Cross-talk of four types of RNA modification writers defines tumor microenvironment and pharmacogenomic landscape in colorectal cancer

Based on the published data, a total of 26 RNA modification “writers” (Table S1), including 3 A-I modification “writers”, 7 m⁶A modification “writers”, 4 m¹A modification “writers”, and 12 APA modification “writers” were included in the current study [13, 19, 24, 28].

To determine the genetic alterations in RNA modification writers in cancer, we assessed the prevalence of non-silent somatic mutations in 26 writers. The mutation frequency of individual writers was relatively low across LAML, PCPG, and UVM cohorts in TCGA, while the COAD cohort showed relatively high mutation frequency of “writers” (Figure S2A). Of the 404 COAD samples, 119 (29.46%) had mutations of RNA modification “writers” (Fig. 1a). Among them, the mutation frequency of ZC3H13 was the highest (7%), followed by PCF11 and KIAA1429, while PABPN1 and NUDT21 did not show any mutations in CRC samples. However, CRC patients with mutations of these “writers” had shorter overall survival than those without mutations (Fig. 1b; log-rank test, p = 0.021), suggested that genetic alteration of “writer” may play functional role in CRC. Next we performed Gene Set Variation Analysis (GSVA) of enrichment analysis [42] using the hallmark gene sets to compare the “writers’“mutation group with the non-mutation group. We found that in the mutation group, more cancer hallmarks related gene sets were enriched, such as KRAS signaling pathway, TNFα signaling pathway, P53 signaling pathway, and IL6/JAK/STAT3 signaling pathway (Figure S2B). IL-6 is a pleiotropic cytokine in immune and inflammatory responses and plays an important role through the JAK3/STAT3/SOCS3 pathway. Moreover, IL-6-mediated dysregulation of the JAK/STAT3/SOCS3 signaling pathway is closely related to the formation of CRC [43, 44]. This suggests that the mutation of “writer” is likely to lead to functional changes that affect the survival prognosis of CRC.

We then examined somatic copy number alterations of these “writers” and found that CSTF1, CPSF1/4, ZC3H13, and KIAA1429 had a widespread frequency of copy number variation (CNV) gain (Fig. 1c). To ascertain whether these genetic variations impacted the expression of RNA modification “writers” in CRC patients, we compared the mRNA alterations of regulators between paired normal and CRC samples and showed that the expression of a majority of the “writers” was significantly increased in CRC (Fig. 1d-g). Compared to the normal colon tissue, RNA modification “writers” with CNV gain (e.g., CSTF1 and CPSF1) were markedly more frequent in CRC tissues (Fig. 1c and g), suggested CNV may be a regulator factor to mRNA expression of “writer”. However, some “writer” showed upregulated mRNA expression but with high frequency of CNV loss. To investigate the discrepancy between CNV values and mRNA expression in tumor sample, we focused on eight writers which showed CNV loss in more than 20% of the samples, and divided the CRC cohort into four groups based on their CNV values, including CNV gain, CNV loss and or non-significant alteration of CNV. Then, we compared mRNA expression of “writer” between these groups (Figure S2C). Indeed, patients with CNV gain showed higher expression than those with CNV loss in these “writers”. METTL14, ADORB1, CPSF2, and PABPN1 showed significant downregulation or non-significant alteration in CNV loss group compared to normal tissues. Tumorigenesis is complex process, CNV changes could not fully explain the differential the expression of “writers” between tumor and normal tissues. Although many of the detected expression alterations of “writers” may be explained by CNVs, the CNV is a partial but not unique factor to regulate mRNA expression [45]. Other features, including DNA methylation and transcription factors, can regulate gene expression [46‐48].

This analysis demonstrated a high heterogeneity of genetic landscape and expression of RNA modification “writers” between normal and CRC samples, indicating that the expression imbalance of RNA modification “writers” has potential roles in the onset and development of CRC.

