Jérôme Waldispühl
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Jérôme Waldispühl has a total of 178 co-authors. Here are the top co-authors by number of shared publications:
- Vladimir Reinharz
Université du Québec à Montréal
12 shared publications - Yann Ponty
LIX, Ecole Polytechnique, Palaiseau, France
9 shared publications - Roman Sarrazin-Gendron
School of Computer Science, McGill University, Canada
8 shared publications Alexander Butyaev
School of Computer Science McGill University Montreal QC Canada
7 shared publications- Mathieu Blanchette
School of Computer Science, McGill University , Montreal, QC H3A 2A7
7 shared publications Vladimir Reinharz
Université du Québec à Montréal
6 shared publications
Latest Publications
- Open Access
- 29 October 2026
- Published by Wiley in Cytometry Part A
Citizen Science Gamers Enable Automated Flow Cytometry Gating Through Machine Learning
Machine learning can greatly accelerate biomedical research but is fundamentally constrained by the scarcity of high‐quality labeled data. We introduce a scalable data‐generation framework based on citizen science and apply it to automated flow cytometry gating, a foundational but subjective step in immunology that limits reproducibility. Using the massively multiplayer online game EVE Online, 839,199 de‐identified player accounts analyzed 52,000 plots, from which stringent quality control produced 31,703 consensus annotations for model training. Leveraging these data, we developed flowMagic, a machine learning framework for automated gating. Evaluated on 92,203 bivariate plots with standardized reference gates spanning 79 immune populations, the template model outperformed the generalized model, achieving 90% and 65% agreement for large and low‐abundance populations, respectively, compared with 78% and 42% for the generalized model. These results show that expert‐provided templates currently provide the highest agreement, while the citizen‐science‐trained generalized model enables gating without gate‐specific templates but remains limited by the absence of expert‐curated examples for complex, low‐abundance or context‐dependent gates. Both model types reproduced biologically meaningful patterns, including neutrophil dynamics in COVID‐19 patients and immune development in newborns. Together, these results establish citizen‐science annotation as a scalable strategy for generating training data that complements, rather than replaces, expert input in automated flow cytometry gating.
- 29 October 2026
- Published by Wiley in Cytometry Part A
Open AccessCitizen Science Gamers Enable Automated Flow Cytometry Gating Through Machine Learning
Machine learning can greatly accelerate biomedical research but is fundamentally constrained by the scarcity of high‐quality labeled data. We introduce a scalable data‐generation framework based on citizen science and apply it to automated flow cytometry gating, a foundational but subjective step in immunology that limits reproducibility. Using the massively multiplayer online game EVE Online, 839,199 de‐identified player accounts analyzed 52,000 plots, from which stringent quality control produced 31,703 consensus annotations for model training. Leveraging these data, we developed flowMagic, a machine learning framework for automated gating. Evaluated on 92,203 bivariate plots with standardized reference gates spanning 79 immune populations, the template model outperformed the generalized model, achieving 90% and 65% agreement for large and low‐abundance populations, respectively, compared with 78% and 42% for the generalized model. These results show that expert‐provided templates currently provide the highest agreement, while the citizen‐science‐trained generalized model enables gating without gate‐specific templates but remains limited by the absence of expert‐curated examples for complex, low‐abundance or context‐dependent gates. Both model types reproduced biologically meaningful patterns, including neutrophil dynamics in COVID‐19 patients and immune development in newborns. Together, these results establish citizen‐science annotation as a scalable strategy for generating training data that complements, rather than replaces, expert input in automated flow cytometry gating.
- Open Access
Adaptive 2.5D base-pairing subgraph search detects RNA small-molecule binding sites
Ribo-LENS is a geometric deep-learning framework for detecting small-molecule binding sites in RNA structures. It is designed to exploit two properties of RNA base-pairing networks: their robustness to conformational fluctuation and the functional signatures they encode. By reasoning directly in the space of base-pairing subgraphs, Ribo-LENS assembles coherent binding sites, in contrast to methods that score residues independently. In extensive experiments, Ribo-LENS is competitive with, and often outperforms, large all-atom co-folding models (AlphaFold3, Chai-1), fine-tuned language models (GerNA-Bind), and structure-based tools (RNAsite), raising mean MCC to 0.380 (versus 0.321 for the state-of-the-art GerNA-Bind). It is strongly robust to apo/holo rearrangement, with its accuracy tracking the base-pairing graph (Spearman ρ = 0.82 with binding-site graph edit distance) rather than backbone displacement ( ρ = −0.15 with RMSD), and depends far less on sequence homology than competing predictors. In an end-to-end, sequence-based virtual screen of the ROBIN assay (∼25,000 compounds), Ribo-LENS guides docking to a small predicted pocket, matching a blind all-atom cavity search (enrichment factors up to 6.1) at a fraction of the search cost; on two SARS-CoV-2 targets its predicted sites align with NMR chemical-shift perturbations. Ribo-LENS turns coarse base-pairing structure into a practical entry point for screening the vast, largely unexplored RNA target space.
