Genetic variation leads to networks of co-regulated transcripts. The implications of these network structures have been discussed extensively, generally with the assumption that such transcriptional networks give rise to corresponding protein networks 
. However, due to limitations in technology, these hypothesized protein networks have not been examined directly, and thus it is not known whether they are driven by underlying transcriptional networks. By measuring protein and transcript levels for 354 genes in a genetically diverse population of yeast segregants, we are now able to address the question “Are protein networks formed primarily on the basis of regulation of their underlying transcripts?” (We use the term “protein networks” to refer to groups of proteins that are co-regulated and not groups of proteins that interact physically or genetically.) Before describing our results, three points are important to consider: First, the magnitude of individual to individual variation in transcript levels for a single gene is generally far less than the magnitude of gene to gene variation in transcript levels within a single individual 
. Second, the demonstration in multiple studies that the correlation between transcript and protein levels for different genes within a single individual is high does not imply that differences in abundance of the same transcripts between different individuals must cause corresponding variation in protein abundance 
. And third, a correlation between transcript and protein networks does not prove a causal relationship between the two.
Protein levels cannot necessarily be inferred from transcript levels because protein levels can be controlled not only by regulating transcripts but also by regulating other steps in protein metabolism, such as translation and protein stability. Thus the degree to which protein levels can be inferred from transcript levels depends on the degree to which the former mode of regulation overwhelms the latter two (). In experimental situations when a gene is placed under a strong promoter like the CMV promoter, a transcript can be elevated 1,000-fold and this generally leads to a striking increase in protein. However, in genetically diverse populations, transcript levels generally do not vary 1,000-fold between individuals; for example, in the population of yeast described in this report, a typical transcript varies just 2.7-fold across 95 individuals. Such modest variation in transcript levels may be buffered such that it causes no variation in protein levels, or regulation of translation and/or protein stability may obscure effects of minor transcriptional variation. Under such circumstances, transcript levels should not be expected to reflect protein levels, and whether such circumstances are the norm or the exception for typical levels of inter-individual variation is not known.
Three possibilities for regulating protein levels.
Several reports have demonstrated that, when comparing different genes whose transcript levels vary over orders of magnitude, high abundance proteins are associated with high abundance transcripts and vice versa 
. There are two important differences between this issue and the issue we address here: First, the magnitude of variation in transcript levels is vastly different in the two situations. Just as 1,000-fold overexpression of a transcript through experimental manipulation is virtually guaranteed to increase the level of the corresponding protein, transcript levels that differ by orders of magnitude between genes are virtually guaranteed to manifest themselves in differences in the corresponding proteins. For example, a 2007 study by Lu et al. of 346 genes in yeast demonstrated a high correlation (R
0.85) between transcript and protein levels, and thus concluded that “… >70% of yeast gene expression regulation [occurs] through mRNA-directed mechanisms” 
. However, if one calculates correlation coefficients with this same data set using a sliding window within which transcript levels vary on average just 3.5-fold, the average correlation drops from 0.85 to 0.36 with almost half of the bins showing correlation coefficients that could easily have been achieved by chance (Figure S1
, Text S1
). Thus, the striking correlation between transcripts and proteins all but disappears when analysis is limited to a range of transcript variation similar to that occurring between individuals. Second, as we will discuss further below, studies like that of Lu et al. involve protein and transcript measurements from a single individual under a single experimental condition, making it impossible to measure gene-specific correlation coefficients (Figure S2
). The point here is not to call into question the solid conclusions of these past reports, but instead to point out that our work addresses a different issue.
In assessing the importance of transcriptional regulation in determining protein levels, it is important to distinguish correlation from causality. Biological pathways that sense physiological conditions can trigger responses that include changes in transcription, translation, and protein stability, often with the same group of genes targeted by more than one of these regulatory mechanisms 
. For example, the TOR pathway, a highly conserved pathway named for the signaling kinase “Target of Rapamycin,” responds to changes in nutritional conditions by increasing both transcription and translation of a group of target genes 
. Thus transcript and protein levels of these genes will be correlated, but if translation has a much larger effect on protein levels than does transcription, this correlation need not reflect causality. More generally, any time a cellular response pathway has both a transcriptional and a post-transcriptional branch, the target genes are expected to show a correlation between transcript and protein levels, but it is only in those cases where the former regulatory mechanism is the dominant one in affecting protein levels that this correlation reflects a predominantly causal relationship between transcript and protein levels ().