Ernie-Switcher: RNA foundation model-guided design of programmable RNA devices in a navigable landscape

Recent advances in biological foundation models have enabled new opportunities for data-driven molecular engineering. However, designing programmable RNA devices remains challenging because of the vast, nonlinear sequence-structure-function landscape. Here, we present Ernie-Switcher, a deep learning framework built upon the RNA foundation model (RNAErnie), comprising three coupled components: (i) integration of RNAErnie-derived sequence-phenotype representations for realistic design generation and filtering; (ii) characterization of RNA landscape navigability as a theoretical basis for sequence optimization; and (iii) incorporation of biophysical constraints to maintain optimization trajectories within biologically plausible bounds.

We validated Ernie-Switcher using RNA toehold switches. Theoretically, the observed lower bound on threshold navigability (86.4%) of the RNA landscape supports the theoretical feasibility of RNA switch sequence optimization. Experimentally, the optimized toehold switches achieved ON/OFF expression ratios of up to 35-fold in Escherichia coli. Collectively, RNAErnie-guided sequence optimization provides a scalable strategy for engineering high-performance RNA devices. To facilitate broad adoption, we deployed Ernie-Switcher as a freely accessible web platform.

The Python code and pre-trained models are available at https://github.com/zhutiblah/Ernie-Switcher and are archived on Zenodo at https://doi.org/10.5281/zenodo.21859277. The RNA toehold switch design platform is available at https://ernie-switcher.bioinformatics-syn.org/.

Supplementary data are available at Bioinformatics online.