Open AccessAdaptive 2.5D base-pairing subgraph search detects RNA small-molecule binding sites
Ribo-LENS is a geometric deep-learning framework for detecting small-molecule binding sites in RNA structures. It is designed to exploit two properties of RNA base-pairing networks: their robustness to conformational fluctuation and the functional signatures they encode. By reasoning directly in the space of base-pairing subgraphs, Ribo-LENS assembles coherent binding sites, in contrast to methods that score residues independently. In extensive experiments, Ribo-LENS is competitive with, and often outperforms, large all-atom co-folding models (AlphaFold3, Chai-1), fine-tuned language models (GerNA-Bind), and structure-based tools (RNAsite), raising mean MCC to 0.380 (versus 0.321 for the state-of-the-art GerNA-Bind). It is strongly robust to apo/holo rearrangement, with its accuracy tracking the base-pairing graph (Spearman ρ = 0.82 with binding-site graph edit distance) rather than backbone displacement ( ρ = −0.15 with RMSD), and depends far less on sequence homology than competing predictors. In an end-to-end, sequence-based virtual screen of the ROBIN assay (∼25,000 compounds), Ribo-LENS guides docking to a small predicted pocket, matching a blind all-atom cavity search (enrichment factors up to 6.1) at a fraction of the search cost; on two SARS-CoV-2 targets its predicted sites align with NMR chemical-shift perturbations. Ribo-LENS turns coarse base-pairing structure into a practical entry point for screening the vast, largely unexplored RNA target space.
- 21 August 2025
- Published by Springer Nature in Nature Computational Science
- Vol. 5 (8) , 606-607
What’s so hard about RNA-targeting drug discovery?
- 21 August 2025
- Published by Springer Nature in Nature Computational Science
- Vol. 5 (8) , 606-607
What’s so hard about RNA-targeting drug discovery?
- Open Access
- 21 March 2025
- Published by Springer Nature in Nature Communications
- Vol. 16 (1) , 1-12
RNAmigos2: accelerated structure-based RNA virtual screening with deep graph learning
RNAs are a vast reservoir of untapped drug targets. Structure-based virtual screening (VS) identifies candidate molecules by leveraging binding site information, traditionally using molecular docking simulations. However, docking struggles to scale with large compound libraries and RNA targets. Machine learning offers a solution but remains underdeveloped for RNA due to limited data and practical evaluations. We introduce a data-driven VS pipeline tailored for RNA, utilizing coarse-grained 3D modeling, synthetic data augmentation, and RNA-specific self-supervision. Our model achieves a 10,000x speedup over docking while ranking active compounds in the top 2.8% on structurally distinct test sets. It is robust to binding site variations and successfully screens unseen RNA riboswitches in a 20,000-compound in-vitro microarray, with a mean enrichment factor of 2.93 at 1%. This marks the first experimentally validated success of structure-based deep learning for RNA VS.
- 21 March 2025
- Published by Springer Nature in Nature Communications
- Vol. 16 (1) , 1-12
Open AccessRNAmigos2: accelerated structure-based RNA virtual screening with deep graph learning
RNAs are a vast reservoir of untapped drug targets. Structure-based virtual screening (VS) identifies candidate molecules by leveraging binding site information, traditionally using molecular docking simulations. However, docking struggles to scale with large compound libraries and RNA targets. Machine learning offers a solution but remains underdeveloped for RNA due to limited data and practical evaluations. We introduce a data-driven VS pipeline tailored for RNA, utilizing coarse-grained 3D modeling, synthetic data augmentation, and RNA-specific self-supervision. Our model achieves a 10,000x speedup over docking while ranking active compounds in the top 2.8% on structurally distinct test sets. It is robust to binding site variations and successfully screens unseen RNA riboswitches in a 20,000-compound in-vitro microarray, with a mean enrichment factor of 2.93 at 1%. This marks the first experimentally validated success of structure-based deep learning for RNA VS.
