In order to validate model design and prediction quality we employed a radical time-split validation. The model was reparameterised based on historical data (ChEMBL database version 23, released in 2017) and used to predict all new (till ChEMBL database version 33), previously unseen dependencies, comprising ~100000 drug-darget pairs. The results were quantified in terms of accuracy (the probability of correct discrimination between active and inactive ligand-target pair) and sensitivity (here, the target-averaged probability of finding a known active compound within the top 20 hits). The results were compared with state-of the art models operating on the same training and validation datasets.