Fig. 7
From: MS2DeepScore: a novel deep learning similarity measure to compare tandem mass spectra
Monte-Carlo dropout provides Tanimoto score predictions, but also the interquartile range as an uncertainty measure (here over 10 × 10 = 100 individual scores). Discarding scores with higher uncertainties (higher IQR, interquartile range) does indeed improve the average prediction performance notably (A), although at the price of lowering the retrieval rate (retrieval rate = fraction of total scores with IQR < threshold B. C shows the root mean squared errors for different IQR threshold and for different Tanimoto score bins