Fig. 3
Observed vs predicted for both case studies. Observed against predicted values on the test set corresponding to a the compound solubility (LogS) dataset (case study 1: QSPR), and b the cyclooxygenase (COX) inhibition dataset (case study 2: PCM). Both a and b were generated with the function CorrelationPlot. The area defined by the blue lines comprises 1 LogS units (a) and 1 pIC50 units (b). Both plots were generated using the predictions on the test set calculated with the Linear Stacking ensembles (Tables 1, 3). Overall, high predictive power is attained on the test set for both datasets, with respective RMSE/\(R_{0}^2\) values of 0.51/0.93 (a), and 0.73/0.63 (b). Taken together, these data indicate that ensemble modelling leads to higher predictive power, although this increase might be marginal for some datasets (b).