Enumeration, conformation sampling and population of libraries of peptide macrocycles for the search of chemotherapeutic cardioprotection agents

Abstract

Peptides are uniquely endowed with features that allow them to perturb previously difficult to drug biomolecular targets. Peptide macrocycles in particular have seen a flurry of recent interest due to their enhanced bioavailability, tunability and specificity. Although these properties make them attractive hit-candidates in early stage drug discovery, knowing which peptides to pursue is non" trivial due to the magnitude of the peptide sequence space. Computational screening approaches show promise in their ability to address the size of this search space but suffer from their inability to accurately interrogate the conformational landscape of peptide macrocycles. We developed an in" silico compound enumerator that was tasked with populating a conformationally laden peptide virtual library. This library was then used in the search for cardio" protective agents (that may be administered, reducing tissue damage during reperfusion after ischemia (heart attacks)). Our enumerator successfully generated a library of 15.2 billion compounds, requiring the use of compression algorithms, conformational sampling protocols and management of aggregated compute resources in the context of a local cluster. In the absence of experimental biophysical data, we performed biased sampling during alchemical molecular dynamics simulations in order to observe cyclophilin" D perturbation by cyclosporine A and its mitochondrial targeted analogue. Reliable intermediate state averaging through a WHAM analysis of the biased dynamic pulling simulations confirmed that the cardio" protective activity of cyclosporine A was due to its mitochondrial targeting. Paralleltempered solution molecular dynamics in combination with efficient clustering isolated the essential dynamics of a cyclic peptide scaffold. The rapid enumeration of skeletons from these essential dynamics gave rise to a conformation laden virtual library of all the 15.2 Billion unique cyclic peptides (given the limits on peptide sequence imposed). Analysis of this library showed the exact extent of physicochemical properties covered, relative to the bare scaffold precursor. Molecular docking of a subset of the virtual library against cyclophilin" D showed significant improvements in affinity to the target (relative to cyclosporine A). The conformation laden virtual library, accessed by our methodology, provided derivatives that were able to make many interactions per peptide with the cyclophilin" D target. Machine learning methods showed promise in the training of Support Vector Machines for synthetic feasibility prediction for this library. The synergy between enumeration and conformational sampling greatly improves the performance of this library during virtual screening, even when only a subset is used.