Transfer Learning Allows Accurate RBP Target Site Prediction with Limited Sample Sizes.

Area of Science:

• Molecular Biology
• Bioinformatics
• Computational Biology

Background:

• RNA-binding proteins (RBPs) are crucial for biological processes.
• RBP dysfunction is linked to diseases like cancer and neurodegenerative disorders.
• Accurate prediction of protein-RNA binding sites is essential for understanding RBP function and disease.

Purpose of the Study:

• To develop and evaluate a novel transfer learning (TL) approach for predicting RNA binding sites.
• To address the challenge of limited experimentally validated binding site data for training deep learning (DL) models.
• To compare the performance of TL against training from scratch (SCR) and existing state-of-the-art methods.

Main Methods:

• Implemented a sophisticated and interpretable DL architecture.
• Trained models using both SCR and TL approaches.
• Benchmarked performance against current state-of-the-art methods.
• Investigated the impact of input features (sequence, evolutionary conservation) and interval sizes.
• Incorporated an attention layer for model interpretability.

Main Results:

• TL significantly enhances model performance, particularly for datasets with limited training data.
• Satisfactory prediction accuracy can be achieved with as few as a few hundred RNA binding sites using TL.
• Integrating both sequence and evolutionary conservation information yields superior performance.
• The attention layer aids in interpreting predictions within a biological context.

Conclusions:

• Transfer learning is a powerful strategy to overcome data limitations in predicting RNA binding sites.
• Combining sequence and evolutionary data improves prediction accuracy.
• The developed model offers enhanced performance and interpretability for RBP binding site prediction.