Fig. 4
From: QSAR-derived affinity fingerprints (part 2): modeling performance for potency prediction
RMSEtest values calculated with models trained on each of the 11 descriptor types considered across the 43 data sets modelled in this study (18 cytotoxicity and 25 protein data sets; Tables 1 and 2). We trained 50 models for each combination of descriptor type and data set, each time holding a different subset of the data as test set. Overall, predictive models were obtained for all descriptor types, and the performance of different descriptor types varied across data sets modelled