Domain-derived knowledge enabled machine learning and functional characterization of cancer-associated RNA methylation sites

Abstract

Background

N6-methyladenosine (m⁶A), the most abundant post-transcriptional modification in eukaryotic mRNA, plays pivotal roles in diverse biological processes. Dysregulation of m⁶A levels has been implicated in numerous human diseases, particularly cancer. Although several computational tools exist for predicting putative m⁶A sites, none have specifically addressed the identification of cancer-associated (or pro-cancer) m⁶A residues at single-base resolution.

Methods

To address this gap, we developed m6A-CAPred, a computational framework for accurate prediction of cancer-associated m⁶A sites at base resolution. Our model was trained on a comprehensive dataset comprising experimentally validated m⁶A sites from 25 cancer cell lines and 23 normal tissue samples, based on a hybrid feature extraction approach integrating both sequence- and curated genome-derived features.

Results

Initial analysis revealed that sequence information alone only provided limited predictive performance. However, by incorporating genomic context features, m6A-CAPred significantly achieved improvement in prediction performance (an average AUROC of 0.885 tested on independent dataset), successfully capturing the distinct characteristics between cancer-associated and normal m⁶A sites. We then applied m6A-CAPred for transcriptome-wide prediction to screen for potential cancer-associated m⁶A sites. The somatic variants derived from 33 types of TCGA cancer projects were extracted for independent validation, and the results showed that cancer-somatic SNP density clearly differentiated the predicted pro-cancer and normal m⁶A sites, further confirming the model's biological relevance. Additionally, the cancer-associated m⁶A sites showed significant enrichment in functional important biological processes and cancer-related pathways.

Conclusions

Overall, we hope that m6A-CAPred will serve as a valuable resource for cancer epitranscriptome research, with potential applications in cancer biomarker discovery. The web server implementing our model is freely available at www.rnamd.org/m6A-CAPred.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12859-026-06402-6.

Introduction

Exploration of RNA epigenetics has led to the discovery of more than 170 types of RNA modification. Among them, N6-methyladenosine (m⁶A), the most prevalent marker in messenger RNA and long noncoding RNA, has been identified as an abundant and dynamically regulated modification. m⁶A was first discovered in poly-A RNA in 1974 [1], and since then, it has been identified in various eukaryotic organisms [2]. Over the past decades, multiple studies have underscored the biological significance of m⁶A modification in various aspects, including but not limited to, regulation of mRNA translation [3], RNA–protein interaction [4], microRNA (miRNA) processing [5], DNA damage response, and regulation of RNA stability [6]. In the context of cancer, m⁶A modification has been demonstrated to be a key factor due to its implications in cancer metastasis [7], abnormal mRNA expression levels [8], and immune evasion [9]. Dysregulation of m⁶A has been reported to play an essential role in tumor proliferation and invasion [10, 11], including colon cancer [8], lung cancer [12], and breast cancer [13]. It is apparent that accurately identifying the locations of cancer-associated (or pro-cancer) m⁶A modifications is crucial for the study and understanding of the downstream effects of RNA modification in cancer biology.

The first whole transcriptome m⁶A profiling approach, m⁶A -seq (MeRIP-seq), was developed in 2012 [14] to locate m⁶A sites. It utilizes antibody-based enrichment of the m⁶A signal to detect and identify regions containing m⁶A modification, with a resolution of approximately 150 nucleotides and applied in many biological contexts [15, 16]. Recent advancements have also been made in integrating ultraviolet cross-linking, enzymatic activity, and domain fusion to achieve high resolution or even base-resolution m⁶A detection. These techniques include miCLIP/m⁶A-CLIP-seq [17], m⁶A-REF-seq [18], and DART-seq [19]. Nevertheless, these alternative approaches require more complex experimental procedures compared to m⁶A-seq and have thus far been utilized in fewer biological contexts.

To date, machine learning has become essential tools in modern bioinformatics [20–24]. A number of epitranscriptome databases [25–31] and computational approaches [32–42] have been developed for large-scale collection and accurate identification of RNA modifications. For m⁶A methylation, pioneer studies such as SRAMP [43] and iRNA toolkits [44–46] were developed based on combination of sequence-derived features. WHISTLE [47] was a high-accuracy predictor that firstly integrating domain knowledge (genomic features) into m⁶A prediction framework. And most recently, the deep learning-based approaches were also proved to be another powerful way for m⁶A prediction [48–50]. These works together have greatly facilitated the development of computational biology [28, 51–54]. However, existing computational models were trained using experimentally validated m⁶A sites as positive data and randomly selected adenine (A) bases from the genome as negative data. Consequently, these models can only predict whether a given nucleotide is methylated or not, without distinguishing between different functional associations of m⁶A modifications.

