Analysis of Multiplexed Flow Cytometric Assays and Toxicogenomic Signatures for Genotoxicity Prediction: A Model Performance and Case Study Approach

Abstract

Genotoxicity assays play a significant role in protecting clinical trial subjects from potential increased risk of genotoxic hazard and cancer during drug development. Traditional genetic toxicology assays typically provide binary outcomes with limited mechanistic insight. This study evaluates machine learning (ML) models based on an in-house implementation of MultiFlow DNA Damage Assay and MicroFlow Assays, and compared these results to previously published Litron assays. Our ML models demonstrated high accuracy, with MultiFlow data achieving 96% balanced accuracy for mode of action (MoA) prediction and 99% for genotoxicity prediction in repeated cross-validation. We collected and interpreted the MicroFlow and MultiFlow data in a dose–response format. The dose–response data enabled us to improve assay inference and model accuracy. In addition, we conducted case studies using toxicogenomic data, including the Toxicogenomic DNA Damage Inducing (TGx-DDI) transcriptomic biomarker and bulk RNA-seq, on a small set of compounds where the MoA is not clearly defined by MultiFlow or MicroFlow. The integration of toxicogenomics provided deeper insights into the molecular mechanisms of genotoxicity, allowing for the identification of specific pathways affected by these compounds. These findings emphasize the importance of careful endpoint selection and data interpretation. Overall, this study enhances the precision of genotoxicity predictions by integrating toxicogenomics, offering a framework for future genotoxicity safety assessments.