• Christopher Burge

We are developing fully interpretable models of RNA splicing and its regulation for improved understanding and various applications. We recently described a model called SMsplice that predicts the splicing patterns of primary transcripts in a variety of animal and plant species, using just core splice site motifs, exon and intron length preferences, and learned scores for splicing regulatory elements (SREs) that act locally on splice sites. This model enables automatic learning of candidate SREs from any organism, and achieves accuracy of 83-86% in fish, insects, and plants and about 70% in mammals. A new direction is the inference of the splicing regulatory activity of a splicing factor from just knockdown/RNA-seq data and a model of its intrinsic binding preferences such as an RNA Bind-n-Seq, RNACompete or SELEX motif, but without using crosslinking data. Application to data from the ENCODE RNA-binding protein dataset and other data yields models that are reproducible across cell lines and species, and which can distinguish direct from indirect regulatory targets and can be used to infer cooperative splicing regulation.

• Philip Fradkin, Philip Fradkin, Vector Institute
• Ruian Shi, Ruian Shi, University of Toronto
• Keren Isaev, Keren Isaev, New York Genome Centre
• Quaid Morris, Quaid Morris, Computational and Systems Biology Program
• Bo Wang, Bo Wang, University Health Network
• Brendan Frey, Brendan Frey, University of Toronto
• Leo Lee, Leo Lee, University of Toronto

In the face of rapidly accumulating genomic data, our understanding of the RNA regulatory code remains incomplete. Recent self-supervised methods in other domains have demonstrated the ability to learn rules underlying the data-generating process such as sentence structure in language. Inspired by this, we extend contrastive learning techniques to genomic data by utilizing functional similarities between sequences generated through alternative splicing and gene duplication. We introduce IsoCLR, a model trained on a novel dataset with a contrastive objective enabling the learning of generalized RNA isoform representations. We validate representation utility on downstream tasks such as RNA half-life and mean ribosome load prediction. Our pre-training strategy yields competitive results using linear probing across 6 tasks, along with up to a two-fold increase in Pearson correlation in low-data conditions. Importantly, our exploration of the learned latent space reveals that our contrastive objective yields semantically meaningful representations, underscoring its potential as a valuable initialization technique for RNA property prediction.

• Frederick Korbel, Frederick Korbel, Max Delbruck Center
• Gabriel Villamil, Gabriel Villamil, Max Delbruck Center
• Ekaterina Eroshok, Ekaterina Eroshok, Max Delbruck Center & Humboldt University
• Uwe Ohler, Uwe Ohler, Max Delbrueck Center & Humboldt University

Regulation of mRNA translation enables rapid and local control of gene expression. As rate-limiting step, translation initiation is primarily controlled by the 5’ untranslated region (5’UTR). In it, regulatory sequence elements including RNA structural motifs and upstream open reading frames (uORFs) dictate the efficiency of translation. A recent convolutional neural network model accurately quantifies the relationship between massively parallel synthetic 5’ UTRs and translation levels, but the underlying sequence determinants remain elusive.
To uncover the input features most important for prediction, feature attribution methods compute importance scores for input features, thus allowing to explain prediction output with respect to its input. Hence, model interpretation can be applied as a tool to uncover functional sequence patterns and generate novel biological hypotheses. Applying model interpretation, we extract representations of regulatory logic, revealing a complex interplay of regulatory sequence elements. Guided by insights from model interpretation, we adapt the model by human reporter data to obtain superior performance.

• Can Cenik, Can Cenik, UT Austin

Characterization of shared patterns of RNA expression between genes across conditions has led to the discovery of novel biological functions and regulatory networks. These RNA co-expression relationships have illuminated the higher-order organization of transcriptomes, yet we currently do not know if patterns of coordination in other gene expression modalities are similarly informative. In particular, translational covariation across cell types have remained unexplored, primarily due to the scarcity of comprehensive translational measurements across a large compendium of biological contexts. Here, we uniformly analyzed 2277 matched ribosome profiling and RNA-seq data from 90 human and 81 mouse tissues and cell lines. We introduce the concept of Translational Efficiency Covariation (TEC), identifying mRNAs that demonstrate coordinated translation patterns across cell types. We demonstrate that TEC is conserved across human and mouse cells and uncover novel gene functions that rely on translational covariation information alone. Moreover, our observations indicate that proteins exhibiting positive covariation at both translational and transcriptional levels are significantly more likely to physically interact. We finally discover TEC patterns indicative of RNA-binding protein (RBP) involvement, suggesting potential mechanisms of shared translational regulation. Our findings establish translational covariations across various conditions as a pervasive and conserved organizing principle of mammalian transcriptomes.

