Deep neural networks for interpreting RNA-binding protein target preferences

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Deep learning models excel at analyzing RNA-binding protein (RBP) binding sites.
• However, the interpretability of these complex models for understanding RBP binding preferences and functions remains limited.

Purpose of the Study:

• To design and validate a multitask and multimodal deep neural network for characterizing in vivo RBP targets.
• To enhance the interpretability of deep learning models for RBP binding site analysis.

Main Methods:

• Developed a multitask and multimodal deep neural network incorporating RNA sequence and binding site region type as input.
• Quantified the contribution of input features to the predictive score for each RBP.
• Trained the model across multiple RBPs simultaneously to mitigate experimental biases.

Main Results:

• The multimodal deep learning model improved prediction performance by integrating sequence and region type data.
• Identified key RNA sequence motifs and transcript context patterns crucial for individual RBP predictions.
• Findings align with known RBP motifs and binding behaviors, offering new functional insights.

Conclusions:

• The developed deep learning approach provides a powerful, interpretable method for characterizing RBP binding sites.
• This work advances the understanding of RBP regulatory functions by uncovering critical sequence and contextual determinants.