RiboDFM: Robust RNA sequence design using backbone-conditioned discrete flow matching

Abstract

Designing RNA sequences that fold into predefined 3D structures is central to RNA engineering, yet it remains challenging due to the scarcity of experimentally determined structures and the complex sequence-structure relationships. Here, we present RiboDFM, a backbone-conditioned discrete flow-matching framework for RNA sequence design. RiboDFM progressively resolves fully masked sequences by integrating a geometric backbone encoder, an auxiliary base-pairing prediction head, and language model representations into a discrete flow sampler. To expand our training data, we supplemented experimental structures with a large-scale collection of trRosettaRNA2-predicted models. Across diverse benchmarks, RiboDFM achieves high sequence recovery and generates sequences with greater tertiary-structure self-consistency than existing methods. RiboDFM also performs strongly on aptamer-focused design tasks. Finally, beyond sequence generation, the nucleotide probabilities derived from RiboDFM' s flow trajectory serve as an intrinsic measure of sequence-structure compatibility. This enables reliable confidence estimation for designed sequences, as well as reference-free, zero-shot assessment of predicted structure models. Together, these results establish RiboDFM as an effective framework for structure-based RNA sequence design.