypes (p<10⁻¹⁵) included two risk factors for overall survival (OS) (IGF2BP2 and IGF2BP3) and four favorable factors for OS (YTHDC1, RBMX, METTL14 and FTO) (online supplemental figure S3). m6Acluster-C3 was characterized by low expression levels of IGF2BP2 and IGF2BP3 and high expression levels of YTHDC1, RBMX, METTL14 and FTO (online supplemental figure S3). Therefore, it was not surprizing that m6Acluster-C3 had the most favorable prognosis (figure 2B).

GSVA was performed to investigate the activity of biological processes among these distinct m⁶A modification patterns. As shown in figure 2C–D and online supplemental table S6, m6Aclusters-C1 and -C3 were markedly enriched in pathways related to immune activation, including the activation of cytokine–cytokine receptor interaction and the chemokine signaling pathway, TCR signaling pathway and BCR signaling pathway. Meanwhile, m6Acluster-C1 showed enrichment in stromal activation pathways such as cell adhesion and ECM receptor interaction. In contrast, m6Acluster-C2 was predominantly associated with the biological process of immunosuppression.

Thorsson et al30 explored the pan-cancer immune landscape and finally identified six immune subtypes (C1–C6) covering 30 cancer types that were assumed to define immune response patterns with implications for further exploration of immunotherapy. Immune subtype C3—characterized by elevated Th17, low to moderate tumor cell proliferation, and lower levels of overall CNVs and aneuploidy than the other immune subtypes—was enriched in most ccRCC patients. Strikingly, the three distinct methylation modification patterns had distinct proportions of the C3 immune subtype, with m6Acluster-C3 having the highest (96.14%), followed by m6Acluster-C1 (90.68%) and C2 (57.27%) (p<0.001) (figure 2A). We then explored the detailed immune characteristics in distinct m⁶A modification patterns. As shown in figure 3A–C, m6Acluster-C1 had a high proliferation rate, ITH, and lower levels of aneuploidy and overall CNVs. m6Acluster-C2 had the highest aneuploidy score and overall CNVs, as well as a high proliferation rate and ITH, and presented a more prominent macrophage signature dominated by M0 macrophages (online supplemental figure S4). m6Acluster-C3 was defined by elevated Th17, low tumor cell proliferation, ITH and lower levels of aneuploidy and overall CNVs.

IMs are essential for cancer immunotherapy with multiple IM agonists and antagonists being investigated in clinical oncology.39 To advance this research, an understanding of their expression in different m⁶A modification patterns is needed. We explored IM gene expression and CNVs among the m⁶A subtypes (figure 3D). IM gene expression varied across m⁶A subtypes and genes with the greatest differences between subtypes (p<10⁻¹⁵) including ADORA2A, CX3CL1, EDNRB, ENTPD1, HMGB1, TNFRSF4, VEGFA and C10orf54 were most highly expressed in m6Acluster-C3 (online supplemental figure S5). CNVs affected numerous IMs and varied across m⁶A subtypes. m6Acluster-C1 and C2 exhibited frequent amplification and deletion of most IM genes.

Spearman’s correlation analysis was performed to explore the specific correlation between each m⁶A regulator and immune cell infiltration. As shown in online supplemental figure S6, there was a widespread correlation between the expression of m⁶A regulators and immune cell infiltration. We focused on the regulator IGF2BP3—an m⁶A reader—demonstrating its association with poor survival in ccRCC patients (figure 4A). We revealed that ccRCC samples with high expression levels of IGF2BP3 demonstrated greater Th2-cell infiltration enrichment in both the TCGA-KIRC dataset and the immunotherapeutic cohort (figure 4B). Consistent with a previous study,40 Th2 cells were associated with negative outcomes in ccRCC patients (figure 4C). Furthermore, we explored whether the expression of IGF2BP3 and Th2-cell infiltration affected the efficacy of anti-PD-1 therapy. In the anti-PD-1 cohort, a trend in impaired survival was observed in patients with high expression levels of IGF2BP3 (figure 4D). As expected, high Th2-cell infiltration was also associated with poor survival with PD-1 blockade (figure 4E).

We applied a methodology (known as m6Ascore) to accurately evaluate the m⁶A modification pattern in individual ccRCC patients. 299 m⁶A subtype-related DEGs (figure 5A, (online supplemental table S6) were identified using the DESeq2 package of R software. Univariate Cox regression analysis was performed to evaluate the prognosis of each gene in the m⁶A subtype-related DEGs; 190 genes conferring significant prognoses were extracted for further PCA to establish the m⁶A gene signature. Changes in the attributes of individual ccRCC patients were visualized with an alluvial diagram which showed that m6Acluster-C2 had the lowest proportion of the C3 immune subtype and was linked to a high m6Ascore (figure 5B).

The correlation between m6Ascore and the known biological processes was analyzed to better demonstrate the features of the m⁶A gene signature (figure 5C, online supplemental table S8). It was shown that m6Ascore was negatively correlated with APM (r = –0.22, p<0.001), but positively correlated with mismatch repair-relevant signatures, including mismatch repair, DNA damage repair and DNA replication. Furthermore, the Kruskal-Wallis test showed a significant difference in m6Ascore among m6Aclusters (figure 5D). Next, the prognostic value of the m6Ascore in patients with ccRCC was explored. The patients were divided into high or low m6Ascore groups, with optimal cut-off values determined by the ‘survminer’ R package. Patients with high m6Ascores demonstrated significant survival impairment (figure 5E). We further performed multivariate Cox regression analysis (with factors related to patient sex, age, grade and TNM status included) to investigate the independent prognostic value of m6Ascore, revealing that m6Ascore serves as an independent prognostic biomarker for ccRCC patients (figure 5F). The prognostic value of m6Ascore in ccRCC was validated in another cohort (91 cases) from the ICGC database (figure 5G). The differences in the distribution of somatic mutations between the high and low m6Ascores in the TCGA-KIRC cohort were explored using the maftools package (figure 5H,I). We also found that high m6Ascore was relatively depleted for PBRM1 mutations (26% vs 44%) (p=0.009).

ICTs (eg, anti-PD-1/PD-L1 therapies) have emerged as a critical breakthrough in the field of tumor therapy. We explored the prognostic value of the m⁶A modification signature in the anti-PD-1 therapy of patients with ccRCC based on one immunotherapeutic cohort. In the anti-PD-1 cohort, the low m6Ascore group presented a markedly prolonged survival (figure 6A,B). In addition, we found that high m6Ascore was relatively depleted for PBRM1 mutations (29% vs 62%) (p<0.001) (figure 6C,D).