Update: We’ve now released the code for RISoTTo. I’ll also be presenting a poster at NeurIPS AI4D3 in San Diego on Dec 6, if you’re around and want to chat about RNA design, come say hi! 🧬
github: https://t.co/9HOM4g44dL
Preprint: https://t.co/RY12JfsT5N
Presenting RISoTTo, a molecular context-aware model for RNA sequence design, building on our work with PeSTo (for protein binding interface prediction) and CARBonAra (for protein design).
📄Preprint: https://t.co/uOlAxWz1Vo
💻Code & more coming soon!
#RNAbiology#AI4Science
Context-aware geometric deep learning for RNA sequence design
1.RISoTTo is a new geometric deep learning model that designs RNA sequences based on their tertiary structure and surrounding molecular context—such as proteins, ligands, ions, and DNA. Unlike previous methods, it doesn’t treat RNA as isolated but as part of a functional environment.
2.The model achieves superior native sequence recovery compared to leading methods. On a 14-RNA benchmark, RISoTTo reaches 62% recovery, outperforming gRNAde (56.8%) and older tools like Rosetta and ViennaRNA.
3.RISoTTo leverages a parameter-free geometric transformer inspired by CARBonAra, originally developed for protein design. It operates on atomic coordinates directly, with no need for feature engineering or structural preprocessing.
4.The model uses a 4-bead coarse-grained representation of the RNA backbone and progressively incorporates geometric and chemical context across 20 transformer layers, handling atomic neighborhoods from 8 to 64 nearest neighbors.
5.Context-awareness is key: sequence recovery improves significantly when molecular context is included—for example, RNA–DNA interaction recovery jumps from 36% to 84%, and RNA–RNA from 35% to 69%.
6.Designs are validated using in silico folding with EternaFold (2D) and AlphaFold 3 (3D), showing strong consistency between generated and native structures. RISoTTo achieves high sc2D MCC and low RMSD in benchmark cases.
7.An in silico case study on the NAD+ riboswitch domain (PDB: 7D7V) shows that RISoTTo-designed sequences maintain the native fold and in several cases improve predicted binding affinity for both NAD+ and the U1A protein.
8.RISoTTo also supports autoregressive generation and sequence imprinting, enabling guided sampling of diverse yet functionally constrained RNA variants—useful for exploring design space while maintaining quality.
9.The architecture generalizes well across data: it was trained on 19.5k RNA subunits (both PDB and trRosettaRNA-predicted structures), with careful augmentation and sequence/structure dissimilarity filtering to avoid overfitting.
10.With molecular context, RISoTTo captures functional interfaces more accurately, making it promising for real-world RNA design tasks such as riboswitch engineering, RNA therapeutics, or CRISPR guide design.
11.Despite its advances, RISoTTo still reflects challenges of RNA design in general, especially limited by the scarcity of high-quality RNA tertiary structure data compared to protein design.
📜Paper: https://t.co/YPn8u4i7fS
#RNA #DeepLearning #StructuralBiology #SyntheticBiology #ComputationalBiology #RNAstructure #InverseFolding #GeometricDeepLearning #TransformerModels #RISoTTo
Presenting RISoTTo, a molecular context-aware model for RNA sequence design, building on our work with PeSTo (for protein binding interface prediction) and CARBonAra (for protein design).
📄Preprint: https://t.co/uOlAxWz1Vo
💻Code & more coming soon!
#RNAbiology#AI4Science
I’m happy to announce my latest paper has been released as a preprint on BioRxiv: Predictions from Deep Learning Propose Substantial Protein-Carbohydrate Interplay. This paper was only able to happen thanks to both @jeffreyjgray and @RonaldSchn37100
https://t.co/2qWKkPlIqH
BREAKING NEWS
The Royal Swedish Academy of Sciences has decided to award the 2024 #NobelPrize in Chemistry with one half to David Baker “for computational protein design” and the other half jointly to Demis Hassabis and John M. Jumper “for protein structure prediction.”
To all Protein Designers!
We @matteodp and the @ISBSIB are working on a database for in silico generated protein structures and need your input that fits our different pipelines. Please fill out our survey and share it: https://t.co/m2NU3Od6rW
#ModelArchive#ProteinDesign
Three fabulous students who worked in our group received their Masters degrees from @iiserkol today. Congratulations, Amardeep, @BibekarParth and @KsagarAbhay ! I am so proud of you.
🇮🇳 Gukesh D wins the #FIDECandidates 2024 and the right to challenge the reigning World Champion 🇨🇳 Ding Liren for the title! 🏆
Congratulations! 👏
📷 Michal Walusza
Conformational ensemble of the NSP1 CTD in SARS-CoV-2: Perspectives fr... https://t.co/O1SdvldXDn
The EMMD method highlights NSP1 conformations incompatible with Ribosomal binding..potential drug targets? @BiophysJ@serbonline
#PadmaAwards
Big day for one of the top new research and education institutes in the country - @IISERPune
Padma Bhushan awarded to Prof. Deepak Dhar (Prof. of Physics)
Padma Shri awarded to Prof. K.N. Ganesh (Founding Director)
Efficient interrogation of transition barriers in a Tyr-Kinase. Do check our new paper based on modified a EMMD platform. @Pallab_Dutta_@ChemPhysChem@iiserkol https://t.co/B9r5Cg6G6y
Our group at IISER Kolkata has a fully funded Ph.D position open. Candidates should have a Masters degree in Science or Engineering and strong analytical abilities. Ref. sec B-1 of the advertisement:
https://t.co/djWOEKHnPa
Message me for further details. RT appreciated!