Fig. 2

Benchmarking the predictive power of ConvNet architectures on cytotoxicity data sets. Mean RMSE values (± standard deviation) on the test set across ten runs for each of the ConvNet architectures explored in this study (AlexNet [57], DenseNet-201 [58], ResNet152 [59] and VGG-19 [60]). Overall, all architectures enabled the generation of models with high predictive power on the test set, with RMSE values in the 0.65–0.96 pIC₅₀ range. However, the extended versions of these architectures that we designed by including 5 fully-connected layers (see Fig. 1) constantly led to increased predictive power on the test set