gRNAde

Designing an RNA sequence that will fold into a desired three-dimensional shape is the central challenge of RNA engineering, underpinning efforts to build aptamers, riboswitches, and catalytic ribozymes. gRNAde tackles this RNA inverse-folding problem: given a target 3D backbone, it generates nucleotide sequences predicted to adopt that fold. Unlike methods that reason only over secondary structure, gRNAde conditions on full 3D coordinates, allowing it to capture non-canonical base pairs and tertiary motifs that planar representations miss.

The model was developed by Chaitanya K. Joshi and colleagues across the MRC Laboratory of Molecular Biology (Philipp Holliger's group) and the University of Cambridge Department of Computer Science. First introduced as a geometric deep learning method at ICLR 2025, gRNAde was extended into a full generative design pipeline and experimentally validated in a December 2025 bioRxiv preprint, where it was shown to design active, novel RNA molecules rather than merely score sequences computationally.

gRNAde fits alongside structure-prediction models such as RhoFold and reactivity predictors such as RibonanzaNet, which it incorporates for in-silico screening. Together these form a closed design-and-filter loop that moves RNA design closer to the generative workflows now common in protein engineering.

#Key Features

• 3D-conditioned generation: The model conditions on backbone coordinates rather than secondary structure alone, letting it represent pseudoknots, non-canonical pairings, and tertiary contacts that drive RNA function.
• Flexible design specifications: Users can constrain generation with a target 3D backbone, a (possibly pseudoknotted) secondary structure, partial sequence motifs, or any combination, enabling motif scaffolding and partial redesign.
• Multi-state modeling: gRNAde reasons over conformational ensembles, reflecting that many functional RNAs adopt multiple states rather than a single rigid structure.
• Integrated in-silico screening: Candidate sequences are filtered with RibonanzaNet (chemical-reactivity prediction) and optional RhoFold structure prediction, prioritizing the most promising designs before synthesis.
• Experimental validation: Designs were tested in the lab, including active RNA polymerase ribozyme variants with high sequence divergence from natural sequences.

#Technical Details

gRNAde is a multi-state geometric graph neural network that encodes the RNA backbone as a geometric graph and decodes sequences autoregressively, functioning as a structure-conditioned sequence language model. It was trained self-supervised on RNA structures from the PDB (≤4 Å resolution) curated via the RNASolo database with an October 2023 cutoff, supporting sequences up to 500 nucleotides and using 75% 3D-coordinate dropout so the model degrades gracefully when only partial structural information is available. On the Eterna OpenKnot benchmark for pseudoknot design, gRNAde reached 95% experimental success, matching human expert performance and substantially exceeding the roughly 70% achieved by competing computational methods. Earlier benchmarking reported native sequence recovery improvements over Rosetta-based inverse folding.

#Applications

gRNAde serves RNA engineers, synthetic biologists, and structural biologists who need sequences for a specified fold. Concrete uses include designing pseudoknotted scaffolds, engineering ribozymes with novel sequences while preserving catalytic function, generating aptamer and riboswitch candidates, and redesigning natural RNAs under partial-sequence constraints. Because it pairs generation with reactivity- and structure-based screening, it slots directly into design-build-test-learn loops, narrowing large sequence spaces to a testable shortlist before wet-lab synthesis.

#Impact

gRNAde brings to RNA the generative, structure-conditioned design paradigm that transformed protein engineering, and is among the first such methods with experimental validation of designed function rather than purely computational benchmarks. Matching human-expert performance on the Eterna OpenKnot competition and producing active divergent ribozymes demonstrates that learned 3D priors can generalize beyond known natural sequences. Released under the MIT license with open code (~308 GitHub stars) and HuggingFace checkpoints and datasets, it provides an accessible foundation for RNA design research, though performance remains bounded by the limited size and resolution of experimental RNA structure data.