Fig. 7

Correlation of the experimental ΔlogD with the Tanimoto similarity for compound pairs in the training set from a single split of the lipophilicity dataset (Top). Correlation of the prediction errors with the Tanimoto similarity for pairs between the test and the training set for the SNN models (Bottom). The left panel refers to the MLP-SNN trained with the similarity-based pairing, the middle to the MLP-SNN with exhaustive pairing, and the right to the Chemformer-SNN with the similarity-based pairing. The red dashed lines indicate the 95% percentile of the distribution