A deep learning framework to predict binding preference of RNA constituents on protein surface

Area of Science:

• Molecular Biology
• Bioinformatics
• Computational Biology

Background:

• Protein-RNA interactions are crucial for post-transcriptional gene regulation.
• Predicting these interactions from protein structures is a significant computational challenge.

Purpose of the Study:

• To develop a deep learning model, NucleicNet, for predicting protein-RNA interactions based on protein structure.
• To assess NucleicNet's ability to predict binding preferences and interaction modes.

Main Methods:

• Utilized a deep learning approach (NucleicNet) to analyze local physicochemical characteristics of protein surfaces.
• Trained and tested the model on diverse RNA-binding proteins, including Fem-3-binding-factor 2, Argonaute 2, and Ribonuclease III.
• Validated predictions against experimental data from structural biology, RNAcompete, Immunoprecipitation Assays, and siRNA Knockdown Benchmarks.

Main Results:

• NucleicNet accurately predicts binding preferences for RNA backbone constituents and bases from protein surface properties.
• The model successfully recovers known interaction modes for challenging RNA-binding proteins.
• Predictions align with experimental results even without in vitro or in vivo assay data.

Conclusions:

• NucleicNet offers a powerful tool for predicting RNA-binding sites and preferences in proteins.
• The model can quantify RNA sequence fitness for specific binding pockets.
• NucleicNet facilitates the discovery of novel RNA-binding proteins and their cognate RNAs.