?
Tackling the challenge of ML-assisted antibacterial activity prediction: One step closer to controlled quaternary ammonium compounds design via neural networks model
The ongoing rise of resistant bacterial pathogens poses a significant threat to current antibacterials' effectiveness putting millions of people's lives at risk. However, modern machine learning (ML) tools promise to tip the scales in the never-ending development of antimicrobial agents' pipelines. Herein we present a novel approach for quaternary ammonium compounds (QACs) antibacterial activity prediction using adaptable Neural Networks model. Recursive Feature Elimination methodology with Random Forest as a base model was implemented for feature selection followed by a Multilayer Perceptron (MLP) model for the regression task with the LeakyReLU activation function as the baseline architecture. Systematic evaluation revealed that architectural complexity does not improve accuracy in antimicrobial activity prediction. Thus, Kolmogorov-Arnold Networks showed substantial underperformance relative to simpler MLP architectures. Through implementation of biology-informing stacking approach ML model was able to predict minimal inhibitory concentration and minimal bactericidal concentration on both Gram-negative E. coli ATCC 25922 and Gram-positive S. aureus ATCC 43300 achieving 61–69 % accuracy and R2 of 0.68 through external validation on novel bis-QACs. Developed ML model can serve as an efficient support tool for antimicrobial drug discovery even with limited datasets.