To gain a comprehensive understanding of the expression pattern of the “writers” involved in tumorigenesis, 1695 CRC samples from eight datasets that contained clinical information (GSE41568, GSE39582, GSE13294, GSE14333, GSE18105, GSE20916, GSE21510, GSE37892) were selected for further analysis (Table S2). Univariate Cox regression showed that 10 of 26 RNA modification “writers’“correlated with CRC prognosis in the GSE39582 dataset (Figure S3A).

To explore the relationship among writers, we calculated pairwise correlations among the expression of 26 writers in CRC and found that positive correlations were more frequent than negative correlations (Fig. 2a). We identified that not only the expression of RNA modification “writers” was remarkably correlated in the same category, but also a significant correlation was present among different types of modification writers. Notably, the expression of ADARB2, ZC3H13, and PABPN1 was negatively correlated with other “writers”. While the expressions of TRMT61B, TRMT6, KIAA1429, TRMT10C, CSTF2/3, CPSF4, and CLP1 were positively correlated to each other (Fig. 2a). Thus, cross-talk among the “writers” may be important for the generation of different RNA modification patterns between individual tumors.

Next, we applied Consensus Clustering based on the expression profiles of 26 selected RNA modification “writers” to classify patients with qualitatively different RNA modification patterns (Table S3). After unsupervised clustering, 727 CRC patients from the combined datasets were identified in Cluster_1, whereas the other 968 patients were identified in Cluster_2 (Fig. 2b). In the prognostic analysis of RNA modification patterns, subtypes revealed a particularly prominent survival advantage in the Cluster_2 modification pattern (Fig. 2c; log-rank test, p = 3.0 × 10^− 5). To identify the biological significance of these distinct RNA modification patterns, we performed GSVA enrichment analysis (Table S5). Cluster_1 was markedly enriched in stromal and carcinogenic activation pathways such as the ECM receptor interaction, TGF-β signaling pathway, and cell adhesion, suggesting that RNA modification “writers” may be associated with tumorigenesis. Cluster_2 was enriched in pathways associated with proliferation and apoptosis, including the activation of the cell cycle, DNA replication, and mismatch repair pathways (Fig. 2c).

A large number of studies have documented the association between TME-infiltrating immune cells and RNA modification. Therefore, we attempted to investigate the functional role of “writers” in TME [31, 32]. To compare the component differences of immune cells among the RNA modification patterns, we used the CIBERSORT method [49], a deconvolution algorithm that uses support vector regression to determine the type of immune cells type in tumors (Table S6-S9). This approach showed that RNA modification “writers” may have a strong correlation with TME cell infiltration (Figure S3B). For example, CSTF1, CSTF3, CPSF3, TRMT6, and TRMT61B were significantly negatively correlated with M2 macrophage differentiation. Differences in TME cell infiltration between the two RNA modification clusters were also analyzed. We observed that the infiltration by M2 macrophages (p = 1.2 × 10^− 16), T regulatory cells (Tregs) (p = 2.2 × 10^− 12), T follicular helper cells (Thf cells) (p = 0.0016), and T gamma delta cells (p = 0.0062) was higher in Cluster_1. The infiltration of activated DCs (p = 9.7× 10^− 5), natural killer cells (p = 0.015), and M1 macrophages (p = 0.0068) were higher in Cluster_2 (Fig. 2e). Overall, the RNA modification Cluster_1 was enriched in immunosuppressive cells, e.g., M2 macrophages and Tregs, indicative of poor prognosis (Fig. 2d). Notably, the distribution of the two types of polarization of macrophages differed significantly between the two clusters. The M2 macrophages were significantly enriched in Cluster_1 (Fig. 2e-f), while M1 macrophages were dominant in Cluster_2. In agreement with this conclusion, the analysis of the expression of macrophage markers indicated that M2 macrophage marker genes IL1R1, FIZ1, TGFB1, IL10, and ARG1 were significantly upregulated in Cluster_1 compared to Cluster_2, while M1 macrophage marker genes IL12A, NOS2, IL23A, and IL15RA were significantly downregulated (Fig. 2g). These suggest RNA modification patterns affected the degree of infiltration by specific immune cell types but did not alter the types of infiltrating immune cells.

