ScholarWorks@성균관대학교

초록

Introduction: We introduce MoltiTox, a novel multimodal fusion model for molecular toxicity prediction, designed to overcome the limitations of single-modality approaches in drug discovery. Methods: MoltiTox integrates four complementary data types: molecular graphs, SMILES strings, 2D images, and C-13 NMR spectra. The model processes these inputs using four modality-specific encoders, including a GNN, a Transformer, a 2D CNN, and a 1D CNN. These heterogeneous embeddings are fused through an attention-based mechanism, enabling the model to capture complementary structural and chemical information from multiple molecular perspectives. Results: Evaluated on the Tox21 benchmark across 12 endpoints, MoltiTox achieves a ROC-AUC of 0.831, outperforming all single-modality baselines. Discussion: These findings highlight that integrating diverse molecular representations enhances both the robustness and generalizability of toxicity prediction models. Beyond predictive performance, the inclusion of C-13 NMR data offers complementary chemical insights that are not fully captured by structure or language-based representations, suggesting its potential contribution to mechanistic understanding of molecular toxicity. By demonstrating how multimodal integration enriches molecular representations and enhances the interpretability of toxicity mechanisms, MoltiTox provides an extensible framework for developing more reliable models in computational toxicology.

키워드