Allosteric prediction via convolutional neural networks and protein structural and dynamical features

Area of Science:

• Biochemistry
• Computational Biology
• Machine Learning

Background:

• Allostery regulates protein function through distal site modulation, vital for cellular processes.
• Predicting allosteric states aids functional annotation and drug development by enabling specific allosteric site targeting.

Purpose of the Study:

• To develop a machine learning approach for predicting protein allosteric functional states.
• To utilize the KRas protein, implicated in cancer, as a model system for this prediction.

Main Methods:

• Employed convolutional neural networks (CNNs), specifically GoogLeNet and ResNet18, for supervised learning.
• Utilized structural and dynamical features (interatomic distances, contact maps, covariance, mutual information) as image-like inputs for CNNs.
• Fine-tuned pretrained CNN architectures to classify KRas into active or inactive states.

Main Results:

• Atomic contact maps proved to be the most effective structural feature for prediction.
• Linearized mutual information surpassed covariance in capturing relevant dynamical correlations.
• Models achieved significant validation accuracy, with atomic contact maps reaching up to 90% accuracy.

Conclusions:

• Deep learning integrating global structural rearrangements and correlated motion can reliably predict protein allosteric states.
• This approach offers a promising framework for understanding allosteric regulation.
• The findings support the development of targeted therapeutics by accurately predicting protein functional states.