Recent trends in RNA informatics: a review of machine learning and deep learning for RNA secondary structure prediction and RNA drug discovery

Abstract

Computational analysis of RNA sequences constitutes a crucial step in the field of RNA biology. As in other domains of the life sciences, the incorporation of artificial intelligence and machine learning techniques into RNA sequence analysis has gained significant traction in recent years. Historically, thermodynamics-based methods were widely employed for the prediction of RNA secondary structures; however, machine learning-based approaches have demonstrated remarkable advancements in recent years, enabling more accurate predictions. Consequently, the precision of sequence analysis pertaining to RNA secondary structures, such as RNA–protein interactions, has also been enhanced, making a substantial contribution to the field of RNA biology. Additionally, artificial intelligence and machine learning are also introducing technical innovations in the analysis of RNA–small molecule interactions for RNA-targeted drug discovery and in the design of RNA aptamers, where RNA serves as its own ligand. This review will highlight recent trends in the prediction of RNA secondary structure, RNA aptamers and RNA drug discovery using machine learning, deep learning and related technologies, and will also discuss potential future avenues in the field of RNA informatics.

INTRODUCTION

The central dogma posits that RNA functions solely as a conduit for the transfer of genetic information from DNA to proteins. Messenger RNAs (mRNAs) perform this role as information carriers. However, a number of exceptions to this paradigm have been discovered, involving RNA molecules participating in a diversity of functions. Transfer RNAs (tRNAs) function in the translation of the triplet codons of mRNAs into amino acids according to the genetic code. Ribosomal RNAs (rRNAs) constitute a primary component of ribosomes and catalyze protein synthesis as ribozymes. Micro RNAs (miRNAs) are involved in RNA silencing and post-transcriptional regulation of gene expression. Small nuclear RNAs (snRNAs) participate in the processing of pre-messenger RNAs within the nucleus. Long noncoding RNAs (lncRNAs), non-protein-coding RNAs with sequences longer than 200 bases, have been demonstrated to have various functions such as gene transcriptional regulation, translational regulation and epigenetic regulation [1]. The diversity of RNA species has been cataloged in databases such as Rfam [2–5] and RNAcentral [6], and the number of RNA species continues to grow.

Many of these functional RNAs execute their functions by adopting tertiary structures that are evolutionarily conserved among RNA species. The experimental determination of RNA tertiary structures can be achieved through techniques such as X-ray crystallography, nuclear magnetic resonance (NMR) and cryo-electron microscopy (cryo-EM); however, these methods are both labor-intensive and cost-prohibitive for high-throughput analysis. As an alternative, RNA secondary structures are often targeted for structural and functional analysis of functional RNAs. An RNA secondary structure is defined as a set of base pairs with hydrogen bonds between two nucleotides, which makes a significant contribution to the tertiary structure in terms of free energy. This means that the folding of RNA is hierarchical in that tertiary interactions can be added without much distortion of the secondary structure [7]. It is well established that RNA secondary structures are also evolutionarily conserved among RNA species. For instance, a multiple sequence alignment of 10 tRNAs extracted from the Rfam database (Figure 1A), in which secondary structures are considered, yields the well-known and evolutionarily conserved cloverleaf shape (Figure 1C). In contrast, a multiple sequence alignment based on sequence identity alone, calculated using Clustal Omega [8], does not preserve the secondary structure at all, as shown in Figure 1B. This suggests that functional RNAs are evolutionarily conserved in their structures, rather than in their sequences and that structure correlates with function. Thus, RNA informatics has extensively studied RNA sequence analysis based on their secondary structures.

Figure 1.

Conserved secondary structures of transfer RNAs. (A) A multiple sequence alignment of 10 tRNAs extracted from the Rfam database is depicted, along with the annotated secondary structure. A string consisting of ‘.’, ‘(’ and ‘)’ at the top of the multiple alignment is the dot-bracket notation that represents the secondary structure. (B) A multiple sequence alignment based on sequence identity was calculated with Clustal Omega [8] for the 10 tRNAs. (C) The conserved secondary structure annotated by the Rfam database was visualized using VARNA [9]. Each colored region of base pairs corresponds to the acceptor stem (blue), D arm (red), anticodon arm (yellow) and T arm (green), respectively.

RNA secondary structure prediction, which has been a subject of study since the 1970s, constitutes a foundational technology in RNA sequence analysis. One of the pioneered works in this field was the Nussinov algorithm [10], which used dynamic programming to calculate the secondary structure with the maximum number of base pairs, under the assumption that the free energy of RNA structures decreases as base pairs are formed, thereby resulting in greater stability. Minimizing free energy using thermodynamic free energy parameters, such as Turner’s free energy parameters [11–13], is a widely used approach for predicting RNA secondary structures [11, 14–18]. However, these thermodynamics-based methods have a drawback in that they are limited by experimental technicalities in increasing the number and improving the quality of the free energy parameters. To address this issue, ‘assisted folding’ with high-throughput experiments involving chemical probing has been frequently employed. Selective 2’-hydroxyl acylation analyzed by primer extension (SHAPE) [19] and dimethyl sulphate (DMS) [20] are examples of high-throughput experiments that can induce chemical modifications on unpaired nucleotides, which can improve the accuracy of secondary structure prediction.

To address the drawback of the thermodynamics-based methods without wet-laboratory experiments, machine learning-based methods have been widely used. These methods learn their parameters from training data that include RNA sequences and their corresponding secondary structures, thus, more accurate models can be constructed by increasing the number of parameters if a sufficient amount of non-biased training data can be obtained. For example, compared with CONTRAfold [21, 22], which has approximately 300 parameters, ContextFold [23] learns a much richer parameter set of approximately 200k parameters from training data, resulting in the ability to predict RNA secondary structures with remarkable accuracy at the time. However, as discussed later, it has been shown that ContextFold’s rich parameterization can easily lead to low prediction accuracy due to overfitting to the training data [24, 25].

