Abstract
Antimicrobial resistance presents a pressing challenge to public health, which requires the search for novel antimicrobial agents. Various experimental and theoretical methods are employed to understand drug-target interactions and propose multistep solutions. Nonetheless, efficient screening of drug databases requires rapid and precise numerical analysis to validate antimicrobial efficacy. Diptool addresses this need by predicting free energy barriers and local minima for drug translocation across lipid membranes. In the current study employing Diptool free energy predictions, the thermodynamic commonalities between selected antimicrobial molecules were characterized and investigated. To this end, various clustering methods were used to identify promising groups with antimicrobial activity. Furthermore, the molecular fingerprinting and machine learning approach (ML) revealed common structural elements and physicochemical parameters in these clusters, such as long carbon chains, charged ammonium groups, and low dipole moments. This led to the establishment of guidelines for the selection of effective antimicrobial candidates based on partition coefficients (logP) and molecular mass ranges. These guidelines were implemented within the Reinforcement Learning for Structural Evolution (ReLeaSE) framework, generating new chemicals with desired properties. Interestingly, ReLeaSE produced molecules with structural profiles similar to the antimicrobial agents tested, confirming the importance of the identified features. In conclusion, this study demonstrates the ability of molecular fingerprinting and AI-driven methods to identify promising antimicrobial agents with a broad range of properties. These findings deliver substantial implications for the development of antimicrobial drugs and the ongoing battle against antibiotic-resistant bacteria.