Programme

Toutes les présentations auront lieu dans la salle PK-1140, située au 201, avenue du Président-Kennedy. L'entrée principale se trouve au coin de Président-Kennedy et Kimberley. Prenez ensuite l'ascenseur jusqu'à la première étage.

Jeudi, 21 septembre 2023

Vendredi, 22 septembre 2023

• Lunch le jeudi: Café Parvis, 433, rue Mayor, Montréal
• Dîner le jeudi soir: Brasserie Galaxie, 1414, rue Clark, Montréal
• Lunch le vendredi: Lola Rosa, 276, rue Sainte-Catherine ouest, Montréal

Conférences

RNA-Puzzles Round V and Perspectives: Assessing Prediction Accuracy in Four Different RNA Types
Chichau Miao, Guangzhou Laboratory
RNA-Puzzles Round V presents the results of a comprehensive assessment of RNA prediction accuracy in four distinct RNA types: RNA module and aptamer, Viral elements, ribozyme, and Riboswitches. This study evaluates the performance of various computational methods and algorithms in predicting the structural features and functionalities of 26 RNA targets. The results demonstrate notable achievements in RNA 3D structure prediction, including enhanced accuracy tertiary structure modeling, and potential functional inferrence. Furthermore, the study underscores the importance of community-wide efforts in advancing computational RNA biology and highlights the challenges that remain in accurately predicting complex RNA structures and functions. In brief, RNA-Puzzles Round V showcases the progress made in accurately predicting the structures and functionalities of diverse RNA types. This comprehensive assessment serves as a valuable resource for researchers and developers of computational tools in their pursuit of understanding RNA biology and designing RNA-based applications. The insights gained from this study pave the way for further improvements in computational RNA biology and the development of novel therapeutic interventions targeting RNA molecules.

Modeling RNA folding with the MELD-NA approach
Alberto Perez, University of Florida
We have developed MELD as a Bayesian inference approach to combine sparse and noisy information with Molecular dynamics simulations. The approach is grounded on physical principles and uses statistical mechanics to process the ensembles of structures generated by the molecular dynamics approach. The method has been quite successful in folding small proteins (up to 100 residues), predicting how peptides bind proteins, and ranking protein-protein complexes. During the RNA-puzzles event we attempted to fold three RNA sequences – achieving a high accuracy prediction for one of them (156 residues in length). Several developments helped to achieve this success: the MELD methodology, the development of an implicit solvent for nucleic acids, improved force fields for RNA simulation, and the accuracy of secondary structure predictions in RNA. We modeled secondary structure predictions as noisy data in which each base could pair up with other bases based on the server predictions – the Bayesian inference approach determines the most likely interpretation of the data and the structures compatible with the data and the force field. At the end, clustering the ensembles and selecting the cluster centroid helped us determine the structures to submit to the RNA puzzles organizers. RNA folding was a new research direction for us and we are both excited about the possibilities of this methodology and working on diagnosing what worked and what did not.

Finding knots and tangles in RNA models
Tomasz Zok, Poznan University
RNA 3D structure prediction methods often rely either on simulation or fragment assembly. Both approaches have pros and cons, but most importantly, they proved in the past to be able to predict a novel fold successfully. These methods, while being fundamentally different, share certain risks and pitfalls. One is the possibility of getting stuck in an unrealistic conformation such as a structural entanglement. Unlike knots in protein structure, which are defined solely upon the trace of the polypeptide backbone, RNA structural entanglements happen when two elementary motifs intersect in 3D space. We can distinguish several different entanglement types based on the kinds of motifs involved and the mutuality of intersection or lack thereof. Some entanglements might be evolutionary conserved (functional) and therefore found in experimentally solved structures – especially when pseudoknots are involved. Others seem to be purely artificial constructs from a botched simulation or fragment-assembly session of an RNA 3D structure prediction. Here we present the method to find and classify structural entanglements and discuss its application to the RNA-Puzzles models.

RNAComposer and mindful modeling of the RNA structure
Marta Szachniuk, Poznan University
RNAComposer is primarily known as an automated web server. However, its advanced functions allow it to be used interactively, e.g. for homology-based modeling or assembling models from user-provided structural elements. During the talk, various scenarios of applying the system to predict the 3D RNA structures will be presented. Their combination with expert input, comparative analysis of models, and routines to select representative output structures form a workflow that the RNApolis group follows in the RNA-Puzzles and CASP challenges. With selected examples, the application of this workflow for the in silico modeling of natural and synthetic RNA molecules targeted in CASP15 will be demonstrated.

