Potentials of Machine Learning in Predicting Key Features of Synthetic Antimicrobial Polymers

As the global rise in antimicrobial resistance calls for new therapeutic strategies, synthetic antimicrobial polymers (SAMPs) have emerged as promising alternatives to host-defense peptides, offering tunable structures and reduced limitations. In this work, we employed machine learning (ML) approaches to elucidate the structure–activity relationships of a library of polyacrylamides systematically varied in (1) amine side-chain chemistry, (2) chain length, (3) cationic amine ratio, and (4) polymer architecture. The library consisted of 23 different polymer designs, 3 of which exhibited low minimum inhibitory concentrations (MIC) against different bacterial strains, and 5 of which caused low red blood cells agglutination. Among the evaluated ML algorithms, regression random forest and gradient boosting regression consistently reproduced feature importance and maintained stable decision-tree structures, with gradient boosting outperforming random forest in predictive power. Gradient boosting achieved RSME values of 20, 6, 13 and 12 μg/ml, respectively, for each modelled MIC of 4 bacterial strains: Pseudomonas aeruginosa PA14, Pseudomonas aeruginosa LESB58, Staphylococcus aureus USA300 and Staphylococcus aureus Newman (total data range 64-513 μg/ml). RSME for modelled hemagglutination was 1 μg/ml. Calculation of feature importances and visualisation with beeswarm and waterfall plots highlighted the contribution of individual polymer features through Shapley additive explanations (SHAP). All bacteria strains considered, the type and percentage of cationic monomer are the most important features determining best-performing SAMP designs. Collectively, our findings demonstrate that boosting-ensemble methods offer consistent robust predictive capability and can serve as effective tools for forecasting the potency and toxicity of future SAMP designs, with potential for application in larger, multi-sourced libraries.