Fig. 3: Comparison of selected models from different data layers and cell types.

a Classification performances obtained from univariate-based models. Each panel reports the test set accuracy estimates (n = 5-fold cross-validation strategy). Data were represented as mean values and 95% confidence intervals. On the X-axis of each plot, the ten single top-performing features of each corresponding dataset grouped with respect to the toxicity labeling used are represented. b Classification performances obtained from multivariate-based models. Each panel reports the mean values of the test set accuracy estimates together with 95% confidence intervals (n = 10 best models). Colors indicate the cell model (THP-1 is represented in red, BEAS-2B is represented in yellow, and mouse lung is represented in turquoise), protein corona (represented in gray) and intrinsic properties (represented in violet and named as phys-chem), while the x-axis labels indicate the specific name of the employed data layer. The labels on the top indicate the classification tasks. (CYT) The testing accuracy of models selected for the cytotoxicity score, (INT) the integrated toxicity score, and (NEU) the in vivo toxicity-based classification task.