Fig. 6
From: Target prediction utilising negative bioactivity data covering large chemical space
AD analysis of the True Positives from the WOMBAT test set. The models perform better overall when the similarity between the test and training sets are high. The area between 0.25 and 0.1 shows the largest variation in performance, when models can have problems distinguishing between activity classes. Similarities below 0.1 consistently predict true positives as inactive. Hence, a cut-off around 0.3 can be employed to give insight into the reliability of the predictions generated by models