• Ori Feldman, Ori Feldman, Ben-Gurion University
• Yaron Orenstein, Yaron Orenstein, Bar-Ilan University

RNA-binding proteins play important roles in various cellular processes. For this reason, researchers have developed experimental assays to measure protein–RNA binding in vivo. However, obtaining these measurements for every protein across various cell types is infeasible due to the high cost and long times of these experiments. Thus, researchers rely on computational methods to predict protein–RNA binding, but so far methods have been limited in their success in predicting RNA binding across cell types. In this work, we present CellRBP, a novel method to accurately and efficiently predict protein–RNA binding across cell types. CellRBP is based on a convolutional neural network that uniquely receives as input cell-type-specific information, such as experimentally measured RNA structure and RNA abundance, which enable the accurate generalization across cell types (Figure 1). We trained CellRBP on 196 of eCLIP experiments and evaluated prediction performance in both cross-validation and across cell types. CellRBP achieved superior performance compared to the state of the art achieving an average AUROC score of 0.889 in cross-validation and 0.772 across cell types, respectively (Figure 2A,B). We interrogated the trained models for the important features they learned using both local and glocal interpretability techniques and discovered known and novel RNA-binding preferences (Figure 2C). CellRBP is expected to help many researchers in predicting protein–RNA binding over various cell types and conditions. CellRBP is available via https://github.com/OrensteinLab/CellRBP.

• Alexander Sasse, Alexander Sasse, University of Washington
• Debashish Ray, Debashish Ray, University of Toronto
• Kaitlin Laverty, Kaitlin Laverty, Memorial Sloan Kettering Cancer Center
• Cyrus Tam, Cyrus Tam, Memorial Sloan Kettering Cancer Center
• Mihai Albu, Mihai Albu, University of Toronto
• Hong Zheng, Hong Zheng, University of Toronto
• Olga Lyudovyk, Olga Lyudovyk, Memorial Sloan Kettering Cancer Center
• Kate Nie, Kate Nie, University of Toronto
• Cedrik Magis, Cedrik Magis, The Barcelona Institute of Science and Technology
• Cedric Notredame, Cedric Notredame, The Barcelona Institute of Science and Technology
• Matthew Weirauch, Matthew Weirauch, Cincinnati Children’s Hospital
• Timothy Hughes, Timothy Hughes, University of Toronto
• Quaid Morris, Quaid Morris, Memorial Sloan Kettering Cancer Center

RNA-binding proteins (RBPs) are key regulators of gene expression. Here, we introduce RBPzoo — a resource of RNAcompete-derived in vitro RNA-binding data for 379 RBPs from 33 diverse eukaryotes. We develop a new method, Joint Protein-Ligand Embedding (JPLE), to map specificity-determining peptides to corresponding RNA motifs for 28,667 RBPs from 690 eukaryotes. We illustrate the broad utility of this resource by inferring post-transcriptional function for 12 eukaryotic RBPs in mRNA stability and reconstructing the evolution of 2,568 RNA motifs. For the latter, we identify a universal set of 19 RNA motifs conserved between plants and metazoa and observe rapid motif evolution arising from whole genome duplications in vertebrate ancestors. RBPzoo represents a powerful resource for the study of gene regulation for any organism with an annotated genome.

• Shaimae Elhajjajy, Shaimae Elhajjajy, University of Massachusetts Chan Medical School
• Zhiping Weng, Zhiping Weng, University of Massachusetts Chan Medical School

RNA-binding proteins (RBPs) are the primary mediators of mRNA regulation, dynamically governing complex processes such as splicing, cleavage, and degradation. Previous studies have shown that structurally diverse RBPs recognize similar motifs but can still bind distinct sites within the transcriptome. While in vitro evidence suggests that motif context plays an important role in RBP binding specificity, the precise underlying mechanisms remain unclear. Despite recent advances in machine learning models to predict RBP binding, current methods are often difficult to interpret and do not categorically investigate motif contexts. Thus, there remains a need for interpretable predictive models to disambiguate the contextual determinants of RBP binding specificity. Here, we present, to the best of our knowledge, the first formulation of the RBP binding prediction task as an NLP-based multiple-instance learning problem. We introduce a novel sequence decomposition strategy to generate entities termed “contexts”, which we use to train and test our deep learning models. We also develop a deterministic motif discovery algorithm that is fast, accurate, and specialized to handle our data structure, recapitulating the motifs of well-characterized RBPs as validation. Importantly, we discover and characterize the in vivo sequence binding contexts for a collection of RBPs. Finally, by integrating motif and context similarity measures with a cross-prediction approach, we propose novel RBP-RBP interaction partners and hypothesize whether these interactions are cooperative or competitive. In summary, we present a comprehensive computational strategy for illuminating contextual determinants of specific RBP binding and demonstrate the implications of our findings in delineating RBP function.

