Caveats to Deep Learning Approaches to RNA Secondary Structure Prediction

RNA structure prediction using deep learning - A comprehensive review

In computational biology, accurate RNA structure prediction offers several benefits, including facilitating a better understanding of RNA functions and RNA-based drug design. Implementing deep learning techniques for RNA structure prediction has led tremendous progress in this field, resulting in significant improvements in prediction accuracy. This comprehensive review aims to provide an overview of the diverse strategies employed in predicting RNA secondary structures, emphasizing deep learning methods. The article categorizes the discussion into three main dimensions: feature extraction methods, existing state-of-the-art learning model architectures, and prediction approaches. We present a comparative analysis of various techniques and models highlighting their strengths and weaknesses. Finally, we identify gaps in the literature, discuss current challenges, and suggest future approaches to enhance model performance and applicability in RNA structure prediction tasks. This review provides a deeper insight into the subject and paves the way for further progress in this dynamic intersection of life sciences and artificial intelligence.

Robust RNA secondary structure prediction with a mixture of deep learning and physics-based experts

A mixture-of-experts (MoE) approach has been developed to mitigate the poor out-of-distribution (OOD) generalization of deep learning (DL) models for single-sequence-based prediction of RNA secondary structure. The main idea behind this approach is to use DL models for in-distribution (ID) test sequences to leverage their superior ID performances, while relying on physics-based models for OOD sequences to ensure robust predictions. One key ingredient of the pipeline, named MoEFold2D, is automated ID/OOD detection via consensus analysis of an ensemble of DL model predictions without requiring access to training data during inference. Specifically, motivated by the clustered distribution of known RNA structures, a collection of distinct DL models is trained by iteratively leaving one cluster out. Each DL model hence serves as an expert on all but one cluster in the training data. Consequently, for an ID sequence, all but one DL model makes accurate predictions consistent with one another, while an OOD sequence yields highly inconsistent predictions among all DL models. Through consensus analysis of DL predictions, test sequences are categorized as ID or OOD. ID sequences are subsequently predicted by averaging the DL models in consensus, and OOD sequences are predicted using physics-based models. Instead of remediating generalization gaps with alternative approaches such as transfer learning and sequence alignment, MoEFold2D circumvents unpredictable ID-OOD gaps and combines the strengths of DL and physics-based models to achieve accurate ID and robust OOD predictions.

Identification of RNA structures and their roles in RNA functions

The development of high-throughput RNA structure profiling methods in the past decade has greatly facilitated our ability to map and characterize different aspects of RNA structures transcriptome-wide in cell populations, single cells and single molecules. The resulting high-resolution data have provided insights into the static and dynamic nature of RNA structures, revealing their complexity as they perform their respective functions in the cell. In this Review, we discuss recent technical advances in the determination of RNA structures, and the roles of RNA structures in RNA biogenesis and functions, including in transcription, processing, translation, degradation, localization and RNA structure-dependent condensates. We also discuss the current understanding of how RNA structures could guide drug design for treating genetic diseases and battling pathogenic viruses, and highlight existing challenges and future directions in RNA structure research.

RNA3DB: A structurally-dissimilar dataset split for training and benchmarking deep learning models for RNA structure prediction

With advances in protein structure prediction thanks to deep learning models like AlphaFold, RNA structure prediction has recently received increased attention from deep learning researchers. RNAs introduce substantial challenges due to the sparser availability and lower structural diversity of the experimentally resolved RNA structures in comparison to protein structures. These challenges are often poorly addressed by the existing literature, many of which report inflated performance due to using training and testing sets with significant structural overlap. Further, the most recent Critical Assessment of Structure Prediction (CASP15) has shown that deep learning models for RNA structure are currently outperformed by traditional methods. In this paper we present RNA3DB, a dataset of structured RNAs, derived from the Protein Data Bank (PDB), that is designed for training and benchmarking deep learning models. The RNA3DB method arranges the RNA 3D chains into distinct groups (Components) that are non-redundant both with regard to sequence as well as structure, providing a robust way of dividing training, validation, and testing sets. Any split of these structurally-dissimilar Components are guaranteed to produce test and validations sets that are distinct by sequence and structure from those in the training set. We provide the RNA3DB dataset, a particular train/test split of the RNA3DB Components (in an approximate 70/30 ratio) that will be updated periodically. We also provide the RNA3DB methodology along with the source-code, with the goal of creating a reproducible and customizable tool for producing structurally-dissimilar dataset splits for structural RNAs.

