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.

D-ORB: A Web Server to Extract Structural Features of Related But Unaligned RNA Sequences.

Identifying the common structural elements of functionally related RNA sequences (family) is usually based on an alignment of the sequences, which is often subject to human bias and may not be accurate. The resulting covariance model (CM) provides probabilities for each base to covary with another, which allows to support evolutionarily the formation of double helical regions and possibly pseudoknots. The coexistence of alternative folds in RNA, resulting from its dynamic nature, may lead to the potential omission of motifs by CM. To overcome this limitation, we present D-ORB, a system of algorithms that identifies overrepresented motifs in the secondary conformational landscapes of a family when compared to those of unrelated sequences. The algorithms are bundled into an easy-to-use website allowing users to submit a family, and optionally provide unrelated sequences. D-ORB produces a non-pseudoknotted secondary structure based on the overrepresented motifs, a deep neural network classifier and two decision trees. When used to model an Rfam family, D-ORB fits overrepresented motifs in the corresponding Rfam structure; more than a hundred Rfam families have been modeled. The statistical approach behind D-ORB derives the structural composition of an RNA family, making it a valuable tool for analyzing and modeling it. Its easy-to-use interface and advanced algorithms make it an essential resource for researchers studying RNA structure. D-ORB is available at https://d-orb.major.iric.ca/.

The DynaSig-ML Python package: automated learning of biomolecular dynamics-function relationships.

The DynaSig-ML ('Dynamical Signatures-Machine Learning') Python package allows the efficient, user-friendly exploration of 3D dynamics-function relationships in biomolecules, using datasets of experimental measures from large numbers of sequence variants. It does so by predicting 3D structural dynamics for every variant using the Elastic Network Contact Model (ENCoM), a sequence-sensitive coarse-grained normal mode analysis model. Dynamical Signatures represent the fluctuation at every position in the biomolecule and are used as features fed into machine learning models of the user's choice. Once trained, these models can be used to predict experimental outcomes for theoretical variants. The whole pipeline can be run with just a few lines of Python and modest computational resources. The compute-intensive steps are easily parallelized in the case of either large biomolecules or vast amounts of sequence variants. As an example application, we use the DynaSig-ML package to predict the maturation efficiency of human microRNA miR-125a variants from high-throughput enzymatic assays. AVAILABILITY AND IMPLEMENTATION: DynaSig-ML is open-source software available at https://github.com/gregorpatof/dynasigml_package.

Sequence-sensitive elastic network captures dynamical features necessary for miR-125a maturation.

The Elastic Network Contact Model (ENCoM) is a coarse-grained normal mode analysis (NMA) model unique in its all-atom sensitivity to the sequence of the studied macromolecule and thus to the effect of mutations. We adapted ENCoM to simulate the dynamics of ribonucleic acid (RNA) molecules, benchmarked its performance against other popular NMA models and used it to study the 3D structural dynamics of human microRNA miR-125a, leveraging high-throughput experimental maturation efficiency data of over 26 000 sequence variants. We also introduce a novel way of using dynamical information from NMA to train multivariate linear regression models, with the purpose of highlighting the most salient contributions of dynamics to function. ENCoM has a similar performance profile on RNA than on proteins when compared to the Anisotropic Network Model (ANM), the most widely used coarse-grained NMA model; it has the advantage on predicting large-scale motions while ANM performs better on B-factors prediction. A stringent benchmark from the miR-125a maturation dataset, in which the training set contains no sequence information in common with the testing set, reveals that ENCoM is the only tested model able to capture signal beyond the sequence. This ability translates to better predictive power on a second benchmark in which sequence features are shared between the train and test sets. When training the linear regression model using all available data, the dynamical features identified as necessary for miR-125a maturation point to known patterns but also offer new insights into the biogenesis of microRNAs. Our novel approach combining NMA with multivariate linear regression is generalizable to any macromolecule for which relatively high-throughput mutational data is available.

CAMAP: Artificial neural networks unveil the role of codon arrangement in modulating MHC-I peptides presentation.

