RiboCouple links RNA backbone generation to sequence design through a learned interaction topology

Abstract

Designing RNAs with prescribed structures is central to RNA engineering. Backbone-first approaches generate geometry before sequence assignment, yet how to transfer relational information to sequence design remains unclear. We introduce RiboCouple, which infers a probabilistic residue-pair state from backbone geometry, recycles pair probabilities through a second pass and transfers them to sequence decoding as soft attention biases. On 550 generated backbones, pair-type decoding achieved joint refolding validity of 45.1 percent versus 33.5 percent for a matched topology-free decoder. In another 110-backbone cohort, replacing aligned pair states with statistics-matched donor states reduced validity from 46.4 percent to 21.4 percent, demonstrating that decoding depends on backbone-aligned relational information. Before external refolding, pair-confidence ranking enriched validity to 59.6 percent in the top 248 and transferred to other generators; on a 94-chain temporally held-out native-backbone benchmark, pair-coupled decoding improved secondary-structure agreement. These findings establish backbone-inferred interaction topology as an intermediate representation connecting backbone generation, sequence decoding, and candidate prioritization.