Predicting improved protein conformations with a temporal deep recurrent neural network.

Area of Science:

• Computational Biology
• Biophysics
• Machine Learning

Background:

• Accurate protein structure prediction remains a challenge, often relying on template-based modeling.
• Template-based methods require experimentally resolved homologous structures, limiting their applicability.
• Molecular dynamics (MD) simulations refine models by exploring conformational landscapes, but identifying optimal states is difficult.

Purpose of the Study:

• To develop a deep recurrent network, DeepTrajectory, for identifying improved conformational states from MD simulations.
• To leverage temporal patterns in MD trajectory data for accurate classification of conformational quality.
• To provide a re-trainable and adaptable deep learning solution for protein structure prediction.

Main Methods:

• A deep recurrent network architecture (DeepTrajectory) was designed to process MD trajectory data.
• The model learns temporal dependencies within features extracted from simulation snapshots.
• Classification categories include improved quality, decreased quality, or no change in conformational state.

Main Results:

• DeepTrajectory achieved high precision in identifying improved conformational states across diverse MD sampling protocols.
• The model demonstrated superior performance compared to state-of-the-art machine learning algorithms that do not account for temporal dependencies.
• The study utilized 904 trajectories from 42 protein systems, totaling over 1.7 million snapshots for training and testing.

Conclusions:

• DeepTrajectory represents the first time-dependent deep learning protocol for MD-based protein structure refinement.
• The model's re-trainable and adaptable nature allows integration with various MD sampling procedures.
• This work highlights the potential of neural networks to learn complex temporal dynamics in the protein folding process.