In our study, by leveraging domain-specific features derived from a comprehensive collection of cancer and normal tissue contexts, m6A-CAPred was trained on cancer-associated m⁶A sites and normal m⁶A residues, based on the difference ratio (DR) calculated to represent the observed difference of a m⁶A site between cancer and normal conditions (see Fig. 1). The model applied a hybrid feature extraction approach integrating both sequence- and curated genome-derived features, achieving an average AUROC of 0.885 tested on independent test datasets and demonstrating its ability to accurately identify pro-cancer m⁶A modifications at single-base resolution. Based on the proposed model, we then conducted a large-scale prediction on ~ 430,000 experimentally validated m⁶A sites to identify potentially cancer-associated m⁶A residues. Independent validation test showed that cancer-somatic SNP density can be clearly differentiated between the predicted pro-cancer and normal m⁶A sites, further confirming the model's biological relevance. Additionally, the cancer-associated m⁶A sites showed significant enrichment in functional important biological processes and cancer-related pathways. To share our findings, we developed a user-friendly web interface for the proposed framework. To summarize, the m6A-CAPred offers (i) a high-accuracy model for in silico identification of cancer-associated m⁶A residues at single-base resolution; (ii) a database holding a total of 111,937 high-confidence experimentally validated m⁶A sites annotated with cancer context labels, which can be extracted for further analysis or model development, and (iii) a web server for high-accuracy prediction of potentially pro-cancer m⁶A sites from user-provided data. The m6A-CAPred is freely accessible at www.rnamd.org/m6A-CAPred.

Fig. 1.

Overall construction of m6A-CAPred. A The dataset used to develop m6A-CAPred comprises high-confidence experimentally validated m⁶A sites and a large array of context-specific MeRIP-seq samples extracted from 25 cancer cell lines and 23 normal tissue samples. B The difference ratio (DR) was calculated to represent the observed difference of a m⁶A site between cancer and normal conditions. C, D The m6A-CAPred integrates genome-derived characteristics with sequence information. The model performance was systemically evaluated through cross-validation and independent testing data. E Functional test and biological validation: Additional analyses including TCGA cancer SNP density test, GO and KEGG analysis established biological relevance

Results

Different sequence encoding approaches used for prediction of pro-cancer m⁶A sites

To try to capture the distinct patterns of pro-cancer m⁶A sites, we applied different sequence-based feature extracting approaches for model development and tested their performances. Firstly, we tried to explore whether there are significant differences in the primary RNA sequences between pro-cancer and normal m⁶A sites. The 41nt RNA sequence centered on each m⁶A site was extracted and encoded using combined sequence-based approaches. We considered a total of three combination: NCP + PSNP, NCP + ND + EIIP, and EIIP + PseEIIP. The Support Vector Machine (SVM) was applied to represent traditional machine learning framework. The results (Table 1) showed that sequence-based information only achieved very limited performances, with the best performance achieved by combination of EIIP and PseEIIP (AUROC of 0.577), suggesting that sequence information alone cannot effectively classify the pro-cancer m⁶A residues from the normal ones.

Table 1.

Performance evaluation using different sequence-based encoding approaches

Encoding approach	Independent test
	Sn (%)	Sp (%)	ACC (%)	MCC	AUROC
NCP + PSNP	58.3	49.1	53.7	0.074	0.550
NCP + ND + EIIP	52.8	56.0	54.4	0.088	0.566
EIIP + PseEIIP	54.1	55.5	54.8	0.096	0.577

We randomly selected 80% of dataset as training dataset and the performance of predictors were evaluated by the rest of 20% of dataset as independent testing data, only data not included in training step was selected for independent testing purpose

