Random Forest Modeling to Predict Small Molecule Accumulation in Gram-Negative Bacteria

Despite extensive efforts over the past ∼60 years to discover new classes of Gram-negative-active antibiotics, the development pipeline remains relatively dry. These failures can be largely ascribed to the complexity of the Gram-negative membranes and an inadequate understanding of physicochemical properties associated with compound permeation, efflux evasion, and overall accumulation in Gram-negative pathogens. Recent work using unbiased accumulation assays and advanced chemical descriptors has utilized machine learning algorithms, particularly random forest modeling, to correlate physicochemical properties of small molecules with accumulation. The ability of a random forest classifier to effectively process large data sets while still providing human-interpretable insight makes this model a valuable tool in accumulation data analysis. Here, we describe in detail a workflow reliant on random forest modeling to probe physicochemical trends linked to small molecule permeation and/or efflux liabilities in Gram-negative pathogens. Applications using random forest modeling have led to guidelines for small molecule accumulation in E. coli and P. aeruginosa that have resulted in multiple new Gram-negative-active antibiotics. Random forest modeling possesses several distinct advantages over other machine learning models for the purposes of identifying physicochemical trends in accumulation, and as such, provides valuable hypothesis-generating information for the study of small molecule accumulation in Gram-negative bacteria.