Prediction of Drug-Induced Nephrotoxicity Using Chemical Information and Transcriptomics Data.

Area of Science:

• Computational toxicology
• Pharmacogenomics
• Machine learning in drug discovery

Background:

• Drug-induced nephrotoxicity prediction is vital in drug development.
• Gene expression data offers mechanistic insights into organ toxicity.
• Current computational models often rely on chemical information.

Purpose of the Study:

• To evaluate gene expression data for nephrotoxicity prediction using machine learning.
• To compare the performance of gene expression-based models against chemical information-based models.
• To identify key genes associated with drug-induced nephrotoxicity.

Main Methods:

• Utilized LightGBM, random forest, support vector machine, and XGBoost for prediction.
• Explored subsets of gene expression data (6000, 9000, 12,000 profiles).
• Applied techniques like optimal probability thresholds, data balancing, and cost-sensitive learning.

Main Results:

• Gene expression-based models achieved high performance, with the best model (GEM20) reaching an AUC of 0.94.
• The top chemical information-based model (CIM19) and an initial gene expression model (GEM9) showed similar AUCs (0.89-0.9).
• SHAP analysis identified potential nephrotoxicity-associated genes: CDC20, RPS6, DDIT4, GAPDH, CCNF, and MRPL12.

Conclusions:

• Gene expression data, particularly when optimally selected and balanced, significantly enhances nephrotoxicity prediction accuracy.
• Machine learning models utilizing gene expression profiles provide a powerful tool for understanding and predicting drug-induced kidney damage.
• Identified genes offer potential biomarkers and mechanistic targets for mitigating drug-induced nephrotoxicity.