RNAcoder: Tokenizing RNA structure with all-atom autoencoders

Protein structure tokens already unlocked several applications: embedded in predictive models, they improve prediction of protein properties and interactions. Exploring their latent space can help reveal unknown protein biology. Expanding this advance to RNA holds significant potential, especially since it could help reconcile atomic-level information with the coarse-grained modeling imposed by data scarcity. Achieving this precision is a key step toward achieving specific RNA-targeting therapeutics. Unlike proteins, RNA folding is heavily driven by base pairing and base stacking, and base pairs are a fundamental and information-rich RNA subunit, along with individual nucleotides. Building on this RNA specificity, we introduce RNAcoder, an all-atom RNA tokenizer extending structural tokenization to base pairs and backbone connections. RNAcoder explicitly tackles modified nucleotides, which are key to RNA function and interactions. The induced latent space recovers known Leontis-Westhof base pair families and chemical modifications semantics in a fully unsupervised manner, validating the biological relevance of the learned embeddings. Plugged into a graph neural network, RNAcoder tokens substantially improve performance over existing representations on small molecule binding site prediction and RNA-protein interface prediction. Our tokens represent a promising direction for RNA structure prediction and analysis. Our code is available anonymously at: https://anonymous.4open.science/r/RNAcoder-233E