Advanced Heterogeneous Network-Based Graph Neural Network Framework for Predicting Anti-CRISPR Protein Sequences.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Anti-CRISPR proteins are key regulators in bacterial-phage interactions, inhibiting CRISPR/Cas systems to ensure phage survival.
• Accurate prediction of anti-CRISPR proteins is vital for advancing phage-host immunity studies and CRISPR/Cas technologies.
• Existing methods analyzing proteins individually may miss crucial sequence similarities and inter-protein relationships.

Purpose of the Study:

• To develop an advanced computational framework for predicting anti-CRISPR proteins.
• To leverage graph neural networks for a more holistic analysis of protein networks.

Main Methods:

• Introduction of PACRGNN, a graph neural network framework.
• Construction of a heterogeneous protein network integrating sequence and structural similarities.
• Utilization of Graph Attention (GAT) and Graph Sample and Aggregation (GraphSAGE) layers to capture topological dependencies.
• Incorporation of six protein feature categories to enhance node representations.

Main Results:

• PACRGNN achieved high performance metrics on the validation set: 0.9577 accuracy, 0.9572 F1-Score, and 0.9876 PRAUC.
• The model outperformed existing methods on an independent test set from the NCBI database (Jan.-Oct. 2024).

Conclusions:

• PACRGNN provides a superior approach for anti-CRISPR protein prediction by considering network topology and diverse features.
• The framework offers significant potential for advancing research in phage-host interactions and CRISPR/Cas systems.