Classifying pro-cancer m⁶A sites using genome-derived information

We next tried to train the classifier by adding genome-derived information, with the integrated model combining sequence-based and 54 additional genomic knowledge (Supplementary Table S1). We found that genomic features greatly enhanced the sequence-based model, achieving an average performance enhancement of approximately 30% (Table 2). Among the tested frameworks, the SVM model combining genomic features with EIIP and PseEIIP achieved optimal performance (AUROC of 0.885 and AURPC of 0.876, Supplementary Figure S1), outperforming other combinations: genomic + NCP + ND + EIIP (AUROC of 0.876) and genomic + NCP + PSNP (AUROC of 0.869) on independent validation datasets. Additionally, comparative analysis across machine learning algorithms revealed SVM as the most robust classifier, showing superior performance to both Generalized Linear Model (GLM) and Naïve Bayes approaches based on 41-nt sequence window (Supplementary Tables S2 and S3). To rigorously evaluate the capability of our model, we performed a chromosome-level split on training and testing data. Specifically, we assigned data from Chr 1–15 to the training set and Chr 16–22 to the independent test set. This strategy ensures that no sites from the same gene (or even the same chromosomal region) are shared between the datasets. The results (Supplementary Table S4) showed that the model achieved an AUROC of 0.859 and an AUPRC of 0.868. Although there is a minor decrease compared to the random split (AUROC of 0.885), the performance remains high and robust. For model interpretation, the SHapley Additive exPlanations (SHAP) analysis was conducted to underscore the critical role of domain-specific genomic features in distinguishing cancer-associated from normal m6A sites (Supplementary Figure S2), thereby validating the reliability of our proposed m6A-CAPred framework.

Table 2.

Performance evaluation using sequence-genomic combined features

Method	fivefold cross validation					Independent test
	Sn (%)	Sp (%)	ACC (%)	MCC	AUROC	Sn (%)	Sp (%)	ACC (%)	MCC	AUROC
Integrated model 1*	78.4	78.5	78.4	0.569	0.872	79.1	79.7	79.4	0.588	0.869
Integrated model 2*	78.8	78.3	78.6	0.571	0.869	78.9	80.1	79.5	0.590	0.876
Integrated model 3*	80.2	80.7	80.4	0.608	0.884	80.8	79.3	80.1	0.602	0.885

Integrated model 1*: NCP + PSNP + genomic feature; Integrated model 2*: NCP + ND + EIIP + genomic feature; Integrated model 3*: EIIP + PseEIIP + genomic feature

Additionally, to investigate whether dataset selection influences model performance, we conducted additional validation using alternative percentile thresholds (top/bottom 5% and 10% outliers) for dataset construction. Our analysis revealed that the model maintains robust performance across all cutoffs (AUROC > 0.82; Table 3). Notably, the predictive accuracy declines proportionally with increasing percentile ranges. These findings suggest three key conclusions: (i) the DR metric effectively identifies the most biologically significant cancer-associated m⁶A sites; (ii) relaxing the threshold expands the gene list without altering the core set of candidates, confirming that the identification of these specific genes is highly stable and not an artifact of a specific cutoff (Supplementary Figure S3); and (iii) our model successfully captures the distinct genomic signatures of these cancer-associated modifications. Overall, the optimal performance at the 2.5% cutoff indicates that this stringent threshold best isolates m⁶A sites with the strongest cancer association while minimizing noise from less significant modifications.

Table 3.

Prediction performance using different percentage of right/left-tail outliers

Cut-off	Sn (%)	Sp (%)	ACC (%)	MCC	AUROC
2.5%	80.8	79.3	80.1	0.602	0.885
5%	74.7	78.9	76.8	0.535	0.843
10%	71.1	75.1	73.4	0.469	0.822

Cut off stands for top/bottom 5% and 10% outliers, calculated by DR ratio

SNP density analysis clearly differentiated the predicted pro-cancer from normal m⁶A sites

To biologically validate the predicted pro-cancer m⁶A sites, we performed a comprehensive analysis of somatic variant density using data from 33 TCGA cancer projects. Leveraging the proposed m6A-CAPred model, we first performed a large-scale prediction on a total of 427,586 experimentally validated m⁶A sites at base-resolution level [55]. We applied different cut-off values (0.3–0.9) for classifying the potentially pro-cancer m⁶A residues and calculated the SNP density within a ± 2 bp flanking window of pro-cancer and normal m⁶A sites, respectively (Table 4). Strikingly, the predicted pro-cancer m⁶A sites exhibited significantly higher SNP density than normal sites across all cutoff values (Table 4), indicating a stronger association with cancer through the disruption of these functional m⁶A methylation sites. Taken together, these results suggest that the m⁶A sites classified into the pro-cancer group using m6A-CAPred are generally more associated with cancer, thereby demonstrating the effectiveness of our newly proposed m6A-CAPred framework.

