A deep learning approach to programmable RNA switches.

Area of Science:

• Synthetic biology
• Computational biology
• RNA engineering

Background:

• Engineered RNA elements are programmable tools for detecting biomolecules.
• Predicting the function of these synthetic RNA components is challenging.
• Deep learning offers potential for enhanced pattern recognition in this field.

Purpose of the Study:

• To investigate Deep Neural Networks (DNNs) for predicting toehold switch function in synthetic biology.
• To develop a data-driven approach for understanding RNA behavior.

Main Methods:

• Synthesized and characterized a large dataset of 91,534 toehold switches in vivo.
• Trained Deep Neural Networks on nucleotide sequences of these RNA switches.
• Compared DNN performance against thermodynamic and kinetic models.

Main Results:

• DNNs trained on nucleotide sequences achieved superior predictive performance (R² = 0.43–0.70) compared to existing models (R² = 0.04–0.15).
• Attention visualizations (VIS4Map) provided insights into RNA switch success and failure modes.
• Demonstrated the capability of DNNs to predict functionality and generate insights.

Conclusions:

• Deep learning approaches are effective for predicting functionality in RNA synthetic biology.
• DNNs offer a powerful tool for designing and understanding engineered RNA elements.
• This work advances the application of AI in RNA engineering and synthetic biology design.