Identification of molecular compounds targeting bacterial propionate metabolism with topological machine learning.

This study demonstrates the transformative potential of machine learning in drug discovery by integrating comparative protein and ligand analysis with novel topological machine learning methods. Our approach sifts through large chemical libraries to identify promising molecular structures for targeting specific proteins with high precision. While many machine learning models have proven effective on benchmark datasets, we apply these techniques to discover compounds targeting methylcitrate dehydratase (AcnD), the second enzyme in the bacterial propionate catabolism pathway. Propionate catabolism is essential in pathogenic bacteria for utilizing host derived lipids and amino acids. Inefficient removal of propionate can lead to toxic accumulation that threatens bacterial survival, making this pathway a potential antimicrobial target. We translate ligand molecular structures into topological vectors and use tailored topological models to prioritize compounds with characteristics consistent with blocking the AcnD active site. Molecular docking simulations indicate that prioritized compounds interact with key amino acid residues critical to AcnD function. Among these, 2-methylidenebutanedioic acid (itaconic acid, itaconate) ranks highly as a potential molecular scaffold for targeting AcnD. Using bacterial growth assays, we find that itaconate at 29.13 mM completely inhibits the growth of Pseudomonas aeruginosa and Acinetobacter baumannii in carbon rich liquid cultures. These findings reinforce itaconate's potential as an antimicrobial metabolite and support the hypothesis that it can disrupt bacterial propionate catabolism, potentially by inhibiting AcnD and promoting the accumulation of toxic intermediates. Overall, our study underscores the value of integrating topology based ligand modeling with comparative sequence structure function analysis and docking to identify molecular scaffolds with favorable geometric fit, energy, and interaction profiles, guiding downstream optimization and experimental validation. Our code is available at (https://github.com/AstritTola/Molecular-Compounds-Targeting).