Comprehensive comparative analysis and identification of RNA-binding protein domains: multi-class classification and feature selection

Abstract

RNA-protein interaction plays an important role in various cellular processes, such as protein synthesis, gene regulation, post-transcriptional gene regulation, alternative splicing, and infections by RNA viruses. In this study, using Gene Ontology Annotated (GOA) and Structural Classification of Proteins (SCOP) databases an automatic procedure was designed to capture structurally solved RNA-binding protein domains in different subclasses. Subsequently, we applied tuned multi-class SVM (TMCSVM), Random Forest (RF), and multi-class ℓ₁/ℓ_q-regularized logistic regression (MCRLR) for analysis and classifying RNA-binding protein domains based on a comprehensive set of sequence and structural features. In this study, we compared prediction accuracy of three different state-of-the-art predictor methods. From our results, TMCSVM outperforms the other methods and suggests the potential of TMCSVM as a useful tool for facilitating the multi-class prediction of RNA-binding protein domains. On the other hand, MCRLR by elucidating importance of features for their contribution in predictive accuracy of RNA-binding protein domains subclasses, helps us to provide some biological insights into the roles of sequences and structures in protein–RNA interactions.

1. Introduction

Regulation of biological processes happens through association and dissociation of macromolecules, i.e., protein, RNA and DNA. Furthermore, functional components of cells are frequently complex assemblies of macromolecules. At the molecular level, RNA-protein complexes play an important role in various cellular processes, such as protein synthesis, gene regulation, post-transcriptional gene regulation, alternative splicing, and infections by RNA viruses. Therefore, it is important to understand the principle of RNA-protein interactions and prediction of RNA-binding proteins is essential in identifying the cellular processes in which RNA-protein complexes are involved.

It is commonly accepted that RNA recognition by proteins is mainly mediated by specific kinds of RNA-binding domains (RBDs) (Morozova et al., 2006; Shulman-Peleg et al., 2008). The RBDs can be classified into different subclasses based on their basic binding motifs, e.g., the KH domain, the double-stranded RNA-binding domain (dsRBD), and the zinc finger motif (Chen and Varani, 2005). Although in recent years, new RBDs have been identified (Parker and Barford, 2006), an increasing amount of evidence on non-coding RNAs suggest that new RBDs will be identified (Lingel and Sattler, 2005).

In order to recognize the RNA functional importance in close relationship with protein in its activities, computational studies of RNA-protein complexes have been significantly increased (Ellis et al., 2007; Jones et al., 2001). Recently, a variety of approaches have been proposed to study RNA-protein interactions (Lunde et al., 2007). Although some interesting results have been obtained, the precise details of the RNA-protein interaction are far from being fully understood. For this reason, it is strongly recommended to develop reliable computational methods to accurately predict RNA-binding proteins and analyze important features in RNA-protein interaction.

Homology-based methods are the most common method to identify the class of unknown proteins at sequence or structure level. These methods are limited by the absence of experimentally annotated homologous proteins in protein databases. Hence it is strongly encouraged to develop computational tools to identify RNA-binding proteins (RBPs) using sequence- and structure-derived features. Most of previous investigations, predict RBPs using sequence- derived features (Han et al., 2004; Shao et al., 2009; Yu et al., 2006). In addition to sequence-based methods, up to now, only one investigation by Shazman and Mandel-Gutfreund (2008) developed a structural-based method to predict RBPs. Shazman and Mandel-Gutfreund developed a multiSVM-based method using four subgroups of features including: (i) Largest patch parameters (such as patch size and patch surface accessibility), (ii) Protein parameters (such as molecular weight) (iii) Cleft/patch parameters (such as the overlap between the largest, second largest, and third largest clefts, and largest patch), and (iv) Parameters related to other surface patches (such as number of residues in the lysine out patch and in the negative patch), to describe the global composition of each protein. Using the jackknife test, they reported a 75.61% accuracy of prediction for three subclasses of RBPs; tRNA-, rRNA-, and mRNA-binding proteins. In comparison with our work, it is limit to three classes of RBPs, and they have done a non-accurate manually data collection. Despite the availability of several methods, identification of RBPs using sequence information with high accuracy is still a major challenge.

Here we present a comprehensive performance evaluation of some state of the art predictor methods on an important problem, i.e., classifying RBDs using sequence- and structure-derived information. Combining a diverse set of features, we developed three different methods including; tuned multi-class SVM (TMCSVM), Random Forest (RF), and Multi-class ℓ₁/ℓ_q -regularized logistic regression (MCRLR). By applying these methods, we have shown that we can classify RBDs based on their RNA target (7S, double-stranded, tRNA, rRNA, or mRNA). In all of five different subclasses of RBPs, no exclusive RNA-binding motif is present. However, in such cases in addition to successful classifying RBPs, we discovered dissimilar sequence and structural features.

2. Materials and methods

2.1. Automatic dataset harvesting

Based on the fact that most of similar works on prediction of RNA-binding proteins, manually collected and annotated datasets, in this work, in order to do a more accurate and automated data harvesting we constructed a dataset of non-redundant RNA-binding protein domains using two main datasets including: (i) Gene Ontology Annotated (GOA) database, available at http://www.ebi.ac.uk/GOA/, which cover ~2.5 million reports of associated protein chains with Gene Ontology (GO) terms, and (ii) 40% non-redundant set of Structural Classification of Proteins (SCOP) 1.75 from ASTRAL website. Based on GO classification, RNA binding root involves 28 leaves. Our first step of automatic procedure was one by one search for RNA binding subclasses GO IDs in GOA database to find associated protein chains to each subclass of RNA binding GO IDs. Briefly, GO is a major bioinformatics tool for the unification of biology. More specifically, one of the aims is annotation of genes and gene products. GO contains three ontologies that describe the molecular functions, biological processes, and cellular components of proteins (Ashburner et al., 2000). For more details and comprehensive discussion we refer to the paper (Chou and Shen, 2006), as well as the discussions as elaborated in (Chou and Shen, 2008). The second step was search across SCOP 1.75 to capture non-redundant RNA binding protein domains in different subclasses (Fig. 1). We eliminated protein domains, which associated to more than one RNA binding subclass.

Figure 1.

