(577b) Discovering Molecules with Selective Membrane Partitioning Using Active Learning and Alchemical Transformation

In this work we integrate coarse-grained molecular dynamics simulation, alchemical free energy calculations, deep representational learning, and Bayesian optimization to discover to discover small organic molecules that selectively permeate into cardiolipin membranes. The phospholipid cardiolipin is exclusively found within the inner mitochondrial membrane in eukaryotic cells and is intimately involved in mediating metabolic and regulatory cascades. Deviations in cardiolipin composition are indicative of a number of pathologies, such as Tangier disease. Identifying molecules that selectively penetrate into membranes composed of cardiolipin to probe the pool of cardiolipin in vivo can serve as a diagnostic tool to identify pathologies in mitochondrial membranes. To effectively screen the space of small organic molecules less than ~450 Da, which are likely to passively permeate without the assistance of membrane transporters, we computationally search the molecular design space using molecular simulation and an active learning-directed molecular design platform. Candidate molecules are evaluated using alchemical transformations to compute the free energy of inserting into lipid bilayers of cardiolipin relative to those of POPG, a common phospholipid found throughout the eukaryotic cells. Deep representational learning is used to embed the discrete molecular design space into a smooth, low-dimensional latent space within which we conduct an active learning search for highly selective molecules using Gaussian process regression and Bayesian optimization. Using this platform, we identify a small number of novel candidate small molecules predicted to have superior selectivity for cardiolipin membranes relative to current state-of-the-art diagnostic drugs.