Fig. 3
From: MS2DeepScore: a novel deep learning similarity measure to compare tandem mass spectra
A Different molecular fingerprints (morgan2, morgan3, RDKit-daylight) and different scoring methods (Tanimoto/Jaccard vs. Dice) lead to very different distributions of scores across the full dataset (15,062 unique InChIKeys hence a total of 15,0622 pairs). Tanimoto scores based on RDKit-daylight fingerprints tend to give higher scores and thereby result in less unbalanced pair labels. The very pronounced unbalanced nature of Tanimoto scores on circular fingerprints (morgan2, morgan3) can partly be circumvented by switching to Dice scores instead. B MS2DeepScore models were trained for each different structural similarity score. RMSEs are here calculated for all spectrum pairs within the test set (3601 spectra) which fall into one of the 10 possible structural similarity score bins (x-axis labels)