Project
Machine Learning for RNA Structure Design
Research Manuscripts
Journal submissions pendingTwo manuscripts developed from my NSERC USRA research on reinforcement-learning-based RNA inverse design. Both are currently working manuscripts being prepared for journal submission.
Oracle Selection in Reinforcement Learning-Based RNA Inverse Design
This project investigates an intuitive question in reinforcement-learning based RNA inverse design: does a potentially more accurate structure predictor necessarily produce a better reward model?
RIDER is a diffusion-based framework for 3D RNA inverse design that uses reinforcement learning to guide sequence generation. I modified its RL pipeline to make the folding oracle interchangeable, allowing different structure-prediction models to be evaluated without changing the rest of the training system. AlphaFold3 was then integrated as an alternative to RhoFold, and the two oracles were compared across the diverse RNA structures in the Das14 benchmark.
The work involved reinforcement-learning fine-tuning, diffusion-model inference, structural similarity evaluation, reward-distribution analysis, and large-scale GPU experiments on high-performance computing clusters via the Digital Research Alliance of Canada (DRAC) and Compute Canada. The results provide insight into how oracle choice affects reward stability, training dynamics, and sequence-design performance.
Secondary-Structure Reward Shaping for RL-Based RNA Inverse Design
A second direction of this research asks whether RIDER can be given a richer training signal by incorporating the computed and discarded RNA secondary structure data back into its reinforcement-learning objective.
RIDER optimizes sequence generation through rewards derived from three-dimensional structural self-consistency. This work extends that framework with secondary-structure-aware reward information, allowing the policy to receive feedback about structural organization at an additional level of RNA representation during fine-tuning.
The project involved modifying RIDER's reward pipeline, integrating secondary-structure evaluation into the existing reinforcement-learning workflow, and studying how the additional objective affects optimization, training behavior, and the quality of generated RNA sequences.