Predicting nucleic acid binding sites by attention map-guided graph convolutional network with protein language...

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• Protein-nucleic acid interactions are vital for gene expression, replication, and transcription.
• Current prediction models use AI, including protein language models and graph neural networks, with structure-based methods offering high accuracy but requiring 3D structures.
• Sequence-based methods are being revisited due to the limitations of structure-based approaches for large-scale sequence data.

Purpose of the Study:

• To develop a novel, highly accurate, sequence-based prediction framework for protein-nucleic acid binding sites.
• To overcome the limitations of structure-based methods by utilizing only protein sequences.
• To improve upon existing sequence-based prediction methods through advanced AI techniques.

Main Methods:

• Proposed a novel framework, ATtention Maps and Graph convolutional neural networks to predict nucleic acid-protein Binding sites (ATMGBs).
• Fused protein language embeddings with physicochemical properties for multiview information.
• Leveraged attention maps from protein language models and employed graph convolutional networks for enhanced feature representation.

Main Results:

• ATMGBs demonstrated significantly improved performance in sequence-based binding site prediction.
• The framework achieved prediction accuracy comparable to structure-based methods.
• Evaluated on multiple independent test sets, confirming robust performance.

Conclusions:

• ATMGBs offers a powerful and accurate sequence-based approach for predicting protein-nucleic acid binding sites.
• The method addresses the challenge of predicting binding sites without requiring 3D protein structures.
• The developed framework advances the field of computational prediction for protein-nucleic acid interactions.