Discovery and Quantification of Long-Range RNA Base Pairs in Coronavirus Genomes with SEARCH-MaP and SEISMIC-RNA

RNA molecules perform a diversity of essential functions for which their linear sequences must fold into higher-order structures. Techniques including crystallography and cryogenic electron microscopy have revealed 3D structures of ribosomal, transfer, and other well-structured RNAs; while chemical probing with sequencing facilitates secondary structure modeling of any RNAs of interest, even within cells. Ongoing efforts continue increasing the accuracy, resolution, and ability to distinguish coexisting alternative structures. However, no method can discover and quantify alternative structures with base pairs spanning arbitrarily long distances - an obstacle for studying viral, messenger, and long noncoding RNAs, which may form long-range base pairs. Here, we introduce the method of Structure Ensemble Ablation by Reverse Complement Hybridization with Mutational Profiling (SEARCH-MaP) and software for Structure Ensemble Inference by Sequencing, Mutation Identification, and Clustering of RNA (SEISMIC-RNA). We use SEARCH-MaP and SEISMIC-RNA to discover that the frameshift stimulating element of SARS coronavirus 2 base-pairs with another element 1 kilobase downstream in nearly half of RNA molecules, and that this structure competes with a pseudoknot that stimulates ribosomal frameshifting. Moreover, we identify long-range base pairs involving the frameshift stimulating element in other coronaviruses including SARS coronavirus 1 and transmissible gastroenteritis virus, and model the full genomic secondary structure of the latter. These findings suggest that long-range base pairs are common in coronaviruses and may regulate ribosomal frameshifting, which is essential for viral RNA synthesis. We anticipate that SEARCH-MaP will enable solving many RNA structure ensembles that have eluded characterization, thereby enhancing our general understanding of RNA structures and their functions. SEISMIC-RNA, software for analyzing mutational profiling data at any scale, could power future studies on RNA structure and is available on GitHub and the Python Package Index.

Discovery and Quantification of Long-Range RNA Base Pairs in Coronavirus Genomes with SEARCH-MaP and SEISMIC-RNA

RNA molecules perform a diversity of essential functions for which their linear sequences must fold into higher-order structures. Techniques including crystallography and cryogenic electron microscopy have revealed 3D structures of ribosomal, transfer, and other well-structured RNAs; while chemical probing with sequencing facilitates secondary structure modeling of any RNAs of interest, even within cells. Ongoing efforts continue increasing the accuracy, resolution, and ability to distinguish coexisting alternative structures. However, no method can discover and quantify alternative structures with base pairs spanning arbitrarily long distances - an obstacle for studying viral, messenger, and long noncoding RNAs, which may form long-range base pairs. Here, we introduce the method of Structure Ensemble Ablation by Reverse Complement Hybridization with Mutational Profiling (SEARCH-MaP) and software for Structure Ensemble Inference by Sequencing, Mutation Identification, and Clustering of RNA (SEISMIC-RNA). We use SEARCH-MaP and SEISMIC-RNA to discover that the frameshift stimulating element of SARS coronavirus 2 base-pairs with another element 1 kilobase downstream in nearly half of RNA molecules, and that this structure competes with a pseudoknot that stimulates ribosomal frameshifting. Moreover, we identify long-range base pairs involving the frameshift stimulating element in other coronaviruses including SARS coronavirus 1 and transmissible gastroenteritis virus, and model the full genomic secondary structure of the latter. These findings suggest that long-range base pairs are common in coronaviruses and may regulate ribosomal frameshifting, which is essential for viral RNA synthesis. We anticipate that SEARCH-MaP will enable solving many RNA structure ensembles that have eluded characterization, thereby enhancing our general understanding of RNA structures and their functions. SEISMIC-RNA, software for analyzing mutational profiling data at any scale, could power future studies on RNA structure and is available on GitHub and the Python Package Index.

sincFold: end-to-end learning of short- and long-range interactions in RNA secondary structure

MOTIVATION

Coding and noncoding RNA molecules participate in many important biological processes. Noncoding RNAs fold into well-defined secondary structures to exert their functions. However, the computational prediction of the secondary structure from a raw RNA sequence is a long-standing unsolved problem, which after decades of almost unchanged performance has now re-emerged due to deep learning. Traditional RNA secondary structure prediction algorithms have been mostly based on thermodynamic models and dynamic programming for free energy minimization. More recently deep learning methods have shown competitive performance compared with the classical ones, but there is still a wide margin for improvement.

RESULTS

In this work we present sincFold, an end-to-end deep learning approach, that predicts the nucleotides contact matrix using only the RNA sequence as input. The model is based on 1D and 2D residual neural networks that can learn short- and long-range interaction patterns. We show that structures can be accurately predicted with minimal physical assumptions. Extensive experiments were conducted on several benchmark datasets, considering sequence homology and cross-family validation. sincFold was compared with classical methods and recent deep learning models, showing that it can outperform the state-of-the-art methods.

