Deep Learning-Based Multimodal Fusion Approach for Predicting Acute Dermal Toxicity.

Area of Science:

• Computational toxicology
• cheminformatics
• machine learning

Background:

• Traditional acute dermal toxicity testing heavily relies on animal studies, which are ethically concerning and resource-intensive.
• There is a growing need for alternative methods to support the 3Rs principle (replacement, reduction, and refinement) in animal testing.
• Predicting chemical toxicity is crucial for pharmaceuticals, pesticides, cosmetics, and industrial chemicals.

Purpose of the Study:

• To develop and validate a reliable and accurate multimodal deep learning framework for predicting acute dermal toxicity.
• To investigate the efficacy of fusing heterogeneous molecular representations for enhanced predictive performance.
• To offer a more ethical and efficient alternative to animal-based toxicity assessments.

Main Methods:

• Proposed a novel deep learning architecture, TriModalToxNet, integrating features from 2D molecular images, SMILES embeddings, and molecular fingerprints.
• Utilized 2D convolutional neural networks, 1D convolutional neural networks, and fully connected neural networks for feature extraction.
• Trained and evaluated models on a dataset of 3845 compounds using stratified 10-fold cross-validation and external validation.

Main Results:

• TriModalToxNet achieved a 95% area under the receiver operating characteristic curve and 91.2% sensitivity in cross-validation.
• The multimodal approach demonstrated superior predictive performance compared to single-modality baseline models (BiModalToxNet).
• External validation confirmed the robustness and generalizability of the TriModalToxNet framework.

Conclusions:

• Multimodal deep learning frameworks integrating diverse molecular representations can significantly improve the accuracy of acute dermal toxicity prediction.
• TriModalToxNet offers a promising computational tool for ethical and efficient chemical safety assessment.
• The developed framework has the potential for integration into regulatory processes and pharmaceutical screening pipelines.