Another approach to further improve prediction accuracy is to incorporate evolutionary information from homologous sequences [26–31]. This approach is potentially very powerful, but it is a different problem from the de novo prediction from single sequences discussed in this review, because this approach incorporates much more information from homologous sequences. In addition, there are few RNA secondary structure prediction algorithms that incorporate evolutionary information in machine learning, and thus this category will not be addressed in this review.

Subsequent tasks include analyses utilizing the results of RNA secondary structure predictions, such as clustering [32, 33], family classification [34, 35], structural RNA detection [36, 37] and degradation prediction [38]. Machine learning techniques are frequently employed in these tasks as well.

With the aforementioned fundamental advancements in RNA informatics, the practical application of RNA research is also progressing at a rapid pace. One such application is RNA drug discovery, which is highly anticipated as a next-generation drug discovery modality. Research on RNA aptamers and mRNA vaccines, in which RNA itself serves as a therapeutic agent, and RNA-targeted drug discovery, in which RNA functions as an alternative drug target to proteins, is of particular interest. Additionally, the application of RNA bioinformatics research utilizing technologies such as machine learning and deep learning is required in these areas, and research is being conducted from various perspectives, including RNA structure.

In this review, we will present recent trends in the prediction of RNA secondary structures utilizing machine learning, deep learning and related technologies. Additionally, we will provide an overview of significant advancements in computational tasks that employ RNA secondary structure information, including the prediction of RNA–protein interactions and RNA drug discovery.

RNA SECONDARY STRUCTURE PREDICTION

What is RNA secondary structure prediction?

An RNA sequence is a sequence composed of four types of nucleotides: adenine (A), cytosine (C), guanine (G) and uracil (U). An RNA secondary structure is defined as a set of base pairs with hydrogen bonds between two nucleotides. Thus, the problem of RNA secondary structure prediction is to predict which two nucleotides in a given RNA sequence form base pairs. Typically, we consider only the Watson–Crick base pairs (A-U and G-C), which are the most prevalent, as well as the wobble base pairs (G-U), which are the next most common, in RNA secondary structure predictions. Other non-canonical base pairs are of great significance for tertiary structures, but they are much more challenging to model computationally, since each base is not guaranteed to form a base pair with at most one other base. An RNA secondary structure can be represented by a string in the dot-bracket notation, where two bases at the corresponding open and close brackets (‘(’ and ‘)’) form a base pair, while a base at the dot (‘.’) does not form a base pair with any base, as shown at the top of the multiple alignment in Figure 1A. A secondary structure that requires only one type of bracket for its dot-bracket notation, resulting in fully nested base pairs, is referred to as a pseudoknot-free secondary structure. Conversely, a substructure consisting of non-nested base pairs, as depicted in Figure 2, where the bases inside the loop form base pairs with the bases outside the loop, is called a pseudoknot. To describe a pseudoknot in the dot-bracket notation, two or more types of brackets (e.g. ‘[’ and ‘]’) are required. RNA secondary structure prediction including pseudoknots has been proven to be NP-hard for optimal solutions with no limitations on the complexity of pseudoknots [39, 40]. Therefore, approximations that restrict the complexity of pseudoknots or introduce heuristics are common approaches.

Figure 2.

An RNA secondary structure with pseudoknots. The arcs connecting the two nucleotides represent base pairs. Since pseudoknot structures are non-nested, the arcs representing base pairs cross each other when the structure is drawn on a 2D plane.

Computational models

De novo computational modeling of RNA secondary structures can be broadly classified into three categories: nearest neighbor models, probabilistic generative models and deep learning models. This subsection provides an overview of these models and their implementations, along with a comparison of the three models. Figure 3 shows a schematic diagram of de novo RNA secondary structure prediction algorithms discussed in this subsection. Additionally, the datasets utilized for constructing models for predicting RNA secondary structures will be discussed.

Figure 3.

A schematic diagram of de novo RNA secondary structure prediction algorithms. Most RNA secondary structure prediction algorithms can be categorized by three aspects: ‘Architecture’, ‘Parameterization’ and ‘Inference’. Rivas et al. [25] added ‘Scoring scheme’ to these aspects, which is uniquely determined by ‘Parameterization’, and thus ‘Scoring scheme’ is omitted in this paper. From the ‘Architecture’ aspect, RNA secondary structure prediction algorithms can be categorized into nearest neighbor models, probabilistic generative models and deep learning models, depending on the RNA computational models. These are further sub-categorized according to their parameter assignment, fine-grainedness, grammatical rules etc. The ‘Parameterization’ aspect classifies RNA secondary structure prediction algorithms into three types, depending on how they find optimal parameter values for the parameter set defined in the ‘Architecture’: wet-lab experiments, machine learning and deep learning. Finally, the ‘Inference’ aspect classifies RNA secondary structure prediction algorithms according to how they use the models determined in the ‘Archtecture’ and ‘Parameterization’ to make secondary structure predictions.