To further characterize the functional role of the two RNA modification patterns identified above, we identified 463 RNA phenotype-related DEGs and performed enrichment analysis. We found that these genes showed enrichment in biological processes, particularly those related to DNA replication, cell cycle, and tRNA metabolic process (Figure S4A). They were also enriched in signaling pathways, particularly the p53 signaling pathway and IL-17 signaling pathway (Figure S4B). To further validate this differential regulation, we performed unsupervised clustering analyses based on the 463 genes related to RNA modification. This analysis classified the patients into two genomic subtypes: gene.cluster_A and gene.cluster_B (Table S3 and S10). Consistent with the clustering of RNA modification patterns (Fig. 2b), 205 of 562 CRC patients were clustered in gene.cluster_A, which is related to Cluster_1, while 357 were clustered in gene.cluster_B, which is related to Cluster_2. In addition, patients in gene.cluster_A had a worse prognosis than patients in gene.cluster_B (Figure S4C; p = 0.0021, log-rank test), similarly to patients in Cluster_1 (Fig. 2d).

Given the heterogeneity and complexity of RNA modifications, we constructed a DEGs-based score model based on these phenotype-related genes to quantify the RNA modification pattern of individual patients with CRC; this model was termed as the WM_Score (“Writers” of RNA Modification_Score; see Methods). We found that WM_Score of Cluster_1 significantly higher than Cluster_2 (Fig. 3b; Wilcoxon test, p < 2.2 × 10^− 16). In consistence, gene.cluster_A had significantly higher WM_Score than gene.cluster_B (Fig. 3c; Wilcoxon test, p < 2.2 × 10^− 16). To assess the effect of the WM_Score on TME, we compared the infiltration of immune cells between the WM_Score-low and -high groups. We found that the infiltration of M2 macrophages, Tregs, Tfh cells, and T gamma delta cells was higher in the WM_Score-high group, and the infiltration of activated DCs and M1 macrophages was higher in the WM_Score-low group (Figure S4D).

We performed overlap analysis of these three different classifiers based on the Wayne diagram and the histogram of frequency distribution. As shown in Figure S4, 668 out of 838 (79.72%) samples in the WM_Score high group overlap with the Cluster_1 sample, 798 out of 857 (93.12%) samples in the WM_Score low group overlap with the Cluster_2 sample (Figure S4E-F). Six hundred eleven out of eight hundred thirty-eight (72.91%) samples in the WM_Score high group overlap with samples in the genecluster_A, and 857 (100%) samples in the WM_Score low group overlap with samples in the gene.cluster_B group (Figure S4G-H). The results suggested that these three computational methods of classification have a high degree of coincidence.

To further assess the clinical relevance of the WM_Score, we divided patients into WM_Score-low and -high group with the cutoff value determined by the survminer package. Patients with low WM_Score demonstrated a prominent survival benefit (Fig. 3d; log-rank test, p < 1.3 × 10^− 6). The AUCs of the time-dependent ROC curves for the WM_Score were 0.64, 066, and 0.64 at, respectively, 3, 6, and 12-months overall survival (Fig. 3h). To examine whether the WM_Score could serve as an independent prognostic factor, we performed multivariate Cox regression analysis using the patient clinical characteristics, including age, gender, and TNM status. We found that WM_Score was a robust and independent prognostic biomarker for evaluating patient outcomes (Fig. 3f; HR = 0.5260, 95% CI 0.3898–0.7097, p < 0.001). The reliability of the WM_Score was validated using 562 samples of CRC patients from the TCGA (Fig. 3e,g; Table S4). Consistent with these findings, the WM_Score-low group had a better overall survival advantage than the WM_Score-high group in univariate (Fig. 3e; log-rank test, p = 0.0025) and multivariate (Fig. 3g; HR = 0.5342, 95% CI 0.3551–0.8037, p = 0.00262) Cox regression analysis. These results imply that the WM_Score can reflect the RNA modification patterns and predict the prognosis of CRC patients.

