Species-Specific Protein Function Prediction in Flavobacterium covae Using Ensemble Machine Learning

Protein function prediction remains a critical challenge in computational biology, particularly for species-specific applications. This study presents a machine learning framework for predicting protein functions in Flavobacterium covae, a Gram-negative bacterium responsible for columnaris disease in channel catfish. We formulate the problem as a multi-label classification task where each protein sequence is associated with multiple Gene Ontology (GO) terms. Our approach integrates four feature groups: homologous sequence information from BLAST searches, essential gene properties from the Database of Essential Genes (DEG), subcellular localization predictions from PSORTb, and physicochemical properties derived from protein sequences. We evaluate three ensemble learning algorithms—Random Forest, XGBoost, and AdaBoost—on a dataset of 69,960 protein sequences with 1,868 GO term categories. Random Forest and XGBoost achieved accuracies exceeding 90%, with XGBoost demonstrating superior performance across all metrics (accuracy: 90.50%, precision: 93.92%, recall: 92.23%, F1-score: 92.67%). The models successfully predicted functions for over 99% of previously unannotated hypothetical proteins, substantially outperforming existing tools like PANNZER. This species-specific approach provides insights into F. covae pathogenicity and demonstrates the efficacy of integrating diverse biological features for protein function prediction in understudied organisms.