- Open Access
Improving the reliability of molecular string representations for generative chemistry
Generative modeling for chemistry has advanced rapidly in recent years, but this surge in popularity raises a foundational question: which molecular representation is best suited for modern machine learning models? Despite not being designed for generative tasks, SMILES remain the most commonly used string-based representation. However, while SMILES follow strict syntactic rules, grammatically correct SMILES strings do not always correspond to valid molecules. SELFIES were introduced as an alternative that addresses this limitation by ensuring that every string of SELFIES tokens represents to a valid molecule. In this study, we comprehensively evaluate the limitations of both SMILES and SELFIES as representations for generative models. We define two key criteria for robust molecular generation: viability, generated strings represent novel, unique molecules with correct valence, and fidelity, the distribution of physicochemical properties from sampled molecules resembles that of the training data. We find that approximately one-fifth of molecules generated using canonical SMILES are invalid, failing the viability criterion. In contrast, all SELFIES-generated molecules are viable, but they deviate significantly from the training distribution, indicating low fidelity. To address these limitations, we develop data augmentation procedures for both representations. While simplifying the SELFIES grammar yields only modest gains in fidelity, our stochastic augmentation method for SMILES, ClearSMILES, significantly improves both viability and fidelity. ClearSMILES simplifies syntax by reducing the vocabulary size and explicitly encoding aromaticity via Kekulé SMILES, making it easier string representations for models to process. Using ClearSMILES, the rate of invalid samples decreases by an order of magnitude, from 20% to 2.2%, and fidelity to the training distribution is also moderately improved. Generative chemistry has seen rapid development recently. However, models based on string representations of molecules still rely largely on SMILES 1 that have not been developed for this context and SELFIES 2 who were introduced to reduce those problems. The goal of this study is to first analyze the difficulty encountered by a small generative model when using SMILES and SELFIES. Our study found that SELFIES and canonical SMILES 3 are not fully reliable representations for a small generative model, i.e. do not ensure concurrently the viability and fidelity of samples. Viable samples represent novel, unique molecules with correct valence, while fidelity is efficient distribution learning of key physico-chemical properties. 4 In fact, 20% of the samples generated using canonical SMILES input representation do not correspond to valid molecules. In contrast, samples generated using SELFIES were all viable but where not able to reproduce as well the distribution of physico-chemical properties as SMILES. As a mitigation strategy for the previously identified problems, we have developed data augmentation procedures for both SELFIES and SMILES. Simplifying the complex syntax of SELFIES yielded only marginal improvements in string stability and overall fidelity to the training set. For SMILES, we developed a stochastic data augmentation procedure called ClearSMILES, which reduces the vocabulary size needed to represent a SMILES dataset, explicitly represents aromaticity via Kekulé SMILES, 3 and reduces the effort required by deep learning models to process SMILES. ClearSMILES reduced the rate of invalid samples by an order of magnitude, from 20% to 2.2%, and improved the fidelity of samples to the training set.