The proposed automatic procedure for dataset harvesting

2.2 Feature generation

In this study a combination of sequence- and structure-derived features were used for prediction of RNA-binding protein domains. Our representation of the protein sequence in this study is a general form of Chou’s pseudo amino acid composition (Chou, 2011). Indeed, to avoid losing many important information hidden in protein sequences, the pseudo amino acid composition (PseAAC) was proposed (Chou, 2001; Chou, 2005) to replace the simple amino acid composition (AAC) for representing the sample of a protein. For a summary about its recent development and applications, see a comprehensive review (Chou, 2009). Ever since the concept of PseAAC was proposes by Chou in 2001, its has rapidly penetrated into almost all the fields of protein attribute prediction (Chen et al., 2009; Ding et al., 2009; Esmaeili et al., 2010; Georgiou et al., 2009; Guo et al., 2011; Hayat and Khan, 2012; Hu et al., 2011; Li et al., 2012; Lin, 2008; Liu et al., 2012; Mei, 2012; Mohabatkar, 2010; Mohabatkar et al., 2011; Nanni et al., 2012; Qin et al., 2012; Qiu et al., 2009; Qiu et al., 2011; Yu et al., 2010; Zhang and Fang, 2008; Zhao et al., 2012; Zou et al., 2011). According to Eq.6 of a recent comprehensive review (Chou, 2011), the form of PseAAC can be generated and formulated as

$$\mathbf{P}={[{\psi }_1{\psi }_2\dots {\psi }_{\text{u}}\dots {\psi }_{\mathrm{\Omega }}]}^{\mathbf{T}}$$ (1)

where T is a transpose operator, while the subscribe Ω is an integer and its value as well as the components ψ1, ψ2, … will be defined by a series of feature extractions as elaborated below.

In addition to sequence-derived features, structure-derived features were generated in this study. Totally, 267 different sequence- and structure-derived features were generated using several information sources, which can be classified into six major subgroups including:

1. Sequence-derived features including: (i) composition of all 20 amino acids (20 features), (ii) composition of amino acids in 9 different physicochemical groups including tiny, small, aliphatic, aromatic, polar, non-polar, charged, acidic, and basic amino acids groups (9 features), (iii) pI, the isoelectric point (1 feature), (iv) molecular weight (1 feature), and (v) number of residues and number of atoms (2 features). This subgroup of features was generated using seqinr package (version 3.0–3) in R environment.

2. Secondary structure features including: (i) composition of all 20 amino acids and composition of amino acids in physicochemical groups, within three different secondary structures, i.e., helix, sheet and random coil (87 features), and (ii) composition of 6 different secondary structures, i.e., H (α-helix), G (3₁₀ helix), E (extended β-strand), B (isolated β-bridge), T (turn), and S (bend) (6 features). Secondary structure parameters in each protein domain were computed using the output of the program DSSP (Kabsch and Sander, 1983). In order to calculate secondary structures structures in three different secondary structures, the six structures were reduced into three classes (H,G → H, E → E, all other states to C).

3. Solvent accessibility features including: composition of all 20 amino acids and composition of amino acids in physiochemical groups, within three different solvent accessibility states, i.e., buried, intermediate, and exposed (87 features). Based on the standard ranges of solvent accessibility values (SAV), three kinds of solvent accessibility states are defined. Buried state, B, is endowed to residues having 0≤SAV≤0.16, intermediate state, I, to residues having 0.16<SAV≤0.36, and exposed state, E, to residues having 0.36<SAV≤1. Solvent accessibility values of residues were computed using ASAView program (Ahmad et al., 2004).

4. Hydrogen bonds features: The hydrogen bond from the backbone CO (i) to the backbone NH (i+N), is expressed by the symbol H-bond (i, i+N). In this study we computed frequencies of H-bond (i, i+N) for N=-5,-4,-3, …, 3, 4, 5. Furthermore, total hydrogen bonds, parallel- and antiparallel hydrogen bonds were computed. The values of these features were divided by length of protein domains. The output of the program DSSP (Kabsch and Sander, 1993) was used to generate these features (13 features).

5. Electrostatic properties features: eight electrostatic properties features including net molecular charge, net molecular charge per atom, overall molecular dipole moment in debyes, net molecular dipole moment per atom, number of positively charged residues, and number of negatively charged residues were calculated using the Protein Dipole Moments Server (http://bip.weizmann.ac.il/dipol/).

6. Patch features: main, second and third patch sizes, main patch’s molecular weight, composition of all 20 amino acids and composition of amino acids in physico-chemical groups, within the main patch were calculated (33 features). In order to extract all continuous positive patches on the proteins surface the PatchFinder algorithm (Stawiski et al., 2003) was used. The patches were sorted based on the number of grid points contained within the patch, and the largest three patches were selected.

2.3. Predictor methods

In this study, we used three different predictor methods including tuned multi-class SVM (TMCSVM), Random Forest (RF), and multi-class regularized logistic regression (MCRLR) to classify RBDs to three and five subclasses. The jackknife test was used to training and testing on databases. Through the jackknife test, one case is removed from the database and training is done using the remaining cases; then testing is done using the removed case. This procedure is repeated until all cases are tested. Although this method is time-consuming, it is more useful for the small databases such as ours. In addition to jackknife we also used self-consistency test to evaluate the prediction results. Both of jackknife and self-consistency are thought to be the most rigorous and objective methods for evaluation of prediction.

Among the independent dataset test, sub-sampling (e.g., 5 or 10-fold cross-validation) test, and jackknife test, which are often used for examining the accuracy of a statistical prediction method (Chou and Zhang, 1995), the jackknife test was deemed the least arbitrary that can always yield a unique result for a given benchmark dataset, as elucidated in (Chou and Shen, 2008) and demonstrated by Eqs.28–32 of (Chou, 2011). Therefore, the jackknife test has been widely recognized and increasingly used by investigators to test the power of various prediction methods (see, e.g., (Chen et al., 2009; Chou et al., 2011; Chou et al., 2012; Ding et al., 2009; Esmaeili et al., 2010; Georgiou et al., 2009; Gu et al., 2010; Jiang et al., 2008; Lin, 2008; Li and Li, 2008; Lin et al., 2008; Lin and Wang, 2011; Li et al., 2012; Mei, 2012; Mohabatkar, 2010; Mohabatkar et al., 2011; Qiu et al., 2010; Wu et al., 2011; Wu et al., 2012; Xiao et al., 2011a; Xiao et al., 2011b; Xiao et al., 2012; Yu et al., 2010; Zeng et al., 2009; Zhang and Fang, 2008; Zhang et al., 2008; Zhou et al., 2007).

2.3.1. Tuned multi-class support vector machine

Basically, support vector machine (SVM) is a kind of learning machines based on statistical learning theory. They have three remarkable characteristics including: the absence of minima, the sparseness of the solution, and implementation using the kernel Adatron algorithm. The kernel Adatron maps inputs to a high-dimensional feature space, and then optimally separates data into their respective classes by isolating those inputs which fall close to the data boundaries. Therefore, the kernel Adatron is especially effective in separating sets of data which share complex boundaries. Because of seeking a global optimized solution and avoiding over-fitting in the SVM training process, dealing with a large number of features is possible. SVMs can only be used for classification, not for function approximation. The theory and algorithms of SVMs can be found in Vapnik (1995, 1998).

