Identifying Protein Features and Pathways Responsible for Toxicity Using Machine Learning and Tox21: Implications for...

Area of Science:

• Computational toxicology and cheminformatics
• Proteomics and systems biology
• Machine learning applications in drug discovery

Background:

• Chemicals pose health risks, necessitating robust toxicity prediction methods.
• Machine learning models excel at predicting toxicity from chemical structures but overlook proteomic data.
• Investigating toxicity-related protein features is crucial for a comprehensive understanding of compound effects.

Purpose of the Study:

• To develop a computational pipeline for predicting critical protein features associated with compound toxicity.
• To analyze toxicity data from the Tox21 dataset using machine learning and the CANDO platform.
• To identify and elucidate biological pathways and protein interactions involved in toxicity.

Main Methods:

• Utilized machine learning models, specifically random forest with SMOTE+ENN resampling, to handle imbalanced toxicity data.
• Applied a computational pipeline within the CANDO platform to analyze the Tox21 dataset (10,000+ compounds, 12 assays).
• Performed enrichment analysis on extracted protein features to identify implicated biological pathways and validated findings for the aryl hydrocarbon receptor (AhR).

Main Results:

• Achieved high performance in predicting toxicity endpoints, with AUCROCs of 0.90 for NR-AhR and 0.92 for SR-MMP.
• Identified key protein features and elucidated upstream/downstream biological pathways correlated with compound toxicity.
• Demonstrated significant relationships between protein-compound interactions, computed pathways, and toxicity endpoints.

Conclusions:

• The developed machine learning pipeline effectively predicts toxicity and identifies crucial proteomic features.
• The study highlights interconnected biological pathways and protein interactions underlying compound toxicity.
• This research advances the understanding of toxicity mechanisms at a proteomic level, aiding therapeutic discovery.