Predicting Active Antimicrobial Compounds Using Machine Learning

Background: Acinetobacter baumannii is a multidrug-resistant (MDR) pathogen rec ognized by the World Health Organization as a critical priority due to its high preva lence in hospital-acquired infections and limited treatment options.

Methods: To address the urgent need for novel therapeutics, artificial intelligence (AI)-based approaches were applied to predict compounds with potential antibacte rial activity against A. baumannii.

Results: The used training set included 11 084 compounds with experimentally deter mined minimum inhibitory concentrations (MICs). The prediction set to be analyzed included 5835 structurally diverse compounds filtered from the ZINC20 database.
Molecular descriptors, MACCS keys, and Morgan fingerprints were generated using RDKit, and a random forest classifier was trained using scaffold-based cross-val idation to classify compounds as active (MIC < 32 µg/mL) or inactive. The model achieved AUROC values of 0.73-0.82 and average precision scores of 0.84-0.90, demonstrating strong predictive performance. Application of the trained model to the prediction dataset identified 375 compounds (6%) as potentially active, includ ing 7 high-confidence candidates (probability > .85).

Conclusion: Scaffold analysis revealed considerable structural diversity among pre dicted compounds, supporting the potential for novel chemotypes. These findings highlight the utility of AI-driven drug discovery workflows for accelerating the iden tification of antibacterial agents targeting MDR A. baumannii.

Cite this article as: Arman İ. Predicting active antimicrobial compounds using machine learning. Trends in Pharmacy 2025, 2, 0018, doi: 10.5152/TrendsPharm.2025.25018.