Fig 1.
Metabolic models provide a context for the analysis of metabolomic data.
A 1. The refinement step denotes the addition of transport and exchange reactions to enable the uptake and secretion of the metabolites detected in the metabolomic profiles of the NCI-60 cell lines [15]. 2. The condition-specific cell line models were generated using minExCard. In total, 120 models (NCI-60 multiplied by 2) were generated from published metabolomic data and the extended metabolic model. 3. The models were analyzed using a set of computational methods. Based on the computational results, the models were divided into different metabolic phenotypes, and drug targets were predicted for each individual model. The approach is applicable to a variety of biomedical applications. An analysis of tumor or patient-specific omics data could be used to stratify disease phenotypes and to predict personalized disease intervention strategies. B Differences in the number of reactions, metabolites, and genes across a large set of models. C Distribution of the number of reactions, metabolites, genes, and exchange reactions among the 120 cell line models.
Fig 2.
A Rank-ordered ATP yields achieved by the models describe a gradual increase rather than clusters around theoretical ATP yields. The spread of ATP yields highlights the metabolic heterogeneity between the 120 models. The cell lines use a mixture of pathways and metabolic fuels for ATP production, which explains that the predicted ATP yield can exceed the theoretical measures. Two major strategies for ATP production can be distinguished based on the ATP yields. The distinction lies in the higher contribution of either phosphoglycerate kinase (green squares) or ATP synthase (red squares) to the total ATP production. B A fine-grained division of the OxPhos models is achieved considering the production strategies of ATP, NADPH, NADH, and FADH2. The table lists the reactions contributing most to ATP, NADH, NADPH, and FADH2 production for each phenotype (I-VIII). C A three dimensional plot of the eight phenotypes with respect to the utilization of glycolysis, the TCA cycle, and oxidative phosphorylation. D Three different oxotypes are distinguished. The distinction between the OxPhos models (blue) is different from the phenotypic classification performed based on the energy and cofactor production strategies depicted above (see also S1 Text). E Six model clusters are distinguished according to each models’ ability to deal with environmental changes. Variations in glucose, glutamine, lactate, and oxygen lead to a distinct stratification of OxPhos models. Fig F in S1 Text shows different perspectives.
Fig 3.
Six model clusters are distinguished according to the models’ robustness towards environmental changes.
Heatmaps display results for one model from each cluster (and subcluster). Lines in the heatmaps indicate the constraints imposed on the exemplified model. Lac = lactose, glc = glucose and gln = glutamine.
Fig 4.
Replicate models of the same cell line predominantly share the same oxotype. Only 11 model pairs have distinct oxotypes (different oxotypes). A tissue pattern becomes apparent for the melanoma cell lines. Melanoma cell lines predominantly have a low oxotype.
Table 1.
Protein abundance data from the human protein atlas supports the predicted tissue pattern.
The footnotes indicate which cell types or cell lines were considered in the hypergeometric tests. The hypergeometric probabilities (1,2,3) are provided in the main text.
Table 2.
RNA expression data from the human protein atlas revealed low levels of VHL and high levels of HIF1α in skin cell lines that are not part of the NCI-60 panel.
The predicted phenotypes are listed for all NCI-60 cell lines in the data set. None of the skin cancer cell lines in the data set was a NCI-60 cell line. CNS = Central nervous system.