Predicting Antimicrobial Class Specificity of Small Molecules Using Machine Learning

While the useful armory of antibiotic drugs is continually depleted due to the emergence of drug-resistant pathogens, the development of novel therapeutics has also slowed down. In the era of advanced computational methods, approaches like machine learning (ML) could be one potential solution to help reduce the high costs and complexity of antibiotic drug discovery and attract collaboration across organizations. In our work, we developed a large antimicrobial knowledge graph (AntiMicrobial-KG) as a repository for collecting and visualizing public <i>in vitro</i> antibacterial assay. Utilizing this data, we build ML models to efficiently scan compound libraries to identify compounds with the potential to exhibit antimicrobial activity. Our strategy involved training seven classic ML models across six compound fingerprint representations, of which the Random Forest trained on the MHFP6 fingerprint outperformed, demonstrating an accuracy of 75.9% and Cohen's Kappa score of 0.68. Finally, we illustrated the model's applicability for predicting the antimicrobial properties of two small molecule screening libraries. First, the EU-OpenScreen library was tested against a panel of Gram-positive, Gram-negative, and Fungal pathogens. Here, we unveiled that the model was able to correctly predict more than 30% of active compounds for Gram-positive, Gram-negative, and Fungal pathogens. Second, with the Enamine library, a commercially available HTS compound collection with claimed antibacterial properties, we predicted its antimicrobial activity and pathogen class specificity. These results may provide a means for accelerating research in AMR drug discovery efforts by carefully filtering out compounds from commercial libraries with lower chances of being active.