Enhancing RNA Foundation Models via Secondary Structure Modelling

Genomic Foundation Models (GFMs) are reshaping our understanding of the code of life, yet their application to Ribonucleic Acid (RNA) is uniquely challenged by its functional dependence on complex structure. Modelling this intricate sequence-structure-function relationship represents a vital problem in computational biology. This thesis aims to resolve three core impediments to progress in structure-aware RNA Foundation Model (FM). First, the interpretability gap, which obscures how models leverage structural information for biological discovery. Second, the fundamental sequence-structure alignment bottleneck, which hinders the sophisticated modelling of RNA for tasks such as design. Finally, the reproducibility crisis, which impedes the rigorous evaluation of competing GFMs. To address these challenges, this thesis presents a multi-scale solution focused on advancing RNA FMs via structure-aware modelling. First, we developed PlantRNA-FM, a high-performance, interpretable foundation model pre-trained with integrated structural information for the plant domain. This model successfully identified new, experimentally validated functional RNA structural motifs, demonstrating the potential of GFMs as engines for structure-based scientific discovery. Second, we propose the OmniGenome model, which fundamentally resolves the sequence-structure alignment problem by learning a robust, bidirectional mapping. The model achieved unprecedented success on the challenging EternaV2 RNA design benchmark, demonstrating a true capability for structure-based generative tasks. Finally, to tackle systemic evaluation issues, we constructed OmniGenBench, a modular, automated benchmarking platform for the rigorous assessment of structure-aware GFMs. Integrating over 31 open-source models and 123 standardised datasets, it provides a transparent and reproducible evaluation standard for the entire field. Taking the above solutions into consideration, this thesis not only enhances the performance and functionality of RNA FMs but also facilitates their transformation from opaque predictors into trustworthy, interpretable tools in structure-aware genomics.<p></p>

Author dc:creator

• Heng Yang (21042086)

• RNA
• Genomics
• Model
• Benchmark

Statement dc:rights

• All rights reserved
• Open Access after 2027-08-02

Identifier

10779/exe.31230154.v1

OAI identifier oai:identifier

oai:figshare.com:article/31230154