Integrating Protein Language Models and Geometric Deep Learning for Peptide Toxicity Prediction

Peptide toxicity prediction is a critical task in biomedical research, influencing drug safety and therapeutic development. Traditional methods, relying on sequence similarity or handcrafted features, struggle to capture the complex relationship between peptide structure and toxicity. In this study, we propose PeptiTox, an advanced deep learning framework that integrates protein language models (PLMs) and geometric deep learning to enhance peptide toxicity prediction. Specifically, ESM2 is employed to extract sequence embeddings, while ESMFold predicts the three-dimensional (3D) peptide structure. The structural information is then transformed into a graph representation, where residues serve as nodes, and interactions between residues form edges. A graph neural network (GNN) is subsequently used to learn peptide representations and classify their toxicity. Experimental results demonstrate that PeptiTox significantly outperforms state-of-the-art models across multiple evaluation metrics. Our findings highlight the importance of integrating sequence and structural knowledge for peptide toxicity prediction, paving the way for safer and more effective peptide-based therapeutics.