Detecting N6-methyladenosine sites from RNA transcriptomes using random forest

Highlights

• N6-methyladenosine (m6A) modifications are one the most frequently occurring RNA post transcriptional modifications. These modifications perform vital roles in different biological processes, including, localization and translation of proteins, X chromosome inactivation, cell stability, microRNA regulation and reprogramming etc.

• An abnormal change in m6A may lead to certain abnormalities, including, cancer, brain related disorders etc. Precise detection of m6A modifications is crucial for the diagnosis and treatment of these diseases.

• Existing methods suffer from the problem of inefficient detection of m6A sites, especially in varied structures of yeast transcriptomes.

• The proposed m6A-pred predictor utilizes a fusion of features including, statistical, and chemical properties of the nucleotides, to precisely predict the presence of m6A sites in RNA sequences.

• The m6A-pred predictor outperforms all the previously reported predictors in terms of accuracy, specificity and Mathew correlation coefficient values.

Abstract

N6-methyladenosine (m6A) modifications are one the most frequently occurring RNA post transcriptional modifications. These modifications perform vital roles in different biological processes, including, localization and translation of proteins, X chromosome inactivation, cell stability, microRNA regulation, and reprogramming etc. Any abnormal change in m6A sites may lead to several abnormalities, including, cancer, brain-related disorders and many other life threatening diseases. Precise detection of m6A modifications is crucial for the diagnosis and treatment of these diseases. Existing methods suffer from the problem of inefficient detection of m6A sites, especially in yeast transcriptomes (due to varied structure) and inability of the computational techniques to capture the encoded information surrounding the m6A sites. In this work, we propose a novel method (called m6A-pred predictor) that utilizes a fusion of characteristics including, statistical, and chemical properties of the nucleotides, to precisely predict the presence of m6A sites in RNA sequences. The fusion of multiple types of features results in a high dimensional vector which is further optimized using an evolutionary algorithm. Finally, the random forest classifier is used to detect m6A sites by using the most discriminative features. The results, benchmarked on yeast transcriptomes, indicate that m6A-pred predictor outperforms all the previously reported predictors, notably, with an accuracy value of 78.58%, specificity value of 79.65% and Matthews correlation coefficient of 0.5717.

Introduction

Among the 150 types of post transcription modifications of cellular RNA, the N6-methyladenosine (denoted as m6A) is the most frequently occurring modification [1]. These modifications are catalyzed by N6-adenosyl methyltransferase complexes including METTL3, METTL14 and WTAP and some others., [2], [3]. The m6A sites are present in both the prokaryotes and eukaryotes, [2]. In a recent study, it is found that m6A modifications control different types of biological processes like localization and translation of proteins and some other essential cellular tasks [4]. Any abnormal change in m6A may lead to certain abnormalities, including, cancer, brain related disorders, and a lot of other diseases [5], [6]. Furthermore, it is also found that m6A is non-randomly distributed across genomes [7]. Thus, identification of m6A sites is important in order to understand many key biological functions.

Traditionally, wet lab experiments are used for providing single-nucleotide resolution map of the m6A sites across human transcriptomes. However, the results of these resolution maps for other species m6A sites are not satisfactory because of false positive and false negative detections [8]. In addition, genome-wide m6A site detection through wet experiments is laborious and costly. Thus, we require computational tools to precisely detect the m6A sites. These computational tools will find m6A sites as well as save experimental time and costs. The benefits of these tools will be a direction for future research in bioinformatics and drug discovery for severe diseases like cancer, brain disorder and other abnormalities. Above aforementioned benefits motivated us to make a predictor for classification of m6A sites.

Due to efficient high-throughput technologies, the m6A genome-wide distributions are existing for different types of species like Homo sapiens, Arabidopsis thaliana, Saccharomyces cerevisiae and likewise other species [7], [9], [9], [10]. Researchers got unprecedented opportunities due to the availability of large volume experimental data. These type of data provide the feasibility for developing computational tools in order to correctly predict m6A sites. Different type of computational methods were proposed according to the nature of data for the identification of m6A sites. For example, in [11], [12] authors proposed two computational tools for yeast specific data. In the first method, the authors used nucleotide chemical property to encode features, while in the second method the authors use pseudo nucleotide composition (PseNC) for encoding the RNA sequence. The first one has low sensitivity on independent dataset, while the second one has low specificity.