Nearest neighbor models

The nearest neighbor model, which has been extensively utilized in the prediction of RNA secondary structures [11, 41–43], decomposes an RNA secondary structure into loop substructures with hairpin loops, stackings, bulge loops, internal loops, multi-branch loops and external loops, depending on the number of closing base pairs, as depicted in Figure 4. Each loop substructure is parametrized with several types of components, characterized by nucleotides in loops, the length of loops and other such factors, which are referred to as the energy parameters. The values assigned to each energy parameter are determined through either experimental methods or machine learning techniques. The free energy of each decomposed loop substructure is computed as the sum of the values of the energy parameters that characterize the loop substructure. The free energy of a given RNA secondary structure can be calculated as the sum of the free energies of the loop substructures that are decomposed from the given RNA secondary structure. Zuker et al. [44] established an efficient algorithm, known as the Zuker algorithm, which is based on the dynamic programming technique to find a secondary structure that minimizes the free energy among all possible secondary structures formed by a given RNA sequence. Many RNA secondary structure prediction methods that model RNA secondary structures without pseudoknots and employ the nearest neighbor model use the Zuker algorithm to find the minimum free energy (MFE) structures. The Zuker algorithm has a computational complexity of in time and in space for an RNA sequence of length . Recently, the LinearFold algorithm [45] has been developed, which finds MFE structures accurately and approximately with computational complexity in both time and space using the beam search technique. The key differences among the implementations of the Zuker algorithm or the LinearFold algorithm are the parametrization of the loop substructures and the determination of each value assigned to their energy parameters.

Figure 4.

The nearest neighbor model decomposes RNA secondary structures into loop substructures. Hairpin loops are closed by a single base pair. Loop substructures that are closed by two base pairs with no unpaired bases are called stackings, those with unpaired bases on one strand are called bulge loops and those with unpaired bases on both strands are called internal loops. Multi-branch loops are loop substructures that are closed by three or more base pairs. Loop substructures closed by no base pairs are called external loops.

The methodology used to determine the energy parameters can be broadly classified into two approaches. The first approach involves determining the free energy parameters through wet-laboratory experiments, which is beyond the scope of this review; for further details, see the reference [46]. Examples of tools that fall under this approach include Mfold/UNAfold [14, 15], RNAfold in the ViennaRNA package [16, 17] and RNAstructure [11, 18]. Turner’s free energy parameters [11–13] are widely used in these thermodynamics-based approaches, and consist of up to approximately 12 700 parameters.

The second approach to determining energy parameters utilizes machine learning techniques to learn them from a large dataset of pairs of RNA sequences and their corresponding secondary structures. CONTRAfold [21, 22] used conditional log-linear models (CLLMs) to train approximately 300 parameters of the nearest neighbor model, resulting in high accuracy in predicting RNA secondary structures with significantly fewer parameters than Turner’s free energy parameters. Since the machine learning-based approach does not depend on wet-lab experiments, it allows for the development of a more comprehensive parametrization. For example, ContextFold [23] employed a fine-grained RNA secondary structure model with a parameter set of more than 200 000, resulting in the state-of-the-art prediction accuracy at the time. However, Rivas et al. [24, 25] pointed out that ContextFold had poor accuracy in predicting secondary structures for families not included in the training data, and was likely to fall into overfitting.

Several hybrid tools that combine both the thermodynamics- and machine learning-based approaches have been developed. SimFold [47, 48] modifies Turner’s free energy parameters to fit training data through the machine learning-based approach using training data including triplets of RNA sequences and their secondary structure as well as their free energies. MXfold [49] combines Turner’s energy parameters with rich-parametrized parameters trained by a max-margin framework, called structured support vector machines. It learns more precise parameters for substructures observed in the training data, reducing overfitting using thermodynamic parameters for unobserved substructures. MXfold2 [50], the successor of MXfold, utilizes deep learning to compute four types of scores for loop substructures: helix stacking scores, helix opening scores, helix closing scores and unpaired region scores, and combines them with Turner’s energy parameters, resulting in highly accurate and robust secondary structure prediction with a reduced risk of overfitting.

The majority of methods developed thus far predict RNA secondary structures based on the MFE under a given set of energy parameters. As the distribution of RNA secondary structures follows the Boltzmann distribution, the MFE is equivalent to the maximum likelihood estimation (MLE), which predicts a secondary structure with the maximum probability under the Boltzmann distribution. An alternative scheme, known as the maximum expected accuracy (MEA) approach, has been proposed, which predicts a secondary structure that maximizes the expected number of correctly predicted base pairs under the Boltzmann distribution, rather than predicting the MFE or MLE structure. The expected number of correctly predicted base pairs is calculated using the McCaskill algorithm [51], which is derived by replacing the ‘min’ operation of the Zuker algorithm with the ‘logsumexp’ operation. This scheme, first proposed in Knudsen et al.[52], is also implemented in various tools, such as CONTRAfold, RNAfold and RNAstructure. Additionally, Hamada et al.[53] have redefined MEA to be more compatible with the accuracy metrics for predicting RNA secondary structures, and developed CentroidFold [53, 54] using Turner’s parameters, CONTRAfold’s parameters and the Boltzmann likelihood parameters by Andronescu et al. [48].

The nearest neighbor model for pseudoknot-free structures was extended by incorporating additional parameters for pseudoknot substructures, and thus nearest neighbor models for pseudoknots, such as the Rivas–Eddy model [55], the Dirks–Pierce model [56] and the Cao–Chen model [57], were developed to model RNA secondary structures that include pseudoknots. Algorithms to compute exact minimum free energies on these models through dynamic programming have been implemented as PKNOTS [55] and NUPACK [56], with a significant computational cost of time for PKNOTS, time for NUPACK, and space for both for the limited complexity of pseudoknotted structures. For more accurate prediction of secondary structures including pseudoknots, further efforts were made to use machine learning techniques to estimate more accurate energy parameters. Andronescu et al. [58] used HotKnots [59], which can rapidly predict pseudoknotted structures through heuristics on the Dirks–Pierce model, to train its parameters from training data using the same methodology as SimFold. In contrast, IPknot [60, 61] utilizes the results of learning the parameters of the nearest neighbor model for pseudoknot-free structures, such as CONTRAfold, from training data and forcibly predicts pseudoknotted structures through a heuristic using integer programming.

