Prediction of RNA Methylation Status From Gene Expression Data Using Classification and Regression Methods

Abstract

RNA N⁶-methyladenosine (m⁶A) has emerged as an important epigenetic modification for its role in regulating the stability, structure, processing, and translation of RNA. Instability of m⁶A homeostasis may result in flaws in stem cell regulation, decrease in fertility, and risk of cancer. To this day, experimental detection and quantification of RNA m⁶A modification are still time-consuming and labor-intensive. There is only a limited number of epitranscriptome samples in existing databases, and a matched RNA methylation profile is not often available for a biological problem of interests. As gene expression data are usually readily available for most biological problems, it could be appealing if we can estimate the RNA methylation status from gene expression data using in silico methods. In this study, we explored the possibility of computational prediction of RNA methylation status from gene expression data using classification and regression methods based on mouse RNA methylation data collected from 73 experimental conditions. Elastic Net-regularized Logistic Regression (ENLR), Support Vector Machine (SVM), and Random Forests (RF) were constructed for classification. Both SVM and RF achieved the best performance with the mean area under the curve (AUC) = 0.84 across samples; SVM had a narrower AUC spread. Gene Site Enrichment Analysis was conducted on those sites selected by ENLR as predictors to access the biological significance of the model. Three functional annotation terms were found statistically significant: phosphoprotein, SRC Homology 3 (SH3) domain, and endoplasmic reticulum. All 3 terms were found to be closely related to m⁶A pathway. For regression analysis, Elastic Net was implemented, which yielded a mean Pearson correlation coefficient = 0.68 and a mean Spearman correlation coefficient = 0.64. Our exploratory study suggested that gene expression data could be used to construct predictors for m⁶A methylation status with adequate accuracy. Our work showed for the first time that RNA methylation status may be predicted from the matched gene expression data. This finding may facilitate RNA modification research in various biological contexts when a matched RNA methylation profile is not available, especially in the very early stage of the study.

Introduction

The structure and function of RNA molecules in cells are regulated with more than 100 chemical modifications, but the specific functions of the majority of those modifications are still enigmatic.¹ While most of the RNA nucleotide modifications are believed to be static due to their covalent attachments, m⁶A is the first type of RNA methylation found to be revisable.² Even though m⁶A is the most prevalent internal mRNA decoration found in eukaryotes and RNA of nuclear-replicating viruses, the physiological function and the prevalence of this post-transcriptional RNA modification remain merely partially revealed. It was reported that m⁶A in nuclear RNA functions as a physiological substrate of fat mass and obesity-associated protein.³ The m⁶A sites were found on many clock gene transcripts and believed to be related to circadian period elongation and RNA processing delay.⁴ More recent studies show that m⁶A is crucial in RNA metabolic processes as it regulates the stability, structure, processing, and translation of RNA. Instability of m⁶A homeostasis would result in flaws in stem cell regulation, decrease in fertility, and risk of cancer.⁵

The transcriptome-wide mapping of m⁶A was available in 2012, due to the invention of methylated RNA immunoprecipitation sequencing method (MeRIP-seq or m⁶A-seq).^6,7 In this method, isolated mRNA is fragmented into ∼100 nucleotides with part of the fragments reserved as untreated input control and the rest precipitated by m⁶A-specific antibodies. Then, both input control and immunoprecipitated (IP) samples are reverse-transcripted to cDNA. After amplification, the cDNA is subjected to high-throughput sequencing. Next, by comparing signal enrichment of IP samples with input controls, the location of m⁶A may be identified. The MeRIP-seq plays a pivotal role in the study of m⁶A. Based on this technique, m⁶A was found to be a ubiquitous modification of mRNA discovered in the mRNAs of more than 7600 genes. The m⁶A is enriched around stop codons, within long internal exons and 3′ untranslated regions (3′UTRs). This clustering tenet was consistent in both human and mouse cells, which suggests that m⁶A may be fundamental in regulating gene expression.^6,7 By investigating the whole-transcriptome m⁶A profiles of 8 types of major human tissues using MeRIP-seq, a positive correlation between m⁶As and gene expression homeostasis was observed.⁵ However, the MeRIP-seq process is laborious and time-consuming, and may take up to 9 days to complete.⁸ Furthermore, the data are still prone to various bias and artifacts.⁹ Although there are already hundreds of MeRIP-seq experiments collected in existing RNA modification databases such as RMBase¹⁰ and MeT-DB,¹¹ they still cover very limited tissue contexts, and a matched RNA methylation profile is usually not available for an arbitrary biological condition.