Table 4.

The SNP density test of TCGA somatic variants around different m⁶A groups

Mutation type	Cut-off	# of predicted pro-cancer m6A sites	# of predicted normal m6A sites	Pro-cancer sites with proximal SNPs (± 2 bp) [n (%)]	Normal sites with proximal SNPs (± 2 bp) [n (%)]	p-value
TCGA somatic variant	0.3	169,680	257,906	17,366 (10.23%)	25,890 (10.04%)	*
	0.4	143,216	284,370	15,174 (11.60%)	28,080 (9.87%)	***
	0.5	121,623	305,963	133,35 (10.96%)	199,09 (9.78%)	***
	0.6	103,077	324,509	116,44 (11.30%)	31,603 (9.74%)	***
	0.7	84,787	342,799	9668 (11.40%)	33,588 (9.80%)	***
	0.8	64,245	363,341	7257 (11.30%)	35,992 (9.91%)	***
	0.9	36,637	390,949	4114 (11.23%)	39,148 (10.01%)	***

^* stands for p-value < 0.05; ** stands for p-value < 0.01; and *** stands for p-value < 0.001

Functional relevance of cancer-associated m⁶A methylomes

To elucidate the mechanistic roles of m⁶A methylation in cancer regulation across diverse tumor types, we systematically investigated the functional relevance of genes harboring cancer-associated m⁶A sites through Gene Ontology (GO) enrichment and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analyses. Our GO enrichment analysis of biological processes (Fig. 2A, B) revealed significant enrichment of cancer-associated m⁶A-CArrying genes in three critical tumorigenesis-related categories: (i) DNA damage repair, including double-strand break repair (p = 1.7901E−07), response to ionizing radiation (p = 7.4738E−06), recombinational repair (p = 4.5649E−06), and DNA integrity checkpoint signaling (p = 2.0641E−06), consistent with emerging evidence that m⁶A modifications affect genomic stability by regulating DNA repair mechanisms in cancer [56, 57]; (ii) cell cycle regulation, encompassing regulation of cell cycle phase transition (p = 4.2919E−05), negative regulation of cell cycle phase transition (p = 1.6307E−05), G2/M transition of mitotic cell cycle (p = 7.2058E−05), and mitotic cell cycle checkpoint signaling (p < 0.001), aligning with the established role of m⁶A in controlling cell cycle progression during malignant transformation [58–60]; and (iii) DNA replication and recombination, involving DNA replication (p = 2.6249E−07), DNA recombination (p = 1.2476E−05), and DNA-templated DNA replication (p = 2.5958E−06), supporting m⁶A involvement in replication processes whose dysregulation drives oncogenesis [61]. Complementary KEGG analysis further demonstrated significant enrichment of these m⁶A-CArrying genes in cancer development pathways, particularly breast cancer (p < 0.001) and cell cycle regulation (p = 4.1451E−06) (Fig. 3C). The complete datasets are available in Supplementary Sheet S1 and S2.

Fig. 2.

Functional relevance of the most cancer-associated m⁶A methylome. A and B The top results of Gene Ontology (GO) enrichment analysis including biological process (BP). C Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis

Fig. 3.

m6A-CAPred online resources. A The m6A-CAPred web-interface consists of two components. B, C A database featuring 111,937 experimentally validated m⁶A sites, annotated with difference ratio and context labels. D Users can also upload their query m⁶A sites with genome coordinates to the online web server. E The returns results indicating whether the predicted m⁶A sites are classified as pro-cancer or normal

The m⁶ACA-Pred online web platform

To facilitate the exploration of cancer-associated epitranscriptomic modifications, we established a comprehensive, open-access online platform (Fig. 3A). This system integrates two core modules: a knowledge base and a predictive web server.

