Differentiable Folding for Nearest Neighbor Model Optimization.

Area of Science:

• Computational Biology
• Biophysics
• Bioinformatics

Background:

• The Nearest Neighbor (NN) model is the standard for RNA secondary structure thermodynamics.
• Current NN models have numerous parameters, making optimization computationally intensive.
• Accurate thermodynamic parameters are crucial for RNA structure prediction and sequence design.

Purpose of the Study:

• To develop an efficient and scalable method for optimizing thermodynamic parameters of RNA folding models.
• To leverage differentiable folding for improved parameter fitting using experimental and structural data.
• To create a significantly improved parameter set for enhanced RNA structure prediction.

Main Methods:

• Utilized differentiable folding to compute gradients of RNA folding algorithms.
• Developed a flexible parameter optimization framework using known RNA structures and thermodynamic data.
• Introduced the RNAometer database of experimentally determined stabilities for RNA model systems.

Main Results:

• Achieved a significantly improved thermodynamic parameter set for RNA folding models.
• Demonstrated superior performance over existing baselines across all evaluation metrics.
• Showcased a >23-fold increase in the average predicted probability of ground-truth RNA sequence-structure pairs.

Conclusions:

• The new parameter optimization framework offers a scalable and efficient approach for RNA modeling.
• This work enables the flexible incorporation of diverse data types and advanced machine learning techniques.
• The findings pave the way for drastically improved RNA structure prediction and design tools.