A Multimodal Geometric Learning Framework for Interpretable Peptide Toxicity Prediction

Peptide therapeutics are promising drug candidates because of their high activity and specificity. However, their potential toxicity poses a major hurdle, making accurate computational toxicity predictions urgently necessary for early safety screening. Existing methods largely rely on sequence data or simple graph representations, thereby limiting their ability to capture higher-order physicochemical interactions. To address this, we propose MBFGNK, an interpretable multimodal geometric learning framework, for peptide toxicity prediction. It integrates sequence representations from pretrained protein language models, 3D structural features via an equivariant graph neural network (EGNN), and physicochemical hypergraph representations capturing global, local, and biologically synergistic residue interactions. A modality attention gate adaptively fuses these heterogeneous data types, while a Chebyshev polynomial-based Kolmogorov-Arnold network (ChebyKAN) models complex nonlinear toxicity patterns. We evaluated MBFGNK using 5-fold cross-validation and an independent test set. The experimental results show that the ensemble achieved an AUC of 0.9687, an MCC of 0.8449, an Acc of 0.9221, and an F1 of 0.9205 on the independent test set, suggesting improved performance relative to previously reported baseline results on the same benchmark. Integrated gradients and modality-level analyses further identified prediction-informative motifs, residues, and spatial regions, providing interpretable evidence consistent with known characteristics of toxic peptides. These results suggest that MBFGNK is an effective and interpretable framework for peptide toxicity prediction and may assist in downstream safety-oriented peptide screening.