Interpretable machine learning unveils key predictors and default values in an expanded database of human in vitro...

The skin is the main route of exposure to plant protection products for operators, workers, residents, and bystanders. Assessing dermal absorption is key for evaluating pesticide exposure. The initial approach to risk assessment involves using default dermal absorption values or applying read-across data from experimental results from different formulations. In this way, to support non-dietary pesticide risk assessment focused but not limited to Brazil, this project evaluated 759 GLP-compliant in vitro human skin dermal absorption studies covering 25 formulation types and 248 active substances at multiple concentrations using interpretable machine learning techniques. Bayesian Additive Regression Trees - BART method indicated that Log Pow and molecular weight have the highest importance when predicting dermal absorption; both parameters exhibit moderate interaction uncertainty within each other and with formulation groups water-based and organic-solvent based and with tested form (concentrates or dilutions). The default values for each formulation group were determined using the upper bound of a non-parametric confidence interval for a specified quantile, with calculations conducted via bootstrapping methods; the proposed values correspond to the upper limit of the 95% confidence interval for the 95th percentile: for concentrates, 10% for organic-solvent based, 4% for water-based and 3% for solid formulations. For dilutions, 42% for organic-solvent based, 37% for water-based and 39% for solid formulations. Organic-solvent based dermal absorption values from experimental data can be used as conservative surrogates for solid and water-based formulations. When no experimental data is available for higher spray dilutions of a given formulation type, a pro-rated correction is proposed to a 2 to 5-fold concentration difference, limited to the respective formulation group default value.