Computer aided drug design and discovery of novel antibacterial compounds: applicative and methodological developments

Over recent decades, drug discovery has decisively transitioned from an experiment-centric endeavor to a data-driven, predictive in silico discipline. Although established computational methods have long complemented traditional workflows, the advent of artificial intelligence has accelerated this transition, positioning computer-aided drug design and discovery (CADD) at the vanguard of biomedical research. This thesis presents the two complementary facets of my doctoral research: the application of established in silico pipelines for the identification of novel bioactive small molecules and the development of innovative computational methodologies to advance drug discovery. A hybrid structure- and ligand-based virtual screening strategy was employed to identify and characterize two structurally novel antibacterial compounds, each introducing a new chemical class. The first, C11, is a benzothiophene-dioxo allosteric inhibitor of the FtsZ protein that exhibits potent activity against Gram-positive pathogens, including clinical isolates, and effectively inhibits and eradicates biofilm formation. The second, RS5430, was identified as a putative allosteric inhibitor of DNA gyrase displaying significant activity against Gram-positive and fluoroquinolone-resistant strains. Both compounds displayed favorable safety profiles but were inactive against Gram-negative bacteria, owing to active efflux. To address this limitation and enhance potency and selectivity, an enhanced-sampling workflow based on Adaptive Steered Molecular Dynamics (ASMD) was employed to predict the potency of a rationally designed derivative. This workflow presented a generally applicable strategy for chemical optimization. Beyond these applications, two methodological innovations are presented. First, a novel computational pipeline integrating cosolvent molecular dynamics simulations with thermodynamic profiling is detailed. Its efficacy in identifying and characterizing high-affinity interaction hotspots within target binding sites is discussed. Second, a novel AI-driven methodology for de novo ligand design is introduced. This approach utilizes a Reinforcement Learning (RL) framework, guided by shape- and pharmacophore-based alignment, to autonomously generate optimized ligands in situ within the binding pocket. The validation and prospective utility of this method are discussed. Notably, this approach does not require a priori training and relies solely on the three-dimensional structure of the molecular target.