Unsupervised learning of collective variables for conformational sampling of cyclic peptides

While many approaches exist for machine learning (ML) collective variables (CVs) using supervised learning techniques, unsupervised methods for CV discovery remain limited. However, these unsupervised approaches are particularly important for systems with numerous metastable states, where labeling conformations is impractical or ambiguous. Developing such methods is also challenging, especially when aiming to connect a large number, sometimes hundreds, of metastable states. In this study, we employed principal component analysis (PCA) and autoencoders to derive CVs and evaluated their performance across cyclic peptides of varying sizes. For cyclic pentapeptides, ML-based CVs performed comparably to the conventional CVs proposed by McHugh et al. However, their effectiveness declined as peptide size increased, likely due to the increasing complexity of the conformational transition pathways. These findings highlight the growing need for the development of more efficient and robust unsupervised ML methods for CV construction.