Fig. 2
Schematic summary of artificial and real urine samples and spectra used for training and test of the predictor. Initially, 3775 artificial urine samples were prepared and their spectra recorded at 300 K. For a subset of 726 samples, selected among those exhibiting high diversity of metabolite concentrations, spectra were also recorded at 302.7 K. These samples/spectra were used to build the predictor (Supplementary Data 1; Supplementary Note 1). The predictor was tested on an independent set of 40 spectra of randomly prepared artificial urine samples. The training of the predictor was repeated by including the 40 additional samples/spectra. An independent set of 120 samples of real urine from 60 individuals was collected and their spectra used as the test set for the predictor (Supplementary Data 3). For 60 of these samples, the concentrations of 11 inorganic ions was independently measured by analytical/clinical methods and used to test the prediction of ion concentrations. Finally, another independent set of 1600 urine spectra17,18,19 was used to test the prediction of 28,983 δ values (Supplementary Data 4)
