Translating from Proteins to Ribonucleic Acids for Ligand-binding Site Detection

Area of Science:

• Computational biology
• Structural biology
• Drug discovery

Background:

• Identifying ligand-binding sites on macromolecular targets is crucial for structure-based drug discovery.
• Deep learning models have demonstrated success in predicting protein ligand-binding sites.
• Predicting binding sites in RNA and RNA-protein complexes remains a challenge.

Purpose of the Study:

• To develop and evaluate deep learning models for predicting RNA and RNA-protein complex binding sites.
• To introduce a novel dataset splitting approach to mitigate structure-related bias in training data.
• To investigate the impact of pre-training neural networks on protein structures before fine-tuning on RNA.

Main Methods:

• Utilized three-dimensional convolutional neural networks for binding site prediction.
• Implemented a specialized dataset splitting strategy to ensure unbiased model training.
• Employed a transfer learning approach, pre-training models on protein structures and then fine-tuning on RNA.

Main Results:

• Models pre-trained on protein structures significantly outperformed models trained solely on RNA structures.
• The developed approach correctly located 71% of known RNA binding sites within 4 Å of their centers.
• The dataset splitting method effectively minimized structure-related bias in the training data.

Conclusions:

• Protein pre-training is a highly effective strategy for enhancing deep learning-based RNA binding site prediction.
• The study presents a promising methodology for identifying druggable sites in RNA targets.
• This work advances the application of deep learning in RNA-focused drug discovery.