Understanding Conformational Transition of Macrocyclic Peptides through Deep Learning

Molecular conformations play a critical role in determining molecular properties such as membrane permeability, binding affinity, and ultimately therapeutic efficacy. Both experimental and computational approaches can characterize conformations within local energy minimum, and calculations of conformational free energy provide insight into why certain conformations are thermodynamically preferred over others. However, focusing solely on sampled conformations provides only a static view of the conformational landscape which may not fully illustrate why molecules result in such a conformational ensemble. Conformational transition between different conformations further explains how and why of the conformations which further inform molecule design. Here, we introduce Internal Coordinate Net (ICoN) version 1 (v1), a deep learning model trained in MD simulation data to learn the underlying physics that governed cyclic peptide conformational dynamics. ICoN-v1 enables the identification of transient conformations and all torsion rotations between local energy minima. By following the minimum-energy pathway (MEP) in the models latent space, ICoN-v1 efficiently generates fully atomistic transition pathways that capture detailed backbone and side-chain interactions governed by concerted torsional rotations. Notably, ICoN-v1 produces smooth transition pathways that are absent from the training data, demonstrating strong generalization beyond the sampled MD conformations. Analysis of the resulting concerted torsional motions and transient states highlights key residues involved at different stages of the transition, providing mechanistic insight that can inform cyclic peptide design and drug discovery.