MHC-I associated peptides (MAPs) play a central role in the elimination of virus-infected and neoplastic cells by CD8 T cells. However, accurately predicting the MAP repertoire remains difficult, because only a fraction of the transcriptome generates MAPs. In this study, we investigated whether codon arrangement (usage and placement) regulates MAP biogenesis. We developed an artificial neural network called Codon Arrangement MAP Predictor (CAMAP), predicting MAP presentation solely from mRNA sequences flanking the MAP-coding codons (MCCs), while excluding the MCC per se. CAMAP predictions were significantly more accurate when using original codon sequences than shuffled codon sequences which reflect amino acid usage. Furthermore, predictions were independent of mRNA expression and MAP binding affinity to MHC-I molecules and applied to several cell types and species. Combining MAP ligand scores, transcript expression level and CAMAP scores was particularly useful to increase MAP prediction accuracy. Using an in vitro assay, we showed that varying the synonymous codons in the regions flanking the MCCs (without changing the amino acid sequence) resulted in significant modulation of MAP presentation at the cell surface. Taken together, our results demonstrate the role of codon arrangement in the regulation of MAP presentation and support integration of both translational and post-translational events in predictive algorithms to ameliorate modeling of the immunopeptidome.

A transcriptome-based approach to identify functional modules within and across primary human immune cells.

Genome-wide transcriptomic analyses have provided valuable insight into fundamental biology and disease pathophysiology. Many studies have taken advantage of the correlation in the expression patterns of the transcriptome to infer a potential biologic function of uncharacterized genes, and multiple groups have examined the relationship between co-expression, co-regulation, and gene function on a broader scale. Given the unique characteristics of immune cells circulating in the blood, we were interested in determining whether it was possible to identify functional co-expression modules in human immune cells. Specifically, we sequenced the transcriptome of nine immune cell types from peripheral blood cells of healthy donors and, using a combination of global and targeted analyses of genes within co-expression modules, we were able to determine functions for these modules that were cell lineage-specific or shared among multiple cell lineages. In addition, our analyses identified transcription factors likely important for immune cell lineage commitment and/or maintenance.

RNA-Puzzles Round IV: 3D structure predictions of four ribozymes and two aptamers.

RNA-Puzzles is a collective endeavor dedicated to the advancement and improvement of RNA 3D structure prediction. With agreement from crystallographers, the RNA structures are predicted by various groups before the publication of the crystal structures. We now report the prediction of 3D structures for six RNA sequences: four nucleolytic ribozymes and two riboswitches. Systematic protocols for comparing models and crystal structures are described and analyzed. In these six puzzles, we discuss (i) the comparison between the automated web servers and human experts; (ii) the prediction of coaxial stacking; (iii) the prediction of structural details and ligand binding; (iv) the development of novel prediction methods; and (v) the potential improvements to be made. We show that correct prediction of coaxial stacking and tertiary contacts is essential for the prediction of RNA architecture, while ligand binding modes can only be predicted with low resolution and simultaneous prediction of RNA structure with accurate ligand binding still remains out of reach. All the predicted models are available for the future development of force field parameters and the improvement of comparison and assessment tools.

Apoptotic endothelial cells release small extracellular vesicles loaded with immunostimulatory viral-like RNAs.

Endothelial cells have multifaceted interactions with the immune system, both as initiators and targets of immune responses. In vivo, apoptotic endothelial cells release two types of extracellular vesicles upon caspase-3 activation: apoptotic bodies and exosome-like nanovesicles (ApoExos). Only ApoExos are immunogenic: their injection causes inflammation and autoimmunity in mice. Based on deep sequencing of total RNA, we report that apoptotic bodies and ApoExos are loaded with divergent RNA cargos that are not released by healthy endothelial cells. Apoptotic bodies, like endothelial cells, contain mainly ribosomal RNA whereas ApoExos essentially contain non-ribosomal non-coding RNAs. Endogenous retroelements, bearing viral-like features, represented half of total ApoExos RNA content. ApoExos also contained several copies of unedited Alu repeats and large amounts of non-coding RNAs with a demonstrated role in autoimmunity such as U1 RNA and Y RNA. Moreover, ApoExos RNAs had a unique nucleotide composition and secondary structure characterized by strong enrichment in U-rich motifs and unstably folded RNAs. Globally, ApoExos were therefore loaded with RNAs that can stimulate a variety of RIG-I-like receptors and endosomal TLRs. Hence, apoptotic endothelial cells selectively sort in ApoExos a diversified repertoire of immunostimulatory "self RNAs" that are tailor-made for initiation of innate immune responses and autoimmunity.

