DeepPBS: Geometric deep learning for interpretable prediction of protein-DNA binding specificity

Area of Science:

• Computational Biology
• Structural Biology
• Genomics

Background:

• Predicting protein-DNA binding specificity is crucial for understanding gene regulation.
• Existing methods face challenges in accurately predicting binding patterns across diverse protein families.

Approach:

• Introduced Deep Predictor of Binding Specificity (DeepPBS), a geometric deep-learning model for predicting protein-DNA binding specificity.
• The model analyzes protein-DNA structures to identify family-specific recognition patterns.
• DeepPBS is applicable to predicted structures and aids in modeling protein-DNA complexes.

Key Points:

• DeepPBS provides interpretable insights, calculating protein heavy atom-level importance scores that correlate with experimental mutagenesis data.
• The model demonstrates fast inference times, suitable for analyzing large datasets like molecular dynamics simulations.
• A case study on the p53-DNA interface highlights the model's interpretability and accuracy.

Conclusions:

• DeepPBS offers a robust foundation for machine-aided studies of protein-DNA interactions.
• The tool can guide experimental design and enhance the understanding of molecular interactions.
• Facilitates advancements in gene regulation research and protein-DNA complex design.