Fig. 2

Detailed implementation of data process logic in MSpectraAI. a Workflow of data transformation from raw file to intensity matrix. Graphs show an example of one sample with original label 0. (i) Raw mass spectra. There are total i scans in this raw file and the original label of each scan is marked with 0 (As the sample is labeled 0). (ii) Feature Swath Extraction. Split windows across m/z dimension (The range between two red dashed lines is referred to as one “window”, j means total window number) and sum all peak intensities in each window (formula (1)). (iii) Intensity matrix. After summation, the intensities in each scan are normalized by dividing the maximum intensity of each scan (formula (2)). Finally, we obtain the intensity matrix and corresponding label matrix. b Leave-one-out cross prediction strategy. In each independent iteration, one single sample as the independent test data set (gold color), and the remaining samples as the training data (grey color). Then we can estimate total performance based on every iteration result. c The computational framework in each iteration. Graphs show an example of kth iteration (the kth sample with original label 1 as test data). In the left dashed box, the deep neural network model is trained by training data sets, then we predict each row (scan) in test data using this model. The predicted labels can be compared with original labels, if same, marked with “√”, otherwise, “×”. In the right dashed box, if more than 50% of the scans in the test sample are predicted correctly, this means the pre-diction for test sample is correct (“√”), otherwise, wrong (“×”)