Premium
Genotoxic mode of action predictions from a multiplexed flow cytometric assay and a machine learning approach
Author(s) -
Bryce Steven M.,
Bernacki Derek T.,
Bemis Jeffrey C.,
Dertinger Stephen D.
Publication year - 2016
Publication title -
environmental and molecular mutagenesis
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1
H-Index - 87
eISSN - 1098-2280
pISSN - 0893-6692
DOI - 10.1002/em.21996
Subject(s) - mode of action , action (physics) , flow cytometry , flow (mathematics) , multiplexing , mode (computer interface) , computational biology , computer science , biology , toxicology , genetics , physics , mechanics , human–computer interaction , telecommunications , quantum mechanics
Several endpoints associated with cellular responses to DNA damage as well as overt cytotoxicity were multiplexed into a miniaturized, “add and read” type flow cytometric assay. Reagents included a detergent to liberate nuclei, RNase and propidium iodide to serve as a pan‐DNA dye, fluorescent antibodies against γH2AX, phospho‐histone H3, and p53, and fluorescent microspheres for absolute nuclei counts. The assay was applied to TK6 cells and 67 diverse reference chemicals that served as a training set. Exposure was for 24 hrs in 96‐well plates, and unless precipitation or foreknowledge about cytotoxicity suggested otherwise, the highest concentration was 1 mM. At 4‐ and 24‐hrs aliquots were removed and added to microtiter plates containing the reagent mix. Following a brief incubation period robotic sampling facilitated walk‐away data acquisition. Univariate analyses identified biomarkers and time points that were valuable for classifying agents into one of three groups: clastogenic, aneugenic, or non‐genotoxic. These mode of action predictions were optimized with a forward‐stepping process that considered Wald test p‐values, receiver operator characteristic curves, and pseudo R 2 values, among others. A particularly high performing multinomial logistic regression model was comprised of four factors: 4 hr γH2AX and phospho‐histone H3 values, and 24 hr p53 and polyploidy values. For the training set chemicals, the four‐factor model resulted in 94% concordance with our a priori classifications. Cross validation occurred via a leave‐one‐out approach, and in this case 91% concordance was observed. A test set of 17 chemicals that were not used to construct the model were evaluated, some of which utilized a short‐term treatment in the presence of a metabolic activation system, and in 16 cases mode of action was correctly predicted. These initial results are encouraging as they suggest a machine learning strategy can be used to rapidly and reliably predict new chemicals' genotoxic mode of action based on data from an efficient and highly scalable multiplexed assay. Environ. Mol. Mutagen. 57:171–189, 2016. © 2016 Wiley Periodicals, Inc.