In this study, we applied the tune function using e1071-package of R environment (version 2.11-1) to develop our multi-class SVM based method. Multi-class SVM in e1071 uses the “one-against-one” strategy, i.e., binary classification between all pairs, followed by voting. On the other hand, the tune function uses Grid Search to find the best functions. Using the tune function through jackknife procedure, it provides as many simulations as the number of cases in databases to select optimum structure each time.

2.3.2. Random Forest

Random forest (RF) was developed by Breiman in 2001 (Vapnik, 1998). The RF classification extends the concept of decision trees and has been successfully used in various biological problems (Dudoit et al., 2002; Statnikov et al., 2008; Jia and Hu, 2011; Kandaswamy et al., 2011; Lin et al., 2011; Pugalenthi et al., 2012; Qiu and Wang, 2011; Shameer et al., 2011). RF is a collection of decision trees instead of one tree, where each tree is trained using a bootstrap sample from the training dataset. The trees are then grown using a randomly selected subset of predictors at each node. After constructing all trees, a new object can then be classified based on the class label with the most votes, where every vote is decided by every tree in the forest. Finally, predictive performance is estimated using the observations left out of the bootstrap sample, termed the out-of-bag (OOB) observations. An appeal of RF is that the forest of trees contains a large amount of information about the relationship between the variables and observations. This information can be used for prediction, clustering, imputing missing data, and detecting outliers. The RF algorithm was implemented by the randomForest (version 4.6-2) R package (Liaw, 2002). We used tune randomForest (tuneRF) function. The number of trees and stepFactor were set to 1000 and 2, respectively. However, there are default values for different features, which are provided by the program and we used in this work.

2.3.3. Multi-class ℓ₁/ℓ_q-regularized logistic regression

A multi-class ℓ1/ℓq-regularized logistic regression model that we used in this study is a generalization of the ℓ1-regularization logistic regression. Development of such strong theoretical guarantees, and great empirical success method is from recent studies in areas such as machine learning, statistics, and applied mathematics (Bach, 2008; Duchi and Singer, 2009; Kowalski, 2009; Negahban et al., 2009; Yuan and Lin, 2006).

The multi-class ℓ1/ℓq-regularized logistic regression is an expression of the form

$$\underset{x}{\mathit{min}}{\sum }_{l=1}^k{\sum }_{i=1}^m{w}_{i\ell }log(1+exp(-y_{i\ell }(x_{\ell }^T{a}_{i\ell }+c_{\ell })))+\lambda {\left|\right|x\left|\right|}_{{\ell }_1/{\ell }_{\text{q}}}$$ (2)

where $a_{i\ell }^T$ indicates vector of size 1 × n, n is number of features for i-th protein domain of the ℓ-th RBDs subclass, w_iℓ is the weight for $a_{i\ell }^T$, y_iℓ is the response of a_iℓ, and c_ℓ is the intercept for the ℓ-th RBDs subclass. To construct multi-class ℓ1/ℓq-regularized logistic regression we used mcLogisticR function of SLEP package (version 4.0) which is written in Matlab. In this function, the elements in y are required to be m × k matrix including elements of 1 or −1 (m is the number of protein domains and k is the number of RBDs subclasses).

3. Results

3.1. Construction of dataset

Constructed dataset cover 7 out of 28 RNA binding subclasses with at least one protein domain member, including 7S RNA binding (10 protein domains), double-stranded RNA binding (16 protein domains), mRNA binding (11 protein domains), rRNA binding (29 protein domains), tRNA binding (16 protein domains), translational factor activity RNA binding (2 protein domains), and single-stranded RNA binding (1 protein domain). The RBDs of our dataset are summarized in Table 1. In construction of our methods we eliminated subclasses with less than 10 protein domain. In addition we constructed methods for prediction of five subclasses (i.e. tRNA-, rRNA-, mRNA-, 7S-, and double-stranded binding domain subclasses) and three subclasses (i.e. tRNA-, rRNA-, and mRNA-binding domain subclasses).

Table 1.

Summarized RNA binding domains in our dataset.