The sequence features that define efficient and specific hAGO2-dependent miRNA silencing guides.

MicroRNAs (miRNAs) are ribonucleic acids (RNAs) of ∼21 nucleotides that interfere with the translation of messenger RNAs (mRNAs) and play significant roles in development and diseases. In bilaterian animals, the specificity of miRNA targeting is determined by sequence complementarity involving the seed. However, the role of the remaining nucleotides (non-seed) is only vaguely defined, impacting negatively on our ability to efficiently use miRNAs exogenously to control gene expression. Here, using reporter assays, we deciphered the role of the base pairs formed between the non-seed region and target mRNA. We used molecular modeling to reveal that this mechanism corresponds to the formation of base pairs mediated by ordered motions of the miRNA-induced silencing complex. Subsequently, we developed an algorithm based on this distinctive recognition to predict from sequence the levels of mRNA downregulation with high accuracy (r2 > 0.5, P-value < 10-12). Overall, our discovery improves the design of miRNA-guide sequences used to simultaneously downregulate the expression of multiple predetermined target genes.

RNA-MoIP: prediction of RNA secondary structure and local 3D motifs from sequence data.

RNA structures are hierarchically organized. The secondary structure is articulated around sophisticated local three-dimensional (3D) motifs shaping the full 3D architecture of the molecule. Recent contributions have identified and organized recurrent local 3D motifs, but applications of this knowledge for predictive purposes is still in its infancy. We recently developed a computational framework, named RNA-MoIP, to reconcile RNA secondary structure and local 3D motif information available in databases. In this paper, we introduce a web service using our software for predicting RNA hybrid 2D-3D structures from sequence data only. Optionally, it can be used for (i) local 3D motif prediction or (ii) the refinement of user-defined secondary structures. Importantly, our web server automatically generates a script for the MC-Sym software, which can be immediately used to quickly predict all-atom RNA 3D models. The web server is available at http://rnamoip.cs.mcgill.ca.

RNA-Puzzles Round III: 3D RNA structure prediction of five riboswitches and one ribozyme.

RNA-Puzzles is a collective experiment in blind 3D RNA structure prediction. We report here a third round of RNA-Puzzles. Five puzzles, 4, 8, 12, 13, 14, all structures of riboswitch aptamers and puzzle 7, a ribozyme structure, are included in this round of the experiment. The riboswitch structures include biological binding sites for small molecules (S-adenosyl methionine, cyclic diadenosine monophosphate, 5-amino 4-imidazole carboxamide riboside 5'-triphosphate, glutamine) and proteins (YbxF), and one set describes large conformational changes between ligand-free and ligand-bound states. The Varkud satellite ribozyme is the most recently solved structure of a known large ribozyme. All puzzles have established biological functions and require structural understanding to appreciate their molecular mechanisms. Through the use of fast-track experimental data, including multidimensional chemical mapping, and accurate prediction of RNA secondary structure, a large portion of the contacts in 3D have been predicted correctly leading to similar topologies for the top ranking predictions. Template-based and homology-derived predictions could predict structures to particularly high accuracies. However, achieving biological insights from de novo prediction of RNA 3D structures still depends on the size and complexity of the RNA. Blind computational predictions of RNA structures already appear to provide useful structural information in many cases. Similar to the previous RNA-Puzzles Round II experiment, the prediction of non-Watson-Crick interactions and the observed high atomic clash scores reveal a notable need for an algorithm of improvement. All prediction models and assessment results are available at http://ahsoka.u-strasbg.fr/rnapuzzles/.

Exploring Alternative RNA Structure Sets Using MC-Flashfold and db2cm.

