Prediction of intracellular metabolic states from extracellular metabolomic data

Abstract

Metabolic models can provide a mechanistic framework to analyze information-rich omics data sets, and are increasingly being used to investigate metabolic alternations in human diseases. An expression of the altered metabolic pathway utilization is the selection of metabolites consumed and released by cells. However, methods for the inference of intracellular metabolic states from extracellular measurements in the context of metabolic models remain underdeveloped compared to methods for other omics data. Herein, we describe a workflow for such an integrative analysis emphasizing on extracellular metabolomics data. We demonstrate, using the lymphoblastic leukemia cell lines Molt-4 and CCRF-CEM, how our methods can reveal differences in cell metabolism. Our models explain metabolite uptake and secretion by predicting a more glycolytic phenotype for the CCRF-CEM model and a more oxidative phenotype for the Molt-4 model, which was supported by our experimental data. Gene expression analysis revealed altered expression of gene products at key regulatory steps in those central metabolic pathways, and literature query emphasized the role of these genes in cancer metabolism. Moreover, in silico gene knock-outs identified unique control points for each cell line model, e.g., phosphoglycerate dehydrogenase for the Molt-4 model. Thus, our workflow is well-suited to the characterization of cellular metabolic traits based on extracellular metabolomic data, and it allows the integration of multiple omics data sets into a cohesive picture based on a defined model context.

Keywords: Constraint-based modeling; Metabolic network; Metabolomics; Multi-omics; Transcriptomics.

Fig. 1

A Combined experimental and computational pipeline to study human metabolism. Experimental work and omics data analysis steps precede computational modeling. Model predictions are validated based on targeted experimental data. Metabolomic and transcriptomic data are used for model refinement and submodel extraction. Functional analysis methods are used to characterize the metabolism of the cell-line models and compare it to additional experimental data. The validated models are subsequently used for the prediction of drug targets. B Uptake and secretion pattern of model metabolites. All metabolite uptakes and secretions that were mapped during model generation are shown. Metabolite uptakes are depicted on the left, and secreted metabolites are shown on the right. A number of metabolite exchanges mapped to the model were unique to one cell line. Differences between cell lines were used to set quantitative constraints for the sampling analysis. C Statistics about the cell line-specific network generation. D Quantitative constraints. For the sampling analysis, an additional set of constraints was imposed on the cell line specific models, emphasizing the differences in metabolite uptake and secretion between cell lines. Higher uptake of a metabolite was allowed in the model of the cell line that consumed more of the metabolite in vitro, whereas the supply was restricted for the model with lower in vitro uptake. This was done by establishing the same ratio between the models bounds as detected in vitro. X denotes the factor (slope ratio) that distinguishes the bounds, and which was individual for each metabolite. (a) The uptake of a metabolite could be x times higher in CCRF-CEM cells, (b) the metabolite uptake could be x times higher in Molt-4, (c) metabolite secretion could be x times higher in CCRF-CEM, or (d) metabolite secretion could be x times higher in Molt-4 cells. LOD limit of detection. The consequence of the adjustment was, in case of uptake, that one model was constrained to a lower metabolite uptake (A, B), and the difference depended on the ratio detected in vitro. In case of secretion, one model had to secrete more of the metabolite, and again the difference depended on the experimental difference detected between the cell lines

Fig. 2

Differences in the use of the TCA cycle by the CCRF-CEM model (red) and the Molt-4 model (blue). The table provides the median values of the sampling results. Negative values in histograms and in the table describe reversible reactions with flux in the reverse direction. There are multiple reversible reactions for the transformation of isocitrate and α-ketoglutarate, malate and fumarate, and succinyl-CoA and succinate. These reactions are unbounded, and therefore histograms are not shown. The details of participating cofactors have been removed. Atp ATP, cit citrate, adp ADP, pi phosphate, oaa oxaloacetate, accoa acetyl-CoA, coa coenzyme-A, icit isocitrate, αkg α-ketoglutarate, succ-coa succinyl-CoA, succ succinate, fum fumarate, mal malate, oxa oxaloacetate, pyr pyruvate, lac lactate, ala alanine, gln glutamine, ETC electron transport chain

Fig. 3

Sampling reveals different utilization of oxidative phosphorylation by the generated models. Different distributions are observed for the CCRF-CEM model (red) and the Molt-4 model (blue). Molt-4 has higher median flux through ETC reactions II–IV. The table provides the median values of the sampling results. Negative values in the histograms and in the table describe reversible reactions with flux in the reverse direction. Both models lack Complex I of the ETC because of constraints arising from the mapping of transcriptomic data. Electron transfer flavoprotein and electron transfer flavoprotein–ubiquinone oxidoreductase both also carry higher flux in the Molt-4 model

Fig. 4

A–K Experimentally determined ATP, NADH + NAD, NADPH + NADP, and GSH + GSSG concentrations, and ROS detoxification in the CCRF-CEM and Molt-4 cells. L Expectations for cellular energy and redox states. Expectations are based on predicted metabolic differences of the Molt-4 and CCRF-CEM models