Can ‘Bacterial-Metabolite-Likeness' Model Improve Odds of ‘in Silico' Antibiotic Discovery?

‘Inductive' QSAR descriptors have been used to develop the series of QSAR models enabling ‘in silico'
distinguishing between antimicrobial compounds, conventional drugs, and druglike substances. The constructed
neural network-based models operating by 30 ‘inductive' parameters have been validated on an extensive
set of 2686 chemical structures and resulted in up to 97% accurate separation of the three types of molecular
activities. The demonstrated ability of ‘inductive' parameters to adequately capture molecular features
determining ‘antibiotic-like' and ‘druglike' potentials have been further utilized to construct a model of
‘Bacterial-Metabolite-Likeness' (BML). The same ‘inductive' descriptors have been used to train a neural
network that could very accurately recognize substances involved into bacterial metabolism (that have been
experimentally identified). When the developed model has been applied to the mixed set of antimicrobials,
drugs, and druglike chemicals (not used for training the BML model), it exhibited a 2−5-fold recognition
preference toward antimicrobial compounds compared to general drugs and an 18- to 45-fold preference
when compared to a druglike substance (depending on the model stringency). These results illustrate immanent
similarity between conventional antimicrobials and native bacterial metabolites and suggest that the developed
BML model can be an effective classification tool for ‘in silico' antibiotic studies.