Contrastive pre-training and 3D convolution neural network for RNA and small molecule binding affinity prediction

Area of Science:

• Computational chemistry
• Drug discovery
• Structural biology

Background:

• RNA molecules are key targets for novel therapeutics due to their diverse structures and functions.
• Accurate prediction of RNA-small molecule binding affinity is essential for advancing RNA-targeted drug discovery.
• Current computational methods for predicting RNA-small molecule binding affinity are limited.

Purpose of the Study:

• To develop a novel computational model for predicting RNA-small molecule binding affinity.
• To leverage deep learning and 3D structural information for improved prediction accuracy.
• To provide a tool for virtual screening in RNA-targeted drug development.

Main Methods:

• Introduction of RLaffinity, a deep learning model utilizing 3D convolutional neural networks (3D-CNN).
• Integration of RNA pocket and small molecule information within a 3D structural context.
• Implementation of a contrastive learning-based self-supervised pre-training model to enhance feature extraction.

Main Results:

• RLaffinity demonstrated superior performance in predicting RNA-small molecule binding affinity compared to baseline methods across all metrics.
• The 3D-CNN effectively captured both global RNA pocket and local nucleotide information.
• Self-supervised pre-training significantly boosted the model's predictive capabilities.

Conclusions:

• RLaffinity is the first deep learning-based method for predicting RNA-small molecule binding affinity from 3D structures.
• The model shows significant potential as a tool for virtual screening in RNA-targeted drug discovery.
• Accurate prediction of binding affinity is achievable through advanced deep learning approaches on structural data.