Fast and Versatile RNA Design via Motif-level Divide-and-Conquer and Structure-level Rival Search

TL;DR

This paper presents a fast, versatile RNA design algorithm that decomposes structures into motifs, combines partial sequences efficiently, and refines designs with rival search, achieving state-of-the-art results across multiple benchmarks.

Contribution

The method introduces a novel divide-and-conquer approach with motif-level decomposition and a rival search strategy for improved RNA design performance.

Findings

Achieves state-of-the-art results on native RNAsolo and Eterna100 benchmarks.

Significantly improves long-structure design, increasing folding probability from 0.18 to 0.39.

Offers an order-of-magnitude speedup over existing methods.

Abstract

RNA design aims to identify RNA sequences that fold into a target secondary structure. This task is challenging in terms of computational efficiency. Most existing methods focus on either minimum free energy (MFE)-based or ensemble-based metrics, leaving a gap for a unified approach that performs well across both. We introduce a fast and versatile RNA design algorithm inspired by our previous work on the undesignability of RNA structures and motifs (i.e., sets of contiguous structural loops). Our approach decomposes a target structure into a tree of sub-targets where each leaf node corresponds to a motif and each internal node corresponds to a substructure. We first design partial sequences for each motif, then these partial sequences are selectively and recursively combined via the cube pruning strategy borrowed from computational linguistics, enabling effective optimization of ensemble-based metrics. Finally, a novel whole-structure rival search further refines sequences to suppress misfolded alternatives and enhance MFE-based performance. Our method is highly efficient and also achieves state-of-the-art results on native RNAsolo structures and the Eterna100 benchmark, excelling in both ensemble- and MFE-based metrics. Additionally, it substantially improves the design of long-structure benchmark derived from 16S rRNA, increasing average folding probability from 0.18 to 0.39 with an order-of-magnitude speedup, demonstrating its effectiveness and scalability. Availability: Source code and data are available at: https://github.com/shanry/FastDesign.