The biological system is a coordinated system. There exists a significant correlation between different types of genetic and epigenetic features that work in harmony to achieve various biological functions. The idea and feasibility of computational predicting omic features have been well documented in previous works,¹² eg, Whitaker et al¹³ predicted the entire human epigenome from DNA motifs via the Epigram pipeline. The possibility of predicting DNA methylome was explored using various approaches.^13-20 Recently, Nath et al²¹ predicted the long non-coding RNA transcriptome with protein-coding genes. In the field of epitranscriptome bioinformatics, although there exists the potential to train the transcriptome-wide prediction model to predict the entire epitranscriptome, most works focused on RNA modification site prediction from a single DNA or RNA sequence.^22,23 Please refer to a comprehensive review.²⁴ Recently, Chen et al²⁵ developed the WHISTLE method and constructed so far the most accurate m⁶A epitranscriptome from sequence and genomic features; however, similar to other approaches, the WHISTLE method provided only a general m⁶A epitranscriptome without considering its dynamics (or condition-specificity). Machine learning predictors were developed for better characterization of the single-based m⁶A profile on sub-regions of mRNA, such as the LITHOPHONE²⁶ and WITMSG²⁷ for the prediction of intronic and lncRNA m⁶A sites. Prediction modeling may also be applied to the functional annotation of m⁶A²⁸ and other types of internal mRNA modifications such as pseudo-uridine²⁹ and m⁷G.³⁰ The deep-m⁶A approach developed by Zhang et al may perform condition-specific quantification of the m⁶A epitranscriptome at base resolution. However, it requires matched epitranscriptome sequencing data (MeRIP-seq), which may not be available in most biological contexts of interests, thus limiting its usage.³¹ Compared with the RNA methylation data, gene expression data are more abundant in public databases and are much easier to obtain. Hence, it is highly desirable to develop in silico approaches to predict the condition-specific RNA methylation status from matched gene expression data.

In this article, we sought to computationally predict the RNA methylation status from gene expression data. We first differentiated the methylated and unmethylated RNA sites using classification methods, then estimated the methylation level using regression methods. Our results suggested that gene expression data can be used to construct predictors of RNA methylation status, which provides a new and easier venue for the pilot studies of RNA methylation under various biological contexts.

Materials and Methods

RNA methylation and gene expression data

MeRIP-seq data of 73 IP samples from different types of mouse cells paired with their matched input controls were collected for this study (Supplement S1). Among them, 58 samples were used for training, and the remaining 15 samples served as the testing set. It is worth noting that the input control of MeRIP-seq data is essentially RNA-seq, which corresponds to gene expression data. For the candidate m⁶A RNA methylation sites, we considered the 102 024 m⁶A sites reported by base-resolution techniques (m⁶A-CLIP and miCLIP) that were collected from the WHISTLE project.²⁵ Reads of MeRIP-seq data in IP samples and input samples are both quantified in terms of reads per million reads mapped (RPKM). Furthermore, we used the M-value, ie, log₂ ratio of reads in IP to reads in input,³² to determine the status of RNA methylation:

$$\mathrm{M}={\log }_2\left(\frac{{\mathrm{RPKM}}_{IP}+0.1}{{\mathrm{RPKM}}_{input}+0.1}\right).$$

For classification analysis, the methylation status of a site is considered positive when the corresponding M-value is greater than 0; otherwise, it is regarded as negative. For regression analysis, we seek to directly predict the absolute amount of methylation from matched gene expression data after ${\log }_2$ transformation. An independent pair of input and IP data obtained from human cells under different experimental conditions (31 samples, 69 433 sites, see Supplement S2) were also tested to evaluate the generalizability of our predicting scheme on human data.

Furthermore, as a contrast to models based on expression level, we also trained Elastic Net-regularized Logistic Regression (ENLR), Support Vector Machine (SVM), and Random Forests (RF) by incorporating sequence-based features, including the presence of purine, amino group and weak hydrogen bonds, and cumulative frequency of nucleotide with window width 41 bp centered by m⁶A.²⁵

The 100 sites with the greatest variation in m⁶A modification level were selected for classification and regression analysis, because these sites were most dynamic and were potentially most responsive to various stimulus and more crucial when studying the context-specificity of m⁶A RNA methylation. For each target site, the gene expression profile of the corresponding site and its 1000 neighbor sites was selected as our predictive features to construct classifiers and regressors. Because of having a lot more features than samples, the prediction model would result in a typical “large p small n” paradigm³³; therefore, both the regression and classification analysis required sparse estimators to avoid overfitting.

