A robust ability to selectively modulate protein–protein interactions (PPIs) would provide a valuable means to control specific biological processes for therapeutic intervention. Unfortunately, owing to their flat and large interfaces, PPIs are challenging targets for traditional small molecule drugs. Cyclic peptides represent a promising solution to target PPIsthey can directly mimic the binding functionalities at the PPIs and have enhanced biostability and bioavailability compared to their linear counterparts. However, there are only ~50 cyclic peptide drugs. Most are simply natural products or their derivatives, rather than deriving from successful de novo cyclic peptide development. A major reason why the design of novel, functional cyclic peptides has proven difficult is the need to simultaneously optimize multiple drug-related properties of cyclic peptides, e.g., binding affinity, water solubility, and membrane permeability. Because cyclic peptides often have ≤12 residues and are connected in a ring, even changing one amino acid can dramatically alter the properties of cyclic peptides. Hence, changing cyclic peptide sequences to optimize for one property often negatively impacts other properties. Machine learning (ML), now widely used to build predictive models for drug properties, holds enormous potential to guide successful cyclic peptide design. Unfortunately, the few attempts at ML models to predict cyclic peptide properties perform quite poorly. The core challenge is that most cyclic peptides, including the current cyclic peptide drugs, adopt multiple conformations in water. It is, therefore, difficult for ML models to decipher how sequence modifications impact the complicated structural ensembles of cyclic peptides, which in turn influence their properties. If we can provide the ML models with this missing structural information, we will greatly improve their performance in predicting cyclic peptide properties. Since no robust experimental methods are available to characterize and quantify the conformations in a cyclic peptide structural ensemble, computational chemistry represents a logical alternative. Although recent work has revealed that explicit-solvent molecular dynamics (MD) simulation is capable of providing high-quality structural predictions of cyclic peptides, it is far too slow to be used at scale. On the other hand, computational methods that provide speedy predictions for cyclic peptide structural ensembles are inaccurate. Our long-term objective is a reliable, robust, and user-friendly platform for the computational design of potent, bioavailable cyclic peptides targeting PPIs. In our first aim, we use explicit-solvent MD results of diverse sequences as training datasets to build novel ML models that can predict cyclic peptide structural ensembles, preserving both accuracy and speed. In our second aim, we use these high-quality structural ensembles as a new descriptor for ML inputs to enable the training of high-per...