No.	Protein Domain	GO term	Class	Fold	Superfamily	Family	Domain	Species
1	d1914a1	7S_RNA_binding	α+β	SRP9/14	SRP9/14	SRP9/14	SRP9	Mouse
2	d1914a2	7S_RNA_binding	α+β	SRP9/14	SRP9/14	SRP9/14	SRP14	Mouse
3	d1hq1a_	7S_RNA_binding	α	SPBD	SPBD	SPBD	SSBP Ffh	EC
4	d1jida_	7S_RNA_binding	α+β	SRP19	SRP19	SRP19	SRP19	Human
5	d1kvva_	7S_RNA_binding	α+β	SRP19	SRP19	SRP19	SRP19	AAF
6	d1lnga_	7S_RNA_binding	α+β	SRP19	SRP19	SRP19	SRP19	AMJ
7	d1ls1a1	7S_RNA_binding	α	FHUDB	D-SRP/SRP receptor G-proteins	D-SRP/SRP receptor G-proteins	SSBP Ffh	TA
8	d1ls1a2	7S_RNA_binding	α/β	P-loop NTP hydrolases	P-loop NTP hydrolases	Nitrogenase iron protein-like	GTPase domain of SSBP Ffh	TA
9	d1qb2a_	7S_RNA_binding	α	SPBD	SPBD	SPBD	SRP54M	Human
10	d1qzxa2	7S_RNA_binding	α	SPBD	SPBD	SPBD	SSBP Ffh	ASS
11	d1di2a_	DS_RNA_binding	α+β	dsRBD-like	dsRBD-like	dsRBD	dsRBD A	XL
12	d1ekza_	DS_RNA_binding	α+β	dsRBD-like	dsRBD-like	dsRBD	Staufen, domain III	DM
13	d1o0wa1	DS_RNA_binding	α	RNase III domain-like	RNase III domain-like	RNase III catalytic domain-like	RNase III ECD	TM
14	d1o0wa2	DS_RNA_binding	α+β	dsRBD-like	dsRBD-like	dsRBD	RNase III, C-terminal domain	TM
15	d1t4oa_	DS_RNA_binding	α+β	dsRBD-like	dsRBD-like	dsRBD	RNase III, C-terminal domain	SC
16	d1uhza_	DS_RNA_binding	α+β	dsRBD-like	dsRBD-like	dsRBD	staufen homolog 2	Mouse
17	d1uila_	DS_RNA_binding	α+β	dsRBD-like	dsRBD-like	dsRBD	ATP-dep RNA helicase A, Dhx9	Mouse
18	d1whna_	DS_RNA_binding	α+β	dsRBD-like	dsRBD-like	dsRBD	Dus2l	Mouse
19	d1whqa_	DS_RNA_binding	α+β	dsRBD-like	dsRBD-like	dsRBD	ATP-dependent RNA helicase A, Dhx9	Mouse
20	d1x47a1	DS_RNA_binding	α+β	dsRBD-like	dsRBD-like	dsRBD	Dgcr8 protein	Human
21	d1x48a1	DS_RNA_binding	α+β	dsRBD-like	dsRBD-like	dsRBD	dsRNA-dependent protein kinase pkr	Mouse
22	d1x49a1	DS_RNA_binding	α+β	dsRBD-like	dsRBD-like	dsRBD	dsRNA-dependent protein kinase pkr	Mouse
23	d2dixa1	DS_RNA_binding	α+β	dsRBD-like	dsRBD-like	dsRBD	Interferon -ids RNA DPK activator A	Human
24	d2dmya1	DS_RNA_binding	α+β	dsRBD-like	dsRBD-like	dsRBD	Spermatid perinuclear RBP	Human
25	d2nuga1	DS_RNA_binding	α	RNase III domain-like	RNase III domain-like	RNase III catalytic domain-like	RNase III ECD	AA
26	d2nuga2	DS_RNA_binding	α+β	dsRBD-like	dsRNA-binding domain-like	dsRBD	RNase III, C-terminal domain	AA
27	d1afwa1	mRNA_binding	α/β	Thiolase-like	Thiolase-like	Thiolase-related	Thiolase	SC
28	d1j1ja_	mRNA_binding	α	α-α superhelix	Translin	Translin	Translin	Human
29	d1kvka1	mRNA_binding	α+β	RP S5 domain 2-like	RP S5 domain 2-like	GHMP Kinase, N-terminal domain	Mevalonate kinase	RN
30	d1kvka2	mRNA_binding	α+β	Ferredoxin-like	GHMP Kinase, C-terminal domain	Mevalonate kinase	Mevalonate kinase	RN
31	d1l5ja1	mRNA_binding	α	α-α superhelix	Aconitase B, N-terminal domain	Aconitase B, N-terminal domain	Aconitase B, N-terminal domain	EC
32	d1l5ja2	mRNA_binding	α/β	The “swivelling” β/β/α domain	LeuD/IlvD-like	LeuD-like	Aconitase B, second N-terminal domain	EC
33	d1l5ja3	mRNA_binding	α/β	Aconitase iron-sulfur domain	Aconitase iron-sulfur domain	Aconitase iron-sulfur domain	Aconitase B, C-terminal domain	EC
34	d1p5fa_	mRNA_binding	α/β	Flavodoxin-like	Class I glutamine AM-like	DJ-1/PfpI	DJ-1	Human
35	d1q67a_	mRNA_binding	β	PH domain-like barrel	PH domain-like	Dcp1	Dcp1	SC
36	d1xlya_	mRNA_binding	α	RNA-binding protein She2p	RNA-binding protein She2p	RNA-binding protein She2p	RNA-binding protein She2p	SC
37	d3gcba_	mRNA_binding	α/β	Cysteine proteinases	Cysteine proteinases	Papain-like	Bleomycin hydrolase	SC
38	d1a32a_	rRNA_binding	α	S15/NS1 RNA-binding domain	S15/NS1 RNA-binding domain	RP S15	RP S15	BST
39	d1diva1	rRNA_binding	α+β	RP L9 C-domain	RP L9 C-domain	RP L9 C-domain	RP L9 C-domain	BST
40	d1dmga_	rRNA_binding	α/β	RP L4	RP L4	RP L4	RP L4	TM
41	d1egaa2	rRNA_binding	α+β	α-lytic protease prodomain-like	Prokaryotic type KH domain	Prokaryotic type KH domain	GTPase Era C-terminal domain	EC
42	d1feua_	rRNA_binding	β	RP L25-like	RP L25-like	RPL25-like	RP TL5 (general stress protein CTC)	TT
43	d1i4ja_	rRNA_binding	α+β	RP L22	RP L22	RP L22	RP L22	TA
44	d1i6ua_	rRNA_binding	α+β	RP S8	RP S8	RP S8	RP S8	AMJ
45	d1iqva_	rRNA_binding	α	Ribosomal protein S7	RP S7	RP S7	RP S7	APH
46	d1loua_	rRNA_binding	α+β	Ferredoxin-like	RP S6	RP S6	RP S6	TT
47	d1n0ua1	rRNA_binding	β	R/I/E factor common domain	Translation proteins	EF	eEF-2, domain II	SC
48	d1n0ua2	rRNA_binding	α/β	P-loop NTP hydrolases	P-loop NTP hydrolases	G proteins	eEF-2, N-terminal (G) domain	SC
49	d1n0ua3	rRNA_binding	α+β	RP S5 domain 2-like	RP S5 domain 2-like	TMC	eEF-2, domain IV	SC
50	d1n0ua4	rRNA_binding	α+β	Ferredoxin-like	EF-G C-terminal domain-like	EF-G/eEF-2 domains III and V	eEF-2	SC
51	d1n0ua5	rRNA_binding	α+β	Ferredoxin-like	EF-G C-terminal domain-like	EF-G/eEF-2 domains III and V	eEF-2	SC
52	d1pkpa1	rRNA_binding	α+β	RP S5 domain 2-like	RP S5 domain 2-like	TMC	RP S5, N-terminal domain	BST
53	d1pkpa2	rRNA_binding	α+β	dsRBD-like	dsRBD-like	RP S5, N-terminal domain	RP S5, N-terminal domain	BST
54	d1rl6a1	rRNA_binding	α+β	RP L6	RP L6	RP L6	RP L6	BST
55	d1rl6a2	rRNA_binding	α+β	RP L6	RP L6	RP L6	RP L6	BST
56	d1seia_	rRNA_binding	α+β	RP S8	RP S8	RP S8	RP S8	BST
57	d1vmba_	rRNA_binding	α+β	Ferredoxin-like	RP S6	RP S6	RP S6	TM
58	d1vqoa1	rRNA_binding	β	SH3-like barrel	TP SH3-like domain	C-terminal domain of RP L2	C-terminal domain of RP L2	AHM
59	d1vqoa2	rRNA_binding	β	OB-fold	Nucleic acid-binding proteins	Cold shock DNA-binding domain-like	N-terminal domain of RP L2	AHM
60	d1wf3a1	rRNA_binding	α/β	P-loop NTP hydrolases	P-loop NTP hydrolases	G proteins	GTPase Era, N-terminal domain	TT
61	d1wf3a2	rRNA_binding	α+β	α-lytic protease prodomain-like	Prokaryotic type KH domain	Prokaryotic type KH domain	GTPase Era C-terminal domain	TT
62	d1whia_	rRNA_binding	β	RP L14	RP L14	RP L14	RP L14	BS
63	d2cqla1	rRNA_binding	α+β	RP L6	RP L6	RP L6	RP L6	Human
64	d2j5aa1	rRNA_binding	α+β	Ferredoxin-like	RP S6	RP S6	RP S6	AA
65	d2v3ka1	rRNA_binding	α/β	α/β knot	α/β knot	EMG1/NEP1-like	EMG1	SC
66	d3bbda1	rRNA_binding	α/β	α/β knot	α/β knot	EMG1/NEP1-like	RBP NEP1	MJ
67	d1dm9a_	SS_ RNA_binding	α+β	α-L RNA-binding motif	α-L RNA-binding motif	Heat shock protein 15 kD	HSP 15 Kd	EC
68	d1d7qa_	TFA_ RNA_binding	β	OB-fold	Nucleic acid-binding proteins	Cold shock DNA-binding domain-like	eIF1a	Human
69	d2if1a_	TFA_ RNA_binding	α+β	eIF1-like	eIF1-like	eIF1-like	eIF- 1 (SUI1)	Human
70	d1a6fa_	tRNA_binding	α+β	RP S5 domain 2-like	RP S5 domain 2-like	RNase P protein	RNase P protein	BSU
71	d1dj0a_	tRNA_binding	α+β	Pseudouridine synthase	Pseudouridine synthase	Pseudouridine synthase I TruA	Pseudouridine synthase I TruA	EC
72	d1fl0a_	tRNA_binding	β	OB-fold	Nucleic acid-binding proteins	Myf domain	EMAP II	Human
73	d1gd7a_	tRNA_binding	β	OB-fold	Nucleic acid-binding proteins	Myf domain	TRBP111 homolog CsaA	TT
74	d1jjca_	tRNA_binding	α+β	Class II aaRS and biotin synthetases	Class II aaRS and biotin synthetases	Class I A-tRNA S- like, catalytic domain	PheRS alpha subunit	TT
75	d1nz0a_	tRNA_binding	α+β	RP S5 domain 2-like	RP S5 domain 2-like	RNase P protein	RNase P protein	TM
76	d1ou5a1	tRNA_binding	α	Poly A PCT region-like	Poly A PCT region-like	Poly A PCT region-like	tRNA CCA-adding E, C-terminal domains	HM
77	d1ou5a2	tRNA_binding	α+β	Nucleotidyltransferase	Nucleotidyltransferase	Poly A polymerase head domain-like	tRNA CCA-adding E, head domain	HM
78	d1pyba_	tRNA_binding	β	OB-fold	Nucleic acid-binding proteins	Myf domain	TRBP111	AA
79	d1r6la1	tRNA_binding	α+β	RP S5 domain 2-like	RP S5 domain 2-like	Ribonuclease PH domain 1-like	Ribonuclease PH, domain 1	PseA
80	d1r6la2	tRNA_binding	α+β	RPH domain 2-like	RPH domain 2-like	Ribonuclease PH domain 2-like	Ribonuclease PH, domain 2	PseA
81	d1rqga1	tRNA_binding	α	ABD of a subclass of class I AA-tRNA-S	ABD of a subclass of class I AA-tRNA-S	ABD of a subclass of class I AA-tRNA-S	MetRS	PA
82	d1rqga2	tRNA_binding	α/β	AN-α hydrolase-like	Nucleotidylyl transferase	Class I A-tRNA S, catalytic domain	MetRS	PA
83	d2c5sa1	tRNA_binding	α/β	AN-α hydrolase-like	AN-α hydrolase-like	ThiI-like	TBP ThiI, N-ter D	BA
84	d2c5sa2	tRNA_binding	α+β	THUMP domain	THUMP domain-like	THUMP domain	TBP ThiI, N-ter D	BA
85	d2iy5a1	tRNA_binding	α	Long alpha-hairpin	tRNA-binding arm	PheRS	PheRS	TT