Open AccessImproving the reliability of molecular string representations for generative chemistry
Generative modeling for chemistry has advanced rapidly in recent years, but this surge in popularity raises a foundational question: which molecular representation is best suited for modern machine learning models? Despite not being designed for generative tasks, SMILES remain the most commonly used string-based representation. However, while SMILES follow strict syntactic rules, grammatically correct SMILES strings do not always correspond to valid molecules. SELFIES were introduced as an alternative that addresses this limitation by ensuring that every string of SELFIES tokens represents to a valid molecule. In this study, we comprehensively evaluate the limitations of both SMILES and SELFIES as representations for generative models. We define two key criteria for robust molecular generation: viability, generated strings represent novel, unique molecules with correct valence, and fidelity, the distribution of physicochemical properties from sampled molecules resembles that of the training data. We find that approximately one-fifth of molecules generated using canonical SMILES are invalid, failing the viability criterion. In contrast, all SELFIES-generated molecules are viable, but they deviate significantly from the training distribution, indicating low fidelity. To address these limitations, we develop data augmentation procedures for both representations. While simplifying the SELFIES grammar yields only modest gains in fidelity, our stochastic augmentation method for SMILES, ClearSMILES, significantly improves both viability and fidelity. ClearSMILES simplifies syntax by reducing the vocabulary size and explicitly encoding aromaticity via Kekulé SMILES, making it easier string representations for models to process. Using ClearSMILES, the rate of invalid samples decreases by an order of magnitude, from 20% to 2.2%, and fidelity to the training distribution is also moderately improved. Generative chemistry has seen rapid development recently. However, models based on string representations of molecules still rely largely on SMILES 1 that have not been developed for this context and SELFIES 2 who were introduced to reduce those problems. The goal of this study is to first analyze the difficulty encountered by a small generative model when using SMILES and SELFIES. Our study found that SELFIES and canonical SMILES 3 are not fully reliable representations for a small generative model, i.e. do not ensure concurrently the viability and fidelity of samples. Viable samples represent novel, unique molecules with correct valence, while fidelity is efficient distribution learning of key physico-chemical properties. 4 In fact, 20% of the samples generated using canonical SMILES input representation do not correspond to valid molecules. In contrast, samples generated using SELFIES were all viable but where not able to reproduce as well the distribution of physico-chemical properties as SMILES. As a mitigation strategy for the previously identified problems, we have developed data augmentation procedures for both SELFIES and SMILES. Simplifying the complex syntax of SELFIES yielded only marginal improvements in string stability and overall fidelity to the training set. For SMILES, we developed a stochastic data augmentation procedure called ClearSMILES, which reduces the vocabulary size needed to represent a SMILES dataset, explicitly represents aromaticity via Kekulé SMILES, 3 and reduces the effort required by deep learning models to process SMILES. ClearSMILES reduced the rate of invalid samples by an order of magnitude, from 20% to 2.2%, and improved the fidelity of samples to the training set.
- Open Access
- 27 August 2024
- Published by Springer Nature in BMC Bioinformatics
- Vol. 25 (1) , 1-14
ARGV: 3D genome structure exploration using augmented reality
Over the past two decades, scientists have increasingly realized the importance of the three-dimensional (3D) genome organization in regulating cellular activity. Hi-C and related experiments yield 2D contact matrices that can be used to infer 3D models of chromosome structure. Visualizing and analyzing genomes in 3D space remains challenging. Here, we present ARGV, an augmented reality 3D Genome Viewer. ARGV contains more than 350 pre-computed and annotated genome structures inferred from Hi-C and imaging data. It offers interactive and collaborative visualization of genomes in 3D space, using standard mobile phones or tablets. A user study comparing ARGV to existing tools demonstrates its benefits.
- 27 August 2024
- Published by Springer Nature in BMC Bioinformatics
- Vol. 25 (1) , 1-14
Open AccessARGV: 3D genome structure exploration using augmented reality
Over the past two decades, scientists have increasingly realized the importance of the three-dimensional (3D) genome organization in regulating cellular activity. Hi-C and related experiments yield 2D contact matrices that can be used to infer 3D models of chromosome structure. Visualizing and analyzing genomes in 3D space remains challenging. Here, we present ARGV, an augmented reality 3D Genome Viewer. ARGV contains more than 350 pre-computed and annotated genome structures inferred from Hi-C and imaging data. It offers interactive and collaborative visualization of genomes in 3D space, using standard mobile phones or tablets. A user study comparing ARGV to existing tools demonstrates its benefits.
- Open Access
- 21 May 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-10
Learning the Game: Decoding the Differences between Novice and Expert Players in a Citizen Science Game with Millions of Players
In recent years, video games have surged in popularity, attracting millions of players across platforms. Citizen science games (CSGs) leverage the processing power of gamers to solve computational and scientific problems. Borderlands Science (BLS) is a mini-game within the mass market game Borderlands 3 that turns multiple sequence alignment (MSA) problems into puzzles. Parallel research demonstrated that BLS players outperformed classical approaches solving small sequence alignment tasks. This study aims to analyze the strategical differences in player solutions in BLS as they gain experience. Through the many collected player solutions from players of different experience level, we gained insights into players’ strategies, differences between expert and non-expert players, and how strategies evolve. We developed a Markov chain trained on solutions from players of different experience levels to understand their actions and outcomes. Results indicate that expert players utilize more gaps and achieve more matches, gradually improving and converging toward unique strategies. Our findings reveal distinct and evolving player strategies. For future citizen science projects, it will be important to consider the identification of player strategies and their evolution over time to improve the game design and data processing.
