The most significant hits identified in this study consist in three candidate loci related to cardiac and body weight, three loci for electrocardiographic (ECG) parameters, two loci for the susceptibility of atrial weight index to iso
, four loci for the susceptibility of SBP to perturbations of the β
-adrenergic system, and one locus for the responsiveness of QTc. An additional 60 loci were suggestive for one or the other phenotype, while 46 were suggestive for one or the other effect of the treatments. Apart from the many apparent correlations with data extracted from the literature and other databases, we are confident that the mapped loci are reliable candidates for two other reasons: first, the hits suggestively associated with cardiac-related traits tagged genomic regions preferentially enriched in genes expressed in the heart (p
0.004); second, a significant fraction (23%; p
0.0003) of the testable candidate markers was replicated with independent publicly-available datasets. To discuss these results further, the following points need to be considered.
One of the major challenges of association studies in general and of in silico mapping in particular is to discern true from false positives without losing (too many) false negatives. By using trait values of eight to ten biological replicates for each phenotype as well as applying a stringent filter to control for missing alleles, conservative thresholds to control for multiple hypotheses testing, and EMMA to correct for population structure and genetic relatedness of the mouse strains, we greatly reduced the number of spurious positive associations. Quality controls, in particular QQ-plots and careful inspections of the patterns of allelic distributions at the top hits further restricted the number of likely false positives.
A significant fraction of the loci identified for the 27 phenotypes was not specific of trait values measured under a single drug treatment but often overlapped across several if not all scans produced for the given trait. This was concordant with our previous observation that robust phenotypes (i.e. those relating to body and to a lesser extent cardiac weight) were marginally affected by ate
. In contrast, most of the loci for ECG values were significant only with respect to a specific drug treatment.
Since the phenotypes were measured in separate cohorts for each drug treatment, individual values of the pharmacological effects of ate
were not available but had to be inferred. This said, it is important to emphasize that measuring pharmacological effects in hundreds of individuals may not be easily implementable for invasive phenotypes such as heart weight. One may argue that the approximations are adding experimental noise to the mapped loci, potentially increasing the number of false positives. However, we note that the vast majority of the hits associated with the effects of iso10 vs ate
treatments (i.e. those for which the values of trait variance were often the largest) were also retrieved when analyzing the reverse ate vs iso10
combinations. Had the computed values been noisy, then these results would have been difficult to replicate. Notwithstanding these limitations, our design produced an attractive list of candidate genes. The identification of at least two members of the hypothalamic-pituitary-thyroid axis as putative modulators of the responsiveness of SBP to perturbations of the β
-adrenergic system appears particularly relevant. Indeed it seems rather unlikely that such a result would emerge by pure chance. As mentioned in the corresponding chapter, the underlying markers seem to discriminate between “SBP-responsive” and “SBP non-responsive” strains for the considered ate
treatments. However, we reported earlier that most of the 22 mouse lines reacted to a ten-fold lower dose of iso
) by reducing SBP 
. Therefore, none of these strains can be considered as a general non-responder. Additional investigations will be required to dissect the underlying mechanisms further.
Unlike other β
undergoes little or no metabolism in the liver and is primarily eliminated unchanged by the kidneys 
. In contrast, iso
is metabolized within minutes into inactive metabolites by the liver catechol-O-methyltransferase (COMT) 
. Since the 22 strains carry a single, identical Comt
haplotype, the pharmacokinetics of both ate
is expected to be only marginally influenced by genetic variation of metabolizing enzymes. This seems well supported by the absence of genes involved in drug metabolism in our list of candidates. Along similar lines, we note that the 22 strains carry the same Adrb1
coding for the main subtype of β
-adrenergic receptors in the heart.