DS_RNA_binding: double-stranded_RNA_binding, SS_RNA_binding: single-stranded_RNA_binding, TFA_RNA_binding: translational_factor_activity_RNA_binding, SRP: Signal recognition particle alu RNA binding heterodimer, SPBD: Signal peptide-binding domain, FHUDB: Four-helical up-and-down bundle, P-loop NTP hydrolases: P-loop containing nucleoside triphosphate hydrolases, R/I/E factor common domain: Reductase/isomerase/elongation factor common domain, RP: Ribosomal protein, Class II aaRS and BS: Class II aaRS and biotin synthetases, Poly A PCT region-like: Poly A polymerase C-terminal region-like, RPH domain 2-like: Ribonuclease PH domain 2-like, ABD of a subclass of class I AA-tRNA-S: Anticodon-binding domain of a subclass of class I aminoacyl- tRNA synthetases, AN-α hydrolase-like: Adenine nucleotide alpha hydrolase-like, TP SH3-like domain: Translation proteins SH3-like domain, D-SRP/SRP receptor G-proteins: Domain of the SRP/SRP receptor G-proteins, Class I glutamine AM-like: Class I glutamine amidotransferase-like, AAF: Archaeon Archaeoglobus fulgidus, TT: Thermus thermophiles, BA: Bacillus anthracis, PA: Pyrococcus abyssi, PseA: Pseudomonas aeruginosa, AA: Aquifex aeolicus, TM: Thermotoga maritime, AMJ: Archaeon Methanococcus jannaschii, MJ: Methanococcus jannaschii, SC: Saccharomyces cerevisiae, BST: Bacillus stearothermophilus, EC: Escherichia coli, BSU: Bacillus subtilis, AHM: Archaeon Haloarcula marismortui, APH: Archaeon Pyrococcus horikoshii, TA:Thermus aquaticus, TM: Thermotoga maritime, HM: Human mitochondrial, RN: Rattus norvegicus, DM: Drosophila melanogaster, XL: Xenopus laevis, ASS: Archaeon Sulfolobus solfataricus; SSBP: Signal sequence binding protein Ffh, RNase III ECD: RNase III endonuclease catalytic domain, PheRS: Phenylalanyl-tRNA synthetase, TBP ThiI, N-ter D: Thiamine biosynthesis protein ThiI, N-terminal domain, MetRS: Methionyl-tRNA synthetase, TRBP: tRNA-binding protein, tRNA CCA-adding E: tRNA CCA-adding enzyme, PheRS: Phenyl-tRNA synthetase, HSP: Heat shock protein, RBP: Ribosome biogenesis protein, EMG: Essential for mitotic growth, RNase III ECD: RNase III endonuclease catalytic domain, Interferon-ids RNA DPK activator A: Interferon-inducible double stranded RNA-dependent protein kinase activator A, Dus: dihydrouridine synthase, TMC: Translational machinery components.

