deepScaffOpt_FCFPL_ECFPL_2Dpp_nolossweights
deepScaffOpt
FCFPL fingerprints (FCFP with 8192 bits), ECFPL (ECFP with 8192 bits) and 2Dpp were all used in this protocol without weights in the loss function.
deepScaffOpt algorithm (still unpublished) trains multiple deep neural netwroks (DNNs) using up to 6 different descriptors (ECFPL, FCFPL, RDK5, 2Dpp, ErgFP, mol2vec). Each DNN is trained using as training set all ligands with binding affinities deposited in CHEMBL for this particular receptor. Only one descriptor is used to train each DNN. Then the best DNNs are selected automatically based on some empirical criteria and are combined to give a Meta-Predictor model. Many Meta-Predictor models are generated, each from a slightly different combination of DNNs. From the ensemble of different Meta-Predictor scorings, only the Meta-Predictor scoring that is closer to the average scores is retained.
No
Yes