Fig. 5
From: MS2DeepScore: a novel deep learning similarity measure to compare tandem mass spectra
Combining MS2DeepScore predictions for spectral pairs corresponding to the same pair of molecules leads to more reliable Tanimoto score predictions. A Individual predictions (red dots) show consistently higher RMSEs than the median of predictions for (pairs of spectra corresponding to) the same compound pair (“all”). We also computed the interquartile range (IQR) of predictions of the same molecule pairs, which can be used to remove high IQR outliers. Retrieval rates after each label indicate the total fraction of scores that fulfilled the given criterion (IQR < x). B When compared to Fig. 2 it is apparent that the high Tanimoto score predictions become notably more reliable when removing scores with large variations of same-InChIKey predictions (here: keep scores with IQR < 0.2)