Accuracy of models generated with various single and paired molecular representations using support vector machine (SVM) during cross-validation (purple heatmap) and testing (blue heatmap)

The top performing standalone fingerprints for each of the 5 ML algorithms

The best and worst performing models using a merged fingerprint for all 5 ML algorithms

Accuracy (%) of models trained with an imbalanced training dataset where the number of BRAF actives is decreased but the number of BRAF inactives is maintained at a fixed number (3600)

Accuracy (%) of models trained with a balanced training dataset where the numbers of BRAF actives and BRAF inactives are both similarly decreased

Recall and precision (%) of models trained with an imbalanced training dataset where the number of BRAF actives is decreased but the number of BRAF inactives is maintained at a fixed number (3600)

Recall and precision (%) of models trained with a balanced training dataset where the numbers of BRAF actives and BRAF inactives are both similarly decreased

Average accuracy for the ‘spiked-in’ “less active”-trained models based on testing with 10 balanced BRAF actives and inactives hold-out test sets

Accuracy of models generated with various single and paired molecular representations using random forest (RF) during cross-validation (purple heatmap) and testing (blue heatmap)

Accuracy of models generated with various single and paired molecular representations using naïve bayes (NBayes) during cross-validation (purple heatmap) and testing (blue heatmap)

Accuracy of models generated with various single and paired molecular representations using k-nearest neighbour (kNN) during cross-validation (purple heatmap) and testing (blue heatmap)

Accuracy of models generated with various single and paired molecular representations using gradient-boosting decision tree (GBDT) during cross-validation (purple heatmap) and testing (blue heatmap)

Average accuracy for the ‘spiked-in’ decoy-trained models based on testing with 10 balanced BRAF actives and inactives hold-out test sets

Accuracy for the ‘spiked-in’ decoy-trained models based on testing with a balanced BRAF actives and decoys hold-out test set