A deep encoder-decoder framework for identifying distinct ligand binding pathways.

Area of Science:

• Computational biology
• Biophysics
• Machine learning

Background:

• Ligand-receptor interactions are crucial for biological activity.
• Molecular Dynamics (MD) simulations model protein-ligand binding but face challenges in analyzing high-dimensional trajectory data.
• Understanding ligand pathways is key to elucidating binding mechanisms.

Purpose of the Study:

• To develop an efficient deep learning framework for analyzing complex ligand-binding pathways.
• To overcome the limitations of traditional methods in identifying distinct ligand trajectories.
• To accurately quantify the free energy and kinetics of ligand binding.

Main Methods:

• Utilized an autoencoder-based deep neural network trained on residue-ligand distances.
• Applied the framework to a T4 lysozyme L99A mutant and its ligand benzene.
• Compared autoencoder performance with linear dimensionality reduction techniques like time-structured independent component analysis.

Main Results:

• The autoencoder successfully identified multiple distinct ligand recognition pathways without user intervention.
• Discovered intermediate ligand locations within protein helices before native pose recognition.
• The derived low-dimensional latent space accurately quantified ligand binding free energy and kinetics.

Conclusions:

• Autoencoder deep neural networks offer an efficient method for discovering complex ligand-binding pathways.
• This approach is particularly effective for systems with transient or low-populated intermediates.
• The framework provides a powerful tool for quantitative analysis of protein-ligand interactions.