Fig. 2

MetOncoFit accurately predicts differential expression in vivo using biochemical and topological properties of metabolic genes. The dot plots show the distribution of the values of each feature for the three classes of genes (upregulated, downregulated, or not differentially expressed in tumors compared to matched normal samples). The prominent diamond is the median value within the distribution, while the lines display the standard deviation from the median. Features are sorted based on their relative importance in predicting differential expression (top 10 shown). The [+], [~], and [−] square panels show the direction of the Pearson correlation value between differentially expressed classes and a given feature (see “Methods” section). Confusion matrices report MetOncoFit performance using 10-fold cross-validation; higher diagonal values indicate higher prediction accuracy of a specific class. Data for the three representative cancers were shown. See Additional file 2: Figures S3–S8 for corresponding data for all nine cancer types. The supplementary website provides data for each gene. Top panel: features predictive of differential expression in breast cancer include NCI-60 gene expression levels, catalytic activity, and flux after gene knockout through arginine and proline metabolism. The topological distances to the biomass component—CMP, are negatively correlated with breast cancer differential expression; enzymes topologically closer to CMP, such as RRM2, were more likely to be upregulated. Middle panel: a similar set of top features found in breast cancer were predictive of differential expression in NSCLC as well. In addition, enzymes topologically closer to the biomass components dGMP and phosphatidic acid were more likely to be upregulated. Bottom panel: in melanoma, the topological distance from ammonia was found to be a top predictor that is negatively correlated with differential expression