• Kristina Sungeun Song, Kristina Sungeun Song, Université de Sherbrooke
• Bernice Yeo, Bernice Yeo, Université de Sherbrooke
• Vivian Seow, Vivian Seow, Université de Sherbrooke
• Laurence Faucher-Giguère, Laurence Faucher-Giguère, Université de Sherbrooke
• Gabrielle Deschamps-Francoeur, Gabrielle Deschamps-Francoeur, Université de Sherbrooke
• Sherif Abou Elela, Sherif Abou Elela, Université de Sherbrooke
• Michelle Scott, Michelle Scott, Université de Sherbrooke

Box C/D small nucleolar RNAs (snoRNAs) are noncoding RNAs crucial for guiding 2’-O-ribose methylation in ribosomal RNA during ribosome biogenesis, primarily through the formation of ribonucleoprotein (snoRNP) complexes with core RNA-binding proteins (RBPs). Additional roles were proposed for box C/D snoRNAs, including the regulation of alternative splicing of protein-coding transcripts, yet few validated examples exist, with unclear mechanisms. Some noncanonical functions are thought to involve interactions with additional RBPs beyond the core snoRNA binders, indicating diverse regulatory roles of snoRNAs by interacting with various RBPs, collectively modulating protein-coding target RNAs. To explore these interactions and their functional implications, we introduce snoFlake, an interaction network of 191 box C/D snoRNAs and 166 human RBPs, showing direct binding interactions and significant overlap of binding sites on shared protein-coding target RNAs, reinforcing their concerted role in gene regulation. Focusing on snoRNAs targeting groups of functionally-related targets, also bound by snoRNA-associated RBPs led to a hub region composed of SNORD22 and U2 and U5-associated splicing factors: SF3B4, PRPF8, EFTUD2 and AQR. SNORD22, PRPF8 and AQR exhibited an enrichment of overlapping binding sites at both 5’ and 3’ splice sites with the highest number of shared protein-coding target RNAs, suggesting their involvement in a splicing regulatory model. Knockdown experiments and differential alternative splicing analysis further highlighted the potential role of the SNORD22 complex in splicing, marking the first snoRNP splicing regulatory complex. This reshapes the understanding of snoRNA biology, emphasizing snoFlake's potential as a foundation for unravelling the impact of snoRNA-RBP interactions in gene regulation.

• Ruiyan Hou, Ruiyan Hou, The University of Hong Kong
• Yuanhua Huang, Yuanhua Huang, The University of Hong Kong

Three-prime single-cell RNA-seq (scRNA-seq) has been widely employed to profile cellular transcriptomes, however, its power of analysing polyadenylation sites (PAS) has not been fully utilised. Here, we present a computational method, scTail, to precisely identify PAS by using reads 1 and quantify its expression by leveraging the reads 2, which enables effective detection of alternative PAS usage. When compared with other methods, PAS detected by scTail are more accurate. With various experimental data sets, we have demonstrated that scTail can accurately identify PAS and the detected alternative PAS usages showed strong specificity in different biological processes.

• Farica Zhuang, Farica Zhuang, University of Pennsylvania
• Danielle Gutman, Danielle Gutman, University of Pennsylvania
• Nathaniel Islas, Nathaniel Islas, University of Pennsylvania
• Bryan Guzmán, Bryan Guzmán, University of North Carolina at Chapel Hill
• San Jewell, San Jewell, University of Pennsylvania
• Daniel Dominguez, Daniel Dominguez, University of North Carolina at Chapel Hill
• Yoseph Barash, Yoseph Barash, University of Pennsylvania