First, the database module hosts a curated collection of 111,937 experimentally validated base-resolution m⁶A sites. Each entry is richly annotated with genomic coordinates, gene details, profiling techniques, and experimental sources (Fig. 3B). A key feature of the interface is the inclusion of cancer versus normal context labels (indicated by visual markers), allowing users to discern context-specific methylation patterns. To enhance data navigability, the database includes an interactive filtering system where users can query sites based on specific Difference Ratio (DR) thresholds or percentile ranks (e.g., Top 2.5% pro-cancer sites) (Fig. 3C). Next, the m6A-CAPred web server allows researchers to perform functional classification on their own data (Fig. 3D). Users can submit genomic coordinates (chromosome, position, and strand) via direct text input or file upload (supporting standard formats such as BED and CSV). Upon submission, the server returns a detailed prediction table encompassing two metrics: a quantitative probability score (ranging from 0 to 1) and a binary classification. Sites yielding a probability score > 0.5 are classified as ‘cancer-related’, while those below this threshold are labeled 'normal' (Fig. 3E). All prediction results and database contents are freely available for download at www.rnamd.org/m6A-CAPred.

Discussion

N6-methyladenosine (m⁶A) has emerged as a pivotal regulator of diverse cellular processes, with its dysregulation increasingly implicated in oncogenesis and cancer progression. While numerous computational tools have been developed for m⁶A site prediction, current methods remain fundamentally limited by their inability to distinguish functionally relevant, cancer-associated m⁶A modifications from background methylation events. To bridge this gap, we developed m6A-CAPred, a machine learning framework that integrates genomic features with sequence information to enable accurate identification of pro-cancer m⁶A sites at single-nucleotide resolution. Our approach demonstrates reasonable performance by specifically capturing the distinct genomic characteristic of oncogenic m⁶A modifications.

Here we present m6A-CAPred, the first computational tool specifically designed to identify cancer-associated m⁶A modifications at single-base resolution, achieving superior accuracy (AUROC of 0.885) by integrating genomic features with sequence information. Alongside the predictive framework, we provide a curated database of 111,937 high-confidence ⁶A sites annotated with cancer-specific labels, enabling systematic exploration of m⁶A’s functional roles in oncogenesis. The accompanying user-friendly web server facilitates large-scale predictions and comparative analyses, empowering researchers to investigate pro-cancer m⁶A sites and accelerate the discovery of potential therapeutic targets. Together, these contributions bridge a critical gap in epitranscriptomics, offering both a robust analytical tool and a foundational resource for studying m⁶A in cancer biology.

Although the current framework already provides a significant advance over existing binary classification tools, several limitations remain to be addressed. First, it is important to note the limitations associated with the use of publicly available MeRIP-seq datasets. Due to the heterogeneity of the data sources, technical variations, such as differences in sequencing depth, antibody efficiency, and library preparation protocols are inevitable. While we applied a uniform processing pipeline, the DR labeling strategy, which depends on the presence or absence of peaks, may still be influenced by these technical factors. For instance, datasets with lower sequencing depth might yield false negatives in peak calling. Therefore, the results should be interpreted with the understanding that technical noise cannot be completely ruled out in large-scale integrative analyses. Second, The cancer group predominantly consists of cell lines, whereas the normal group comprises tissues. We acknowledge that in vitro cell cultures possess distinct transcriptomic and epigenetic landscapes compared to in vivo tissues, potentially introducing confounding factors related to cell-line adaptation rather than tumorigenesis. Some of the identified DR signals may partially reflect sample-type disparities. Consequently, while the model demonstrates robust performance, future studies would significantly benefit from the integration of matched patient-derived tumor and adjacent normal tissue samples. Utilizing such clinical datasets would minimize sample-type bias and better assist in modeling the complex regulatory networks of the cancer epitranscriptome.

The practical advantage of m6A-CAPred lies in its implementation as an openly accessible, user-friendly web server that offer identification of putative cancer-associated m⁶A modifications. This platform enables researchers to: (i) efficiently screen transcriptome-wide m⁶A sites for cancer relevance, (ii) compare predictions against our curated database of 111,937 annotated sites, and (iii) prioritize functional sites for experimental validation. By providing base-resolution predictions of pro-cancer m⁶A residues, we hope the proposed model could advances our understanding of m⁶A functional mechanisms in cancer biology and may facilitate the discovery of novel therapeutic targets.

Furthermore, we acknowledge that model accuracy could benefit from integration of newer experimental datasets as they become available. To further advance the capabilities of m6A-CAPred: (i) integrating multi-omics data (including transcriptomic, proteomic, and epigenomic profiles) could significantly enhance prediction accuracy by capturing the complex regulatory networks governing cancer-associated m⁶A modifications, (ii) developing tissue-specific and cancer subtype-specific models would improve clinical relevance by addressing tumor heterogeneity across different malignancies. Together, these advancements would transform m6A-CAPred into a precision medicine tool capable of guiding biomarker discovery and targeted therapeutic development across diverse cancer contexts.

