Cyclic peptides have gained interest as potential therapeutics due to their ability to target specific protein-protein interactions and be membrane-permeable. Understanding the sequence-structure relationship of cyclic peptides would greatly benefit their rational design. However, cyclic peptides tend to adopt multiple conformations in solution, and it remains challenging to use experimental techniques such as solution NMR to delineate their structural ensembles: i.e., the different structures a cyclic peptide adopts and the associated populations. Alternatively, molecular dynamics (MD) simulations can be used to provide such information. However, MD simulations are computationally expensive and not applicable for large-scale screening. Our group has developed the StrEAMM (Structural Ensembles Achieved by Molecular Dynamics and Machine Learning) computational platform and applied it to predict structural ensembles of head-to-tail cyclized pentapeptides and hexapeptides. However, head-to-tail cyclized peptides can be challenging to synthesize due to low yield and complicated reaction workup and product isolation. Furthermore, head-to-tail cyclized peptides are not compatible with screening techniques like mRNA display. Here, we expand the StrEAMM method to thioether-linked cyclic peptides, a popular scaffold in mRNA display. The trained graph neural network models are able to provide fast and simulation-quality structural ensembles for thioether-linked cyclic peptides. Using these models, we identify four thioether-linked cyclic pentapeptides that are predicted to be the best-structured and subsequently experimentally synthesize and characterize them by solution NMR. We observe general agreement between the predicted structures and the NMR results. Ultimately, we envision that StrEAMM-thioether models can work synergistically with the current mRNA platform to streamline the resource-intensive process of drug discovery and design of cyclic peptides.
Ho et al. (2026) studied this question.