Tox-GAN

Tox-GAN generates toxicogenomic transcriptomic profiles using a deep generative adversarial network to simulate gene expression across doses and treatment durations for toxicological analysis and biomarker development.

Key Features:

• Deep Generative Adversarial Network (GAN) Framework: Employs a GAN architecture to simulate gene activities and expression profiles across various doses and treatment durations in toxicogenomics (TGx).
• Transcriptomic Data Generation: Derives new transcriptomic profiles from existing animal study results and chemical structures without additional experiments.
• High-Fidelity Data Generation: Generated profiles show intensity similarity of 0.997 ± 0.002 and fold change similarity of 0.740 ± 0.082 compared to real gene expression profiles.
• Mechanistic Consistency (Gene Ontology): Produces gene expression outputs with over 87% agreement in Gene Ontology between generated and real data.
• Biomarker Concordance: Demonstrates high concordance in biomarkers between real and generated datasets in both predictive performance and specific biomarker genes.
• Cross-Dataset Validation: Models constructed using Open TG-GATES data generate transcriptomic profiles that align with those reported in DrugMatrix.
• Chemical-Based Read-Across Utility: Facilitates prediction of toxicological effects based on chemical structures to support chemical-based read-across.

Scientific Applications:

• Toxicological Mechanism Inference: Uses generated transcriptomic profiles to infer molecular mechanisms underlying toxicity, supported by Gene Ontology agreement.
• Biomarker Development and Validation: Supports discovery and validation of gene expression–based biomarkers with high concordance to real datasets.
• Read-Across and Chemical Hazard Prediction: Enables chemical-based read-across for hazard prediction by linking chemical structures to predicted transcriptomic responses.
• Reduction of Additional Animal Experiments: Provides synthetic transcriptomic data from existing studies to reduce the need for further animal testing.

Methodology:

Tox-GAN employs a deep generative adversarial network trained on Open TG-GATES data to simulate gene activities and generate transcriptomic profiles across doses and treatment durations, with validation against DrugMatrix.