Predicting RNA-small molecule interactions
Shi-Jie Chen, University of Missouri
The intricate 3D structures and functional roles of RNAs render RNA molecules ideal targets for therapeutic drugs. The rational design of RNA-targeted drugs requires accurate computational modeling of interactions between RNA and small molecules. Here, we present recently developed physics-based approaches for computationally predicting RNA-small molecule binding. Additionally, we describe comparisons with other existing models used for predicting RNA-ligand binding poses and conducting virtual screening of RNA-targeted small molecules.

New and improved tools for the RNA sequence-structure relationship
Craig Zirbel, Bowling Green State University
We will give an overview of tools developed by the BGSU RNA group to study the relationship between RNA 3D structure and sequence. A python version of FR3D (Find RNA 3D) is now available. Basepair annotations have been systematically reviewed and improved. Modified nucleotides are mapped to standard nucleotides so their basepairs can be annotated. We introduce new tools to map positions in RNA 3D structures to columns of multiple sequence alignments, using Rfam sequences aligned by Infernal. We present visual diagnostics which show that the overall quality of the alignments is high, even in non-Watson-Crick regions. We demonstrate how to retrieve sequence variants for basepairs and motifs seen in 3D structures.

Solving RNA’s Puzzling Puzzle Pieces
François Major, Université de Montréal
In this presentation, I focus on an innovative approach to decipher the complex 2D structural landscape of RNA molecules, as generated by a prediction algorithm, in particular mcff, a dynamic programming implementation of the MC-Fold algorithm. Moving away from the classic 3D exploration commonly employed in RNA puzzles, the method I propose involves intensive searching in 2D space. This strategic shift accelerates the identification of nucleotide interaction networks and pivotal structural motifs, thereby facilitating a comprehensive understanding of their functional implications. It should be noted that a correctly predicted 2D structure serves as an effective springboard for an accurate and meaningful exploration of its 3D space.

RNA 3D structure prediction à la AlphaFold
Helena Rivas, Harvard University
I will discuss the challenges and opportunities of adapting the technologies implemented by AlphaFold for the prediction of RNA 3D structure. The amount of 3D structural data for RNA is orders of magnitude smaller than that existing for proteins (from 100,000 protein structures deposited in the Protein Data Bank, to approximately 2,000 RNA structures, once identical sequences are removed), and the training of models with large number of parameters may be compromised by the reduced amount of data. Alignments and attention on protein alignments are a crucial component of the AlphaFold alignment transformer technology (the evoformer). Method such as R- scape and CaCoFold have demonstrated that there is a large amount of structural information contained in RNA structural alignments ready to be harvested by transformer methodologies based on alignment attention. Regarding the geometry component, the number of atoms per nucleotide in the RNA backbone is larger than the number of atoms per amino acid, thus the characterization of all atom’s positions requires more torsion angles, and may be subjected to larger errors. I will discuss different strategies to address this issue. Finally, while the protein 2D structures (alpha helix or beta sheet) is local and captured by the torsion angles, there is no equivalent for structured RNAs. On the other hand, RNA 3D structure is heavily directed by Watson-Crick base pair interactions (WC pairs). WC base pairs follow specific substitution patterns to be captured by the alignment transformer (the evoformer), and we advocate for the addition of pair representations and loss functions directed to capture this RNA-specific base pair information.

Multiscale refinement and reconstruction of RNA structures
Simon Poblete, Universidad San Sebastian
Helix-based models can significantly accelerate the conformational sampling of RNA structures provided their secondary structure. They can quickly move large helices or domains and propose tertiary contacts between them. However, reconstructing an all-atom representation of their predictions requires the assembly of fragments which can produce many unphysical effects, including clashes, broken bonds, and topological artifacts. We introduce the SPlit and conQueR (SPQR) package for refining and exploring RNA structures at the nucleotide level. We show how to quickly fix these issues before an all-atom refinement, using specific energy terms and virtual sites. Moreover, the search for tertiary contacts, given the proximity of loops and helices, is benchmarked against X-ray structures by searching the possible conformations consistent with these interactions. We also present and suggest criteria to consider for the introduction of three-dimensional restraints, which in the present case, use the ERMSD metric and require the definition of a cutoff distance between the nucleotides involved. Altogether, SPQR with these tools can propose well-refined structures up to thousands of nucleotides and be used as a refinement tool for the RNA-Puzzles predictions.

