RibonanzaNet

RibonanzaNet (also called RNet) is an RNA foundation model from the Das Lab at Stanford University, built on the Ribonanza dataset — chemical-mapping measurements on roughly two million diverse RNA sequences, totaling on the order of hundreds of millions of nucleotide-level measurements. The data were collected through "dual crowdsourcing": citizen-scientist sequence design via the Eterna platform and a public Kaggle competition (the Stanford Ribonanza RNA Folding challenge) that solicited and prospectively benchmarked deep learning models. The work was posted to bioRxiv in February 2024 and published in Cell in 2024.

Chemical mapping probes how accessible each nucleotide is in solution, providing an experimental readout that correlates with base pairing and structure. RibonanzaNet distills the winning ideas from the Kaggle challenge — together with the earlier RNAdegformer architecture — into a single self-contained network that predicts these reactivity profiles directly from sequence, without relying on pre-computed base-pairing matrices or external secondary-structure algorithms.

Because it is a pretrained model that captures general RNA structural signal, the fixed checkpoint can be applied to new problems without retraining, or fine-tuned for specialized tasks. A recent application paper (Townley, Kladwang, Baker, Das and colleagues) uses RibonanzaNet to guide de novo design of RNA pseudoknots, illustrating its role as a reusable structural oracle for RNA engineering.

#Key Features

• Trained on experimental chemical mapping at scale: Learns from ~2 million RNA sequences and hundreds of millions of reactivity measurements (2A3 and DMS profiles), rather than from limited 3D structure data.
• Dual-crowdsourced data: Combines Eterna citizen-scientist sequence design with a public Kaggle challenge that prospectively benchmarked model submissions.
• Self-contained sequence-to-structure: Predicts reactivity and structure directly from sequence, with no dependence on external base-pairing or folding algorithms.
• Reusable, fine-tunable checkpoint: A fixed pretrained model that can be applied zero-shot to guide RNA design, or fine-tuned for secondary structure, hydrolytic degradation, and experimental dropout.
• Open inference: The rnet-inference repository provides automatable, self-contained inference under an MIT license, with weights distributed via a Git submodule and on Kaggle.

#Technical Details

RibonanzaNet is a deep network combining Transformer encoder layers with 1D convolutions, adapted from the RNAdegformer design. It maintains both a sequence-level representation and a pairwise representation, exchanging information between them through outer-product-mean and triangular multiplicative updates — mechanisms reminiscent of structure-prediction networks — which lets structural signal flow bidirectionally rather than depending on pre-computed inputs. The exact parameter count of the base model is not prominently reported; the publicly cited "~100M-parameter" figure refers to the later RibonanzaNet2 successor, so the base size should be treated as unspecified. When fine-tuned, RibonanzaNet reaches state-of-the-art results on auxiliary benchmarks, including secondary-structure prediction (mean F1 around 0.89 on a PDB-derived test set and ~0.94 on CASP15 targets), improved pseudoknot recovery, and RNA degradation modeling that surpasses prior OpenVaccine competition winners. There is currently no HuggingFace model card; the canonical distribution is the GitHub rnet-inference repository with its RibonanzaNet-Weights submodule and Kaggle.

#Applications

RibonanzaNet serves RNA biologists and designers who need fast, experimentally grounded structural predictions. It supports secondary-structure and chemical-reactivity prediction, modeling of RNA hydrolytic degradation (relevant to mRNA-vaccine stability), and prediction of experimental sequence dropout. As a pretrained oracle, it can guide de novo design tasks such as constructing RNA pseudoknots, and its fine-tunable backbone provides a starting point for new RNA property-prediction models without retraining from scratch.

#Impact

RibonanzaNet demonstrated that large-scale crowdsourced chemical-mapping data, paired with a competitive benchmarking process, can yield a general-purpose RNA foundation model that outperforms specialized prior methods across several tasks. By releasing both the Ribonanza dataset and self-contained inference code, the Das Lab established a reusable resource for the RNA structure community, and the model has since been extended (RibonanzaNet2, ~100M parameters) and applied to downstream RNA design problems. Its main limitations are that it is rooted in chemical-reactivity rather than direct 3D structure, and that documentation is spread across GitHub and Kaggle rather than a single model card.