BitterGNN: An Explainable Graph-Based Framework for Bitter Peptide Classification

Abstract

Bitter peptides play an important role in nutrition, sensory science, and pharmaceutical studies. However, identifying and classifying them is still difficult because their structural differences are often subtle, and available data is limited. This paper introduces BitterGNN, a graph-based learning framework created to improve the prediction of bitter peptides by capturing relationships that traditional feature-based models usually miss. The workflow begins with cleaning the data by correcting outliers, followed by using Recursive Feature Elimination with Cross-Validation (RFECV) to find the most meaningful descriptors. These selected features are then used to build a mutual knearest neighbor graph, allowing a GraphSAGE encoder to learn both local and global interactions among peptide samples. The proposed model achieves strong performance, with an accuracy of 0.9844, precision of 0.9848, recall of 0.9844, and a Cohen's kappa of 0.9688. Automated hyperparameter tuning helps reach these results. The AUC of 0.9993 shows that the model can clearly separate the classes. To make the predictions more transparent, LIME and gradient-based attribution highlight which biological traits influence the model's decisions. Overall, BitterGNN demonstrates that combining graph-based representations with feature selection is an effective way to improve peptide classification.