Classifying protein kinase conformations with machine learning

Area of Science:

• Structural biology
• Computational biology
• Pharmacology

Background:

• Protein kinases are crucial in cell signaling and are significant drug targets.
• Kinase activity is regulated by conformational changes, particularly in the catalytic domain's activation loop and DFG motif.
• Accurate structural annotation of kinases is vital for targeted drug design but requires scalable and interpretable methods.

Purpose of the Study:

• To develop and validate interpretable machine learning models for automated annotation of protein kinase structures.
• To classify kinase structures based on their active/inactive states and DFG motif conformations (DFG-in, DFG-out, other).
• To identify key structural features driving kinase conformational states and assess the accuracy of predicted structures from tools like AlphaFold2.

Main Methods:

• Collected and curated a diverse dataset of protein kinase catalytic domain sequences and structures.
• Clustered structures based on DFG conformation and manually annotated them for training.
• Developed ensemble decision tree models, initially using 1692 structural variables, to classify kinase states and DFG conformations.

Main Results:

• The active/inactive classification model achieved 99.9% accuracy on 3289 structures.
• The DFG conformation model achieved >99.8% accuracy on 8826 structures.
• Identified key structural variables, primarily near the activation loop, that are critical for classification, providing insights into conformational preferences.

Conclusions:

• Interpretable machine learning models offer a robust and automated approach for large-scale structural annotation of protein kinases.
• These models accurately classify kinase conformations, aiding in the understanding of signaling pathways and drug targeting.
• Analysis of AlphaFold2-predicted structures revealed discrepancies in DFG-in proportions compared to the Protein Data Bank, highlighting the utility of these models for evaluating predicted structures.