Cross-target transfer enables small molecule binding RNA generation

Riboswitches, aptamer biosensors and ligand-gated gene circuits all depend on RNA that recognises a chosen small molecule, and each is still built by selecting binders experimentally, one target at a time. However, identifying RNA sequences that recognize a desired compound remains challenging because of the vast RNA sequence space and the limited amount of experimentally characterized RNA–compound interaction data. We introduce RIFM-Mol, which casts compound-conditional RNA design as an additional interaction modality within a model already trained on RNA interaction data. RIFM-Mol generates candidate binders directly from a SMILES string, requiring no target structure, docking or iterative optimisation, and is trained and evaluated on drug-like chemical matter. Across two held-out benchmarks, the generated sequences achieve predicted binding comparable to experimentally validated binders and far above natural non-binders, while remaining novel and folding more stably than shuffled controls. Generation is genuinely compound-driven because sequences designed for the same compound are markedly more similar to one another (62\%) than to those designed for different compounds (41\%), a separation that does not emerge when the same task is approached by fine-tuning a sequence-only RNA foundation model. Code and data will be made publicly available upon acceptance.