CRC can be divided into four consensus molecular subtypes (CMSs), CMS1–4, with distinct molecular features [50]. They include CMS1 (microsatellite instability) with immune cell infiltration, hypermutated state, unstable microsatellites, and strong immune activation, CMS2 (canonical) with marked activation of WNT and MYC signaling, CMS3 (abnormal metabolism) with evident metabolic dysregulation, and CMS4 (mesenchymal) with prominent TGF-β activation, invasion of stromal cells, and angiogenesis [50]. We have shown that Cluster_1 is accompanied by the activation of EMT, TGF-β, and VEGF signaling pathways (Fig. 2c). Additionally, we calculated the EMT score based on the expression of 25 epithelial marker genes and 52 mesenchymal marker genes [51] and found a positive correlation between the WM_Score and EMT score (Figure S5A; Rs = 0.29; p = 1.7 × 10^− 8). Moreover, the EMT score was significantly higher in the WM_Score-high group than in the WM_Score-low group in the TCGA-COAD/READ cohort (Figure S5B; p = 5.2 × 10^− 5).

To examine the association between the WM_Score and CMS subtypes, we compared the WM_Scores of different CMS subtypes in four GEO datasets (GSE39582, GSE13294, GSE14333, GSE20916) and the TCGA Cohort, respectively (Table S3–4). We found that WM_Scores varied significantly among CMS subtypes, with the CMS4 group showing the highest score (Fig. 4a-b). Also, the distribution of CMS subtypes was significantly different between the WM_Score-high and -low groups. The CMS4 subtype was more frequent in the WM_Score-high group, while the CMS1/2 subtype was predominant in the WM_Score-low group (Figure S5C-D). We further analyzed the signaling pathways characteristic of the different CMS subtypes. The signaling pathways activated mostly in the CMS4 samples were WNT, TGF-β, VEGF, and EMT signaling pathways, while the characteristic signaling pathways activated in CMS1/2 were those related to the cell cycle and proteasome (Figure S5E-F). Consistently, in both cohorts, the enrichment score of CMS4-related signaling pathways was significantly higher in the WM_Score-high group, while the enrichment score of CMS1/2-related signaling pathways was significantly higher in the WM_Score-low group (Fig. 4c-d).

CMS subtypes are reflected in tumor progression and clinical outcome. CMS4 tumors tend to be diagnosed at more advanced stages (Figure S5G) and are associated with worse overall survival (Figure S5H; log-rank test, p = 0.01). We further documented that the WM_Score is different among tumor stages and is higher in more advanced CRC (Fig. 4e), suggesting the involvement of parameters comprising the WM_Score in tumor progression. Although improved treatment strategies involving surgery and chemo- and radiotherapy have increased the overall survival rates in patients with early stages of CRC, 40–50% of CRC patients present with metastasis either at the time of diagnosis or as a recurrent disease after a curative therapy [52]. Most CRC patients with distant metastasis are not suitable candidates for conventional treatment and exhibit a poor 5-year survival rate of < 10% [53]. The correlation between the WM_Score and CMS4 subtype and between the WM_Score and CRC stage suggests that the WM_Score might be associated with patient survival by reflecting tumor metastasis. WM_Score was significantly higher in metastatic than in non-metastatic CRC patients (Fig. 4f). These results suggest that the WM_Score correlates closely with CMS subtypes, and a high WM score may indicate a poor prognosis by being associated with the activation of EMT, TGF-β, and other signaling pathways mediating tumor metastasis.

