Fig. 1

Overview of the workflow used in this study. We initially assembled and modelled 1360 data sets from ChEMBL database using RF (Random Forest) and conformal prediction. Of these, 440 displayed high predictive power in cross validation (q² > 0.5) and on external molecules (R² _{test set} > 0.6), and hence, were selected to build QAFFP fingerprints for 18 cytotoxicity and 25 protein target data sets (Tables 1 and 2) assembled from ChEMBL database as well. To benchmark the predictive signal of QAFFP, we compared the performance of RF models trained on QAFFP against models generated using Morgan2 fingerprints or 1-D and 2-D physicochemical descriptors across a diverse set of 43 bioactivity data sets (Tables 1 and 2)