Probabilistic generative models

The use of stochastic context-free grammars (SCFGs) as a probabilistic generative model for modeling RNA secondary structures without pseudoknots was first proposed by Eddy and Durbin [62] and Sakakibara et al. [63]. A variant of this approach, known as covariance models, has been applied to the popular RNA homology search tool, Infernal [64]. Building covariance models requires highly accurate RNA multiple sequence alignments, which enable accurate and robust homology searches due to the evolutionary information provided by the alignment. Pfold [52, 65] is a method for RNA secondary structure prediction that utilizes simple context-free grammars. Dowell et al. [66] compared nine lightweight grammars for RNA secondary structure prediction, including the Pfold grammar. Sato et al. [67] proposed a method for learning RNA grammars with appropriate complexity using a non-parametric Bayesian approach. TORNADO [24] is a flexible framework that can describe a variety of RNA grammars, allowing SCFGs to emulate the nearest neighbor model with Turner’s parameters or CONTRAfold, and demonstrated prediction accuracy comparable with their counterparts.

Since RNA secondary structure prediction including pseudoknots is beyond the capacity of context-free grammars, context-sensitive grammars, such as tree-adjoining grammars [68, 69] and multiple context-free grammars [70], are alternatively used for predicting pseudoknotted structures. However, due to their large computational complexity, it is impractical to use them for secondary structure prediction including pseudoknots.

Deep Learning models

Deep learning techniques have been leveraged to achieve groundbreaking advancements in a plethora of fields, including the life sciences, and have been applied to the prediction of RNA secondary structures. A majority of deep learning-based methods for RNA secondary structure prediction make no assumptions about the structures themselves, such as the nearest neighbor model and probabilistic generative models. Instead, these methods perform secondary structure prediction by solving multiple binary classification problems for all combinations of two bases in a given RNA sequence, determining whether each of the two bases form a base pair or not. In order to address the constraints that RNA secondary structures must satisfy, such as the restriction that each base can only form a base pair with at most one other base, methods such as E2Efold [71] and UFold [72] utilize linear programming, while Akiyama et al. [73] employ integer programming, originated from IPknot [60, 61]. SPOT-RNA [74] did not aim to solve such constraints, instead attempting to predict base triplets, and employed ensemble of multiple networks with different hyperparameters to mitigate overfitting. UFold, on the other hand, aimed to reduce overfitting by utilizing data augmentation, through the random generation of a large number of artificial RNA sequences and their secondary structures predicted by CONTRAfold, as additional training data.

Unlike other deep learning-based approaches, MXfold2 [50] employs deep learning to infer the energy of decomposed loop substructures within the nearest neighbor model, subsequently utilizing the Zuker algorithm to predict RNA secondary structures. To mitigate overfitting, MXfold2 introduces thermodynamic regularization, ensuring that the energy of the secondary structure calculated by MXfold2 does not deviate significantly from the free energy calculated using Turner’s parameters.

It has been acknowledged that the utilization of deep learning for RNA secondary structure prediction can easily result in overfitting, owing to the high number of parameters that require training. This implies that the accuracy of secondary structure prediction for structurally dissimilar RNA sequences from the training data is not particularly high if no efforts are taken to prevent overfitting [50, 73, 75, 76]. For example, E2Efold [71] was not designed against overfitting, and in their benchmark experiments, the training and test datasets were created by randomly splitting a single dataset. Consequently, overfitting could not be detected because structurally similar sequences were included in training and test datasets. This resulted in very low prediction accuracy of E2Efold for families not included in the training data, which is unfortunately not practical.

Comparison of the three computational models

Table 1 summarizes the de novo RNA secondary structure prediction tools presented in this review that are currently available. The nearest neighbor model is based on the knowledge of RNA secondary structures, which has been extensively studied for a long time. To date, the mainstream approach for predicting RNA secondary structures has been to conduct wet-lab experiments or employ machine learning techniques to determine the energy parameters of the nearest neighbor model. Probabilistic generative models, on the other hand, provide a framework for describing RNA structure modeling through the use of formal grammars. As demonstrated by Rivas et al. [24], the nearest neighbor model can be articulated by SCFGs, with prediction accuracy that is comparable with that of its nearest neighbor model counterpart. However, to date, no RNA grammar has been developed that surpasses the prediction accuracy of the nearest neighbor model.

Table 1.

List of de novo RNA secondary structure prediction tools presented in this paper that are currently available.

	Model	TM	ML	MEA	PK	Ref.	Year	Related URL
RNAstructure	NN	✓				[11, 18]	1999	https://www.urmc.rochester.edu/rna/
PKNOTS	NN	✓			✓	[55]	1999	https://github.com/EddyRivasLab/PKNOTS
Mfold/UNAfold	NN	✓				[14, 15]	2003	http://www.unafold.org/
RNAfold	NN	✓				[16, 17]	2003	https://www.tbi.univie.ac.at/RNA/
NUPACK	NN	✓			✓	[56]	2003	https://www.nupack.org/
CONUS	PG		MLE			[66]	2004	http://eddylab.org/software/conus/
HotKnots	NN	✓	CG		✓	[58, 59]	2005	https://www.cs.ubc.ca/labs/algorithms/Software/HotKnots/
CONTRAfold	NN		CLLM	✓		[21, 22]	2006	http://contra.stanford.edu/contrafold/
SimFold	NN	✓	CG, BL			[47, 48]	2007	https://www.cs.ubc.ca/labs/algorithms/Projects/RNA-Params/
CentroidFold	NN	✓		✓		[53, 54]	2009	https://github.com/satoken/centroid-rna-package
ContextFold	NN		MM			[23]	2011	https://www.cs.bgu.ac.il/~negevcb/contextfold/
IPknot	NN	✓		✓	✓	[60, 61]	2011	https://github.com/satoken/ipknot
TORNADO	PG		MLE	✓		[24]	2012	https://github.com/EddyRivasLab/tornado
MXfold	NN	✓	MM			[49]	2018	https://github.com/mxfold/mxfold
LinearFold	NN	✓				[45]	2019	https://github.com/LinearFold/LinearFold
SPOT-RNA	DL		DL		✓	[74]	2019	https://github.com/jaswindersingh2/SPOT-RNA
E2Efold	DL		DL		✓	[71]	2020	https://github.com/ml4bio/e2efold
MXfold2	NN	✓	DL+MM			[50]	2021	https://github.com/mxfold/mxfold2
EternaFold	NN		multitask	✓		[77]	2022	https://github.com/eternagame/eternafold
Ufold	DL		DL		✓	[72]	2022	https://github.com/uci-cbcl/UFold
NeuralFold	DL		DL		✓	[73]	2022	https://github.com/keio-bioinformatics/Neuralfold

