Predicting the functional state of protein kinases using interpretable graph neural networks from sequence and...

Area of Science:

• Biochemistry and Structural Biology
• Computational Biology and Bioinformatics

Background:

• Protein kinases regulate crucial cellular functions and are key targets for treating diseases like cancer.
• Existing structural data for kinases is extensive, but automated methods to link structure to function are limited, hindering drug discovery.
• Developing efficient, automated techniques is vital for advancing structural biology and therapeutic development.

Purpose of the Study:

• To develop an interpretable graph neural network (GNN) framework for classifying protein kinase functional states (active/inactive) using only tertiary structure and amino acid sequence.
• To leverage graph neural networks and Graph Grad-CAM for automated identification of functionally critical residues and contacts without manual input.
• To utilize the interpretable framework for analyzing structural variations across kinase sub-classes and enhancing protein engineering.

Main Methods:

• Implementation of a graph neural network (GNN) framework utilizing protein tertiary structure and amino acid sequence for kinase state classification.
• Application of Gradient-weighted Class Activation Mapping for graphs (Graph Grad-CAM) to identify structurally important residues and contacts.
• Utilizing Grad-CAM maps as vector embeddings to discern subtle structural differences among kinase sub-classes in the Protein Data Bank (PDB).

Main Results:

• The GNN models achieved high accuracy (>97%) in classifying kinase structures into active and inactive states.
• Graph Grad-CAM successfully identified functionally critical motifs, including the conserved DFG and HRD motifs of the hydrophobic spine, consistent with existing literature.
• The framework effectively differentiated subtle structural variations among kinase sub-classes, demonstrating its utility in high-throughput analysis.

Conclusions:

• The developed interpretable GNN framework provides an automated and accurate method for determining protein kinase structure-function relationships.
• This approach accelerates the identification of critical functional sites, aiding in the design of targeted small molecule therapies.
• The framework holds significant potential for engineering novel proteins and advancing drug discovery efforts in kinase-related diseases.