Machine Learning-Guided Design of High-Performance Antifungal Polymers against Candida albicans

This study leverages machine learning (ML) to accelerate the development of antifungal synthetic polymers. An in-house curated data set was used to train a Random Forest binary classification model, which identified five critical features for predicting antifungal activity against the fungal pathogen Candida albicans. This model shows that antifungal polymers should contain a hydrophilic composition of at least 30%, a calculated partition coefficient (cLogP) of 0 to +0.5, hydrophobic composition limited to 20%, degree of polymerization (DP) of 18 or less, and cationic composition limited to 50% for the polymer to have an MIC(90) of 64 μg/mL or less. Based on these insights, a library of polymers was synthesized via reversible addition-fragmentation chain transfer (RAFT) polymerization using 2-(butylthiocarbonothioylthio)-propanoic acid (BTPA) as the RAFT agent, exploring combinations of these features. A polymer exhibiting all five feature characteristics for an effective antifungal polymer presented an MIC(90) of 32 μg/mLlower than the 64 μg/mL Class 1 "good" polymer threshold. After optimizing the polymer for significant antifungal activity by fulfilling the five features of the antifungal impact, we improved its biocompatibility. We did this by replacing the BTPA RAFT agent with 2-dodecyl thiocarbonothioylthio-propanoic acid (DTPA), which has a longer, more hydrophobic end-group. This modification resulted in a polymer with an MIC90 of 16 μg/mL and a hemolysis concentration (HC50 exceeding 2000 μg/mL, yielding a selectivity index greater than 125aligning with the most competent polymer involved in the ML data set. This work validates the efficiency of ML-guided design for developing advanced antifungal polymers with enhanced potency and biocompatibility.