Classification and regression methods

For classification analysis, ie, to differentiate methylated m⁶A sites and unmethylated m⁶A sites, we used ENLR, SVM, and RF for classification, as they were reported to be promising classifiers in previous DNA methylation predicting tasks.^13,14,19 The ENLR model minimizes an objective function consisting of negative log-likelihood of logistic regression along with both $l_1$- and $l_2$-penalty to obtain a sparse generalized linear model by shrinking the coefficients of less-informative sites into zeros.³⁴ There are 2 parameters to be learned, one is the overall weight of the penalty and the other is the weight between $l_1$- and $l_2$-penalty. SVM maps features of interest into a higher dimensional space via kernel functions and generates a decision hypersurface that maximizes the margin between examples from different categories.³⁵ We chose radial basis function as our kernel and tuned the inverse kernel width for the radial basis kernel function as well as the cost regularization parameter which controls the smoothness of the fitted function. RF is an ensemble learning algorithm that generates a specified number of decision trees and cast predictions based on the majority of the votes from an individual tree.³⁶ For RF, we optimized the number of variables randomly sampled as candidates at each split.

For regression analysis, ie, to estimate directly the absolute RNA methylation level of m⁶A site, we considered the Elastic Net (EN) family with objective function:

$$\begin{array}{c}F\left(\beta \right)=\frac{1}{2s}||y-{\beta }^{t}X{||}_2^2+\frac{\lambda \left(1-\alpha \right)}{2}||\beta {||}_2^2+\lambda \alpha ||\beta {||}_1,\#,\end{array}$$

where s is the sample size, y is the response, β is the coefficient to be estimated, X is the covariate, ${||\cdot ||}_1$ is the $l_1$-norm and ${||\cdot ||}_2$ is the $l_2$-norm. Note that when $\alpha =1$, this model reduces to lasso; when $\alpha =0$, this model reduces to ridge regression. All analysis was implemented in R. The scripts and processed data used in this project are publicly downloadable on GitHub (https://github.com/xvehao/m6Aprediction).

Results and Discussions

Predict condition-specific m⁶A sites with classification analysis

For each site, we constructed ENLR, SVM, and RF models as classifiers, ran a 5-fold cross-validation to tune the parameters of each model with the caret R package,³⁷ and then tested their performance on the testing samples. To evaluate the performance of classifiers, we generated the receiver operating characteristic (ROC) curve (true positive rate vs false positive rate under different threshold) and computed the area under the ROC curve (AUC) as the metric. As shown in Figure 1A, positive classification results were achieved from all the 3 classifiers tested with their AUCs substantially higher than the classifiers based on sequence features (highest median AUC = 0.54 by ENLR, see Supplement S3), which suggested that it is indeed possible to predict context-specific m⁶A RNA methylation sites from matched gene expression data. The medians of AUC of 3 methods across test samples were very close, but the result of SVM had a narrower AUC spread and the highest overall accuracy.

Figure 1.

Model performance on the mouse test set: (A) All 3 classifiers have a median AUC over 0.8 in the identification of condition-specific m⁶A sites. Among them, SVM achieved the best performance as it has the highest AUC and the narrowest AUC spread. (B) Both median PCC and SCC between the predicted value by ENLR and the actual value are above 0.7, which implies a strong linear and monotonic relation between the predicted and actual values. AUC indicates area under ROC; ENLR, Elastic Net-regularized Logistic Regression; PCC, Pearson correlation coefficient; RF, Random Forests; SCC, Spearman correlation coefficient; SVM, Support Vector Machine.

To further investigate the expression level of which kinds of genes were determinant in predicting m⁶A status, we examined the biological meaning of ENLR model because ENLR can perform feature selection automatically due to its sparse property. We performed gene site enrichment analysis, with DAVID,³⁸ using those sites ever selected as predictors by ENLR as input and all sites as background. Three functional annotation terms were found to be statistically significant with Family Wise Error Rate less than 0.05: phosphoprotein, SRC Homology 3 (SH3) domain, and Endoplasmic Reticulum (ER). This result further attached potential biochemical significance to ENLR model as all 3 terms were found to be closely related to the m⁶A pathway in literature.