We created an accelerated version of MC-Fold called MC-Flashfold that allows us to compute large numbers of competing secondary structures including noncanonical base pairs. We visualize the base pairs in these sets using high quality intuitive dot plots and arc plots. Our new tools allow us to explore RNA dynamics by visualizing the competing structures in free energy bands. Here we describe how to use these tools to generate dot plots that reveal the postulated anti-terminator stem in the E. coli trp operon leader sequence. These plots show the anti-terminator hairpin loop during transcription and as a minor population of the full-length leader sequence. This is a case of switching RNA structure that had been originally postulated based on short dyad inverted repeats. Other switching RNA sequences can be analyzed by using our method.

Structural dynamics control the MicroRNA maturation pathway.

MicroRNAs (miRNAs) are crucial gene expression regulators and first-order suspects in the development and progression of many diseases. Comparative analysis of cancer cell expression data highlights many deregulated miRNAs. Low expression of miR-125a was related to poor breast cancer prognosis. Interestingly, a single nucleotide polymorphism (SNP) in miR-125a was located within a minor allele expressed by breast cancer patients. The SNP is not predicted to affect the ground state structure of the primary transcript or precursor, but neither the precursor nor mature product is detected by RT-qPCR. How this SNP modulates the maturation of miR-125a is poorly understood. Here, building upon a model of RNA dynamics derived from nuclear magnetic resonance studies, we developed a quantitative model enabling the visualization and comparison of networks of transient structures. We observed a high correlation between the distances between networks of variants with that of their respective wild types and their relative degrees of maturation to the latter, suggesting an important role of transient structures in miRNA homeostasis. We classified the human miRNAs according to pairwise distances between their networks of transient structures.

MiRBooking simulates the stoichiometric mode of action of microRNAs.

In eucaryotes, gene expression is regulated by microRNAs (miRNAs) which bind to messenger RNAs (mRNAs) and interfere with their translation into proteins, either by promoting their degradation or inducing their repression. We study the effect of miRNA interference on each gene using experimental methods, such as microarrays and RNA-seq at the mRNA level, or luciferase reporter assays and variations of SILAC at the protein level. Alternatively, computational predictions would provide clear benefits. However, no algorithm toward this task has ever been proposed. Here, we introduce a new algorithm to predict genome-wide expression data from initial transcriptome abundance. The algorithm simulates the miRNA and mRNA hybridization competition that occurs in given cellular conditions, and derives the whole set of miRNA::mRNA interactions at equilibrium (microtargetome). Interestingly, solving the competition improves the accuracy of miRNA target predictions. Furthermore, this model implements a previously reported and fundamental property of the microtargetome: the binding between a miRNA and a mRNA depends on their sequence complementarity, but also on the abundance of all RNAs expressed in the cell, i.e. the stoichiometry of all the miRNA sites and all the miRNAs given their respective abundance. This model generalizes the miRNA-induced synchronistic silencing previously observed, and described as sponges and competitive endogenous RNAs.

RNA-Puzzles Round II: assessment of RNA structure prediction programs applied to three large RNA structures.

This paper is a report of a second round of RNA-Puzzles, a collective and blind experiment in three-dimensional (3D) RNA structure prediction. Three puzzles, Puzzles 5, 6, and 10, represented sequences of three large RNA structures with limited or no homology with previously solved RNA molecules. A lariat-capping ribozyme, as well as riboswitches complexed to adenosylcobalamin and tRNA, were predicted by seven groups using RNAComposer, ModeRNA/SimRNA, Vfold, Rosetta, DMD, MC-Fold, 3dRNA, and AMBER refinement. Some groups derived models using data from state-of-the-art chemical-mapping methods (SHAPE, DMS, CMCT, and mutate-and-map). The comparisons between the predictions and the three subsequently released crystallographic structures, solved at diffraction resolutions of 2.5-3.2 Å, were carried out automatically using various sets of quality indicators. The comparisons clearly demonstrate the state of present-day de novo prediction abilities as well as the limitations of these state-of-the-art methods. All of the best prediction models have similar topologies to the native structures, which suggests that computational methods for RNA structure prediction can already provide useful structural information for biological problems. However, the prediction accuracy for non-Watson-Crick interactions, key to proper folding of RNAs, is low and some predicted models had high Clash Scores. These two difficulties point to some of the continuing bottlenecks in RNA structure prediction. All submitted models are available for download at http://ahsoka.u-strasbg.fr/rnapuzzles/.

