While the "AlphaFold moment" revolutionized protein biology, RNA remains a significantly more difficult challenge for structural biology. The core of this difficulty is a massive data imbalance: while public repositories contain hundreds of thousands of protein structures, high-resolution RNA structures represent less than one percent of that volume. As Di Liu notes in the development of RhoFold+, this scarcity creates a bottleneck that prevents deep learning models from learning a generalized "structural grammar" of RNA.
To overcome this, researchers are moving beyond simple sequence-to-structure mapping, instead using geometric deep learning, generative flow matching, and transfer learning to squeeze every bit of utility from the limited experimental data available.
The Data Bottleneck
The disparity between protein and nucleic acid data is the defining constraint of the field. Duy Duong-Tran and Shuhua Yang point out that protein structures outnumber RNA structures by a factor of twenty-five. This scarcity is compounded by the inherent flexibility of RNA; unlike the relatively stable folds of proteins, RNA molecules are dynamic and fluid, making them harder to capture via X-ray crystallography or cryo-EM.
This lack of data leads to a "memorization" problem. Marko Ludaic argues in Limits of deep-learning-based RNA prediction methods that current state-of-the-art models often succeed by recognizing motifs they saw during training rather than generalizing to novel folds. When models encounter structural motifs that represent a negligible fraction of the training data, they typically fail to generalize the necessary geometric requirements.
Geometric Intuition and 2.5D Graphs
To make models more data-efficient, researchers are encoding physical and geometric priors directly into the model architectures. Instead of treating RNA as a 1D string of nucleotides, tools like RNAglib, developed by Carlos Oliver and colleagues, represent RNA as "2.5D" multi-relational graphs. This representation captures the complex spatial relationships and non-canonical base pairs that define RNA tertiary structure.
Taking this further, Lei Fan and Guang Yang introduced HyperRNA, which uses hypergraphs to model higher-order, multi-atomic interactions. By explicitly modeling the physics of how multiple atoms interact simultaneously, these models can learn more from fewer examples than a standard neural network could.
Generative Models for De Novo Design
The field is also shifting from predicting structures to designing them—a task known as inverse folding. Arian Rokkum Jamasb and Chaitanya K. Joshi developed gRNAde, a geometric deep learning pipeline that designs sequences for specific 3D shapes. Unlike previous methods like Rosetta, which could take hours, these graph-based models can design valid sequences in milliseconds.
Recent breakthroughs have applied "flow matching" and diffusion—the tech behind image generators—to RNA backbones. RNA-FrameFlow adapts SE(3) flow matching to the flexible geometry of RNA, allowing for the de novo generation of entirely new RNA structures. Similarly, RiboDiff by Xiner Li and others uses a joint diffusion process to co-design sequences and structures simultaneously, leveraging pre-trained molecular priors to compensate for the lack of raw RNA data.
Transfer Learning from the Protein World
If RNA data is scarce, why not borrow from the protein world? Yiren Jian demonstrated in CoT-RNA that models trained on massive protein datasets can "transfer" their structural intuition to RNA. By translating RNA sequences into a protein-like "language," the model can apply the folding principles learned from tens of thousands of protein structures to the data-poor domain of RNA.
This is complemented by the rise of massive RNA foundation models. AIDO.RNA, a 1.6-billion-parameter model pre-trained on 42 million sequences, uses the abundance of sequence data to build a latent understanding of RNA function, which can then be fine-tuned for 3D structure tasks where structural data is the limiting factor.
Beyond Isolated Structures
Finally, researchers are realizing that RNA doesn't exist in a vacuum. The CARD model, developed by Zixun Zhang and Zhen Li, shifts the focus from isolated RNA to its protein-binding context. By fusing RNA tertiary structures with protein features, models can better predict how these molecules interact in the complex environment of the cell, moving us closer to the automated design of high-affinity RNA-protein therapeutics.