Fig. 6

a Ideal y-probabilities as a function of the delta of PRF versus RF error margins across emulated train-test standard deviations. Overall, results shown here for a threshold of pChEMBL value of 5 (10 µM) highlight the most optimal PRF probability estimates were observed in cases when standard deviation in the test set most closely resembled that in the training set. It can also be seen that the largest benefit in terms of error margin for the PRF (lower values on the y-axis) are observed toward the midpoint of the ideal ∆y scale, particularly for higher training set standard deviations. This is when the original RF weights the marginal cases equivalent in distinguishing between activity classes. b Effect of Sphere Exclusion (SE) on the error margin between models with and without SE data across different emulated test-train standard deviations. Overall results show that there is no clear advantage of including or excluding SE data when there is no SD in the data. When SD is greater or equal to 0.2, there is a statistically significant difference and hence the inclusion of SE data reduces performance of PRF