Noncoding regions of C. elegans mRNA undergo selective adenosine to inosine deamination and contain a small number of editing sites per transcript.

ADARs (Adenosine deaminases that act on RNA) "edit" RNA by converting adenosines to inosines within double-stranded regions. The primary targets of ADARs are long duplexes present within noncoding regions of mRNAs, such as introns and 3' untranslated regions (UTRs). Because adenosine and inosine have different base-pairing properties, editing within these regions can alter splicing and recognition by small RNAs. However, despite numerous studies identifying multiple editing sites in these genomic regions, little is known about the extent to which editing sites co-occur on individual transcripts or the functional output of these combinatorial editing events. To begin to address these questions, we performed an ultra-deep sequencing analysis of 4 Caenorhabditis elegans 3' UTRs that are known ADAR targets. Synchronous editing events were determined for the long duplexes in vivo. Furthermore, the validity of each editing event was confirmed by sequencing the same regions of mRNA from worms that lack A-to-I editing. This analysis identified a large number of editing sites that can occur within each 3' UTR, but interestingly, each individual transcript contained only a small fraction of these A-to-I editing events. In addition, editing patterns were not random, indicating that an editing event can affect the efficiency of editing at subsequent adenosines. Furthermore, we identified specific sites that can be both positively and negatively correlated with additional sites leading to mutually exclusive editing patterns. These results suggest that editing in noncoding regions is selective and hyper-editing of cellular RNAs is rare.

Computational identification of RNA functional determinants by three-dimensional quantitative structure-activity relationships.

Anti-infection drugs target vital functions of infectious agents, including their ribosome and other essential non-coding RNAs. One of the reasons infectious agents become resistant to drugs is due to mutations that eliminate drug-binding affinity while maintaining vital elements. Identifying these elements is based on the determination of viable and lethal mutants and associated structures. However, determining the structure of enough mutants at high resolution is not always possible. Here, we introduce a new computational method, MC-3DQSAR, to determine the vital elements of target RNA structure from mutagenesis and available high-resolution data. We applied the method to further characterize the structural determinants of the bacterial 23S ribosomal RNA sarcin-ricin loop (SRL), as well as those of the lead-activated and hammerhead ribozymes. The method was accurate in confirming experimentally determined essential structural elements and predicting the viability of new SRL variants, which were either observed in bacteria or validated in bacterial growth assays. Our results indicate that MC-3DQSAR could be used systematically to evaluate the drug-target potentials of any RNA sites using current high-resolution structural data.

Autism-related deficits via dysregulated eIF4E-dependent translational control.

Hyperconnectivity of neuronal circuits due to increased synaptic protein synthesis is thought to cause autism spectrum disorders (ASDs). The mammalian target of rapamycin (mTOR) is strongly implicated in ASDs by means of upstream signalling; however, downstream regulatory mechanisms are ill-defined. Here we show that knockout of the eukaryotic translation initiation factor 4E-binding protein 2 (4E-BP2)-an eIF4E repressor downstream of mTOR-or eIF4E overexpression leads to increased translation of neuroligins, which are postsynaptic proteins that are causally linked to ASDs. Mice that have the gene encoding 4E-BP2 (Eif4ebp2) knocked out exhibit an increased ratio of excitatory to inhibitory synaptic inputs and autistic-like behaviours (that is, social interaction deficits, altered communication and repetitive/stereotyped behaviours). Pharmacological inhibition of eIF4E activity or normalization of neuroligin 1, but not neuroligin 2, protein levels restores the normal excitation/inhibition ratio and rectifies the social behaviour deficits. Thus, translational control by eIF4E regulates the synthesis of neuroligins, maintaining the excitation-to-inhibition balance, and its dysregulation engenders ASD-like phenotypes.