A Deep Learning Model for RNA-Protein Binding Preference Prediction Based on Hierarchical LSTM and Attention Network

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Identifying protein-binding regions in RNA sequences is critical for understanding gene regulation.
• Conventional and existing deep learning methods struggle to effectively learn these important sequence features.

Purpose of the Study:

• To develop an improved method for identifying functionally important regions in RNA sequences, specifically those involved in protein binding.
• To leverage attention mechanisms and LSTM for enhanced feature extraction and importance evaluation in RNA sequences.

Main Methods:

• Utilized Long Short-Term Memory (LSTM) networks to extract correlation features within RNA sequences.
• Implemented an attention mechanism to weigh the importance of different RNA sequence sites.
• Optimized model performance through hyperparameter experiments, including k-mer length, stride window, sentence length, and optimization functions.
• Investigated the impact of k-mer vector length and RNA structure data on model performance.

Main Results:

• The proposed method demonstrated superior performance compared to existing approaches.
• Hyperparameter optimization identified optimal settings for k-mer parameters and the attention mechanism.
• Model performance was sensitive to variations in k-mer vector length and related parameter settings.
• The inclusion of RNA structure data further enhanced model performance.

Conclusions:

• The combined LSTM and attention mechanism approach effectively identifies important RNA sequence regions for protein binding.
• The study provides insights into optimal parameter settings for k-mer based RNA sequence analysis.
• This method offers a significant advancement in RNA sequence analysis for biological applications.