Establishment of interpretable cytotoxicity prediction models using machine learning analysis of transcriptome...

Area of Science:

• Computational biology
• Pharmacology
• Toxicology

Background:

• Cytotoxicity, measured by cell viability, is vital for assessing drug safety in vitro.
• Predicting cytotoxicity aids early-stage drug development by identifying potential safety concerns.

Purpose of the Study:

• To develop accurate machine learning models for predicting drug cytotoxicity and cell viability.
• To interpret these models and identify key genes involved in cytotoxicity.
• To enable high-throughput screening of safe substances.

Main Methods:

• Integrated cellular transcriptome and cell viability data.
• Employed machine learning algorithms: SVM, RF, XGBoost, LightGBM.
• Utilized ensemble methods: voting and stacking.
• Validated models across diverse cell lines.

Main Results:

• Achieved high prediction accuracy for 50% and 80% cell viability with AUROC of 0.90 and 0.84.
• Models demonstrated robust performance across various cell lines.
• Identified Cytotoxicity Signature (CTS) genes for the first time.
• Enabled analysis of mechanisms for narrow therapeutic index (NTI) compounds.

Conclusions:

• The developed models offer superior accuracy for cytotoxicity prediction compared to previous studies.
• These models facilitate efficient and accurate screening of high-safety substances.
• The identified CTS genes provide new insights into drug mechanisms of action, particularly for NTI compounds.