Abstract
Bacteriocins offer a promising solution to antibiotic resistance, possessing the ability to target a wide range of bacteria with precision. Thus, there is an urgent need for a computational model to predict new bacteriocins and aid in drug development. This work centers on constructing web-based predictive models using the XGBoost machine learning algorithm, based on the physicochemical properties, structural characteristics, and sequence profiles of protein sequences. We employed correlation analyses, cross-validation, and hypergraph-based techniques to select features. Cross-validated feature selection (CVFS) partitions the dataset, selects features within each partition, and identifies common features, ensuring representativeness. On the contrary, hypergraph-based feature evaluation (HFE) focuses on minimizing hypergraph cut conductance, leveraging higher-order data relationships to precisely utilize information regarding feature and sample correlations. The XGBoost models were built using the selected features obtained from these two feature evaluation methods. We also analyzed the feature contributions directly from the best model using SHapley Additive exPlanations (SHAP). Our HFE-based approach achieved 99.11% accuracy and an AUC of 0.9974 on the test data, overall outperforming the CVFS-based feature evaluation method and yielding results comparable to existing approaches. The most influential features are related to solvent accessibility for buried residues, followed by the composition of cysteine. Our web application, accessible at https://shiny.tricities.wsu.edu/bacteriocin-prediction/, offers prediction results, probability scores, and SHAP plots using both cross-validation- and hypergraph-based methods, along with previously implemented approaches for feature selection.