Fig. 5. Applicability domains of each modelling technique. The figure illustrates the enhanced extrapolation into chemical space facilitated by the inclusion of heterogeneous descriptors. While all models give similar CCR performance of >0.85 when tested on the chemical compounds that are the most similar to the training data, the models built on chemical descriptors only give the worst performance on less similar chemicals with a CCR of <0.6 on the least similar chemicals. In contrast, the models built from all three data domains are the best performing on the molecules that are the most dissimilar to the training set. Abbreviations: as for Fig. 3.