Haloperidol-induced catalepsy is an ideal phenotype for integrative analysis of gene networks, genetic diversity and behavior. The mechanism of drug action is well established (blockade of D2
-like dopamine receptors), the striatum is the major target for the extra-pyramidal response and the phenotype is highly heritable 
. Differences in response among inbred mouse strains are not due to differences in D2
receptor density, the relative amounts of the long and short receptor isoforms or haloperidol pharmacokinetics 
. The genetic differences are not haloperidol specific; for example, the D2
receptor antagonist, raclopride, shows the same inbred strain distribution pattern 
; all of the selected lines differed significantly in their response to raclopride (data not shown). However, the genetic differences are not paralleled by differences in response to D1
receptor antagonists, e.g., SKF 23390 
From a translational perspective, comparison of mouse populations with different genetic backgrounds is an essential first step. Before any genetic or molecular mechanisms can generalize to humans, they must be shown to be similar in genetically distinct subpopulations of the same species. Broadly speaking, a systems-biology approach was used to detect mechanisms that transcend a specific genetic background. Key to this approach was the inclusion of a selection from the genetically diverse HS-CC founders. As noted by 
and confirmed by genome-wide next-generation sequencing 
intercrosses and four-way and eight-way HS populations formed from standard mouse laboratory strains will capture only a fraction of Mus musculus
genetic diversity; F2
and standard HS populations are actually more similar than different. In contrast, the eight strains used to form the HS-CC, which includes three wild-derived strains, capture approximately 90% of the available allelic diversity. However, despite the marked difference in diversity, the basal HS-CC striatal coexpression network is similar albeit not identical to that found in F2
and HS4 animals 
The WGCNA is one of several graph/network approaches for analyzing gene coexpression structure. We recognize that other approaches and indeed subtle variations in the implementation of the WGCNA can yield different results 
. The WGCNA aligns strongly with two principles. The first is that coexpressed genes are likely to share biological functions; importantly, this principle allows putative annotation of genes and noncoding RNAs with no known function. In our case, we used an unsigned network 
, which implies that both positively and negatively correlated genes are coexpressed and often assigned to the same module. The second principle is that the network structure is scale-free and follows a power law distribution; this implies that the network is best described by a few highly connected hub genes, with most of the nodes sparsely connected 
. WGCNA was used to show that it was at the level of network structure and not differential expression that one could discriminate the nonhuman primate and human brain transcriptome and the regional brain differences in the human transcriptome 
. Importantly from a translational prospective, recent work has demonstrated that eQTL in combination with coexpression network analysis can identify novel candidate genes related to schizophrenia 
. The WGCNA has been used to dissect the transcriptome into modules associated with specific cell types (neurons, oligodendrocytes, astrocytes and microglia), specific organelles and synaptic functions 
; our data replicated this observation. Recent work 
has also illustrated that the WGCNA is sufficiently sensitive to detect behaviorally relevant differences in network structure even among an inbred strain.
Data previously reported 
were the first to show that selection for a behavioral phenotype had marked effects on WGCNA generated network structure. In this example, the HS4 was specifically created to assess selection for haloperidol response; of the founder strains, two (D2 and C) are haloperidol responsive and two (B6 and LP) are non-responsive 
. For the gene coexpression analysis of the HS4 selected lines, the construction of the consensus network from the responsive and non-responsive lines was key; the data illustrated that the module showing the most significant effects of selection was one enriched in genes associated with intracellular signaling and locomotor behavior (similar to the “Pink” module of the current study). The consensus network used in the current study was somewhat different from that reported 
, in that a larger number of transcripts were included in the analysis and data were collapsed across three different selections. An important observation 
was that selection had only modest effects on differential gene expression. This point was extended in the current study and further it was found that there was no overlap in differentially expressed genes among the three selections. These data align with previous observations that haloperidol-response quantitative trait loci (QTL) are genotype dependent 
. The consensus network approach revealed that selection for haloperidol response, regardless of genetic background, disrupted similar aspects of network structure; the key disruptive features were changes in intramodular connectivity. Three modules (“Grey60”, “Pink” and “Green”) were consistently affected. But this commonality was only detectable at the module level; connectivity changes among individual transcripts were founder population unique. Thus, within the striatal transcriptome the common selection element is the gene module and not individual gene connectivity. It is of interest to note that the specificity of the analysis was largely generated by the F2
data; the more complex crosses recruited additional modules such that in the HS-CC sample, nearly all of the modules within the striatal gene network were affected.
The question arises as to whether or not the apparent relationship between genetic diversity and the extent of module involvement is a general rule: does selection recruit a more complex biology in the more complex crosses? The data presented here cannot address this issue; replicate selections from the HS4 and HS-CC populations would be needed to eliminate the effects due to random allele fixation.
The module-centric/genotype-dependent view of selection, heritability and behavior could be generalized to other phenotypes and other populations under certain conditions. While publically available and behaviorally relevant datasets are generally too small to accurately construct gene networks, a meta-analysis approach 
could address this issue provided the behavior is measured under similar conditions. It has been argued 
that it is gene-environmental interactions that are key to understanding the natural variation in behavior, which is the basis of selective breeding. Our data provide an example of holding the environment “constant”, varying genetic background and observing selection-induced changes in network structure. But in real-world situations, e.g., a genome wide association study of natural variation, neither the background nor the genotype is held constant. Thus, the possibility of aligning phenotypic variation with connectivity becomes more difficult.
Given this complexity and the fact that module(s) specifically and networks generally may be the key to understanding heritability, how best to investigate the gene-behavior relationship, especially with a translational perspective? Here, we simply offer one approach. Some behaviors, especially those that substantially involve subcortical circuitry, have a large number of relevant isomorphic animal models. With these models in hand, it is possible to use network techniques to detect key modules; the translational goal is to determine which genes within the key module(s) are targets for manipulation and behavioral modification. We argue that the inclusion of the HS-CC or a related population such as the Diversity Outcross 
is key to the translational perspective. The genomic granularity and allelic diversity allowed a reasonably accurate alignment of changes in connectivity with specific genomic regions that segregate with selection (). In addition and only in the HS-CC population, there were eight transcripts with significant differential expression and significant changes in connectivity in the key catalepsy modules (Dataset S6, in File S1
). This process significantly reduced the number of targets identified for manipulating haloperidol response. By definition (because they are members of a key coexpression module), they have a role in haloperidol response.
Another target driven approach is to focus on the “hub” genes. From the scale-free network perspective, hub disruption will have the greatest affect on modular connectivity. It is not necessary that the hub targets be genes that are affected by selection although focusing on these genes provides insight into the causes of natural variation. In the three selection consensus modules, there are a number of hub genes previously shown to significantly affect locomotor behavior and in some cases haloperidol response. In addition to Drd2
(“Pink”), these include Rgs9, Pde10a
(all “Pink”), Tcf4
(“Green”) and Lrrk2
(“Grey 60”) 
. It is also of interest to note that two histone demethylases (Jmjd1a
) are key hub genes (“Grey60” and “Pink”, respectively), perhaps suggesting an epigenetic strategy for modifying haloperidol response.
Overall, the data presented here argue that a systems biology approach is needed, at least in some contexts, to investigate the relationships between genes and behavior. The results from the three independent selections illustrated that at the gene transcript level, neither differential expression, nor differential connectivity in one selection population predicted similar results in another selection population; however at the module level, connectivity changes did overlap. Extracting module-dependent information requires relatively large gene expression datasets, which in turn facilitate an accurate assessment of the covariance structure. This assessment can be improved by substituting RNA-Seq for microarray based strategies