Fig. 3

Performance of the (Glucose,Formate) classifier. a Colour map of the relative mutual information between the classifier prediction and the reference classification as a function of the formate and glucose thresholds. b Scatter plot of human serum samples as a function of the formate and glucose concentrations, colour coded by the sample subtypes. The horizontal and vertical lines represent the best glucose and formate threshold, respectively. c ROC plot for the glucose + formate-based classifier, focusing on the cancer class. The brown area reports the FPR and TPR for different formate and glucose thresholds and the red symbol the corresponding cross-validation values (symbol for median and error bars for observed range excluding the 5% lowest and largest values). d, e Comparison of the different classifiers according to their TPR (d) and FPR (e). Corrected indicates accounting for obesity incidence. Error bars indicate the range excluding the lowest 5% and highest 5% values, boxes the range excluding the lowest 25% and highest 25% values, and the line within the box the median