RiboFlow: Conditional De Novo RNA Co-Design via Synergistic Flow Matching

Ribonucleic acid (RNA) binds to molecules to achieve specific biological functions. While generative models are advancing biomolecule design, existing methods for designing RNA that target specific ligands face limitations in capturing RNA's conformational flexibility, ensuring structural validity, and overcoming data scarcity. To address these challenges, we introduce RiboFlow, a synergistic flow matching model to co-design RNA structures and sequences based on target molecules. By integrating RNA backbone frames, torsion angles, and sequence features in an unified architecture, RiboFlow explicitly models RNA's dynamic conformations while enforcing sequence-structure consistency to improve validity. Additionally, we curate RiboBind, a large-scale dataset of RNA-molecule interactions, to resolve the scarcity of high-quality structural data. Extensive experiments reveal that RiboFlow not only outperforms state-of-the-art RNA design methods by a large margin but also showcases controllable capabilities for achieving high binding affinity to target ligands. Our work bridges critical gaps in controllable RNA design, offering a framework for structure-aware, data-efficient generation.

Recent work has laid foundations for RNA design. Tools like RNAiFold [13] and gRNAde [24] generate sequences matching predefined secondary or tertiary structures, while RNA-FrameFlow [3], MMDiff [31], and RNAFlow [32] focus on backbone generation. Yet, designing RNA for smallmolecule targeting remains an open problem due to three gaps: (1) the inherent conformational flexibility of RNA requires the simultaneous and consistent consideration of both its structure and sequence [15]; (2) existing models lack explicit conditioning on ligand geometry, limiting their ability to capture RNA-ligand binding dynamics; and (3) the scarcity of RNA-ligand structural data restricts the scalability and generalizability of data-driven approaches.

To bridge these gaps, we identify three key challenges. First, ensuring generated RNAs satisfy both binding specificity and biophysical validity requires co-designing sequence and structure in a 39th Conference on Neural Information Processing Systems (NeurIPS 2025). synergistic framework. Second, modeling the structural flexibility of RNA, especially its torsion angles and backbone dynamics, while maintaining sequence-structure compatibility, remains a complex problem involving both geometric and thermodynamic considerations. Third, the absence of large-scale, standardized RNA-ligand interaction datasets hinders training robust generative models.

We try to address these challenges. For model design, we introduce RiboFlow, a synergistic flow matching model for de novo RNA discrete sequence and continuous structure co-design. By conditioning on ligand geometry and leveraging RNA backbone frames, torsion angles, and sequence features, RiboFlow models conformational flexibility while enforcing sequence-structure consistency. A novel co-design pre-training strategy is proposed to further enhance geometric awareness by distilling structural priors from RNA crystal structures. To overcome data scarcity, we introduce RiboBind, a comprehensive dataset of RNA-ligand complexes systematically curated from the PDB database, comprising 1,591 RNA-ligand complexes and 3,012 RNA-ligand pairs. Our contributions are: (i) Task Formulation: We propose a first-of-its-kind synergistic flow-matching framework for ligand-conditioned de novo RNA design. The model incorporates torsion angle and backbone frame modeling, enabling sequence-structure co-design for specified ligands while offering controllable ligand-binding specificity. (ii) Dataset: We present RiboBind, a large standardized RNA-ligand interaction benchmark, enabling data-driven RNA discovery. (iii) Evaluation: We develop a multi-faceted pipeline assessing structural validity and binding affinity (via docking and scoring). Experimental results demonstrate that RiboFlow outperforms state-of-the-art RNA design methods by a large margin (e.g., achieving a 2.2-fold improvement in the AF3 binding metric and a 50% increase in validity), but also showcases controllable capabilities for achieving high binding affinity to target ligands. We anticipate this work advancing RNA design toward structure-aware, ligand-conditioned design, with promising applications in therapeutics and synthetic biology.

2 Related Work

Currently, RNA design can be broadly categorized into two main approaches: sequence-based and structure-based methods. Sequence-based design primarily aims to address the RNA inverse folding problem, which involves designing an RNA sequence that folds into a desired RNA structure. While early efforts [13,46,9,34] largely focus on RNA secondary structure information, recent studies [38,24,22,44] begin to explore the use of RNA 3D structural information to guide sequence design. On the other hand, structure-based design, still in its early stages, encompasses approac