RNA G-quadruplexes (rG4) are RNA secondary structures known to play an important role in gene regulation. Despite their importance, the effects of genetic variants on rG4 formations and functions remain unexplored. To address this challenge, we introduce G4mer, a deep learning model that predicts transcriptome-wide rG4 formations using high throughput RT-stop experimental data, rG4-seeker. While computational methods have been developed to predict whether rG4s are likely to form on a given sequence, we show that G4mer outperforms other state-of-the-art models, especially for non-canonical rG4 that do not confer to the consensus GGG-{N-1:7}(3)-GGG motif. Additionally, G4mer offers a computational approach to study the effect of variants on rG4 formation and the association of these variants with diseases. With G4mer, we map variants in the 5’ and 3’ untranslated regions that are predicted to alter rG4 formations. Then using the Penn Medicine BioBank, we identify those associated with diseases such as breast cancer. By carefully interpreting the learned G4mer model, we identify rG4 length as a significant factor that deviates between experimental data and human rG4s. Finally, we validate the effect of disease-associated rG4-altering variants on protein expression using dual luciferase assay, and assess the effect of variants on structure formation using Circular Dichroism and RT-stop assays. These experiments point to a potential interplay between structure and sequence motifs affecting downstream gene translation. Overall, our work offers a compelling framework for detecting and validating the functional effects of rG4-altering variants that are significantly associated with diseases.

• Jérôme Waldispühl

RNAs constitute a vast reservoir of mostly untapped drug targets. Structure-based virtual screening (VS) methods are key to massively screen molecular targets and identify promising candidate molecules binding. However, this strategy does not scale well with the size the small molecule databases and the number of potential RNA targets. Furthermore, this approach is also hampered by the scarcity of RNA 3D structural data.
In this talk, we show that using an augmented classification of RNA base pairs combined with graph machine learning methods enable us to design a new class of algorithms for screening RNAs and promising molecular compounds. We describe a data-driven VS pipeline that deals with the unique challenges of RNA molecules through coarse grained modeling of 3D structures and heterogeneous training regimes. We demonstrate strong prediction and generalizability of our framework and discuss further expansion of this platform.

• Frederic Runge, Frederic Runge, University of Freiburg
• Jörg K.H. Franke, Jörg K.H. Franke, University of Freiburg
• Daniel Fertmann, Daniel Fertmann, University of Freiburg
• Rolf Backofen, Rolf Backofen, University of Freiburg
• Frank Hutter, Frank Hutter, University of Freiburg

RNA design is a key technique to achieve new functionality in fields like synthetic biology or biotechnology. Computational tools could help to find such RNA sequences but they are often limited in their formulation of the search space. In this work, we propose partial RNA design, a novel RNA design paradigm that addresses the limitations of current RNA design formulations. Partial RNA design describes the problem of designing RNAs from arbitrary RNA sequences and structure motifs with different design goals. By separating the design space from the objectives, our formulation enables the design of RNAs with variable lengths and desired properties, while still allowing precise control over sequence and structure constraints at individual positions. Based on this formulation, we introduce a new algorithm, libLEARNA, capable of efficiently solving different constraint RNA design tasks. A comprehensive analysis of various problems, including a realistic riboswitch design task, reveals the outstanding performance of libLEARNA and its robustness.

• J. White Bear, J. White Bear, McGill University
• Gregoire De Bisschop, Gregoire De Bisschop, IRCM
• Eric Lecuyer, Eric Lecuyer, IRCM
• Jérôme Waldispühl, Jérôme Waldispühl, McGill University

RNA are known to be highly flexible and take on multiple conformations to perform various tasks and binding in vivo. This makes structural analysis more challenging than with larger, lower entropy molecules. Structure probing with chemical reagents has been a key tool for developing deeper understandings of secondary RNA structure. Traditional probing methods use an averaged mutational profile to detect modifications and infer secondary structural features using reverse transcription. However, this requires shortened segment lengths which can obfuscate key structural information. Moreover, mutational profiles do not express alternative conformations or indicate optimality and conserved features that may be equally viable or relevant to function. Indeed, the averaged profile may be suboptimal. In our study, we use long read nanopore technology to directly sequence RNA, with the reagent acetlyimadizole (AcIm). Our software, Dashing Turtle (DT), can identify AcIm modifications at high resolution, inferring secondary structural features, and determining diverse conformations. DT applies a unique deconvolution of large RNA samples and examines both conservation and dominance across the sample, potentially yielding optimal conformations. Identifying dominant conformations may further lead to a better understanding of key RNA hybridization strategies that can only be observed in transitional interactions and co-modifications. Additionally, DT can identify conserved states across these conformations that may have key functional implications. Furthermore, our method has the potential for application to time or phase-based strategies that can help us understand intermediate structures that play key roles in binding or other in vivo activities for both drug delivery and pathogenesis.

• Michelle Scott