Skip to main content

Advertisement

Fig. 8 | Journal of Cheminformatics

Fig. 8

From: KekuleScope: prediction of cancer cell line sensitivity and compound potency using convolutional neural networks trained on compound images

Fig. 8

Absolute errors in prediction. Relationship between the absolute errors in prediction for the same test set instances calculated with ConvNets (x-axis) and RF (y-axis) models trained on the same training set instances. The predictions generated by each model differ in > 2 pIC50 units in some cases, and are moderately correlated (R2 in the 0.58–0.65 range; P < 0.05). Note that most of the instances are located in the lower-left quadrant (bins coloured in blue), thus indicating that the absolute errors in prediction for most instances (i.e., those instances in the diagonal in the plots shown in Fig. 7) are low and correlated for the two modelling strategies. This is expected given the high predictive power of the models (Fig. 4) and the correlation of the predictions (Fig. 7)

Back to article page