RNA3DB: A structurally-dissimilar dataset split for training and benchmarking deep learning models for RNA structure prediction

With advances in protein structure prediction thanks to deep learning models like AlphaFold, RNA structure prediction has recently received increased attention from deep learning researchers. RNAs introduce substantial challenges due to the sparser availability and lower structural diversity of the experimentally resolved RNA structures in comparison to protein structures. These challenges are often poorly addressed by the existing literature, many of which report inflated performance due to using training and testing sets with significant structural overlap. Further, the most recent Critical Assessment of Structure Prediction (CASP15) has shown that deep learning models for RNA structure are currently outperformed by traditional methods. In this paper we present RNA3DB, a dataset of structured RNAs, derived from the Protein Data Bank (PDB), that is designed for training and benchmarking deep learning models. The RNA3DB method arranges the RNA 3D chains into distinct groups (Components) that are non-redundant both with regard to sequence as well as structure, providing a robust way of dividing training, validation, and testing sets. Any split of these structurally-dissimilar Components are guaranteed to produce test and validations sets that are distinct by sequence and structure from those in the training set. We provide the RNA3DB dataset, a particular train/test split of the RNA3DB Components (in an approximate 70/30 ratio) that will be updated periodically. We also provide the RNA3DB methodology along with the source-code, with the goal of creating a reproducible and customizable tool for producing structurally-dissimilar dataset splits for structural RNAs.

Tailor made: the art of therapeutic mRNA design

mRNA medicine is a new and rapidly developing field in which the delivery of genetic information in the form of mRNA is used to direct therapeutic protein production in humans. This approach, which allows for the quick and efficient identification and optimization of drug candidates for both large populations and individual patients, has the potential to revolutionize the way we prevent and treat disease. A key feature of mRNA medicines is their high degree of designability, although the design choices involved are complex. Maximizing the production of therapeutic proteins from mRNA medicines requires a thorough understanding of how nucleotide sequence, nucleotide modification and RNA structure interplay to affect translational efficiency and mRNA stability. In this Review, we describe the principles that underlie the physical stability and biological activity of mRNA and emphasize their relevance to the myriad considerations that factor into therapeutic mRNA design.

AliNA - a deep learning program for RNA secondary structure prediction

Nowadays there are numerous discovered natural RNA variations participating in different cellular processes and artificial RNA, e. g., aptamers, riboswitches. One of the required tasks in the investigation of their functions and mechanism of influence on cells and interaction with targets is the prediction of RNA secondary structures. The classic thermodynamic-based prediction algorithms do not consider the specificity of biological folding and deep learning methods that were designed to resolve this issue suffer from homology-based methods problems. Herein, we present a method for RNA secondary structure prediction based on deep learning - AliNA (ALIgned Nucleic Acids). Our method successfully predicts secondary structures for non-homologous to train-data RNA families thanks to usage of the data augmentation techniques. Augmentation extends existing datasets with easily-accessible simulated data. The proposed method shows a high quality of prediction across different benchmarks including pseudoknots. The method is available on GitHub for free (https://github.com/Arty40m/AliNA).

In-silico study of antisense oligonucleotide antibiotics

Background

The rapid emergence of antibiotic-resistant bacteria directly contributes to a wave of untreatable infections. The lack of new drug development is an important driver of this crisis. Most antibiotics today are small molecules that block vital processes in bacteria. To optimize such effects, the three-dimensional structure of targeted bacterial proteins is imperative, although such a task is time-consuming and tedious, impeding the development of antibiotics. The development of RNA-based therapeutics has catalyzed a new platform of antibiotics-antisense oligonucleotides (ASOs). These molecules hybridize with their target mRNAs with high specificity, knocking down or interfering with protein translation. This study aims to develop a bioinformatics pipeline to identify potent ASO targets in essential bacterial genes.

Methods

Three bacterial species (P. gingivalis, H. influenzae, and S. aureus) were used to demonstrate the utility of the pipeline. Open reading frames of bacterial essential genes were downloaded from the Database of Essential Genes (DEG). After filtering for specificity and accessibility, ASO candidates were ranked based on their self-hybridization score, predicted melting temperature, and the position on the gene in an operon. Enrichment analysis was conducted on genes associated with putative potent ASOs.

