Training Neural Network Models Using Molecular Dynamics Simulation Results to Efficiently Predict Cyclic Hexapeptide...

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Machine learning

Background:

• Cyclic peptides are promising therapeutics, but their de novo design is difficult.
• Current cyclic peptide drugs are often natural products, limiting design flexibility.
• Understanding cyclic peptide structural ensembles in solution is crucial for rational drug design.

Purpose of the Study:

• To evaluate the effectiveness of machine learning models in predicting cyclic peptide structural ensembles.
• To improve upon existing methods for cyclic peptide structure prediction, particularly for larger cyclic peptides.
• To explore the utility of advanced neural network architectures for capturing complex molecular interactions.

Main Methods:

• Utilized molecular dynamics simulations to generate training data for machine learning models.
• Developed and applied the StrEAMM (Structural Ensembles Achieved by Molecular Dynamics and Machine Learning) method.
• Compared linear regression models with convolutional neural networks (CNNs) and graph neural networks (GNNs) for prediction accuracy.

Main Results:

• Linear regression models showed good performance for cyclic pentapeptides (R² = 0.94) but struggled with cyclic hexapeptides (R² = 0.47).
• Including (1,4) interactions improved linear models for hexapeptides (R² = 0.75).
• CNNs and GNNs achieved high accuracy for both cyclic pentapeptides (R² = 0.97) and hexapeptides (R² = 0.91), capturing nonlinear interactions.

Conclusions:

• Advanced neural network architectures significantly enhance the prediction of cyclic peptide structural ensembles compared to traditional methods.
• The findings demonstrate the potential of machine learning, particularly deep learning, to overcome challenges in cyclic peptide drug design.
• This work provides a powerful computational tool for the rational design of novel cyclic peptide therapeutics.