CPconf_score: A Deep Learning Free Energy Function Trained Using Molecular Dynamics Data for Cyclic Peptides.

Area of Science:

• Computational chemistry
• Structural biology
• Bioinformatics

Background:

• Characterizing cyclic peptide (CP) structures, particularly small ones with cis-peptide bonds, is vital for designing bioactive molecules but remains challenging.
• Accurate conformational prediction is essential for understanding CP function and guiding rational design.

Purpose of the Study:

• To develop a novel deep learning model for predicting the conformational free energies of cyclic peptides.
• To assess the model's accuracy in identifying near-native conformations compared to experimental structures and existing prediction tools.

Main Methods:

• High-temperature molecular dynamics (high-T MD) simulations were used to generate conformational ensembles for 250 cyclic peptides.
• The point-adaptive k-nearest neighbors (PAk) method was employed to estimate free energies of sampled conformations.
• A SchNet-based deep learning model (CPconf_score) was trained on the simulation data to predict conformational free energies.

Main Results:

• CPconf_score accurately predicted near-native conformations for 41 out of 50 cyclic peptides tested, achieving a backbone RMSD < 1.0 Å compared to crystal structures.
• The developed model significantly outperformed established tools like HighFold (12 CPs) and Rosetta (19 CPs) in predicting accurate cyclic peptide structures.
• The study demonstrates the potential of deep learning in advancing cyclic peptide structure prediction.

Conclusions:

• CPconf_score offers a highly accurate and efficient method for predicting cyclic peptide conformations.
• This advancement facilitates the rational design of novel bioactive cyclic peptides.
• The approach provides a valuable tool for researchers in drug discovery and peptide science.