3.2. ANOVA analysis for feature selection

In order to consider the effect of number of features on performance of methods, ANOVA was used to select significantly different features between three and five RNA-binding protein domain subclasses. Table 2 and Table 3 have shown 10 top features with the lowest p-values. From ANOVA results, RNA binding subclasses show an obvious difference in sequence- and structure-based features. Figure 2 shows difference of shape, size of RBDs, size of main patch and frequency of two important charged amino acids, i.e., Arg and Lys, in five different RBD subclasses. In addition reduced models were constructed using selected features with significant level of <.05, which were 45 and 102 features in three and five subclasses, respectively.

Table 2.

Indicating ten top selected features between three RNA-binding domain subclasses by using ANOVA analysis.

Number	Feature	P-value
1	Number of Arg in intermediate regions	2.0E-04
2	Molecular weight of RBDs	5.0E-04
3	Number of Ser in buried regions	7.0E-04
4	Number of Cys in main patch	0.001
5	Number of basic amino acids in sequence	0.0012
6	Number of Glu in sheet	0.0026
7	Number of charged amino acids in sequence	0.003
8	Number of Arg in exposed regions	0.0034
9	Number of charged amino acids in sheet	0.0039
10	Isoelectric point	0.0041

Table 3.

Indicating ten top selected features between five RNA-binding domain subclasses by using ANOVA analysis.

Number	Feature	P-value
1	Dipole	4.00E-10
2	Total number of residues	8.00E-10
3	Total number of atoms	1.40E-09
4	Total number of negative residues	7.40E-09
5	Total number of positive residues	4.93E-08
6	Number of Lys in main patch	8.92E-08
7	Molecular weight of RBDs	8.55E-07
8	RM	2.18E-06
9	Number of small amino acids in main patch	1.50E-05
10	Number of Ala in buried regions	4.62E-05

Figure 2.

Diversity of features between five different RBDs. (A) sample of 7S RBDs (d1914a1), (B) sample of rRNA RBDs (d2v3ka1), (C) sample of double_stranded RBPs (d1ekza_), (D) sample of mRNA RBDs (d1afwa1), and (E) sample of tRNA RBDs (d1a6fa_). The gray region represents the main patch, blue represents Arg amino acids, and green represents Lys amino acids.

3.3. Tuned multi-class support vector machine analysis

We used a tune function to select optimized structure of TMCSVM through jackknife and self-consistency tests. The most important parameter of TMCSVM topology is kernel function which was searched for the best one among four different kernel functions, i.e., linear, polynomial, radial, and sigmoid. Table 8 and 9 show the highest performance obtained by TMCSVM in overall. TMCSVM and reduced-TMCSVM show the highest rate of 79.31% in prediction of rRNA BDs subclass in comparison with the other methods in five subclasses prediction and also SVM shows the highest rate of 50% for prediction of tRNA BD subclass in three subclasses prediction. Our results confirm that although TMCSVM is a machine learning method, dealing with a large number of features is possible because of seeking a global optimized solution and avoiding over-fitting in the SVM training process. However, obtained results in three subclasses prediction, emphasize that this ability is diminished to limited range of features/samples ratio.

Table 8.

Results of self-consistency and jackknife tests in prediction of five subclasses

Test	Method	Rate of correct prediction for each RBD subclasses					Overall rate of accuracy
		7s (%)	Double-stranded (%)	mRNA (%)	rRNA (%)	tRNA (%)
Self-consistency	RF	100	100	100	100	100	100
	Reduced-RF	100	100	100	100	100	100
	SVM	100	100	100	100	100	100
	Reduced-SVM	100	100	100	100	87.50	97.56
	MCRLR	100	100	100	100	100	100
	Reduced-MCRLR	100	100	100	100	100	100
Jackknife	RF	60.00	81.25	63.64	44.83	37.50	54.88
	Reduced-RF	70.00	81.25	81.82	44.83	37.50	58.54
	SVM	40.00	87.50	54.55	79.31	50.00	67.07
	Reduced-SVM	50.00	87.50	45.45	79.31	43.75	65.84
	MCRLR	60.00	100	54.55	65.52	31.25	63.41
	Reduced-MCRLR	50.00	87.50	45.45	65.52	43.75	60.98

Table 9.

Results of self-consistency and jackknife tests in prediction of three subclasses

Test	Method	Rate of correct prediction for each RBD subclasses			Overall rate of accuracy
		mRNA (%)	rRNA (%)	tRNA (%)
Self-consistency	RF	100	100	100	100
	Reduced-RF	100	100	100	100
	SVM	100	100	100	100
	Reduced-SVM	100	100	100	100
	MCRLR	100	100	100	100
	Reduced-MCRLR	100	100	100	100
Jackknife	RF	81.82	62.07	43.75	60.71
	Reduced-RF	81.82	62.07	62.50	66.07
	SVM	63.64	82.76	37.50	66.07
	Reduced-SVM	63.64	79.31	62.50	71.43
	MCRLR	54.55	82.76	62.50	71.43
	Reduced-MCRLR	18.18	79.31	37.50	55.36

3.4. Random Forest analysis

R randomForest package was used to construct RF for prediction of RBD subclasses. In order to optimize performance of RF, we defined cutoffs based on distribution of RBDs in subclasses, i.e., number of RBDs in each subclass divided by total number of RBDs. Obtained results reveal that although RF can predict all of RBDs correctly through self-consistency, performance of jackknife test drastically reduced (Table 8 and 9). However, reduced-RF shows the highest rate in prediction of 7S RBDs subclass (70%), and mRNA RBDs subclass (81.82%) in comparison with the other methods in five subclasses prediction. In addition, RF and reduced-RF show the highest rate of 81.82% for prediction of mRNA BD subclass in three subclasses prediction. Furthermore, from obtained results, it is obvious that number of features in RF training is an important issue and it is independent of number of subclasses. Indeed, RF is over-fitting prone when we train it using large number of features.

3.5. Multi-class ℓ₁/ℓ_q-regularized logistic regression

We ran a MCRLR method on the dataset in five and three subclasses using jackknife and self-consistency. MCRLR provides useful information about preferred and avoided features in each one of RNA binding subclasses. tRNA BD subclasses shows some preferred and avoided with higher average values in comparison with the other subclasses in three- and five subclasses through jackknife and self-consistency procedures (Tables 4–7). Our results confirm previous reported unique properties of tRNA BPs by Shazman and Mandel-Gutfreund (2008).

Table 4.

Indicating average weights of ten top important features computed by MCRLR in the jackknife procedure using all of features for prediction of five RNA-binding domain subclasses. Positive values show preference of the features in related subclasses and negative values show avoidance of the features in related subclasses.

