Fig. 3: Overview of Tanimoto similarities (denoted with Sim in the figure) and NLLs (negative log-likelihoods) for a ChEMBL-series consisting of five compounds.
From: Exhaustive local chemical space exploration using a transformer model
To the top the ChEMBL-series consisting of five compounds and displayed as a graph, where the nodes depict molecules and the edges depict the NLL of the molecular transformation and the Tanimoto Similarity (Sim in the plot) between the two connected nodes (molecules). Note that the NLL is not symmetric between a compound pair. The arrows on top of NLL denote the direction of the molecular transformation from the source molecule. If any generic pair of molecule in the graph s and t are connected then \({\overrightarrow{NLL}}\) denotes \(-\log p(t| s)\) and \({\overleftarrow{NLL}}\) denotes \(-\log p(s| t)\). The lower-left plot depicts the correlation between Tanimoto similarity (evaluated on fingerprints with counts) and the NLL for the molecular transformation for all the possible pairs contained in the 200 compound series. The red dots denote all the pairs of the above series. The lower-right plot illustrates the maximum NLL reachable with different beam sizes. The fill color surrounding the solid lines represents the standard deviation, which is calculated based on the beam the 200 compound series.
