Graph-based deep learning approach for high-throughput protein-DNA interaction scoring.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Deep Learning

Background:

• Accurate quantification of protein-DNA interactions (PDIs) is essential for understanding biological processes and drug design.
• The flexibility of nucleic acids presents a challenge for structural determination and training predictive models.
• Existing scoring functions (SFs) struggle with the complexity of PDI complexes.

Purpose of the Study:

• To develop a novel deep learning-based scoring function (SF) for predicting protein-DNA interactions (PDIs).
• To address the limitations of current methods in handling nucleic acid flexibility and large interaction interfaces.
• To create a robust and generalizable tool for PDI prediction in research and therapeutic design.

Main Methods:

• Developed PDIScore, a deep learning SF utilizing a comprehensive graph representation for nucleotide flexibility.
• Employed a scalable GraphGPS architecture with BigBird linear global attention for large interfaces.
• Integrated Mixture Density Networks (MDNs) to model residue-nucleotide distance distributions.
• Trained on a dataset of ~7000 protein-nucleic acid complex structures.

Main Results:

• PDIScore demonstrated superior performance across screening, docking, and ranking tasks compared to existing methods.
• Achieved excellent screening power (e.g., AUROC=0.82) and high docking success rate (48.94% top1).
• Case studies highlighted PDIScore's ability to elucidate biological mechanisms and identify key interaction sites.

Conclusions:

• PDIScore is a robust and generalizable deep learning tool for predicting protein-DNA interactions.
• It significantly advances the capabilities for PDI quantification, aiding biological research.
• Offers potential for accelerating therapeutic design by improving PDI prediction accuracy.