- 21 May 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-10
Open AccessLearning the Game: Decoding the Differences between Novice and Expert Players in a Citizen Science Game with Millions of Players
In recent years, video games have surged in popularity, attracting millions of players across platforms. Citizen science games (CSGs) leverage the processing power of gamers to solve computational and scientific problems. Borderlands Science (BLS) is a mini-game within the mass market game Borderlands 3 that turns multiple sequence alignment (MSA) problems into puzzles. Parallel research demonstrated that BLS players outperformed classical approaches solving small sequence alignment tasks. This study aims to analyze the strategical differences in player solutions in BLS as they gain experience. Through the many collected player solutions from players of different experience level, we gained insights into players’ strategies, differences between expert and non-expert players, and how strategies evolve. We developed a Markov chain trained on solutions from players of different experience levels to understand their actions and outcomes. Results indicate that expert players utilize more gaps and achieve more matches, gradually improving and converging toward unique strategies. Our findings reveal distinct and evolving player strategies. For future citizen science projects, it will be important to consider the identification of player strategies and their evolution over time to improve the game design and data processing.
- Open Access
- 15 April 2024
- Published by Springer Nature in Nature Biotechnology
- Vol. 43 (1) , 76-84
Improving microbial phylogeny with citizen science within a mass-market video game
Citizen science video games are designed primarily for users already inclined to contribute to science, which severely limits their accessibility for an estimated community of 3 billion gamers worldwide. We created Borderlands Science (BLS), a citizen science activity that is seamlessly integrated within a popular commercial video game played by tens of millions of gamers. This integration is facilitated by a novel game-first design of citizen science games, in which the game design aspect has the highest priority, and a suitable task is then mapped to the game design. BLS crowdsources a multiple alignment task of 1 million 16S ribosomal RNA sequences obtained from human microbiome studies. Since its initial release on 7 April 2020, over 4 million players have solved more than 135 million science puzzles, a task unsolvable by a single individual. Leveraging these results, we show that our multiple sequence alignment simultaneously improves microbial phylogeny estimations and UniFrac effect sizes compared to state-of-the-art computational methods. This achievement demonstrates that hyper-gamified scientific tasks attract massive crowds of contributors and offers invaluable resources to the scientific community.
- 15 April 2024
- Published by Springer Nature in Nature Biotechnology
- Vol. 43 (1) , 76-84
Open AccessImproving microbial phylogeny with citizen science within a mass-market video game
Citizen science video games are designed primarily for users already inclined to contribute to science, which severely limits their accessibility for an estimated community of 3 billion gamers worldwide. We created Borderlands Science (BLS), a citizen science activity that is seamlessly integrated within a popular commercial video game played by tens of millions of gamers. This integration is facilitated by a novel game-first design of citizen science games, in which the game design aspect has the highest priority, and a suitable task is then mapped to the game design. BLS crowdsources a multiple alignment task of 1 million 16S ribosomal RNA sequences obtained from human microbiome studies. Since its initial release on 7 April 2020, over 4 million players have solved more than 135 million science puzzles, a task unsolvable by a single individual. Leveraging these results, we show that our multiple sequence alignment simultaneously improves microbial phylogeny estimations and UniFrac effect sizes compared to state-of-the-art computational methods. This achievement demonstrates that hyper-gamified scientific tasks attract massive crowds of contributors and offers invaluable resources to the scientific community.
- Open Access
- 30 January 2024
- Published by Oxford University Press (OUP) in Bioinformatics
- Vol. 40 (2)
PERFUMES: pipeline to extract RNA functional motifs and exposed structures
Motivation: Up to 75% of the human genome encodes RNAs. The function of many non-coding RNAs relies on their ability to fold into 3D structures. Specifically, nucleotides inside secondary structure loops form non-canonical base pairs that help stabilize complex local 3D structures. These RNA 3D motifs can promote specific interactions with other molecules or serve as catalytic sites. Results: We introduce PERFUMES, a computational pipeline to identify 3D motifs that can be associated with observable features. Given a set of RNA sequences with associated binary experimental measurements, PERFUMES searches for RNA 3D motifs using BayesPairing2 and extracts those that are over-represented in the set of positive sequences. It also conducts a thermodynamics analysis of the structural context that can support the interpretation of the predictions. We illustrate PERFUMES’ usage on the SNRPA protein binding site, for which the tool retrieved both previously known binder motifs and new ones. Availability and implementation: PERFUMES is an open-source Python package (https://jwgitlab.cs.mcgill.ca/arnaud_chol/perfumes).
