An Interpretable Convolutional Neural Network Framework for Analyzing Molecular Dynamics Trajectories: a Case Study...

Area of Science:

• Computational Biology and Biochemistry
• Structural Biology
• Machine Learning in Bioinformatics

Background:

• Molecular dynamics (MD) simulations are crucial for understanding biomolecular systems but analyzing vast conformational data to identify functional states and key residues remains challenging.
• The 'black-box' nature of traditional deep learning methods limits their interpretability in analyzing complex MD trajectories.
• G-protein-coupled receptors (GPCRs) are critical drug targets, and understanding their diverse active states and activation mechanisms is essential.

Purpose of the Study:

• To develop an interpretable deep learning framework for automatic identification of diverse active states from MD trajectories of GPCRs.
• To address the limitations of black-box models by providing interpretable insights into the functional mechanisms.
• To identify important residues that regulate distinct GPCR activities and activation pathways.

Main Methods:

• Introduction of a pixel representation to avoid information loss in conformational structure representation.
• Development of an interpretable convolutional neural network (CNN)-based deep learning model (ICNNMD) for feature extraction and classification.
• Implementation of a local interpretable model-agnostic explanation interpreter to identify key residues contributing to classification results.

Main Results:

• Achieved over 99% classification accuracy for three diverse GPCR systems with distinct active states.
• Successfully identified important residues involved in regulating biased activities of GPCRs, providing insights into their activation mechanisms.
• Demonstrated the model's capability as a general tool for analyzing MD trajectories across various biomolecular systems.

Conclusions:

• The ICNNMD model provides an effective and interpretable approach for analyzing MD simulations and identifying functional states and critical residues in GPCRs.
• The interpretability feature aids in elucidating complex activation mechanisms and biased signaling pathways of GPCRs.
• The developed framework has broad applicability for analyzing MD trajectories in diverse biomolecular research areas.