Neural network model / ECFP
RDkit (2018.03.4), Keras (2.2.2)
None
The Ligand-Target-Affinity dataset was downloaded from the BindingDB website. The BACE-1 ligands were pulled out using our in-house scripts, and the binding affinities were adopted according to the preference of kd > ki > IC50. Others such as kon and koff were not adopted. After merging similar entries and removing large ligands with more than 20 rotatable bonds, the size of ligands in the BACE-1 ligand dataset was reduced to 6754. This dataset was then divided to the training set and the validation set, with a 80:20 ratio. After comparing the error metrics of the validation set predictions using varied molecular descriptors/fingerprints and machine learning methods, a full-connected three-layer neural network trained on the ECFP4 fingerprints generated via RDkit was considered to be the best performer, and used to make the predictions for the target ligands. The neutral network model also showed better performance than other machine learning models such as random forest.
Yes
Yes