Advancing toxicity AI-based prediction with multilevel systems biology: a case study on genotoxicity

Area of Science:

• Computational toxicology
• Genomics
• Machine learning

Background:

• Chemical diversity poses challenges for health and environmental risk assessment.
• Toxicity prediction is hindered by heterogeneous cellular responses to chemical exposures.
• Integrating multimodal data is crucial for predicting individual health effects.

Purpose of the Study:

• To develop a multimodal deep learning framework, GenotoxNet, for enhanced genotoxicity prediction.
• To systematically integrate chemical structures, in vitro assay data, and transcriptomics data.
• To improve the prediction of chemical-induced genotoxicity by capturing cellular heterogeneity and mechanistic complexity.

Main Methods:

• GenotoxNet framework utilizes multimodal deep learning.
• Integration of chemical structures, high-throughput in vitro assay data, and transcriptomics data.
• Model performance evaluated using AUCROC on internal and external test sets.

Main Results:

• GenotoxNet achieved an AUCROC of 0.891 ± 0.017 on the internal test set.
• The multimodal approach outperformed single-modality prediction models.
• The model demonstrated strong performance on an external chemical set.

Conclusions:

• GenotoxNet offers a robust framework for predicting genotoxicity.
• The model facilitates mechanistic interpretation by aligning features with adverse outcome pathways (AOPs).
• This approach supports preventive strategies and regulatory decisions for hazardous chemicals.