A Gaussian processes approach facilitates the identification of robust biomarkers for exposure to complex pesticide mixtures 5

In this study, we obtained transcriptomic data from the springtail Folsomia candida under two binary mixture exposures in a grid design and finely resolved for stress intensity. The mixtures either had the same or different presumed mechanisms of toxic action. The mixtures consisted of either two neonicotinoids insecticides, imidacloprid and clothianidin, or imidacloprid and the azole fungicide cyproconazole, respectively. We analyzed the data using a custom-made statistical framework based on Gaussian Processes (GP) models to meet common challenges of mixture toxicity transcriptomic data, and analyzed the two datasets in conjunction without the need for batch effect correction. Overall design: The following concentrations were chosen for the neonicotinoid mixture: imidacloprid: 0.00, 0.08, 0.16, 0.24, 0.32, and 0.40 mg kg-1 dry soil combined with 0, 0.048, 0.096, 0.144, 0.192, 0.240 mg of clothianidin kg-1 dry soil. For the imidacloprid-cyproconazole mixture, the same concentrations of imidacloprid were chosen, and 0, 50, 100, 150, 200, and 250 mg of cyproconazole kg-1 dry soil. Concentrations of all pesticides were equal to 0, ?, ?, ?, ?, and 1 TU (Toxic Unit = Effect Concentration reducing juvenile counts by 50 %). Only conditions 1:0, 0:1, and 1:1 (TU:TU) were used for the highest exposures. All other concentrations were combined in a full-factorial design, resulting in 28 conditions per experiment and 168 samples over the two experiments. See the supplementary file for the concentration of each pesticide in mg / kg dry soil or in TU. TU values are wrapped in quotation marks to ensure data analysis software will import them as character strings, e.g., "0", "1/5", and "2/5".