No.	Feature/Average Value
	7s RBD subclass	Double -stranded RBD subclass	mRNA RBD subclass	rRNA RBD subclass	tRNA RBD subclass
1	Number of Met in MP/1.517	Number of basic AAs in MP/0.967	Molecular weight/1.373	Number of His in MP/1.913	Number of Arg in IR/2.004
2	Number of Met in RC/1.274	Number of Glu in MP/0.900	Number of Ser in BR/1.176	Number of Arg in IR/-1.555	Number of His in sheet/2.003
3	Number of Cys in sheet/0.998	Dipole/0.900	Number of Cys in MP/1.142	Number of Ile in MP/1.499	Number of Pro in MP/-1.583
4	Number of His in seq./-0.920	Number of Glu in IR/0.891	Number of Ser in IR/-1.094	Number of Leu in helix/-1.481	Number of Ile in MP/-1.576
5	Number of Met in ER/0.990	Number of Tyr in MP/0.831	Isoelectric point/-1.007	Number of Met in RC/-1.458	Number of Asp in BR/1.371
6	Number of Glu in MP/-0.886	Number of Ala in BR/0.811	Number of Cys in RC/0.969	Number of Ser in RC/1.330	Number of Ile in IR/-1.335
7	Number of Tyr in helix/0.811	Number of Gln in RC/-0.799	Number of charged AAs in seq./-0.949	Number of Tyr in MP/-1.304	Number of Met in sheet/1.255
8	Number of Glu in IR/0.759	Number of Arg in MP/-0.779	Number of Asp in ER/0.934	Number of Val in helix/1.280	Number of Phe in ER/-1.226
9	Frequency of antiparallel HB/-0.753	Number of Lys in MP/0.752	Number of Arg in ER/-0.931	Number of His in ER/-1.269	Number of Phe in IR/1.166
10	Second patch size/-0.729	Number of Asp in MP/0.728	Number of Asp in IR/0.920	Number of Cys in RC/-1.258	Number of Gln in IR/1.114

MP: main patch, RC: randomcoil, ER: exposed regions, IR: intermediate regions, HB: hydrogen bond, AAs: amino acids, BR: buried regions, Seq.: Sequence.

Table 7.

Indicating average weights of ten top important features computed by MCRLR in the jackknife procedure using selected features by ANOVA analysis for prediction of three RNA-binding domain subclasses. Positive values show preference of the features in related subclasses and negative values show avoidance of the features in related subclasses.

No.	Feature/Average Value
	mRNA RBD subclass	rRNA RBD subclass	tRNA RBD subclass
1	Number of Glu in seq./-3.365	Number of Ala in helix/-4.742	Number of Asn in helix/-5.94
2	Number of Cys in helix/-3.347	Number of Glu in seq./3.999	Number of Leu in seq./-3.765
3	Number of Asp in seq./1.950	Number of Leu in seq./2.571	Number of Trp in seq./-2.978
4	Number of Gln in helix/1.759	Number of Asn in helix/2.481	Number of Ala in helix/2.913
5	Number of Glu in helix/-1.621	Number of Cys in helix/1.900	Number of Gly in helix/2.386
6	Number of Val in seq./1.620	Number of Trp in seq/1.837	Number of aliphatic AAs in seq./2.048
7	Number of Leu in seq./1.570	Number of Lys in seq./-1.712	Number of Lys in seq./1.773
8	Number of Ala in helix/1.525	Number of Ile in seq./-1.618	Number of Ile in seq./1.567
9	Number of nonpolar AAs in seq./1.187	Number of Pro in seq./1.551	Number of Pro in helix/1.402
10	Number of Trp in seq./1.051	Number of nonpolar AAs in seq./-1.355	Number of Met in helix/-1.241

The results of jackknife and self-consistency tests, which shown in Table 8 and 9, are obtained according to the output of the model. High performance measures of MCRLR model through self-consistency confirm usefulness of defined features in prediction of RBPs subclasses. Results of jackknife tests show that performance of reduced-MCRLR drastically decreased especially in three subclasses prediction using selected features. Rationale for decrease of MCRLR performance is restriction of shrinkage ability using limited number of features (N=45 for prediction of three subclasses). Indeed, ℓ1/ ℓq-regularized constrains the total weight allocated to a set of features, with the end result that some features received zero weight. Additionally, MCRLR shows the highest rate in prediction of double-stranded RBDs in comparison with the other methods.

4. Discussion

Knowledge regarding how bio-macromolecules interact with each other is essential in the understanding of cellular processes. In this study, we investigated interaction of protein and RNA as an important interaction in various cellular processes.

According to a recent comprehensive review (Chou, 2011), to establish a really useful predictor for a protein system, we need to consider: construct a valid benchmark dataset, formulate the protein samples with an effective mathematical expression, develop a powerful algorithm to operate the prediction, evaluate the anticipated accuracy of the predictor, and establish a user-friendly web-server, respectively.

From previous reports, it is mentioned that the aminoacyl tRNA synthetases, and bacterial factors, which mimic tRNA BPs have highly negatively charged surface (Tworowski et al., 2005; Nakamura and Ito, 2003). But there is no more information about variation in feature distribution in different RBPs. In this study, in addition to multi-class classification of RBDs we tried to do feature selection. In addition to comparable prediction accuracy with TMCSVM, a clear variety of feature distributions was elucidated by using MCRLR. For example, our results demonstrate exciting diversity in distribution of Lys and Arg, two important charged amino acids in interaction and catalytic reaction, in different RBDs subclasses. From our data in tRNA BD subclass, Lys is preferred in sequence and Arg is preferred in intermediate regions with high scores (Tables 4–7). In mRNA BDs subclass, Arg is preferred in sequence, and in double-stranded RBD subclass, Lys is preferred in main domain while Arg is avoided in main domain. In addition, in rRNA BD subclass, Lys is avoided in sequence and it seems that Arg is preferred to be on surface as it has been determined with negative value of being in intermediate regions. In 7S RBD subclass, Arg and Lys have not been selected among top preferred or avoided residues. From our results we can understand that Lys and Arg in tRNA BDs, Arg in mRNA BDs, Lys in double-stranded RBDs, and exposed Arg in rRNA BDs are possibly important in RNA-protein interaction and catalytic reaction of RBDs. Figure 2 illustrates distribution of Arg and Lys in main patches of different RBDs subclasses.

These results emphasize that the tRNA BDs have unique local and global properties that can be utilized for identifying novel proteins possibly involved in tRNA processing. Moreover, it is worth to mention that the size of secondary patch show positive average value in tRNA BDs subclasses and it means secondary patch may have specific properties as mentioned by Shazman and Mandel-Gutfreund (2008). Growth of 3D solved protein databases will be helpful to discover more details about RBDs.

In this study we developed a first of its kind in silico approach for analysis and prediction of RBDs subclasses in three and five subclasses using RF, TMCSVM and MCRLR. In overall, TMCSVM outperforms the other methods, although tuning of SVM is time consuming. On the other hand, MCRLR show some advantages including fast training, report of more important features for RBD prediction, and detection of avoided and preferred features in each subclass.

