Investigating the Generalizability of the MultiFlow ® DNA Damage Assay and Several Companion Machine Learning Models with a Set of 103 Diverse Test Chemicals.

The in vitro MultiFlow DNA Damage assay multiplexes p53, γH2AX, phospho-histone H3, and polyploidization biomarkers into 1 flow cytometric analysis (Bryce, S. M., Bernacki, D. T., Bemis, J. C., and Dertinger, S. D. (2016). Genotoxic mode of action predictions from a multiplexed flow cytometric assay and a machine learning approach. Environ. Mol. Mutagen. 57, 171-189). The work reported herein evaluated the generalizability of the method, as well as several data analytics strategies, to a range of chemical classes not studied previously. TK6 cells were exposed to each of 103 diverse chemicals, 86 of which were supplied by the National Toxicology Program (NTP) and selected based upon responses in genetic damage assays conducted under the Tox21 program. Exposures occurred for 24 h over a range of concentrations, and cell aliquots were removed at 4 and 24 h for analysis. Multiplexed response data were evaluated using 3 machine learning models designed to predict genotoxic mode of action based on data from a training set of 85 previously studied chemicals. Of 54 chemicals with sufficient information to make an a priori call on genotoxic potential, the prediction models' accuracies were 79.6% (random forest), 88.9% (logistic regression), and 90.7% (artificial neural network). A majority vote ensemble of the 3 models provided 92.6% accuracy. Forty-nine NTP chemicals were not adequately tested (maximum concentration did not approach assay's cytotoxicity limit) and/or had insufficient conventional genotoxicity data to allow their genotoxic potential to be defined. For these chemicals MultiFlow data will be useful in future research and hypothesis testing. Collectively, the results suggest the MultiFlow assay and associated data analysis strategies are broadly generalizable, demonstrating high predictability when applied to new chemicals and classes of compounds.