Alignment-driven Cross-Graph Modeling for 3D RNA Inverse Folding

Abstract

The 3D RNA inverse folding task aims to design RNA sequences that can fold into a target tertiary structure. This task is complicated by the inherent structural complexity and flexibility of RNA molecules. Considering that conserved RNA structures often exhibit sequence similarity, we present AlignIF, a pipeline designed to generate RNA sequences conditioned on the target RNA 3D structure and its multiple structure alignment (MStA), explicitly incorporating structural evolutionary conservation. AlignIF introduces a novel MStA backbone encoder designed to model multiple geometric graphs derived from MStA, which can better capture meaningful information from MStA structures. Additionally, AlignIF uses a random-ordered autoregressive sequence decoder to generate candidate RNA sequences. In our constructed benchmark dataset, AlignIF achieves a perplexity of 2.32 and a recovery rate of 57. 08%, respectively, a relative improvement of 17.8% and 11.8% over the state- of-the-art methods. In addition, AlignIF excels in diversity and ranking capability. We further demonstrate the added value of the incorporated MStA structures along with the proposed encoder architecture.