Researchers from Tel-Aviv University and the University of California, San Diego, have developed a computational method that integrates gene- and protein-expression data with a metabolic network model to predict tissue-specific metabolism in human tissues.
The approach builds upon the first manually reconstructed human metabolic network, which UCSD’s Bernhard Palsson and colleagues published last year, but it enables, for the first time, the prediction of tissue-specific, post-transcriptional regulatory effects — a capability that could have far-reaching consequences, according to its developers.
“We used [the Palsson model] and tried to predict how it works in different contexts in the body,” Tomer Shlomi of Tel-Aviv University told BioInform. The Palsson group “did the hard work, collecting all the evidence, all the reactions,” he said. “We just activated the model in different contexts.”
This effort has essentially added an analytical layer to the original Palsson model, tacking tissue-specific behavior onto the topology of the human metabolic network. “We developed a computational method that takes this model and tells you how it works in different tissues,” Shlomi said.
Shlomi, Palsson, and colleagues published the method in the Aug. 17 advance online edition of Nature Biotechnology.
According to the authors, the approach promises a number of applications, including “the classification of tissue-specific gene-expression measurements from either healthy or sick individuals, based on the predicted metabolic behavior that they induce.”
Another “compelling” potential application, they write, “would be the prediction of disease and tissue-specific biomarkers that could be identified using biofluid metabolomics.”
Shlomi said that other researchers can use the approach to integrate a genome-scale metabolic network model with tissue-specific gene-expression and protein-abundance data. This method can help scientists study their data sets, visualize them, and it can lead to scientific discovery, too, he said. “What’s important [about this study] is the concept of integrating high-throughput data in the model,” he said.
In the Nature Biotech paper, the authors describe how they used the method to predict metabolic activity in 10 human tissues — liver, brain, heart, kidney, lung, pancreas, prostate, spleen, thymus, and skeletal muscle. They found that around one-fifth of the metabolic genes expressed in these tissues are post-transcriptionally regulated, Shlomi said.
“It tells you that relying on gene expression alone is problematic because in 20 percent of the cases gene expression alone is not really indicative of whether the enzymes are really activated or not,” he said.
“What’s important is the concept of integrating high-throughput data in the model.”
While the scientific community is aware that gene-expression data does not tell the entire metabolic story, Tomer said, “What is surprising is that we can identify these cases and show that our method predicts the activation and inactivation in the correct context, the correct tissue.”
Other research groups may be now able to build on top of this work, giving it more layers of complexity, and helping to put their own data into a larger network-based context, said Shlomi. A clinician-scientist researching obesity, for example, might discover an enzyme tentatively implicated in the disorder in a series of blood samples. “Then our method and analysis can help find out in which tissue specifically this enzyme is working and in what processes it is involved,” he said.
The method relies on constraint-based modeling, or CBM, an approach that has mainly been used to study microbial metabolism. This modeling approach “requires very few a priori data on enzymatic reactions and successfully predicts metabolic behavior on an organism level,” Shlomi said.
The authors explain in the paper that, for each tissue type, they used an optimization method called a mixed integer linear programming problem in order to identify a steady-state flux distribution that satisfied several constraints, such as thermodynamic constraints, while maximizing the number of reactions with activity consistent with their expression state.
The method relies on enzyme expression levels to infer tissue-specific metabolic flux. “We treat tissue-specific variations in enzyme-expression levels not as the final determinants of enzyme activity, but as cues for the likelihood that the enzyme in question supports metabolic flux in its associated reaction(s),” the note.
Microbes to Man
Constraint-based modeling has been used for many years in microbial metabolic modeling, but its use in studying human metabolism is still relatively new.
In a paper published in the Proceedings of the National Academy of Sciences last year, Palsson and colleagues reconstructed the first human metabolic network map based on genomic data and legacy data taken from the literature. A total of 1,496 ORFs, 2,004 proteins, 2,766 metabolites and 3,311 metabolic and transport reactions were transformed into an in silico model called Recon 1.
Palsson and co-authors used the model to interpret the effects of gastric bypass surgery on skeletal muscle metabolism and revealed “subtle but consistent overall patterns of expression change.”
Noting that genome-scale microbial metabolic reconstructions have been used to “successfully perform systems analysis,” they wrote in the PNAS paper that they expect global human metabolic reconstruction to “enable significant dimensions in the study of human systems biology.”
They acknowledged, however, that arriving at this goal would require the integration of transcriptomic, fluxomic, proteomic, and metabolomic data. “Recon 1 provides the context for integration and analysis of these data into predictive models,” the PNAS paper stated.
As Shlomi explained, the Palsson model is a collection of all possible human metabolic reactions and needs to be computationally adapted to answer specific biological questions, including how metabolism operates in different contexts.
“What we show on top of the network is which part of the network is really activated based on the expression data, and you can see what the added value of our method on top of the expression data is,” said Shlomi.
In the Nature Biotech paper, the team explains that it first applied the method to predict the metabolic behavior in yeast based on gene expression data and the growth of yeast cultures in various media. The predicted metabolic activity to measured metabolic flux in the central carbon metabolism revealed a precision of 0.71 and a recall of 0.89. Without using the computational method, the predicted flux obtained solely from expression data was 0.83 with a recall of 0.61, indicating, as the scientists wrote, “markedly lower accuracy.”
The team next built on the Palsson model by determining the activity states of 644 genes, along with gene- and protein-expression measurements, in 10 human tissues. They found that many genes predicted to be active in a certain tissue were not highly expressed there, nor were the genes predicted to be inactive showing low expression values. This indicated that “this shows the considerable amount of additional information obtained by integrating expression data with the metabolic network to infer metabolic gene activity.”
“What we predict is supposed to be the actual metabolic flow based on this expression data, but it is not quite the same [as the expression data],” Shlomi said. “This superimposing of the method on the network reveals what the real metabolic behavior is.”
The scientists also created visualizations of their data, which can be explored by using the open source Cytoscape network visualization software.
Humans are Hard
Shlomi noted that it is difficult to model obesity and metabolic diseases, a task that requires comprehensive and accurate models of various cell types, tissues, and organs, while accounting for the dynamic nature of these systems. Tissue metabolism comprises thousands of biochemical reactions that modify the chemical properties of thousands of small molecules. “The problem is that we currently know of only a subset of these reactions and small molecules,” he said.
Even for known reactions, the rates at which they operate are often unknown, as are the factors that regulate their function and their response to environmental and genetic factors. In addition, in more complex organisms, tissues differ in their metabolic objectives. “A major task of systems biology is to use this partial knowledge, which is, of course, constantly extended and improved by biologists, to derive models that can provide novel, valuable insights on the function of biological systems,” Shlomi said.
Shlomi and his colleagues believe that their new method opens the door to computationally investigating metabolic disorders using expression data. Building on the network from the Palsson lab, it promises to predict flux that is consistent with the expression data, he said, which lays the foundation for the rapid development of metabolic models that are human tissue-specific and can help researchers study metabolic disorders computationally.
Shlomi is about to start a professorship at Technion, the Israel Institute of Technology, in the computer science department, where he plans to continue the computational analysis of metabolism in the context of large-scale biological networks.
“I intend to start with simulations of metabolic alterations that result from relatively ‘simple’ enzymatic disorders, compared to complex multi-factorial diseases that involve multiple genetic and environmental factors, such as obesity and diabetes, toward the identification of diagnosis biomarkers and in the long run, possible drug-targets,” he said.
The team’s datasets are available here.