- 30 January 2024
- Published by Oxford University Press (OUP) in Bioinformatics
- Vol. 40 (2)
Open AccessPERFUMES: pipeline to extract RNA functional motifs and exposed structures
Motivation: Up to 75% of the human genome encodes RNAs. The function of many non-coding RNAs relies on their ability to fold into 3D structures. Specifically, nucleotides inside secondary structure loops form non-canonical base pairs that help stabilize complex local 3D structures. These RNA 3D motifs can promote specific interactions with other molecules or serve as catalytic sites. Results: We introduce PERFUMES, a computational pipeline to identify 3D motifs that can be associated with observable features. Given a set of RNA sequences with associated binary experimental measurements, PERFUMES searches for RNA 3D motifs using BayesPairing2 and extracts those that are over-represented in the set of positive sequences. It also conducts a thermodynamics analysis of the structural context that can support the interpretation of the predictions. We illustrate PERFUMES’ usage on the SNRPA protein binding site, for which the tool retrieved both previously known binder motifs and new ones. Availability and implementation: PERFUMES is an open-source Python package (https://jwgitlab.cs.mcgill.ca/arnaud_chol/perfumes).
- Open Access
RNAmigos2: Fast and accurate structure-based RNA virtual screening with semi-supervised graph learning and large-scale docking data
RNAs constitute a vast reservoir of mostly untapped drug targets. Structure-based virtual screening (VS) methods screen large compound libraries for identifying promising candidate molecules by conditioning on binding site information. The classical approach relies on molecular docking simulations. However, this strategy does not scale well with the size of the small molecule databases and the number of potential RNA targets. Machine learning emerged as a promising technology to resolve this bottleneck. Efficient data-driven VS methods have already been introduced for proteins, but these techniques have not yet been developed for RNAs due to limited dataset sizes and lack of practical use-case evaluation. We propose 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 using synthetic data augmentation and RNA-centric self supervision. We report strong prediction and generalizability of our framework, ranking active compounds among inactives in the top 2.8% on average on a structurally distinct drug-like test set. Those predictions are sensitive, but robust to pockets alterations, opening the door to its use on binding site detection methods outputs. Our model results in a ten thousand-times speedup over docking techniques while obtaining higher performance. Finally, we deploy our model on a recently published in-vitro small molecule microarray experiment with 20,000 compounds and report a mean enrichment factor at 1% of 2.93 on four unseen RNA riboswitches. To our knowledge, this is the first experimental evidence of success for structure-based deep learning methods in RNA virtual screening. Our source code and data, as well as a Google Colab notebook for inference, are available on GitHub. 1
Open AccessRNAmigos2: Fast and accurate structure-based RNA virtual screening with semi-supervised graph learning and large-scale docking data
RNAs constitute a vast reservoir of mostly untapped drug targets. Structure-based virtual screening (VS) methods screen large compound libraries for identifying promising candidate molecules by conditioning on binding site information. The classical approach relies on molecular docking simulations. However, this strategy does not scale well with the size of the small molecule databases and the number of potential RNA targets. Machine learning emerged as a promising technology to resolve this bottleneck. Efficient data-driven VS methods have already been introduced for proteins, but these techniques have not yet been developed for RNAs due to limited dataset sizes and lack of practical use-case evaluation. We propose 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 using synthetic data augmentation and RNA-centric self supervision. We report strong prediction and generalizability of our framework, ranking active compounds among inactives in the top 2.8% on average on a structurally distinct drug-like test set. Those predictions are sensitive, but robust to pockets alterations, opening the door to its use on binding site detection methods outputs. Our model results in a ten thousand-times speedup over docking techniques while obtaining higher performance. Finally, we deploy our model on a recently published in-vitro small molecule microarray experiment with 20,000 compounds and report a mean enrichment factor at 1% of 2.93 on four unseen RNA riboswitches. To our knowledge, this is the first experimental evidence of success for structure-based deep learning methods in RNA virtual screening. Our source code and data, as well as a Google Colab notebook for inference, are available on GitHub. 1