Poster abstracts
Poster number 87 submitted by Yuntong Zou
Million-Scale Functional Profiling Enables Machine Learning for RNA Function Prediction and Design in ZTP Riboswitches
Yuntong Zou (nterdisciplinary Biological Sciences Program, Northwestern University, Evanston, IL, USA), Skyler L. Kelly (Department of Biological Sciences, The University at Buffalo, Buffalo, NY, USA), Eric J. Strobel (Department of Biological Sciences, The University at Buffalo, Buffalo), Julius Lucks (Department of Chemical and Biological Engineering, McCormick School of Engineering and Applied Science, Evanston, IL, USA)
Abstract:
Accurate function prediction and rational design of RNA remain central challenges in synthetic biology since RNA activity depends on complex relationships among sequence, structure, and biophysical context. Recent advances in machine learning for RNA modeling have been limited by the scarcity of large-scale datasets with quantitative RNA functional measurements; a challenge that high-throughput experimental assays are now beginning to overcome. Here, we leverage Transcription Elongation Complex display (TECdisplay) to obtain ZTP riboswitch sequence-function measurements at up to the million-sequence scale, providing a foundation for machine learning of RNA function.
Using this dataset, we evaluated multiple sequence representations and model architectures in two complementary tasks: predicting RNA function from sequence and designing candidate sequences for desired functional outputs. Initial results suggest that compact predictive models can achieve useful accuracy for riboswitch function prediction, and that lightweight generative models can produce candidate sequences with biologically plausible characteristics at target functional levels. The in-vivo experimental validation further supports the reliability and predictive accuracy of the model. These findings indicate that relatively efficient architectures may capture important functional constraints in this system and provide a practical starting point for design-oriented RNA modeling. Together, this work provides a large-scale RNA functional dataset and a machine learning framework for RNA function prediction and design, with clear potential to incorporate structural and biophysical priors, knowledge distillation of pretrained RNA foundation models to build more advanced RNA models. In this context, ZTP riboswitches profiled by TEC Display offer a useful benchmark for investigating how RNA machine learning models can achieve both high precision within a specific system and useful generalizability across related RNAs and downstream tasks.
Keywords: RNA Biology, Synthetic Biology, Machine Learning