Figure 2.

Workflow of the aggregated Mondrian CP set up for the development of the models for the biological assays and the in vivo endpoints. The aggregated CP framework included 20 random splits in calibration and proper training data sets, on which individual RF models were trained, and the resulting p-values per test compound were afterward averaged. The feature selection step was implemented with a lasso model and only included in the development of the in vivo toxicity CP models (in vivo toxicity CP models without feature selection were also trained for comparison).