Skip to main content
. 2019 Jun 19;11:41. doi: 10.1186/s13321-019-0364-5

Fig. 1.

Fig. 1

KekuleScope framework. a We collected and curated a total of 8 cytotoxicity data sets from ChEMBL version 23. b Compound Kekulé representations were generated for all compounds and used as input to the ConvNets. c We implemented extended versions of 4 commonly used architectures (e.g., VGG-19-bn shown in the figure) by including five additional fully-connected layers to predict pIC50 values on a continuous scale. d The generalization power of the ConvNets was assessed on the test set, and compared to RF models trained using Morgan fingerprints as covariates