A complex tissue such as the brain cannot be completely modeled without taking into account the fact that it is spatially organized into different regions and that it consists of multiple different cell types.

A few years ago, researchers from the laboratory of Bernhard Palsson, at the University of California, San Diego, reported a genome-scale reconstruction of the human metabolic network, and subsequent refinements of this model included tissue-specific information as well. But cell-specific information was still lacking. In a recent paper, Nathan Lewis, Palsson and colleagues have now taken this work to a still finer-grained level, modeling metabolic coupling between different cell types in the human brain. They examined metabolic coupling between astrocytes and three different types of neurons—glutamatergic, GABAergic and cholinergic neurons.

To do this, they used the existing human metabolic network (named Recon 1) as the context for cell-specific information curated from the literature. “Basically, we used proteomic and transcriptomic data to help us understand what metabolic pathways are active in the brain,” says Lewis. “Then we parsed this out into individual cell types and coupled the cells with known transporters that go between the cell types.” This needed both manual curation of cell type–specific information from the literature, as well as the use of databases like the Human Protein Reference Database, the Human Protein Atlas and the Human Proteome Organization (HUPO) brain proteome project. What resulted were mathematical models that describe metabolic interactions between the defined cell types and that can be used, in simulations, to make predictions about systems-level properties of the brain.

The fluxes in the models, the researchers report, were consistent with what is known from experimental measurements, giving confidence that they are reasonably accurate representations of reality. What is more, the models could predict the empirically known protection of GABAergic neurons from cell death in early stages of Alzheimer's disease, as compared to glutamatergic or cholinergic neurons. These differences could be described in terms of the response to changes in central metabolic enzymes in individuals with the disease.

As Palsson points out, genome-scale reconstructions are integrations of hard biological data and are thus particularly useful for generating mechanistic understanding. “Most analysis [of systems-level data] is statistically driven, which means it gives you clusters and correlations and a lot of inference-type information,” he says. But when you add 'omics' data to a genome-scale mathematical model, as the researchers did in this study, “you suddenly start to see the mechanisms at work, for example, that underlie the generation of an expression profile,” he says. In their analysis of cell type–specific metabolic changes in Alzheimer's disease, for instance, Lewis, Palsson and colleagues used the model to generate precise, molecular and testable hypotheses about the underlying basis for the difference between the neuronal types. Thus these models provide a mechanistic 'context' for high-throughput data 'content'.

One of the reasons it has been possible to build predictive models for metabolic networks is that there is a great deal known about the function of the gene products involved (these number about 1,500 open reading frames in Homo sapiens and about 1,300 in Escherichia coli), but there are more and more such data becoming available for other biological processes, such as transcriptional regulation, as well. Coupled with a shift in emphasis from data generation to data integration and analysis, we may be heading for a time when models much more routinely lead to a mechanistic understanding of biological processes.