Bacteriocin prediction through cross-validation-based and hypergraph-based feature evaluation approaches

Area of Science:

• * Computational biology
• * Bioinformatics
• * Machine learning in drug discovery

Background:

• * Antibiotic resistance is a growing global health threat.
• * Bacteriocins show promise as targeted antimicrobial agents.
• * Predictive models are needed to accelerate bacteriocin discovery and drug development.

Purpose of the Study:

• * To develop and validate web-based computational models for predicting bacteriocins.
• * To compare feature selection methods, including cross-validated feature selection (CVFS) and hypergraph-based feature evaluation (HFE).
• * To identify key protein features influencing bacteriocin activity.

Main Methods:

• * Construction of XGBoost machine learning models using protein sequence data.
• * Feature selection using CVFS and HFE techniques.
• * Analysis of feature importance using SHapley Additive exPlanations (SHAP).
• * Development of a publicly accessible web application for bacteriocin prediction.

Main Results:

• * The HFE-based XGBoost model achieved 99.11% accuracy and an AUC of 0.9974 on test data.
• * The HFE method outperformed the CVFS method and matched existing approaches.
• * Key predictive features include solvent accessibility of buried residues and cysteine composition.

Conclusions:

• * Computational models, particularly those employing HFE, can effectively predict bacteriocins.
• * The developed web application facilitates bacteriocin discovery and aids antibiotic drug development.
• * Understanding feature contributions enhances the interpretability of predictive models.