On the principle, it is well acknowledged that computational association mapping in inbred mice provides high-resolution mapping. However, the power of this strategy greatly depends on the size and genetic diversity (recombination) of the chosen strains, the complexity of the trait studied, and the frequency of the underlying variants 
. A panel such as our GRP is well suited to detect loci of strong genetic effects for traits of moderate complexity but has little power to uncover variants with small effect size. This is well illustrated by the relatively modest number of hits identified, that rarely exceeded a single putative locus in any genome-wide scan. Our attempts at analyzing modular phenotypes, that combined multiple traits and/or treatment conditions, did not improve this general outcome (data not shown). Thus, not only the measured traits but also the effects of the drugs are likely governed by a complex architecture of genetic variants. Our results also fit with the notion that the whole strategy is well designed to map common variants, with MAFs of the top markers typically ranging from 0.3 to 0.5.
High-resolution mapping of the mouse phylogeny recently established that the genomes of classical inbred lines were overwhelmingly of Mus musculus domesticus
origin, with limited contributions from Mus musculus musculus
and Mus musculus castaneus
. In the present set of strains, this translates into genomes being on average of 94.5% M. m. domesticus
origin, while the remaining fraction is composed of an estimated 5.2% M. m. musculus
and 0.3% M. m. castaneus
haplotypes, respectively. There is also a strong bias toward multiple lines sharing the same M. m. musculus
haplotype in some regions 
. Reduced genetic diversity precludes the identification of the majority of the genetic determinants existing in natural populations and likely accounts for a large part to the moderate number of loci identified. Conversely, it may further explain the apparent large portion of the phenotypic variance explained by our candidate loci, even for otherwise naturally complex traits.
At this stage, it would be unreasonable to expect that the identified loci are the causal variants and thus should be confirmed by genotyping. Indeed, the mouse HapMap (ca 80¢000 SNPs) does not contain all of the genetic variation between the strains, which, according to the latest figures, amounts to ca 4 mio SNPs per classical inbred genome 
. Because linkage disequilibrium (LD) regions in these mice span in the range of a megabase or more 
, it is likely that hundreds of linked variants exist in any implicated region. For similar reasons, discussing the identity of putative candidate genes in more details than what is provided with the results would be too premature and clearly speculative.
Even in inbred mice, mutations naturally occur at a certain rate. For the reasons mentioned above, we did not perform confirmatory genotyping and thus cannot formally exclude that one or the other marker mapped in the phenotyped animals differs from that reported in the HapMap database. While chances that such events would precisely affect the reported candidates are low, we also reasoned that the amount of genetic differences between a phenotyped animal and the genotyped animal of the same strain is anyway several orders of magnitude smaller than the genetic differences between the strains. Moreover, since many of the reported loci are supported by more than a single SNP, it seems unlikely that multiple mutation events (or genotyping errors) would simultaneously affect multiple consecutive variants in any given strain. For these reasons, we believe that natural mutation events have little impact on our general conclusions.
In summary, we have three complementary pieces of evidence in favor of the correctness of our results: first, looking at external data overlapping with our studies, we generally observed consistent behavior. Second, among the candidate genes associated with the cardiac-related traits, there is a significant enrichment for genes expressed in the heart. Third, we have accumulated a number of associations pointing to genes that have previously been implicated in cardiac functioning or whose molecular function makes such a role very plausible. Nevertheless, it is important to emphasize that further independent studies will be needed to replicate and refine the loci we identified, to map the underlying causative genes and variants, and to functionally validate them. Future experiments may include more comprehensive expression profiling across the strains and treatments, as well as properly designed replication studies. To improve statistical power and mapping resolution, one may also consider complementing the panel of inbred strains with well-chosen informative lines such as those of the Hybrid Mouse Diversity Panel 
, the Collaborative Cross 
or wild-derived strains 
Knowing which genes are involved in drug response is a key step to understand the biology of the measured traits and associated diseases. Translating these findings into humans can potentially provide valuable insights into the nature of the genes and pathways that modulate inter-individual response variability to therapies with β-blockers or β-agonists. A more detailed understanding of these determinants and their relationships will ultimately lead to new or improved strategies for personalized treatment in patients suffering from cardiovascular disorders.