Results

A total of 45,628 ASOs were generated from 348 unique essential genes in P. gingivalis. A total of 1,117 of them were considered putative. A total of 27,273 ASOs were generated from 191 unique essential genes in H. influenzae. A total of 847 of them were considered putative. A total of 175,606 ASOs were generated from 346 essential genes in S. aureus. A total of 7,061 of them were considered putative. Critical biological processes associated with these genes include translation, regulation of cell shape, cell division, and peptidoglycan biosynthetic process. Putative ASO targets generated for each bacterial species are publicly available here: https://github.com/EricSHo/AOA. The results demonstrate that our bioinformatics pipeline is useful in identifying unique and accessible ASO targets in bacterial species that post major public health issues.

Automatic recognition of complementary strands: lessons regarding machine learning abilities in RNA folding

Introduction: Prediction of RNA secondary structure from single sequences still needs substantial improvements. The application of machine learning (ML) to this problem has become increasingly popular. However, ML algorithms are prone to overfitting, limiting the ability to learn more about the inherent mechanisms governing RNA folding. It is natural to use high-capacity models when solving such a difficult task, but poor generalization is expected when too few examples are available. Methods: Here, we report the relation between capacity and performance on a fundamental related problem: determining whether two sequences are fully complementary. Our analysis focused on the impact of model architecture and capacity as well as dataset size and nature on classification accuracy. Results: We observed that low-capacity models are better suited for learning with mislabelled training examples, while large capacities improve the ability to generalize to structurally dissimilar data. It turns out that neural networks struggle to grasp the fundamental concept of base complementarity, especially in lengthwise extrapolation context. Discussion: Given a more complex task like RNA folding, it comes as no surprise that the scarcity of useable examples hurdles the applicability of machine learning techniques to this field.

Machine learning modeling of RNA structures: methods, challenges and future perspectives

The three-dimensional structure of RNA molecules plays a critical role in a wide range of cellular processes encompassing functions from riboswitches to epigenetic regulation. These RNA structures are incredibly dynamic and can indeed be described aptly as an ensemble of structures that shifts in distribution depending on different cellular conditions. Thus, the computational prediction of RNA structure poses a unique challenge, even as computational protein folding has seen great advances. In this review, we focus on a variety of machine learning-based methods that have been developed to predict RNA molecules' secondary structure, as well as more complex tertiary structures. We survey commonly used modeling strategies, and how many are inspired by or incorporate thermodynamic principles. We discuss the shortcomings that various design decisions entail and propose future directions that could build off these methods to yield more robust, accurate RNA structure predictions.

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

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.

Machine learning for RNA 2D structure prediction benchmarked on experimental data

Since the 1980s, dozens of computational methods have addressed the problem of predicting RNA secondary structure. Among them are those that follow standard optimization approaches and, more recently, machine learning (ML) algorithms. The former were repeatedly benchmarked on various datasets. The latter, on the other hand, have not yet undergone extensive analysis that could suggest to the user which algorithm best fits the problem to be solved. In this review, we compare 15 methods that predict the secondary structure of RNA, of which 6 are based on deep learning (DL), 3 on shallow learning (SL) and 6 control methods on non-ML approaches. We discuss the ML strategies implemented and perform three experiments in which we evaluate the prediction of (I) representatives of the RNA equivalence classes, (II) selected Rfam sequences and (III) RNAs from new Rfam families. We show that DL-based algorithms (such as SPOT-RNA and UFold) can outperform SL and traditional methods if the data distribution is similar in the training and testing set. However, when predicting 2D structures for new RNA families, the advantage of DL is no longer clear, and its performance is inferior or equal to that of SL and non-ML methods.

Sequence similarity governs generalizability of de novo deep learning models for RNA secondary structure prediction

Making no use of physical laws or co-evolutionary information, de novo deep learning (DL) models for RNA secondary structure prediction have achieved far superior performances than traditional algorithms. However, their statistical underpinning raises the crucial question of generalizability. We present a quantitative study of the performance and generalizability of a series of de novo DL models, with a minimal two-module architecture and no post-processing, under varied similarities between seen and unseen sequences. Our models demonstrate excellent expressive capacities and outperform existing methods on common benchmark datasets. However, model generalizability, i.e., the performance gap between the seen and unseen sets, degrades rapidly as the sequence similarity decreases. The same trends are observed from several recent DL and machine learning models. And an inverse correlation between performance and generalizability is revealed collectively across all learning-based models with wide-ranging architectures and sizes. We further quantitate how generalizability depends on sequence and structure identity scores via pairwise alignment, providing unique quantitative insights into the limitations of statistical learning. Generalizability thus poses a major hurdle for deploying de novo DL models in practice and various pathways for future advances are discussed.