In addition, RF shows the worst accuracy among three predictor methods which means RF is prone to over-fitting especially when large numbers of features are fed into it. In conclusion, we used two types of predictor methods including: (1) MCRLR as a statistical method and (2) RF and TMCSVM as machine learning methods. Statistical methods are commonly accepted and popularity of these models may be attributed to the interpretability of model parameters and ease of use, although they suffer from their specific limitations. For example, statistical methods use linear combinations of independent variables and, therefore, are not the best adept at modeling grossly nonlinear complex interactions as has been demonstrated in biological systems. On the other hand, machine learning methods are rich and flexible nonlinear systems that show robust performance in dealing with noisy or incomplete data and have the ability to generalize from the input data. They may be better suited than other modeling systems to predict outcomes when the relationships between the variables are complex, multidimensional, and nonlinear as found in complex biological systems. Although machine learning methods can give high prediction accuracy, some problem may be raised in their training. For example in this study we showed that RF as a well-known machine learning method is not well suited for our problem and is prone to over-fitting. “black box” nature, and the empirical nature of model development are other disadvantages of machine learning methods (Tu, 1996).

5. Conclusion

A great challenge in classifying ligand binding proteins (such as RBDs) is to be able to identify to which ligand it will bind. For this purpose, we applied three different predictor methods to classify RNA-binding domains using a large number of sequence and structural features, which was trained on three and five different subclasses of known RBDs classified according to their RNA target. From our results TMCSVM shows the highest prediction accuracy in compare with other methods. Overall, the results we obtained are encouraging, reinforcing the idea that combination of sequence and structural properties of protein domains can give clues to the protein’s interacting partner.

It is important to note that subclassification of the RBDs to three and five subclasses using our multiclass approach is only possible –given the prior knowledge that the protein domain binds RNA. Indeed we have to mention that requiring known protein domains as RNA binding is a limitation of such predictor models.

Finally, our results showed that, in addition to multi class prediction, biological diversity of RBD’s subclasses would be interpretable using state-of-the-art methods like ℓ1/ℓq-regularized logistic regression.

Since user friendly and publicly accessible web-servers represent the future direction for developing practically more useful predictors (Chou and Shen, 2009), we shall make efforts in our future work to provide a web-server for the method presented in this paper.

Table 5.

Indicating average weights of ten top important features computed by MCRLR in the jackknife procedure using selected features by ANOVA analysis for prediction of five RNA-binding domain subclasses. Positive values show preference of the features in related subclasses and negative values show avoidance of the features in related subclasses.

No.	Feature/Average Value
	7sRBD subclass	Double-stranded RBD subclass	mRNA RBD subclass	rRNA RBD subclass	tRNA RBD subclass
1	Number of Cys in sheet/2.029	Number of Glu in IR/1.675	Number of Ser in IR/-2.051	Number of small AAs in helix/3.773	Number of Ile in MP/-4.351
2	Number of Met in MP/2.011	Number of Arg in MP/-1.394	Number of Arg in ER/-2.034	Number of Ser in RC/3.652	Number of Arg in IR/4.083
3	Number of Met in RC/1.653	Number of Tyr in MP/1.230	Number of Cys in MP/1.976	Number of Ile in MP/3.378	Number of Small AAs in helix/-3.625
4	Frequency of antiparallel HB/-1.54	Dipole/1.160	Number of Ser in BR/1.733	Number of acidic AAs in ER/3.219	Number of Small AAs in RC/-3.118
5	Number of His in seq./-1.370	Number of Gln in RC/-1.070	Molecular weight/1.554	Number of Ser in IR/3.030	Number of His in seq./3.027
6	Number of Met in ER/1.300	Number of Phe in MP/1.062	Number of small AAs in BR/1.486	Number of Tyr in MP/-2.954	Number of Phe in IR/2.643
7	Number of Glu in IR/1.292	Number of Aromatic AAs in MP/1.047	Number of Arg in IR/1.482	Number of Pro in sheet/2.7428	Number of tiny AAs in sheet/2.288
8	Number of Leu in sheet/1.235	Number of Ser in IE/1.036	Number of Glu in IR/-1.363	Number of Met in RC/-2.742	Frequency of i+3 HB/2.091
9	Number of Gln in RC/1.231	Number of basic AAs in BR/-0.989	Isoelectric point/-1.319	Number of Arg in IR/-2.588	Number of Small AAs in MP/-1.816
10	Second patch size/-1.1781	Number of Met in RC/0.981	Number of Ile in MP/1.317	Number of Glu in IR/-2.458	Number of Phe inRC/-1.790

MP: main patch, RC: randomcoil, ER: exposed regions, IR: intermediate regions, HB: hydrogen bond, AAs: amino acids, BR: buried regions, Seq.: Sequence.

Table 6.

Indicating average weights of ten top important features computed by MCRLR in the jackknife procedure using all of features for prediction of three RNA-binding domain Subclasses. Positive values show preference of the features in related subclasses and negative values show avoidance of the features in related subclasses.

No.	Feature/Average Value
	mRNA RBD subclass	rRNA RBD subclass	tRNA RBD subclass
1	Number of charged AAs in seq./-1.337	Number of Arg in IR/-2.794	Number of Arg in IR/2.129
2	Number of Ser in IR/-1.132	Number of Glu in sheet/1.556	Number of Pro in MP/-1.699
3	Number of Arg in seq./1.131	Number of Phe in helix/-1.278	Number of Pro in IR/-1.581
4	Number of Pro in seq./0.953	Number of charged AAs in seq./1.109	Number of Phe in IR/1.339
5	Number of Glu in sheet/0.949	Number of Phe in MP/1.091	Number of Asn in IR/-1.183
6	Number of Ser in BR/0.898	Number of Ser in IR/1.021	Number of Val in helix/-1.055
7	Number of Phe in IR/-0.782	Second patch size/-1.015	Number of Ile in IR/-1.033
8	Number of Arg in ER/-0.769	Number of Cys in helix/0.954	Number of Phe in helix/0.921
9	Number of Arg in IR/0.753	Number of His in sheet/-0.930	Number of Cys in helix/-0.790
10	Number of Cys in RC/0.712	Number of Cys in RC/-0.789	Second patch size/0.766

MP: main patch, RC: randomcoil, ER: exposed regions, IR: intermediate regions, AAs: amino acids, BR: buried regions, Seq.: Sequence.

Highlights.

1. Multi-class prediction of RNA-binding protein domains.

2. We compared prediction accuracy of three different state-of-the-art predictor methods.

3. In addition to successful classifying RBPs, we discovered dissimilar sequence and structural features using ℓ1/ℓq-regularized logistic regression.

4. Our method could be applied to identify novel RNA-binding proteins with unique folds.