Another group of researchers, developed a predictor named pRNAm-PC by using chemical property based auto-covariance and auto cross-variance features in pseDNC [13]. The proposed technique has two problems, first, the feature vector dimension is very high and secondly, the accuracy they reported is considerably low. Another predictor tool namely, SRAMP was proposed for mammalian m6A sites classification [14]. Their work was inspired from the works of [11], [12] which primarily uses binary encoding scheme for feature extraction. The specificity of the proposed work is good but sensitivity is very low.

Subsequently, some researchers in [15] developed another predictor MethyRNA for the prediction of m6A site in Homo sapiens and Mus musculus datasets. While the performances of these proposed techniques are good for finding m6A sites in mammalian transcriptomes, but their results are not good to identify m6A sites accurately in yeast transcriptome [14], [15]. The use of these features result in inefficient detection of m6A sites in yeast transcriptomes due to different structure from other species and inability of the computational techniques to capture the encoded information surrounding the m6A sites. Moreover, the information around yeast m6A sites are not well studied [14].

Some authors introduced heuristic based nucleotide physical-chemical property selection algorithm for m6A sites identification. The results of this technique are good from previous techniques in term of sensitivity but specificity is still low [16]. Recently, further improvements were carry out by Chen et al. [17], using ensemble support vector machines. In their work, they used PseKNC (pseudo K-tuple Nucleotide Composition), along with motif and gapped k-mer based techniques for feature extraction. The final output is generated by using majority voting strategy. The accuracy is good but time complexity is high due to high dimensional feature vectors as well ensemble classifiers.

In [18], the authors introduced a new predictor named imethyl-STTNC. They used split trinucleotide composition and split tetranucleotide composition features for RNA sequence encoding. Multiple classifiers were used for the detection of m6A sites in RNA sequences. The support vector machine achieved high performance on these features. The reported accuracy is good for Homo sapiens dataset but not satisfactory for yeast dataset. Recently, in [19], a new predictor called BERMP has been introduced for classification of m6A sites of multiple species. The proposed predictor’s accuracy is good for mammalian and plant species. In this work, the authors used ENAC (Enhanced Nucleic Acid Composition) as feature encoding scheme. They used deep learning and Random forest based ensemble classifier for prediction, while the prediction accuracy still did not improve for yeast species RNA sequences.

Most recently, M6AMRFS was introduced for the prediction m6A sites in multiple species namely, Saccharomyces cerevisiae, Homo sapiens, Musculas and Arabidopsis Thaliana, [20]. The authors used local position specific dinucleotide frequency composition and binary encoding to encode features. They used Extreme Gradient boost (XG boost) classifier. However, its performance in identification of m6A sites in yeast transcriptome remained average, and further enhancements are still needed.

A predictor named DeepM6ASeq was proposed in [21] for classification m6A sites. They used CNN for the detection of m6A sites. But the predictor is only evaluated for mammalian data. In recent time, another predictor WHISTLE, [22] was developed. The authors extracted sequence derived features and genome derived features from human mature messenger RNA sequences and full transcript sequences. They did not test their predictor on yeast data. Mostly recently, a predictor SICM6A [23] was developed for cross species m6A data classification. However, the accuracy of the mentioned predictor is good for all species, but in yeast transcriptomes, its results are nearly equal to [19].

From the extensive literature review above, we conclude that, the feature representation techniques used, are not satisfactory for the identification of m6A sites in yeast species. Moreover, apart from statistical and chemical properties of nucleotides in transcriptomes, another aspect that is not widely explored is the fusion of these characteristics surrounding the m6A sites. In this work, we propose a novel method (called m6A-pred predictor) that utilizes a fusion of characteristics including, statistical, and chemical properties of the nucleotides, to precisely predict the presence of m6A sites in RNA sequences. The proposed method will overcome the problem of inefficient detection of m6A sites in yeast transcriptomes (due to varied structure) and inability of the computational techniques to capture the encoded information surrounding the m6A sites. The extracted hybrid features are usually high dimensional. To reduce feature dimensionality, we also explore the feature importance through feature selection methods.

The main contributions of this study are as follow:

• Combination of chemical properties based features and statistical based features are extracted from RNA to capture sounding information of m6A sites in more efficient way.

• Wrapper based feature selection method is used to select important features and remove redundant and unnecessary features.

The remaining part of the paper is organized as follows: in Section 2, we detail the materials and method used to develop the proposed m6A-pred predictor. We benchmark our results in Section 3, on the intriguing RNA transcripts of Saccharomyces cerevisiae species. Lastly, we conclude the study in Section 4.