The column labeled ‘Model’ indicates the category to which the method belongs (NN: nearest neighbor model, PG: probabilistic generative model, DL: deep learning-based model). The ‘TM’ column indicates whether the method uses thermodynamic parameters. The column labeled ‘ML’ indicates whether the method can train its parameters using machine learning, and if so, which training method is used (MLE: maximum likelihood estimation, CG: constraint generation, CLLM: conditional log-linear models, BL: Boltzmann likelihood, MM: max-margin framework, DL: deep larning). The ‘MEA’ column indicates whether the method predicts secondary structure by default (✓) or optionally () with the maximum expected accuracy. The column labeled ‘PK’ indicates whether the method can predict pseudoknotted structures.

Conversely, full deep learning methods, with the exception of MXfold2, do not rely on knowledge of RNA secondary structures, thereby allowing for a high degree of freedom in model construction. This can lead to improved fitting of the training data, and thus high prediction accuracy for RNA sequences with structures similar to those in the training data. However, this also increases the risk of overfitting and poor prediction accuracy for structurally dissimilar sequences. The problem of overfitting is a prevalent issue not only in deep learning but also in other machine learning techniques with rich parametrization; it is particularly acute in deep learning as models can easily be scaled to an enormous number of parameters [75, 78].

Datasets for building models

Frequently employed benchmark datasets for RNA secondary structure prediction are summarized in Table 2, which include RNA STRAND dataset [79], Archive II dataset [80] and RNAStralign dataset [81]. These benchmark datasets were constructed by compiling RNA sequences with known and reliable secondary structures from various databases such as Comparative RNA Web (CRW) Site [83], tmRNA database [84], Sprinzl tRNA Database [85], RNase P Database [86], SRP Database [84] and others. These benchmark datasets, however, are limited in their diversity of RNA secondary structures, containing only 8–10 RNA families. More recently, a more comprehensive dataset, bpRNA-1m dataset [82], has been constructed by incorporating RNA sequences with secondary structure annotations from Rfam 12.2 [4] in addition to RNA sequences derived from the aforementioned databases, comprising 102 318 sequences from approximately 2600 RNA families.

Table 2.

Benchmark datasets commonly used in recent machine learning- and deep learning-based methods.

	# fam.	# seq.	Length	Ref.
RNA STRAND	172	4666	4 ~ 4381	[79]
TORNADO dataset				[24]
TrainSetA	11	3166	10 ~ 734
TestSetA	11	697	10 ~ 768
TrainSetB	22	1094	27 ~ 237
TestSetB	22	430	27 ~ 244
Archive II	10	3975	28 ~ 2968	[80]
RNAStralign	8	30 451	30 ~ 1851	[81]
bpRNA-1m	2600	102 318	2 ~ 4381	[82]
bpRNAnew	1500	5401	33 ~ 489	[50]

# fam.: the number of families. # seq.: the number of sequences.

In many previously conducted benchmark experiments for RNA secondary structure prediction methods, these datasets have been randomly partitioned into training and test data for cross-validation. This means that the test data may not contain highly homologous sequences to those in the training data, but structurally similar sequences from the same families.

Rivas et al. [24, 25] have highlighted that accuracy evaluations utilizing ‘sequence-wise cross-validation’ cannot detect overfitting, and subsequently established TrainSetA, TestSetA, TrainSetB and TestSetB. TrainSetA and TestSetA were compiled from literature sources [21, 47, 48, 66, 85, 87], while TrainSetB and TestSetB, comprising 22 families with 3D structure annotations, were extracted from Rfam 10.0 [3]. The sequences in Train/TestSetB share less than 70% sequence identity with the sequences in TrainSetA. It is worth mentioning that the literature-based TrainSetA and Rfam-based TestSetB share no families and are structurally dissimilar, while TrainSetA and TestSetA exhibit some structural similarity. By conducting ‘family-wise cross-validation’, in which training is performed using TrainSetA and subsequently testing with TestSetB, Rivas et al. [24, 25] demonstrated that ContextFold, which exhibits an extremely high prediction accuracy in sequence-wise cross-validation, is prone to overfitting. Similar trends have been reported for deep learning-based methods by Szikszai et al. [75].

To further perform family-wise cross-validation on models trained from a bpRNA-1m-based dataset, Sato et al. [50] constructed bpRNAnew dataset from approximately 1500 families newly discovered and registered with Rfam since the release of Rfam version 12.2 (January 2017), on which bpRNA-1m dataset is based.

In general, machine learning and deep learning algorithms operate under the assumption that training data are randomly and uniformly sampled from the population. However, this assumption is often invalidated, even in large datasets such as bpRNA-1m dataset, as novel RNA families continue to be discovered to this day. If prior knowledge of the family of a sequence of interest is available, a homology search tool such as Infernal [64] can be more accurate in predicting its structure. However, when no knowledge of the family is available, it is necessary to make RNA secondary structure predictions solely from the sequence in order to perform structural and functional analyses. Therefore, it is crucial to evaluate prediction accuracy in such situations through the use of family-wise cross-validation, avoiding methods with significantly lower accuracy for unknown families.