Conclusion

In summary, we present m6A-CAPred, a computational framework designed for the identification of cancer-associated m⁶A modifications at single-base resolution. By integrating sequence-based features with domain-specific biological knowledge, our model achieved robust predictive performance (AUROC of 0.885) in distinguishing pro-cancer m⁶A sites from normal ones. Leveraging large-scale transcriptomic data and validation using TCGA somatic variants, we further demonstrated that predicted pro-cancer m⁶A sites exhibit distinct molecular signatures, supporting their biological relevance in cancer. We hope our work will contribute to the discovery of potential RNA methylation-based therapeutic targets.

Materials and methods

Benchmark dataset

m6A-CAPred was proposed to predict the pro-cancer m⁶A methylation sites at base-resolution level. The positive dataset (P) and negative dataset (N) were all high-confidence experimentally validated m⁶A sites collected from m⁶A-Atlas database (with record time > 2, a total of 111,937 sites) [55]. The dataset P (pro-cancer m⁶A sites) and N (normal m⁶A sites) were further classified by checking whether they localized in m⁶A-enriched regions from 25 cancer cell lines and 23 normal tissue samples collected from m⁶A-TSHub [62] (Supplementary Sheet S3). It may be worth noting that, the m⁶A sites that did not occur in either cancer or normal conditions were identified as background noise and were consequently excluded from further experiments. Specifically, the difference ratio (DR) was calculated to represent the observed difference of a m⁶A site between cancer and normal conditions:

1

where P_Cancer represents the percentage of occurrence of an m⁶A site in the cancer cell lines, while P_Normal represents that of normal-condition samples. The DR ranges from minus 0.5 to 0.5, with 0.5 indicating the most cancer-associated m⁶A sites and minus 0.5 indicating m⁶A sites totally observed in normal samples. A DR of zero represents no difference between cancer and normal contexts. Figure 4 shows the overall distribution of DRs of all high-confidence m⁶A sites, with a majority of m⁶A sites showing no significant differences between cancer and normal conditions.

Fig. 4.

The difference ratio (DR) was calculated to represent the observed difference of a m⁶A site between cancer and normal conditions. Only a small number of m⁶A sites show differences under these two contexts and can be further classified into positive and negative training datasets

Based on the DR value, the positive (P) and negative (N) datasets for model training and testing were selected by identifying the m⁶A residues most associated with cancer and normal conditions using the Two-tailed test. The pro-cancer m⁶A sites (P) were defined as the top 2.5% of right-sided (cancer-associated) m⁶A sites with a DR > 0.165. Conversely, the normal m⁶A sites (N) were identified as the top 2.5% of left-sided (normal) m⁶A sites with a DR < -0.38. Specifically, a limited number of base-resolution m⁶A sites were selected as positive (cancer-associated, 2,660 sites) and negative (normal, 2,959 sites) datasets. The negative sites were randomly chosen to maintain a 1:1 P-to-N ratio. For performance evaluation, 80% of the dataset was randomly selected as training data, while the remaining 20% was used for independent testing. To test the SNP density around the predicted pro-cancer m⁶A sites, we collected a total of 2,264,915 cancer somatic variants from 33 different human cancer types in the Cancer Genome Atlas (TCGA) database (version v35) [63]. The detailed information of the SNP datasets can be found in Supplementary Sheet S4.

Sequence-derived feature

In our study, five different sequence encoding methods were employed to compose three combinations, including Nucleotide Chemical Property (NCP), position-specific nucleotide propensity (PSNP), Nucleotide Density (ND), Electron–ion interaction potential (EIIP) and pseudo-EIIP (PseEIIP).

