Fig. 3

Performance comparison of the six different molecular representations in discriminating active ligands from inactives using a similarity search. Evaluation scheme of the similarity search using the six molecular representations. a The scheme of filtering steps for preparing the benchmark dataset for similarity search. b The performance for the targets of 5 or more actives and inactives (Condition I) and c the targets after filtering structurally similar active-active pairs (Tanimoto coefficient < 0.4 by ECFP4) (Condition II). The performance was measure as AUROC where active-active pairs were regarded as positives, and active-inactive pairs as negatives. The similarity was measured by Tanimoto coefficient for ECFP4/ECFP6 & MHFP6, weighted Tanimoto coefficient for APM (i.e. Ruzicka similarity), cosine similarity for graph2vec and the modified Tanimoto similarity for ErG as in the original study. [27]. (* p < 0.05, ** p < 0.005 and *** p < 0.0005 by t-test)