WM_Score is an assessment model based on the expression of 26 RNA modification “writers”, which regulate post-transcriptional modifications, including RNA transport, localization, translation, and other biological processes. To further assess the power of WM_Score in the interpretation of transcriptional and post-transcriptional events, we focused on RNA modification “writers”-related processes (e.g., APA, RNA editing). Given that transcripts processed by APA have a short 3’UTR, thus tolerating the regulation of miRNAs [54], we hypothesized that RNA modification patterns were associated with distinct miRNA characteristics under the action of APA modification “writers”. We first analyzed differences in miRNAs expression among different RNA modification patterns in the TCGA-COAD/READ cohort. Thirty-nine miRNAs significantly differentially expressed between WM_Score-high and -low groups were screened out, and enrichment analysis of the signaling pathways of their target genes was performed (Table S11). The miRNA-targeted genes were significantly correlated with cell cycle, apoptosis, EMT, PI3K-Akt, and other signaling pathways, and the targeted genes were differentially expressed between the two groups. Thirteen out of 14 miRNA-targeted-DEGs in the EMT signaling pathway were highly expressed (Fig. 5a). In contrast, target genes of a group of miRNAs with lower expression in the WM_Score-low group were upregulated and were significantly enriched in the cell cycle and apoptosis signaling pathway. Specifically, 48 out of 55 miRNA-targeted DEGs in the cell cycle, 24 out of 30 miRNA-targeted DEGs in apoptosis, 25 out of 33 miRNA-targeted DEGs in the mTOR signaling pathway showed enrichment in the WM_Score-low group (Fig. 5a). These analyses suggested that the WM_Score is associated with miRNA expression and the regulation of signaling pathways.

To explore the functional role of RNA modification writers, we analyzed the APA and A-I editing events of each gene in the TCGA-COAD/READ cohort to observe the post-transcription characteristics. We identified the genes with the differences in APA between different RNA modification patterns and compared the survival associated with these genes to determine whether the length of 3’UTR affects the survival of CRC patients (Figure S6A; Table S12–13). Most of the genes with shortening APA events were enriched in the WM_Score-low group and were associated with shorter survival (Fig. 5b and S6A). Akt was associated with the telomerase-reverse transcriptase catalytic subunit HEATR3, which may be activated by Akt in melanoma [55, 56]. A rescue study indicated that LINC00958 regulates the proliferation, motility, and EMT of oral squamous cell carcinoma cells through YBX2 [57]. HEATR3 (Diff = − 0.22; p = 7.61 × 10^− 7) and YBX2 (Diff = − 0.19; p = 5.67 × 10^− 6) transcripts exhibited statistically significant shortening, which was associated with worse survival of CRC patients (Fig. 5c; HEATR3: log-rank test, p = 0.0033; YBX2: log-rank test, p = 3.0 × 10^− 4). We raised the possibility that in the WM_Score-high group, due to the shortening of HEATR3 and YBX2, which, in turn, shorten the 3’UTR, miRNA may not be able to target the corresponding gene, resulting in the activation of gene expression and contributing to the initiation and development of CRC.

We identified the genes with differences in A-to-I editing between WM_Score-high and -low groups in 159 TCGA-READ/COAD samples (Table S14–15). The A-to-I editing differential genes identified were less due to relatively low read coverage of RNA-seq data in CRC, with fewer informative editing sites [58]. An increasing amount of evidence shows that ADAR1-mediated A-to-I editing can modify the 3’UTR region of mRNA, affecting the binding of miRNAs [59]. The genes with lower A-to-I editing rate were enriched in the WM_Score-low group and were associated with worse survival of CRC patients (Fig. 5d and S6B). NUP43 is a stable component of the Nup107–160 complex, which may regulate the malignant transformation of aged cells by enabling the translocation of the phosphorylated extracellular signal-regulated kinase (ERK) to the nucleus [60]. Reduction in NUP107 attenuates growth factor signaling in senescent eukaryotic cells, triggering their apoptosis [61]. The potassium channel KCNE3 is a VEGFA-inducible gene selectively expressed in vascular endothelial cells [62]. KCNE3 is suppressed in tumors responding to the VEGFA blockade [63]. The A-to-I editing of NUP43 (Diff = − 0.33; p = 8.0 × 10^− 6) and KCNE3 (Diff = − 0.15; p = 0.0026) is associated with shorter survival time of CRC patients (Fig. 5e; NUP43: log-rank test, p = 0.024; KCNE3: log-rank test, p = 0.0077). The difference in the rate of A-to-I editing of these genes between WM_Score-high and -low groups may be regulated by miRNA via the editing of the 3’UTR regions, thus affecting the occurrence and development of CRC.

