Structure prediction of cyclic peptides by molecular dynamics + machine learning.

Area of Science:

• Computational chemistry
• Biophysics
• Machine learning

Background:

• Cyclic peptides are crucial therapeutics, but their conformational flexibility impacts cell permeability.
• Current methods struggle to predict the full structural ensembles of cyclic peptides, especially those with broad conformational distributions.
• Understanding these ensembles is key for rational drug design.

Purpose of the Study:

• To develop a rapid and accurate method for predicting the complete structural ensembles of cyclic peptides.
• To enable the design of novel cyclic peptide therapeutics by overcoming limitations in conformational prediction.

Main Methods:

• Utilized molecular dynamics (MD) simulation data from cyclic pentapeptides to train machine learning (ML) models.
• Developed a novel technique, StrEAMM (Structural Ensembles Achieved by Molecular Dynamics and Machine Learning), for structure prediction.
• Validated ML model predictions against extensive explicit-solvent MD simulations.

Main Results:

• Achieved predictions of structural ensembles with accuracy comparable to multi-day MD simulations.
• The StrEAMM method predicts structures for hundreds of thousands of cyclic peptide sequences in under one second per peptide.
• Demonstrated significant speed improvement (seven orders of magnitude) without sacrificing accuracy.

Conclusions:

• StrEAMM offers the first efficient method for predicting complete cyclic peptide structural ensembles.
• This accelerates the rational design of cyclic peptide therapeutics, particularly those with complex conformational landscapes.
• The approach significantly enhances the speed and scope of cyclic peptide structure prediction.