Dresden 2026 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

BP: Fachverband Biologische Physik

BP 18: Focus session: Integrative Structural Modeling

BP 18.5: Talk

Wednesday, March 11, 2026, 12:15–12:30, BAR/0106

RNA 3D Folding Using Diffusion Models and Agentic Tree Search — •Arunodhayan Sampathkumar and Danny Kowerko — Professorship of Media Informatics, Technische Universität Chemnitz

Accurate prediction of RNA three-dimensional structure is essential for understanding RNA function and guiding RNA-based therapeutics. However, RNA folding remains challenging due to complex tertiary interactions, context-dependent base pairing and limited structural data, especially for short sequences with fewer than 80 nucleotides. To address this, we introduce a multi-model RNA structure prediction framework that combines an RNA-adapted Protenix model, a Boltz diffusion sampler and a template-based modeling (TBM) baseline together with GAN-augmented training for underrepresented short RNAs. Protenix offers accurate local geometry, Boltz supplies diverse global conformations and TBM contributes strong constraints when templates exist. All models are evaluated on the Kaggle RNA 3D public and private test sets. Protenix reaches TM-scores of 0.48/0.46, Boltz 0.41/0.40 and TBM 0.61/0.57. We further introduce an agentic tree-search ensemble that selects and refines conformations using consensus scoring and RMSD-based diversity. This ensemble significantly improves performance, achieving TM-scores of 0.68 (public) and 0.63 (private). Our results demonstrate that integrating generative models, template signals and agentic search yields more accurate RNA structures and improves robustness for novel RNA families.

Keywords: RNA 3D structure prediction; GAN augmentation; diffusion models; template-based modeling; agentic ensemble search