To further understand the effects of the WM_Score on drug response, we assessed the association between the WM_Score and the response to drugs in cancer cell lines. Using the Spearman correlation analysis, we identified 42 significantly correlated pairs between WM_Score and drug sensitivity in the Genomics of Drug Sensitivity in Cancer (GDSC) database [64] (Fig. 6a,Table S16). Among them, 24 pairs showed that drug sensitivity correlated with the WM_Score, including the MEK inhibitor Selumetinib (Rs = − 0.29, p = 1.31 × 10^− 15), mTOR inhibitor LJI308 (Rs = − 0.24, p = 8.21 × 10^− 10), and EGFR inhibitor AZD3759 (Rs = − 0.26, p = 4.12 × 10^− 13). Eighteen pairs exhibited drug resistance correlated with the WM_Score, including cell cycle checkpoint kinase inhibitor AZZD7662 (Rs = 0.23, p = 9.9 × 10^− 11) and Bcl-2 inhibitor AZD5991 (Rs = 0.24, p = 1.19 × 10^− 9). Further, we analyzed the signaling pathways of the genes targeted by these drugs. We found that drugs whose sensitivity was associated with WM_Score-high were mostly targeting MAPK, mTOR, and VEGF signaling pathways. In contrast, the drug whose sensitivity was associated with WM_Score-low were targeting apoptosis and cell cycle signaling pathway (Fig. 6b). Together, these results imply that RNA modification patterns are correlated with drug sensitivity. Thus, the WM_Score may be a potential biomarker for establishing appropriate treatment strategies.

A major effort has been made to identify biomarkers to predict the response to immunotherapy, including tumor mutation burden (TMB) and the expression of PD-L1 protein [65‐67]. Considering that WM_Score appears to be associated with the immune microenvironment of the tumor (Figure S4D), we examined the power of the WM_Score to predict the response of patients to ICB therapy. This analysis was based on two immunotherapy cohorts (Fig. 6c-i). We found that patients with WM_Score-low exhibited significant clinical benefits and had a markedly prolonged overall survival in both anti-PD-L1 cohorts, including the IMvigor210 cohort [68] (Fig. 6c; log-rank test, p = 0.022) and the bladder cancer cohort [69] (Fig. 6i; log-rank test, p = 0.0065). The 348 patients of the IMvigor210 cohort exhibited different degrees of response to anti-PD-L1 blocker, including complete response (CR), partial response (PR), stable disease (SD), and progressive disease (PD). The CR patients showed the lowest WM_Score than patients with other types of responses (Fig. 6d). The chi-squared test performed between WM_Score-low and -high groups also showed significantly better therapeutic outcomes in WM_Score-low patients (Fig. 6e; p = 0.014). We analyzed the WM_Score of the three immune subtypes of IMvigor210, including “immune inflamed”, “immune excluded”, and “immune desert” [70] (Fig. 6f), and observed that the WM_Score of the “immune inflamed” type was lower than in the other two groups. In addition, the TMB and neoantigen burden were significantly higher in the WM_Score-low group than in the WM_Score-high group (Fig. 6g-h), which may explain, at least in part, the survival advantage and the greater benefit of the ICB treatment in the WM_Score-low group.

We have also found that the activation of M2 macrophages and TME of stromal cells was significantly higher in WM_Score-high tumors, and these processes may mediate the immune tolerance of tumors. This result suggested that WM_Score-high tumors may represent “cold tumors” characterized by resistance to immunotherapy. In summary, the performed analyses suggest that the established “writers” of RNA modification score model might improve the selection of drugs for CRC and the prediction of response to anti-PD-L1 immunotherapy.