APPLICATIONS OF RNAs TO THERAPEUTICS

Currently, the majority of pharmaceuticals consist of small molecules that target disease-associated proteins in small molecule drug discovery. However, the limitations of this approach are well known, and there is a pressing need for alternative drug modalities. To address these limitations, synthetic and natural RNAs have garnered significant attention as potential drug candidates and targets, respectively. This section outlines the use of machine-learning and deep-learning techniques in the discovery and development of RNA-based drugs, including the use of synthetic RNAs as drugs and natural RNAs as targets for drug discovery. We will also discuss the importance of RNA structures in these approaches (Figure 5B).

Figure 5.

(A) Schematic representation of the RNA aptamer development process and (B) the significance of RNA structure in each RNA therapeutic approach. (A) The basic and exploratory research comprises multiple stages, encompassing HT-SELEX, candidate identification from the HT-SELEX output, assessment of the selected candidates’ activity and sequence truncation/optimization. (B) The importance of RNA structure (right) in three RNA-based therapeutics (left) that have been highlighted in this review.

RNA drug discovery—making RNAs into drugs

RNA aptamers

RNA aptamers are single-stranded RNA molecules that range in length from 20 to 50 bases and form specific three-dimensional structures based on their nucleotide sequence. These structures allow RNA aptamers to fit into the shape of target substances, such as disease-related proteins, and act as drugs; it is therefore important to consider RNA structures in aptamer design. RNA aptamers have several advantages over traditional drugs, including high affinity and specificity, the ability to be designed for a wide range of target molecules, including membrane proteins, and low immunogenicity. As a result, RNA aptamers are considered to be potential next-generation drugs. As of February 2022, only one RNA aptamer, Macugen® (pegaptanib), has been approved for the treatment of age-related macular degeneration.

The process of creating RNA aptamers follows a similar path to traditional drug development, comprising several stages such as basic and exploratory research, preclinical study, clinical study, application and production and distribution (Figure 5A). Many computational studies on RNA aptamers focus on the initial stage of basic and exploratory research, which is further divided into several steps. The first step involves obtaining candidate aptamer sequences using a technique called SELEX (Systematic Evolution of Ligands by EXponential enrichment) [88]. This process involves repeatedly binding and amplifying RNA sequences that bind strongly to the target from a pool of random sequences, enriching for aptamers with high binding activity. High-throughput sequencing techniques, known as HT-SELEX (High-Throughput SELEX), enable comprehensive measurement of the sequence information in the enriched sequence pool at each round, generating a large amount of sequence data in each round of HT-SELEX.

A number of computational approaches have been proposed to improve the identification of aptamers from HT-SELEX data [89–91], including sequence/structure clustering-based methods [92–94], motif-based methods [95, 96], scoring-based methods [97, 98] and so forth. Here, we review some of these approaches that utilize machine learning and deep learning and discuss their effective utilization.

Bashir et al. [99] proposed a machine learning (ML)-guided Particle Display methodology (MLPD), which integrates physical experiments and machine learning. They used particle display (PD) to partition aptamer libraries according to affinity and used these data to train machine learning models. This method allows for the in silico prediction of aptamer affinity, and the authors were able to successfully identify novel aptamers with enhanced affinity.

RaptGen [100] employs a combination of a variational autoencoder (VAE) and a profile hidden Markov model (HMM) to effectively model aptamers with local motifs that contain substitutions and indels. The latent spaces learned by RaptGen are used for several purposes: (i) visualizing SELEX data and generating novel aptamers, (ii) optimizing aptamers using Bayesian optimization with additional information, such as detailed affinity scores obtained through surface plasmon resonance experiments and (iii) designing shortened (truncated) aptamers, which is realized using shorter profile HMM in RaptGen model. While the current version of RaptGen does not take structural information of RNA into account, this can be addressed using a profile stochastic context grammar (SCFG; see the previous section) instead of a profile HMM.

Di Gioacchino et al. [101] created a restricted Boltzmann machine (RBM) model using SELEX data for thrombin, which is a probabilistic generative model capable of generating novel aptamer sequences (similar with RaptGen). They demonstrated that the log-likelihood of sequences correlates with their fitness (i.e. binding ability to the target).

Recently, Andress et al. [102] proposed a method called Daptev, which combines a deep generative model (VAE) and molecular simulation (molecular docking). As both data-driven and simulation-based approaches (considering tertiary structures of aptamers) can be useful for in silico aptamer design, this may be a promising approach.

Note that the above-mentioned approaches, such as RaptGen and MLPD, assume a target protein with experimental data (e.g. SELEX). More general (and difficult) problem setting that predicts pairs of aptamer-protein is introduced. AptaNet [103] uses a multi-layer perceptron (MLP) to predict aptamer–protein pairs, taking a pair of RNA and amino acid sequences as input. The training dataset for this model was compiled from Aptagen and Aptamer Base and consists of 850 positive and 2554 negative instances. In Torkamanian-Afshar et al. [104], a classifier for aptamer–protein pairs was trained using the sequential and structural properties of known aptamer–protein complexes, utilizing positive and negative data from RPINBASE [105]. This classifier was then used to select target-binding RNA sequences as a potential biomarker for aminopeptidase N (CD13). These sequences were utilized as the starting population for a genetic algorithm (GA) to generate new aptamers that exhibit higher selectivity for binding to CD13 compared with the original ones.

In contrast, a wealth of data on RNA sequences that bind to various natural RNA-binding proteins (RBPs) have been accumulated, and research utilizing these data through machine learning and deep learning methods is ongoing [106]. For example, Yamada et al. [107] proposed a BERT (Bidirectional Encoder Representations from Transformers)-based model for predicting RNA–protein interactions with biological implications, and Kashiwagi et al. [108] introduced a max-margin model for predicting residue-level contact in RNA–protein interactions. These studies on natural RBPs could potentially inform the design of artificial RNA aptamers that target proteins; for further details, see the reference [109].

