Accelerating antibiotic discovery by leveraging machine learning models: Application to identify novel inorganic complexes

Abstract

New antibiotics are required to combat the emergence of drug-resistant bacteria. S. aureus is a Gram-positive pathogen that often displays multidrug resistance. Through conventional screening approaches, the discovery of new antibiotics against S. aureus has proven to be challenging. Molecular property prediction of novel antibiotics candidates by machine-learning (ML) methods has increased the rate at which such molecules are identified. The bottleneck of the existing approaches relies on the structure similarities for the existing antibiotics. Then the question about discovering and developing new unconventional antibiotic classes has challenged preconceptions about the scope and applicability of the existing methods. Herein, we developed an ML approach that predicts the minimum inhibitory concentration (MIC) of Re-complexes towards two S. aureus strains (ATCC 43300 - MRSA and ATCC 25923 - MSSA). In our framework, we tailored a Multi-layer Perceptron (MLP) by inherently accounting for the structure features of the Re- complexes to develop a prediction model for antimicrobial activity assessment. Although our approach is demonstrated with a specific example based on the rhenium carbonyl complexes, the predictive model can be readily adapted to other candidate metal complexes. The developed model emphasizes the application of machine learning in the de novo design of a novel generation of antibiotic molecules with targeted activity against a challenging pathogen.