RiboC2F: Pose-First Coarse-to-Fine Flow Matching for Protein-Conditioned RNA Co-Design

Computational design of ribonucleic acid (RNA) is essential for therapeutic discovery and synthetic biology. Designing RNAs that bind specified protein targets is bottlenecked by a scarcity of paired-complex 3D structural data. Existing approaches focus on importing structural priors from external biomolecular datasets, leaving where to localize the complex signal itself unresolved. We propose RiboC2F, a coarse-to-fine flow-matching framework that concentrates this signal first and foremost on the global binding pose. RiboC2F runs a two-stage cascade over a shared trunk, where a coarse stage commits the binding pose and a fine stage generates the full RNA design. The coarse pose enters as a conditioning signal rather than a fixed initialization, avoiding cross-stage error accumulation. On PRI30k, RiboC2F delivers the strongest docking among published baselines, with competitive intrinsic validity and multi-sample diversity. Probing further shows that the fine stage exploits the coarse pose's direction but is robust to pointwise coordinate corruption, confirming pose direction as the operative coarse-signal axis.