The workflow of our study was shown in Figure S1A. Public gene expression data and complete clinical annotations from the same sequencing platform were retrieved in Gene-Expression Omnibus (GEO) and the Cancer Genome Atlas (TCGA) database. mRNA expression, miRNA expression, somatic mutation, SCNAs, and clinical data, including tumor stage, histology subtype, gender and overall survival times were obtained from TCGA database (https://​portal.​gdc.​cancer.​gov/​). Eight GEO colorectal cancer cohorts (GSE41568, GSE39582, GSE13294, GSE14333, GSE18105, GSE20916, GSE21510, GSE37892) and TCGA-COAD/READ cohort were included for further analysis. The data information is summarized in Table S2. The “ComBat” algorithm of sva Package [85] was used to correct the batch effect caused by non-biotechnological bias. The data was analyzed using R (version 3.6.2) and R Bioconductor packages.

We collected datasets with immunotherapy. The IMvigor210 cohort and bladder cancer with anti-PD-L1 cohort were included in the study of the relationship between WM_Score and immunotherapy prognosis. The IMvigor210 cohort [68]: Intervention treatment of advanced urinary tract transitional cell carcinoma (atezolizumab, anti-PD-L1 antibody). IMvigor210 cohort with the expression of the data and detailed clinical notes are downloaded from http://​research-pub.​gene.​com/​IMvigor210CoreBi​ologies. Expression and clinical information of bladder cancer with anti-PD-L1 cohort downloaded from https://​doi.​org/​10.​5281/​zenodo. 546,110 [69].

Unsupervised clustering algorithm was applied to cluster analysis of RNA-modified “writers” in 1695 colorectal cancer samples. RNA modification “writer” consists of 7 m⁶A modification enzymes (METTL3, METTL14, WTAP, RBM15, RBM15B, ZC3H13, and KIAA1429), 4 m¹A modification enzymes (TRMT61A, TRMT61B, TRMT10C, and TRMT6),12 APA modification enzymes (CPSF1-4, CSTF1/2/3, PCF11, CFI, CLP1, NUDT21, and PABPN1) and 3 A-I modification enzymes (ADAR, ADARB1, and ADARB2). Unsupervised clustering was applied to detect the robust clustering of colorectal cancer. We used the Consensus-Clusterplus package for the above steps, and conduct 1000 repetitions to ensure the stability of the classification [86, 87].

In order to study the differences of RNA modification patterns in biological processes, we used “GSVA” R package to conduct GSVA enrichment analysis [42]. The gene set “c2.cp.kegg.v7.1” and “ h.all.v7.2 “ for GSVA analysis was downloaded from the MSigDB database. The clusterProfiler R Package was used to functionally annotate 26 RNA modification enzyme genes [88].

We use CIBERSORT algorithm (https://​cibersort.​stanford.​edu/​) to quantify the relative abundance of 22 types of immune cells in colorectal cancer with parameters as follows: the input mixture matrix is our gene expression matrix, the input of gene signature reference for 22 immune cell types from Newman et al. [49], 100 times for permutation test, and RNA-seq data without quantile normalization, while microarray data with quantile normalization.

We obtained epithelial-to-mesenchymal transition gene signatures from Mak et al [51], including 25 epithelial and 52 mesenchymal marker genes. The EMT score for each sample was estimated as \( \sum \limits_i^N\frac{M^i}{N}-\sum \limits_j^n\frac{E^j}{n} \), as described in a previous study [51], where M and E represent the expression of the mesenchymal gene and epithelial gene, respectively, and N and n respectively represent the number of mesenchymal genes and epithelial genes.

The transcription profiles for about 1000 cancer cell lines, drug response measurements as AUC for antitumor drugs in cancer cell lines, and targets/pathways of drugs are downloaded from Genomics of Drug Sensitivity in Cancer (GDSC, http://​www.​cancerrxgene.​org/​downloads) [64]. We performed Spearman correlation analysis to calculate the correlation between drug sensitivity and WM_Score, and considered |Rs| > 0.2 and used Benjamini and Hochberg adjustment for FDR and considered an FDR < 0.05 as significant correlation.

Spearman and distance correlation were used to calculate the correlation coefficient of RNA modification “writers” expression. Wilcoxon test was used to compare the differences. Receiver operating characteristic (ROC) curve was used to verify the validity of the model.