MoltiTox: a multimodal fusion model for molecular toxicity prediction

Area of Science:

• Computational toxicology
• Drug discovery
• cheminformatics

Background:

• Single-modality models for molecular toxicity prediction have limitations.
• Integrating diverse molecular data can enhance predictive accuracy and robustness.

Purpose of the Study:

• Introduce MoltiTox, a novel multimodal fusion model for molecular toxicity prediction.
• Overcome limitations of single-modality approaches in drug discovery.

Main Methods:

• MoltiTox integrates molecular graphs, SMILES strings, 2D images, and 13C NMR spectra.
• Utilizes modality-specific encoders (GNN, Transformer, 2D CNN, 1D CNN) and an attention-based fusion mechanism.

Main Results:

• Achieved a ROC-AUC of 0.831 on the Tox21 benchmark across 12 endpoints.
• Outperformed all single-modality baselines in toxicity prediction.

Conclusions:

• Integrating diverse molecular representations enhances robustness and generalizability of toxicity prediction.
• 13C NMR data provides complementary chemical insights for mechanistic understanding.
• MoltiTox offers an extensible framework for reliable computational toxicology models.