Optimized ensemble learning with multi-feature fusion for enhanced anti-inflammatory peptide prediction

Anti-inflammatory peptides (AIPs) represent promising therapeutic agents, however, their discovery via traditional experimental methods is constrained by low throughput and high costs. Current computational prediction methods face persistent challenges, including dataset imbalance, inadequate feature integration, and limited predictive accuracy. To address these limitations, we assembled a high-quality benchmark dataset by integrating sequences from the latest constructed datasets. After evaluating feature selection methods including Least Absolute Shrinkage and Selection Operator (LASSO), Principal Component Analysis (PCA), Autoencoder and Random Forest (RF), we finally selected the 377-dimensional features ranked by RF. For classification algorithms, we implemented three deep learning architectures (Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNN), Deep Neural Networks (DNN)), as well as five conventional machine learning classifiers including eXtreme Gradient Boosting (XGBoost), Random Forest (RF), Adaptive Boosting (AdaBoost), Gradient Boosting Decision Tree (GBDT), and Light Gradient Boosting Machine (LightGBM). Comparative analysis demonstrated that a soft voting strategy integrating predictions from the five ensemble classifiers achieved superior performance, surpassing state-of-the-art predictors. Further sequence composition analysis revealed significant enrichment of positively charged residues (e.g., Lysine, Arginine) in AIPs, whereas non-AIP sequences were predominantly characterized by hydrophobic residues. This finding provides a molecular basis for the rational design of novel anti-inflammatory peptides.