RNA Contact Prediction by Data Efficient Deep Learning
Alexander Schug, Forschungszentrum Jülich
On the molecular level, life is orchestrated through an interplay of many biomolecules. To gain any detailed understanding of biomolecular function, one needs to know their structure. Yet the structural characterization of many important biomolecules and their complexes - typically preceding any detailed mechanistic exploration of their function- remains experimentally challenging. For proteins, the richness of labeled training data enables highly successfully deep-learning approaches. A direct transfer of methods to RNA, however, is hampered by the simple lack of such massive data. The limited available data, however, can still be used to predict spatial adjacencies (”contact maps”) as a proxy for 3D structure. Statistical physics driven methods, such as direct coupling analysis (DCA), can work on a single multiple sequence alignment. Shallow and deep learning driven approaches can take advantage of all available data for RNA and further increase accuracy. Our current model, BARNACLE, combines the utilization of unlabeled data through self-supervised pre-training and efficient use of the sparse labeled data through an XGBoost classifier. We observe a considerable improvement over both the established baseline DCA and a shallow neural network architecture.

rna-tools: a Swiss army knife for RNA 3D structure modeling workflow and application to understand RNA-protein molecular machines
Marcin Magnus, Harvard University
Significant improvements have been made in the efficiency and accuracy of RNA 3D structure prediction methods in recent years; however, many tools developed in the field stay exclusive to only a few bioinformatic groups. To perform a complete RNA 3D structure modeling analysis as proposed by the RNA-Puzzles community, researchers must familiarize themselves with a quite complex set of tools. In order to facilitate the processing of RNA sequences and structures, we previously developed the rna-tools package. However, using rna-tools requires the installation of a mixture of libraries and tools, basic knowledge of the command line and the Python programming language. To provide an opportunity for the broader community of biologists to take advantage of the new developments in RNA structural biology, I developed https://rna-tools.online. The web server provides a user-friendly platform to perform many standard analyses required for the typical modeling workflow: 3D structure manipulation and editing, structure minimization, structure analysis, quality assessment, and comparison. Furthermore, I will discuss the application of various tools aimed at improving our comprehension of RNA dynamics within the context of RNA-protein molecular machines, specifically illustrating this with examples from splicing and group II introns.

Predicting 3D RNA structure and dynamics using Discrete Molecular Dynamics and machine learning
Nikolay Dokholyan, Penn State College of Medicine
We present an integrative approach to model RNA structure and dynamics using physics, machine learning, and experiments. We utilize discrete molecular dynamics (DMD) simulations and a physics-based Medusa force field to guide the simulations of RNA molecules. We utilize experimental approaches to reduce the search for the native basin of the RNA molecule. We further implement deep machine learning approaches to predict molecular dynamics outcomes. Our integrative approach allows the robust prediction of the native state of RNA molecules.

Accelerating cryoEM for screening, solving, and evaluating RNA tertiary structures
Grace Nye, Stanford
In recent years, single-particle cryoEM has emerged as a promising technique for resolving the three-dimensional structure of RNA molecules. Experimentally determining the tertiary structure of RNA molecules is challenging because these molecules are typically smaller than many of their protein counterparts and contain a larger degree of inherent flexibility and heterogeneity. As methods for this technique are continually evolving, we set out to evaluate our current capabilities through RNA Puzzle 39, a viral cloverleaf RNA, by experimentally resolving the structure of the RNA within the time limits of the challenge. Within the three week time limit, we were able to resolve partial maps of regions of the structure. We lay out here the expertise and the resources required to resolve the structure. On our first grid preparation attempt, within an hour of collection, we were able to clearly see that this RNA was a highly ordered molecule, suggesting that cryoEM should be useful as a tool to quickly screen for atomically ordered RNAs. However, more time and care is required to determine RNA tertiary structures from cryoEM data, and due to the poor signal-to-noise ratio of small constructs, we are currently limited to lower resolution than can be obtained with crystallography. Since the challenge, we have been able to perform additional analysis on the cryoEM data, in particular examining what information can be learned by comparing and contrasting the cryoEM data and partial maps with both the crystal structure and predictions. Finally, we pilot the use of raw single particle cryoEM 2D particles to assess predictions and reweight ensembles of predictions based on the posterior probability of the atomic model given the data. CryoEM is a powerful tool for quickly determining the order of a given RNA but presents complex challenges in order to resolve robust models of the tertiary structures of small RNA molecules.

RNA structure prediction: Overcoming a multi-front challenge
Yaoqi Zhou, Shenzhen Bay Laboratory
RNA structure prediction is considered as a similar problem to protein structure prediction because both predict a structure from a sequence. As a result, some expected a quick solution given the success of AlphaFold in protein structure prediction. However, recent top four performance of RNA structure prediction in CASP 15 were all from traditional energy-based techniques, indicating that AI-driven RNA structure prediction has an issue of over-training. In this talk, we will discuss how RNA structure prediction faces the challenge from multiple fronts and what should we do about them.