As previously mentioned, beneficial machine learning techniques for in silico aptamer design have been developed in recent years and are anticipated to aid in the expansion of aptamer drug discovery in the future. Furthermore, the refinement of RNA structure prediction contributes significantly to the development of highly effective RNA aptamers.

mRNA vaccines

Since 2020, the development of coronavirus disease 2019 vaccines such as BNT162b2 (Pfizer/BioNTech) and mRNA-1273 (Moderna) has been active, and the drug discovery modality of mRNA medicine has garnered significant attention [110–112]. mRNA vaccines have also been proposed as a potential therapeutic approach for cancer [113]. To facilitate the rapid production of mRNA vaccines, computational design of mRNA sequences is crucial, involving the comprehensive design of 5’ untranslated region (UTR), coding sequence (CDS) and 3’UTR sequences.

The sequence of the 5’UTR is closely correlated with translation efficiency. A study by Sample et al. [114] developed a convolutional neural network (CNN) model to predict Mean Ribosome Load (MRL), a measure of ribosome association, for given 5’UTR sequence. The authors used MRL measurements from 280 000 random 5’UTRs as a training dataset for the model. Utilizing this CNN model and a GA, the authors were able to generate 5’UTR sequences with a desired MRL value. These comprehensive data will be beneficial for 5’UTR design in the development of mRNA vaccines.

The sequence of CDS also impacts the abundance of translation. iCodon [115] is a tool designed to optimize coding regions that contain synonymous codon substitutions in order to increase mRNA stability and protein expression (e.g. designing high-expression reporters) or de-optimize sequences containing synonymous codon substitutions (e.g. designing sequences with reduced expression). The prediction model of mRNA stability is proposed in Medina-Muñoz et al. [116]. Zhang et al. [117] proposed an efficient method, named LinearDesign, for mRNA design by reducing it to a problem in computational linguistics. The optimal mRNA is analogous to finding the most probable sentence among similar-sounding alternatives. The algorithm takes 11 min for the Spike protein and can optimize stability and codon usage concurrently.

According to Leppek et al. [118], a method for optimizing the structure, stability and translation of mRNA through combinatorial means was introduced. Initially, viral and cellular UTRs mined from literature were procured, followed by structure-informed CDS design in which Eterna (crowdsourced) [77] and the LinearDesign were utilized as efficient design tools.

RNA-targeted drug discovery—making RNAs into drug target

Another challenge in drug discovery is the depletion of potential drug targets. Currently, disease-related proteins are the primary targets for drug discovery. Recent research has shown that lncRNAs play a vital role in a variety of intracellular regulatory processes in eukaryotes, including humans [119, 120]. Moreover, many lncRNAs have been found to be associated with serious diseases such as cancer and neurodegenerative disorders, making them potential new drug targets [121–123].

A strategy for RNA-targeted drug discovery is to design small molecules (i.e., traditional drugs) that bind to RNA structures in lncRNAs [124–126] and ribo-switches [127], indicating that the consideration of RNA structures is essential in this kind of researches. Although there exists limited studies for RNA–target drug discovery using machine learning [128], we will introduce a few studies in the following.

RNAmigos [129] is a tool that constructs and encodes network representations of RNA structures and predicts potential ligands for novel binding sites. It employs a graph convolutional neural network (GCN) to represent RNA structures as an Augmented Base Pairing Network (ABPN), including both canonical and non-canonical base-pairs. The training data were sourced from the RNA–ligand pairs in the RCSB PDB Data Bank [130].

A recent study by Yazdani et al. [131] that analyzed data from screening experiments suggests that there may be a correlation between the properties of RNA and the properties of small molecule ligands that bind to RNA. Using machine learning methods to analyze their own library of RNA-bound small molecules, the authors found that general chemical properties of RNA-bound small molecule compared with protein-bound small molecules and FDA-approved drugs.

Stefaniak et al. [132] developed AnnapuRNA, a machine-learning statistical scoring function, to accurately predict the structure of RNA-ligand complexes. Their program utilizes a coarse-grained representation for both the RNA and small molecule ligands involved in the interaction. On the other hand, Chhabra et al. [133] used a distance-dependent fingerprint to characterize the binding pose of a ligand in an RNA binding pocket (RNAPosers). They trained a machine-learning algorithm using data from 80 experimentally determined RNA–ligand complexes and used it to score docking poses.

Grimberg et al. [134] sought to design novel small molecule inhibitors that would bind to the RNA hairpin within the ribosomal peptidyl transferase center (PTC) of Mycobacterium tuberculosis through the use of computational optimization models integrating CNNs with classical machine learning regression and decision tree models, using approximately 800 training data points [135]. Upon synthesizing the 10 small molecules identified by these computational means, functional validation was conducted, revealing that four of the molecules were potent inhibitors targeting hairpin 91 in the ribosomal PTC of M. tuberculosis, thereby inhibiting translation. This study demonstrates the potential for optimizing RNA-binding drugs with sufficient training data.

It should be noted that all of the aforementioned methods assume targeted RNA elements of relatively small size. However, it is crucial to identify specific RNA elements (such as structures, modifications and binding sites of other biomolecules) that are suitable for drug targeting, as lncRNAs are lengthy and the location of functional elements can be challenging to determine. To this end, various approaches have been proposed, including infoRNA [136]. While machine learning-based approaches in this direction are limited, these approaches may prove useful in identifying functional elements in lncRNAs in the future if sufficient data are available.

