ORIGAMI: Orientation-Aware Graph Neural Network for Assessing Multimeric Interfaces of Protein Complex Structures.

Area of Science:

• Computational structural biology
• Deep learning applications
• Protein structure prediction

Background:

• Assessing the quality of computationally predicted multimeric protein structures is a significant challenge.
• Existing graph neural network methods for interface assessment overlook crucial 3D geometric and orientational features.

Purpose of the Study:

• To introduce ORIGAMI, an orientation-aware graph neural network designed for high-accuracy assessment of protein complex interfaces.
• To leverage both scalar and vector representations for symmetry-aware and SO(3)-equivariant geometric operations.

Main Methods:

• Developed ORIGAMI, an orientation-aware graph neural network incorporating scalar and 3D vector node representations.
• Implemented symmetry-aware and SO(3)-equivariant operations to capture fine-grained orientational relationships at protein-protein interfaces.
• Trained ORIGAMI to estimate the interface Local Distance Difference Test (iLDDT) score.

Main Results:

• ORIGAMI demonstrated superior performance on interface quality assessment benchmarks, including CASP challenges.
• Achieved significant gains in the CASP16 interface-level evaluation compared to baseline methods.
• Showcased robust cross-metric generalization by accurately predicting superposition-based DockQ scores.

Conclusions:

• ORIGAMI represents a significant advancement in assessing the quality of predicted protein complex structures.
• The method's orientation-aware approach enhances accuracy and generalization capabilities.
• ORIGAMI provides a valuable tool for computational structural biology research.