Fig. 6

Error-based calibration plots for the LS-NN uncertainties based on a GCNN model with a 9k+1k training set consisting of Crippen’s logP training data as well as the GCNN model trained on 145k+5k samples. Each bin contains 250 of the test samples. a NN_10k \(\rho _{rank}=-0.04\) b NN_150k \(\rho _{rank}=0.17\)