The first 2 terms may refer to the formation of a methyltransferases (MTases) complex comprising MTase-like 3 (Mettl3) and MTase-like 14 (Mettl14), called Mettl3-Mettl14 complex. This complex efficiently catalyzes methyl group transfer by using Mettl3 as catalytic core and Mettl14 as an RNA-binding platform. Moreover, Mettl3-Mettl14 complex exhibits a much higher catalytic activity than either Mettl3 or Mettl14 alone in vitro. The Mettl3 MTase domain and the Mettl14 MTase domain are connected by the N-terminal α-helical motif (NHM) and by the C-terminal motif (CTM) with a phosphoserine via a salt bridge. Phosphorylation therefore plays an important regulatory role in MTase binding as it ensures the 2 MTases are connected tightly, and their extensive interaction network is difficult to disrupt.³⁹ SH3 domain is important in salt bridge formation between the conserved acidic residue in the SH3 domain and the favored arginine residue, either N-terminal or C-terminal to the Pro-X-X-Pro motif,⁴⁰ which is exactly the structure of Mettl3-Mettl14 complex. Moreover, the change in ER also alters the m⁶A modification. ER stress responses are found to contribute to differential m⁶A modification,⁴¹ and SIALKBH2, an active m⁶A demethylase, is found to be located in the ER.⁴² After analyzing the mechanism of ENLR model, we conjecture that the gene expression level of sites regulating phosphoprotein, SH3 domain, and ER might influence the formation of Mettl3-Mettl14 complex and demethylation of m⁶A in ER, and therefore are associated with the m⁶A level.

Predict condition-specific m⁶A level with regression analysis

In the regression analysis, we ran a 5-fold cross-validation to tune the associated parameters with mean squared error (MSE) as the metric. Furthermore, we computed both Pearson correlation coefficient (PCC) and Spearman correlation coefficient (SCC) between predicted methylation value and the MeRIP-seq data to evaluate the performance of our regressor (Figure 1B). While PCC measured the linearity between 2 data sets, SCC accessed the monotonic relationship between them. Both PCC and SCC were within −1 to 1, and a higher absolute value of PCC or SCC implied higher concordance between 2 data sets. The median of both PCC and SCC across different samples was above 0.7, suggesting a strong linear relationship between the predicted value and the actual value, while the median PCC and SCC obtained by model based on sequence-based features were respectively, 0.49 and 0.43 (Supplement S3). We identified 1 outlier in the boxplot with PCC −0.1 and SCC −0.15, which corresponds to Mettl3 knocked-out mouse embryonic stem cell. The abnormal performance of our regressor on this particular sample was likely to be caused by the removal of Mettl3, which was reported to lead to near-complete depletion of m⁶A on mRNA,⁴³ thus rendering the association between gene expression level and m⁶A methylation level completely different from that in other types of cells.

Additional investigation on human data

To further investigate whether our predicting scheme could be generalized to different species, we tested our classifiers with independent input and IP data obtained from different types of human cells under heterogeneous experimental conditions (26 samples as training set, 5 as testing). The AUC of each model on test set was shown by Table 1, from which we can see that the gene expression data can serve as covariates for effectively predicting m⁶A methylation status in human cells as well, with SVM being the best classifier.

Table 1.

Performance of classifiers (AUC) on the human test set.

Sample	Classification method
	ENLR	SVM	RF
1	0.90	0.91	0.91
2	0.82	0.86	0.89
3	0.82	0.88	0.87
4	0.85	0.91	0.76
5	0.77	0.80	0.71
Average	0.83	0.87	0.83

Abbreviations: AUC, area under ROC curve, ROC, receiver operating characteristic; ENLR, Elastic Net-regularized Logistic Regression; RF, Random Forests; SVM, Support Vector Machine.

All 3 classifiers have an average AUC above 0.8. SVM achieved the best performance in the identification of condition-specific m⁶A sites.

The bold values in the table highlight the best performance achieved in each sample.

Conclusions

We implemented both regression and classification to predict RNA m⁶A methylation from gene expression data. With the positive results, we showed, for the first time, that condition-specific m⁶A methylation status may be predicted from gene expression data. Gene Site Enrichment Analysis on important sites by ENLR further suggested that those sites related to phosphoprotein, SH3 domain and ER be determinant in m⁶A methylation status prediction. Therefore, we recommend that features related to those 3 functional terms should be considered in regressors or classifiers in future m⁶A predicting study. We are optimistic that the accuracy may be further improved with efforts.

Nevertheless, our work suffers from the 3 main limitations. First, as the data samples were collected from different public data sets provided by different laboratories under heterogeneous experimental conditions, the systematic error arising in each experiment is unavoidable. Different experiment reagents and protocols (such as polyA selection and ribo-minus in RNA library construction) may lead to a substantial difference between samples. As a consequence, the accuracy of prediction would be undermined. The prediction models could be trained later using the technical independent quantification and the inference methods on MeRIP-Seq. Second, as MeRIP-Seq data are prohibitively difficult to obtain, it is possible that the limited samples in hand may not be representative of all mouse/human cells. Therefore, larger-scale MeRIP-seq data from different tissues and developmental stages are needed for the development of prediction tools with higher accuracy in future studies. Third, while we considered in our prediction only the gene expression data in the form of RNA-seq, it is reasonable to assume that the accuracy may be further improved with other matched omic data, such as proteomic data from LC/MS or DNA methylation data.