Development of machine learning-based approaches to develop cytotoxicity flags for Tox21/ToxCast assays
Results of in vitro assays used in the Tox21 and ToxCast high-throughput screening (HTS) programs may be confounded by overt cell stress and cytotoxicity, such that a decrease in viable cells could erroneously be attributed to a chemical’s mechanistic effects. Integration of cytotoxicity assessment with assay endpoints can bolster confidence in the interpretation of assay outcomes. However, many chemicals tested in HTS assays lack directly relevant cytotoxicity data needed to ensure overt toxicity does not confound mechanistic outputs. Chemicals may also vary in their potency for eliciting cytotoxicity across cell types and time trajectories. NIEHS scientists are investigating applying multiple machine-learning algorithms to predict chemical- and cell type-specific cytotoxicity concentrations to provide context for flagging nonspecific in vitro chemical-elicited bioactivity using cytotoxicity assays included in Tox21 and ToxCast. Cell type- and time point-specific machine-learning models were developed to predict chemical concentrations likely to induce cytotoxicity to provide context for assays without concurrent cytotoxicity data. After being further refined, the predictive models will be integrated with bioactivity data to provide context and bolster confidence in assay outcome interpretation for identifying specific vs. nonspecific/cytotoxicity-confounded bioactivities. Concurrent cytotoxicity readouts from Tox21/ToxCast assays were mapped for 492 assay endpoints, allowing direct comparison of bioactivity potency against cytotoxicity. These comparisons will be integrated into future versions of concentration–response visualizations for the curated HTS data in the Integrated Chemical Environment Curve Surfer tool. This project was described in an abstract accepted for a poster presentation (Tedla et al.) at the 2024 SOT annual meeting.