CHALLENGES AND OPPORTUNITIES

The accuracy of RNA secondary structure prediction has considerably improved in recent years due to the utilization of machine learning and deep learning techniques. One potential avenue for further enhancement in prediction accuracy is the incorporation of evolutionary information from homologous sequences, which can be achieved through methods such as common secondary structure prediction from multiple sequence alignments [26, 27], probabilistic consistency transformation of base-pairing probabilities from homologous sequences [28, 29], MSA transformers [137] and the utilization of pre-trained large language models, such as BERT [107, 138, 139].

Additionally, high-throughput experiments such as selective 2’-hydroxyl acylation analyzed by primer extension (SHAPE) [19] and dimethyl sulphate (DMS) [20], which can stochastically induce chemical modifications on unpaired nucleotides, have been shown to improve the accuracy of secondary structure predictions. The incorporation of pseudo-free energy calculated using the reactivity of chemical probing from high-throughput experiments has also been shown to substantially enhance the accuracy of thermodynamics-based RNA secondary structure prediction [140]. However, despite the potential benefits, few machine learning methods have been developed to predict RNA secondary structures from RNA sequences with chemical reactivity due to a lack of a large amount of training data including not only RNA sequences and their structure, but also their chemical reactivities. EternaFold [77] augmented the accuracy of secondary structure prediction by refining the parameters of the nearest neighbor model via multitask learning with high-throughput experimental data that lack secondary structure annotations, thereby demonstrating the potential of using high-throughput experimental data in machine learning-based RNA secondary structure prediction. Currently, the accuracy of RNA secondary structure prediction is still insufficient for long RNA sequences longer than 500 bases. One of the reasons for this is that the number of long RNAs with known secondary structures is small, and models that can handle long sequences cannot be sufficiently trained by machine learning or deep learning. If secondary structure prediction models can be trained from high-throughput experimental data, which are easily available even for long sequences, the accuracy of secondary structure prediction for long sequences is expected to improve.

RNA modifications play a significant role in various biological processes including splicing, translation, cell development and disease [141, 142]. In mRNA vaccines, all uridines are modified to N1-methylpseudouridines, which enables them to bypass the Toll-like receptors (TLRs) that detect RNA viruses and thus produce viral proteins [143]. Due to the need for modified free energy parameters and potential alteration of base pairing partners, the development of RNA secondary structure prediction methods that can consider RNA modifications is limited [144]. However, as RNA modifications are more prevalent in vivo than previously thought, and the increasing demand for mRNA vaccine stability prediction and other applications make the development of highly accurate RNA modification-aware RNA secondary structure prediction by machine learning an urgent task. However, this is a challenging task due to the scarcity of data of RNA sequences containing modified bases with secondary structures available.

It is highly demanding to establish high-throughput methods for determining RNA 3D structures, not only for RNA structural and functional analysis, but also for RNA drug discovery and RNA-targeted drug discovery. AlphaFold2 [145] has achieved highly accurate protein 3D structure prediction comparable with experimental structure determination. Inspired by AlphaFold2, similar deep learning approaches have been applied to tackle RNA 3D structure prediction and have been reported to perform well on their datasets [146–148]. However, in the competition for RNA 3D structure prediction held at the most recent CASP 15 (https://predictioncenter.org/casp15/), these deep learning-based RNA 3D structure prediction methods were not at all comparable with conventional approaches. The number of 3D structures registered in Protein Data Bank (PDB) [149] is 173 649 for proteins, but only 1682 for RNAs (December 2022). Therefore, highly accurate RNA 3D structure prediction without falling into overfitting is presumed to be challenging with fully deep learning-based approaches like AlphaFold2.

CONCLUSION

In recent years, there has been a growing interest in structure-based RNA analysis, as it is believed that the function of many RNAs is closely related to their structures. In this paper, we have reviewed the latest advancements in RNA secondary structure prediction, which is a fundamental technique for structure-based RNA analysis, particularly in methods that utilize machine learning and deep learning. We have also discussed the use of machine learning and deep learning in RNA drug discovery and RNA-targeted drug discovery, which are among the most notable applications of structure-based RNA analysis in recent times. It is important to note that compared with proteins, there are orders of magnitude fewer known samples of RNA structures and interactions with other molecules. Therefore, when applying machine learning and deep learning techniques to analyze RNAs, it is essential to be cognizant of the fact that the training data may be small, biased or both, and to implement various strategies to enhance generalization capabilities.

Key Points

• In this review, we have outlined the field of RNA secondary structure prediction, focusing particularly on methods that utilize machine learning and deep learning.

• It is important to note that, in order to maintain the prediction accuracy of these methods, the test data used for benchmarking must be carefully constructed to detect any potential overfitting.

• Fundamental technologies of RNA informatics are applicable to the development of RNA-based therapeutics.

• We provided a review of RNA drug discovery and RNA–target drug discovery, in which various machine learning and deep learning approaches are effectively utilized.

Author Biographies

Kengo Sato is a professor at Tokyo Denki University, Japan. He received his Ph.D. in Computer Science from Keio University, Japan, in 2003. His research interests include bioinformatics, computational linguistics and machine learning.

Michiaki Hamada is a professor at Waseda University in Japan, and is concurrently appointed at National Institute of Advanced Industrial Science and Technology (AIST) and Nippon Medical University. He obtained his Ph.D. in Science from the Tokyo Institute of Technology in 2009. His research interests encompass RNA bioinformatics, drug discovery and artificial intelligence.

FUNDING

This work was partially supported by a Grant-in-Aid for Scientific Research (B) (No. 22H03689) from the Japan Society for the Promotion of Science (JSPS) to K.S., and by JST CREST (grant nos. JPMJCR1881 and JPMJCR21F1), AMED (Nos. 21479280, JP22ama121055) and JSPS KAKENHI (Nos. 22H04925, 20H00624, 17K20032) to M.H.