RNA-X

RNA-X, developed by the Cicek Lab at Bilkent University and released as a bioRxiv preprint in November 2025, is described by its authors as the first RNA interaction foundation model. Where most prior RNA models focus on structure prediction or sequence representation in isolation, RNA-X is built around the problem of molecular interaction: how a given RNA sequence engages with a partner molecule. This framing lets the model treat the design of functional RNAs—aptamers, regulatory elements, and binders—as a conditional generation task defined by a target.

The model is a masked language model pretrained on more than 100 million RNA-target interactions, learning joint representations of RNA molecules and the partners they bind. Critically, the same pretrained model can be conditioned on proteins, DNA, or other RNA molecules as the target, enabling conditional design of RNA sequences against all three target classes without retraining a separate model for each. The authors further demonstrate zero-shot design against novel targets that were not part of the training interactions.

This positions RNA-X alongside the broader wave of biological foundation models (ESM for proteins, RNA structure models such as RhoFold, and interaction-aware co-folding systems) while staking out interaction-conditioned RNA design as its distinct contribution. Because design is driven by the target rather than by a fixed task head, a single model spans a design space that previously required bespoke, target-specific pipelines.

#Key Features

• Interaction-centric foundation model: Pretrained on 100M+ RNA-target interactions, RNA-X learns joint RNA-and-target representations rather than modeling RNA sequence or structure in isolation.
• Multi-target conditional design: The same model conditionally designs RNA sequences against protein, DNA, and RNA targets, and can design a single RNA to bind multiple targets simultaneously—without retraining per target type.
• Zero-shot generalization: The authors demonstrate design against novel targets not seen during pretraining, indicating the learned interaction representations transfer beyond the training distribution.
• Masked language modeling backbone: A masked-LM pretraining objective over interaction data provides the self-supervised signal that underpins both representation learning and conditional generation.

#Technical Details

RNA-X is a masked language model trained end-to-end on more than 100 million RNA-target interactions to learn joint representations of RNA and target molecules. The authors report that the trained model can design an RNA sequence to target RNA, DNA, and proteins, and can produce a single RNA molecule that simultaneously binds multiple targets. Specific architectural hyperparameters (layer count, hidden dimensions, total parameter count) and quantitative benchmark figures are not detailed in the preprint abstract or the public repository documentation at the time of writing. The GitHub repository (ciceklab/RNA-X) references checkpoint and prediction weight paths but does not spell out a public weight-download mechanism, so reproducing the model's generative capabilities may require contacting the authors. The repository states a CC BY-NC-SA 2.0 license for academic use, with commercial use directed to the corresponding authors; the preprint itself is distributed under CC BY-NC-ND.

#Applications

RNA-X targets RNA design problems where a functional sequence must be tailored to a specific molecular partner: designing aptamers and binders against protein targets, RNA sequences that engage DNA or other RNAs, and multi-specific RNAs intended to bind several targets at once. Researchers in synthetic biology, RNA therapeutics, and molecular diagnostics could use such a model to propose candidate sequences computationally before committing to wet-lab synthesis and binding assays, with zero-shot conditioning offering a route to explore targets for which no curated interaction data exist.

#Impact

As a self-described first RNA interaction foundation model, RNA-X frames RNA design as a unified, target-conditioned generation problem rather than a collection of separate task-specific tools, which is a notable conceptual contribution to the RNA modeling landscape. Its real-world impact remains to be established: as a recent (November 2025) preprint, its design claims await peer review and independent experimental validation, and the absence of a published model card, data card, clear benchmark numbers, and a documented weight-release mechanism currently limits reproducibility and external adoption.