Fig. 1: Similarity (x-axis) and NLL (y-axis) of generated target compounds with penicillin as the source compound (depicted on top).
From: Exhaustive local chemical space exploration using a transformer model
NLL represents the negative log-likelihood. Top and bottom scatter plots refer to \({\mathcal{D}}\) (fingerprints without counts) and \({{\mathcal{D}}}^{c}\) (fingerprints with counts), respectively. The first two scatter plots from the left show the Tanimoto similarity against the NLL for the λ = 0 and λ = 10 models. λ is the hyperparameter controlling the strength of the regularization term. The CORRELATION, i.e., the Pearson correlation coefficient between the Tanimoto similarity and the NLL, is always better for λ = 10 models. The green line in each scatter plot is the linear fitting of the data points. The plot to the right shows the distribution of the Tanimoto similarity of the generated compounds for the two models estimated for 100 samples.
