Machine Learning Prediction of Antibacterial Activity of Block Copolymers

As the problem of antibiotic resistance continues to escalate, there is an immediate need for fresh antimicrobial strategies. Traditional antibiotic development processes are time consuming and can require many years, from hit identification to drug approval. While machine learning (ML) has proven useful in predicting the properties of nanomaterials, current techniques often rely on extensive data sets and identifying structural similarities to existing antibiotics. This presents a challenge for alternative classes of antimicrobial agents such as polymers. To tackle this issue, we created an ML classification model to forecast the antibacterial activity of polymers. We evaluated several models on an imbalanced data set consisting of 157 polymers, of which 27 were antibacterial active polymers (positive cases). The decision tree model achieved the highest performance, with an accuracy of 0.91 and a recall of 1, properly identifying all positive cases. In the course of the SHapley Additive exPlanations (SHAP) analysis, we discovered that the calculated partition coefficient ( c log P ) is the most significant feature in forecasting antibacterial activity, followed by the composition of the cationic and hydrophobic monomers. Our analysis indicates that controlling the values of these features could help design better antimicrobial polymers. For instance, the high antibacterial activity necessitates the c log P values of polymers that are neither too high nor too low, aligning with the previous findings (P. T. Phuong et al., Biomacromolecules, 2020, 21, 5241–5255).