Decoding GuaB: Machine Learning-Powered Discovery of Enzyme Inhibitors Against the Superbug Acinetobacter baumannii

Abstract

Background/Objectives: GuaB, which is known as inosine 5′-phosphate dehydrogenase (IMPDH), is an enzymatic target involved in the de novo guanine biosynthetic pathway of the multidrug-resistant (MDR) Acinetobacter baumannii. GuaB has emerged as a potential therapeutic target to cope with increasing antibiotic resistance. Here, we used machine learning-based virtual screening as a verification technique to find potential inhibitors possessing different chemical scaffolds, using structure-based drug design as a discovery platform. Methods: Four machine learning models, built based on chemical fingerprint data, were trained, and the best models were used for virtual screening of the ChEMBL library, which covers 153 active molecules. Molecular dynamics (MD) simulations of 200 ns were carried out for all three compounds in order to explain conformational changes, evaluate stability, and provide validation of the docking results. Post-simulation analyses include principal component analysis (PCA), bond analysis, free-energy landscape (FEL), dynamic cross-correlation matrix (DCCM), radial distribution function (RDF), salt-bridge identification, and secondary-structure profiling, etc. Results: For molecular docking, the screened compounds were used against the GuaB protein to achieve proper docked conformation. Upon visual examination of the best-docked compounds, three leads (lead-1, lead-2, and lead-3) were found to have better interaction with the GuaB protein in comparison to the control. The mean RMSD scores between the three leads and the control were between 2.54 and 2.89 Å. In addition, the three leads as well as the control were characterized for pharmacokinetic features. All three leads met Lipinski’s Rule 5 and were thus drug-like. PCA and FEL analyses showed that lead-2 exhibited improved conformational stability, identified as deeper energy minima, whereas RDF and DCCM analyses revealed that lead-2 and lead-3 exhibited strong local structuring and concerted dynamics. In addition, lead-2 displayed a very rich hydrogen-bonding network with a total of 460 frames possessing such interactions, which is the highest among the complexes investigated here. Based on entropy calculations and the maximum entropy method of gamma–gram, lead-1 proved to be the most stable one with the lowest binding free-energy. Conclusions: This study provides an integrated machine learning-based virtual screening pipeline for the identification of new scaffolds to moderate infections associated with AMR; however, in vitro validation is still required to assess the efficacy of such compounds.