Predicting RNA sequence-structure likelihood via structure-aware deep learning.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• RNA functionality is critically dependent on its sequence and structure.
• Accurate RNA sequence-structure pair evaluation is vital for researchers.
• Existing machine learning models face challenges with feature selection and characterization scope.

Purpose of the Study:

• To develop advanced deep learning models for evaluating RNA sequence-structure pairs.
• To address limitations in feature engineering for RNA characterization.
• To enhance the efficacy of RNA design processes through improved modeling.

Main Methods:

• Developed NU-ResNet, a convolutional neural network model encoding RNA sequence-structure information into a 3D matrix.
• Developed NUMO-ResNet, building on NU-ResNet, incorporating 2D folding motifs extracted via an automated method.
• Evaluated model performance on independent datasets and through 10-fold cross-validation.

Main Results:

• Both NU-ResNet and NUMO-ResNet demonstrated superior performance compared to existing literature models on independent test datasets.
• Models showed robust performance across different RNA families, indicating strong generalization ability.
• The incorporation of nucleotide-level and structural motif features enhanced predictive accuracy.

Conclusions:

• Introduced NU-ResNet and NUMO-ResNet, novel deep learning models for RNA sequence-secondary structure evaluation.
• These models represent advancements in data-driven approaches for RNA research.
• Proposed a new method for encoding RNA sequence-structure pairs, facilitating better RNA design.