Wuhan University
Abstract
RNA has emerged as an important programmable molecule in synthetic biology, therapeutics, and biotechnology. Its three-dimensional structure is closely linked to biological functions such as regulation, specific molecular recognition, and catalysis, making RNA a focus of growing interest in biomanufacturing and healthcare. Developing methods for de novo RNA design based on three-dimensional structures is therefore essential for investigating RNA function and advancing RNA molecular engineering.
Several models for three-dimensional RNA design have recently been proposed. Most use deep neural networks to learn mappings between RNA structures and naturally occurring sequences. However, the scarcity of experimentally determined RNA structures and the many-to-many relationships between sequences and structures limit the accuracy and generalizability of existing models. Their potential for experimental applications also remains unclear.
Here, we present DS3dRNA, a de novo RNA design model that incorporates higher-order interactions. Built on the TriRNASP framework for coarse-grained, many-body statistical potentials, DS3dRNA combines Markov chain Monte Carlo sampling with GPU parallelization to efficiently explore sequence space through hierarchical population sampling, screening, mutation, and optimization of local thermodynamic parameters.
Rigorous benchmarking uses structural datasets from CASP15 and CASP16, together with recently released CASP17 structures. The results show that DS3dRNA offers clear performance advantages in both single-state and multistate design across sequence recovery, macro-averaged F1 score, and the RMSD and TM-score of AlphaFold 3-predicted structures relative to their targets. We also experimentally validated functional RNAs designed de novo with DS3dRNA, including the twister ribozyme and the Mango-II fluorogenic aptamer.
Keywords: RNA 3D structure; de novo design; higher-order interactions; multistate design
English translation of the author-provided oral-presentation abstract.
