JAX-RNAfold: scalable differentiable folding.

Area of Science:

• Computational Biology
• Bioinformatics
• RNA Structure Prediction

Background:

• Differentiable folding is a novel approach for RNA design, optimizing probabilistic sequence representations using gradient descent.
• Existing methods face memory limitations, restricting RNA sequence lengths to under 50 nucleotides.

Purpose of the Study:

• To present JAX-RNAfold, a software package significantly enhancing differentiable RNA folding algorithms.
• To enable scalable RNA design for sequences up to 1,250 nucleotides using a single GPU.

Main Methods:

• Developed a drastically improved differentiable folding algorithm implemented in JAX.
• Optimized the algorithm to reduce memory overhead associated with differentiating the expected partition function.
• Packaged the algorithm into an open-source software, JAX-RNAfold, installable as a Python package.

Main Results:

• JAX-RNAfold demonstrates scalability to 1,250 nucleotides on a single GPU, a 25-fold increase over previous methods.
• The software allows differentiable folding to be integrated into deep learning pipelines.
• Facilitates complex RNA design tasks, including mRNA design with adaptable objective functions.

Conclusions:

• JAX-RNAfold overcomes the scalability limitations of prior differentiable RNA folding techniques.
• The software provides a powerful tool for advanced RNA design and analysis in computational biology.
• Enables new possibilities for deep learning applications in RNA sequence and structure optimization.