The first encoding method NCP categorizes nucleotides into three groups based on their distinct structural chemical properties [64]. Primarily, the number of rings in the nucleotides is considered: adenosine and guanosine contain two rings, while cytidine and uridine consist of only one ring. Secondly, adenosine and cytidine are characterized by the presence of an amino group, whereas guanosine and uridine feature a keto group. Lastly, cytidine and guanosine demonstrate stronger hydrogen bonding than adenosine and uridine. Utilizing these properties, the i-th nucleotide in sequence S can be encoded as a vector Si = (xi, yi, zi):

2

PSNP refers to the variation in the frequency of nucleotides at specific positions in RNA sequences between positive and negative datasets. By calculating the frequency of the appearances of A, C, G, and T at the i-position respectively, two matrices with 4 × 41 dimensions, namely Z_plus and Z_minus, were obtained. Z_plus was derived from the sequence of all positive data, while Z_minus was derived from the sequence of all negative data. In this context, the Position-Specific Nucleotide Propensity (PSNP) matrices were denoted as Z_PSNP:

3

ND represents the distribution and cumulative frequency of nucleotides at each position. The density of the i^th nucleotide is calculated as the number of nucleotides of the same type appearing before the (i + 1)^th position, divided by i. For the sequence 'AGAUUCA', the density of 'A' is 1 (1/1), 0.67 (2/3), and 0.43 (3/7) at the 1st, 3rd, and 7th positions, respectively.

The EIIP values of nucleotides was originally proposed by Nair et al., in 2006 [65]. Specifically, each nucleotide is encoded as a numeric value that represents its electron–ion interaction potential (Supplementary Table S5). Additionally, the pseudo-EIIP (PseEIIP) was calculated by multiplying the sum of the numeric value of tri-nucleotides by their frequency in a given sequence. In the following section, we will explore which combination of encoding strategies yields the best performance for model development.

Genome-derived feature

Genome-derived feature guided by domain characteristics have been encoded as an effective feature type that contributing to the performance of prediction models in classifying modified or unmodified RNA residues [66]. In our work, we try to capture the distinct patterns between pro-cancer and normal m⁶A modification sites. Specifically, 54 domain (genomic) features were extracted for both pro-cancer and normal base-resolution m⁶A sites. These genomic properties including dummy variables (1: overlapped; 0: no overlapped) indicating overlapped regions (such as CDS, 5'UTR), counting of adjacent input site and neighboring A, region length, conservation score (PhastCons v3.7.2 and fitCons v3.7.1), RNA secondary structures predicted by RNAfold package (v2.5.1) [4, 67, 68], and distance to regions’ 5’/3’ ends. The ‘TxDb.Hsapiens.UCSC.hg38.knownGene’ (v3.16.0) annotation file was used to extract the corresponding human genomic regions. Please refer to Supplementary Table S1 for more details about the genomic features considered in the m6A-CAPred model.

Machine-learning approach used for model construction

Support Vector Machine (SVM) has been widely used in previous prediction model with satisfactory performance. We used the R language interface of LIBSVM [69] to build the final m6A-CAPred model, the radial basis function was set as kernel, with other parameters using the default setting.

Performance evaluation

For performance evaluation, we applied the following evaluation metrics. In general, Receiver Operating Characteristic (ROC) curve (sensitivity against 1-specificity) and the area under the ROC curve (AUROC) were used as the primary performance evaluation metrics. In addition, we also calculated sensitivity (Sn), specificity (Sp), Matthew's Correlation Coefficient (MCC), and overall accuracy (ACC) as additional indicators for evaluating the reliability of the model. A fivefold cross-validation was applied on training datasets, while the testing datasets was used for independent testing. Only the m⁶A sites that do not include as training step were selected for independent testing purpose.

4

5

6

7

Among them, TP represents true positive, while TN represents true negative; FP stands for the number of false positive, and FN stands for the number of false negative.

Website construction

m6A-CAPred web interfaces were constructed by using HyperText Markup Language (HTML), Hypertext Preprocessor (PHP), and Cascading Style Sheets (CSS). All metadata was stored using MySQL tables. To present statistical diagrams, EChars was exploited.

Author contributions

Y.X., G.T. and W.Z. initialized the study and supervised the project; Z.C., J.H. and Y.L. presented the computational framework; X.W. and J.M. constructed the web-interface; X.Y. performed the GO and KEGG analysis; Z.C. and G.T. drafted the manuscript; Y.X., G.T. and W.Z. revised the manuscript; All authors have read and agreed to the published version of the manuscript.Among these authors, Z.C.,Y.L and J.M. contributed equally to this work.

Funding

Project of Key research and development Plan of Xuzhou Science and Technology Bureau (KC23278).

Data availability

The datasets used in this study are listed in the Supplementary Sheet S3-S4 and, collected in our online database: www.rnamd.org/m6A-CAPred.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.