Integrating chemical structure and high-throughput transcriptomics for mechanistically interpretable Tox21...

Area of Science:

• Computational toxicology
• Systems biology
• Pharmacogenomics

Background:

• Current computational toxicology models often lack biological interpretability, relying solely on chemical structures.
• Integrating mechanistic information with chemical screening is crucial for advancing predictive toxicology.

Purpose of the Study:

• To develop and validate a multimodal modeling framework for predicting chemical activity across 41 Tox21 assay endpoints.
• To integrate chemical fingerprints with high-throughput transcriptomic (HTTr) dose-response profiles for enhanced predictive accuracy and mechanistic insight.

Main Methods:

• Utilized TempO-Seq for HTTr profiling in MCF-7, U-2 OS, and HepaRG cells exposed to ToxCast compounds.
• Employed gradient-boosted decision trees and nested cross-validation for model training and performance evaluation.
• Applied SHapley Additive exPlanations (SHAP) for feature attribution analysis to understand prediction drivers.

Main Results:

• Achieved robust performance (mean AUPRC > 0.75) for 13 out of 41 assays, covering nuclear receptor signaling, stress response, and xenobiotic metabolism.
• Demonstrated that model predictions are influenced by both chemical structural motifs and specific transcriptional programs.
• SHAP analysis confirmed mechanistic relationships between chemical structure, nuclear receptor biology, and cellular adaptive responses.

Conclusions:

• The multimodal framework successfully integrates chemical structure and HTTr data for accurate and mechanistically grounded bioactivity predictions.
• This approach advances computational toxicology towards more transparent and biologically informed predictions.
• The findings highlight the potential of combining transcriptomic signatures with chemical data for improved predictive toxicology.