Fig. 2

a Partitioning of initial data into training, testing and validation datasets. For each isoform, the best N molecules were sampled in the “active” class, and the worst N in the “inactive” class, according to flexible bioactivity thresholds. Molecules not sampled in the training or testing dataset, and with bioactivities lower than 20 nM or higher than 100 nM were used as a validation dataset. b Workflow for machine learning experiments: 75% of the initial dataset was subjected to a tenfold cross validation (training phase) and used to train models against a previously unseen 25% (testing phase). Finally, the whole initial dataset was used to train models to predict the validation dataset (validation phase)