Cellular metabolism—the process by which nutrients are converted into energy, macromolecular building blocks, and other small organic compounds—depends upon the expression of genes encoding enzymes and their regulators. Well-characterized transcriptional regulatory circuits such as the lac and trp operons in E. coli and the galactose utilization system in S. cerevisiae illustrate how the concentration of metabolites such as tryptophan or galactose can modulate gene expression. In addition, changes in gene expression can lead to increases or decreases in the concentrations of enzymes and regulatory proteins, thereby affecting concentrations of intracellular metabolites. While individual cases of mutual regulation by metabolites and gene products have been and continue to be described, identifying the full scope of these interactions is important for improving rational control of metabolism to meet therapeutic and bioengineering objectives. Clinical scientists, for instance, may be interested in developing novel treatments that control blood glucose levels in diabetic patients, or that fight cancer by disrupting metabolism in tumor cells. This line of inquiry is also relevant to bioengineers seeking to increase the production of small molecules (such as biofuels or flavor molecules) by knocking out or overexpressing individual genes.
The simultaneous measurement of metabolite and transcript concentrations is one method that has begun to show promise for identifying gene products and small molecules involved in the same biological processes 
. A number of studies 
have followed the behavior of specific secondary metabolites of interest such as volatile signaling molecules 
or compounds with pharmaceutical properties 
, as well as transcripts, in response to genetic or biochemical perturbations. The further refinement of high-throughput experimental technologies such as mass spectrometry has enabled recent studies to measure many functional classes of metabolites together with a large proportion of the transcriptome 
. For example, one recent ground-breaking study collected extensive data on metabolite, protein, and transcript levels in E. coli
following the disruption of genes in primary carbon metabolism or changes in growth rate, and concluded that metabolite concentrations tended to be stable with respect to these perturbations 
. Another study 
compared transcript and metabolite concentrations in S. cerevisiae
under two different growth conditions, and using a novel computational method in which known metabolic pathways were divided into smaller pathways termed “reporter reactions,” the authors observed that when two different growth conditions were compared, the majority of the reporter reactions showed changes in transcript concentrations, with fewer revealing significant alterations in metabolite levels. Such methods, which make inferences based on comprehensive reconstructions of biochemical pathways in an organism, represent valuable tools for analyzing metabolomic and transcriptional data together. However, there is still a need for approaches that are designed to answer the problem of identifying novel interactions between specific gene products and metabolites that include both enzymatic and regulatory relationships.
Of prime importance to the problem of finding gene–metabolite relationships from data is the question of whether functionally-related metabolites and transcripts do indeed show coherent patterns of concentration changes that can be used to make valid predictions. Studies aimed at addressing this question have relied on computing correlation coefficients between profiles of transcript and metabolite concentrations, which can then be ranked 
or used to co-cluster the metabolomic and transcriptomic data 
. However, it is possible that other types of regulation, such as post-translational protein modifications and feedback inhibition, could be more predominant in the aggregate than transcriptional regulation 
. Accordingly, a major limitation with these computational techniques is that the extent to which transcripts and metabolites are co-regulated is not known. The proportion of strong gene–metabolite correlations that are due to chance or indirect effects, as opposed to enzymatic or regulatory relationships, has also not been determined by previous investigations.
In part due to these concerns, previous work has come to contradictory conclusions about the extent of coordination between metabolite and transcript concentrations. Some qualitative evidence has been provided for the claim that transcripts and metabolites are substantially co-regulated 
, including the comparison of clustering patterns in each data set 
, and examples of coherent correlations between biosynthetic enzymes and their products 
. In contrast, other studies contend that transcript and metabolite profiles tend to behave differently 
, and some have argued that correlative approaches are not specific enough to draw conclusions about which genes and metabolites are functionally related (such that the expression of a gene product controls the concentration of a metabolite, or vice versa) 
Indeed, observed correlations within metabolic networks often confound straightforward interpretations. Metabolic networks, unlike transcriptional or protein-interaction networks, consist of molecular species which chemically interconvert. As a result, metabolites that are only distantly related in terms of the underlying pathways can show high levels of correlation 
. This is especially true in the case of global perturbations (e.g., nutrient starvation, diurnal cycles) which affect many different branches of metabolism at once 
. It is therefore likely that the interpretation of correlations between transcript and metabolite concentrations will depend on contextual factors, such as the branch of metabolism being studied or the experimental perturbation under which the correlations were observed.
In order to examine these questions further, we conducted a systems-level investigation of the metabolome and transcriptome of S. cerevisiae, in which we measure the dynamic responses of metabolites and transcripts to two nutrient deprivations. We examine whether transcripts and metabolites are co-regulated in general, and demonstrate the existence of a strong trend for correlated genes and metabolites to participate in related biological processes. We also demonstrate that the correlations observed for related gene–metabolite pairs are dramatically different depending on the type of metabolite and the perturbation to which the cells are subjected, and we develop a Bayesian algorithm capable of accounting for these dependencies. When applied to our experimental data, this algorithm makes gene–metabolite interaction predictions that are significantly more precise and complete than those made by correlation alone.