Fig. 4

Comparison of the similarity-based pairing with exhaustive pairing to train the MLP-SNN (left), and Chemformer-SNN (right). The number of reference compounds was chosen based on the smallest RMSE as 10 for the lipophilicity, 6 for the freesolv and 7 for the ESOL dataset for the similarity-based pairing, and 8 for the lipophilicity, 10 for the freesolv and 19 for ESOL for exhaustive pairing. For the Chemformer-SNN it was 7, 11 and 7. The error bar indicates the standard deviation from the tenfold cross validation. For the random pairing, each compound was paired with 50 randomly selected compounds as a surrogate of exhaustive pairing to the Chemformer-SNN