QSAR Modeling and Deep Learning-Based Prediction of Antimicrobial Activity of β-Lactam Derivatives Using Molecular Fingerprints

Antimicrobial resistance poses a critical global health challenge, necessitating the accelerated discovery of novel antibacterial agents. This study presents a quantitative structure–activity relationship (QSAR)-based multiclass classification framework for predicting the antimicrobial activity of β-lactam, azetidinone, and thiazolidinone derivatives. A chemically diverse library of over 220 compounds was constructed through combinatorial scaffold expansion guided by structure–activity relationship principles, with activity classified as Inactive, Moderate, or Active based on minimum inhibitory concentration (MIC) values. Molecular features were encoded using a fused descriptor set comprising Morgan fingerprints (radius 2 and 3), MACCS structural keys, and eight physicochemical descriptors, yielding a 1,199-dimensional feature vector. Class imbalance was addressed via SMOTE applied exclusively to the training set. Multiple machine learning models were developed and compared, including Random Forest, Gradient Boosting, XGBoost, a stacking ensemble, and a deep neural network with residual connections, batch normalization, and dropout regularization. Hyperparameter optimization was performed using randomized search with stratified 5-fold cross-validation. Model performance was evaluated using accuracy, weighted F1-score, balanced accuracy, and multiclass ROC-AUC. Visualization strategies including t-SNE, PCA, and feature importance analysis confirmed meaningful chemical space organization and robust structure–activity discrimination across all classifiers.