We developed a novel network model that integrates genetic, transcription and protein-protein interaction information to pinpoint App as a key insulin regulatory molecule in pancreatic islet tissue. The computational model we developed has several unique features.
Instead of pursuing cis-regulating genetic factors, it focused on networks of genes that were trans-regulated. The goal was not to identify the genetic factors whose variation at DNA level would lead to changes in circulating insulin. Instead, the model identifies networks of genes showing transcriptional changes as result of variation in the genetic factors. This is based on the assumption that the disease phenotype is at least partially mediated by these transcriptional changes. Genes identified by this approach could also have a more direct link to the disease phenotype compared to the upstream genetic factors. The model also simultaneously considered multiple loci, which enabled the study of the interactions between trans-regulated gene modules. As it is extremely common for complex disease phenotype traits to map to multiple loci, it is clear that we need models considering the joint effects of multiple loci. Ideally such models should not only be meaningful in the mathematical terms, but also provide biological insight to the possible mechanisms. Although the linear regression model indicated a joint regulation of the insulin trait, it did not generate any hypotheses on how the joint regulation occurred biologically.
Compared to other network models, such as co-expression network 
, ARACNE 
, and Bayesian network 
, which focus on grouping co-expression of individual genes, our method focuses on dissecting potential mechanisms of integrating information from multiple co-expression modules. By considering the protein-protein interactions across the two groups of genes, it is possible to actually identify potential molecular mechanisms involved in joint regulation. Although currently the protein-protein interaction dataset we compiled may be rather incomplete, hundreds of genes were connected by these interactions. This makes prioritizing genes for experimental validation a more important task compared to finding out what could have been missed due to incomplete protein interaction information. To prioritize the key nodes in the disease network, we developed the novel scoring system in the context of the protein interaction network. As we posit that proteins their function by interacting with their neighbors in the network, the TIE score gives a weighted estimation on how strongly the intensity of these interactions correlates with the phenotype. A gene with high TIE score suggests that the intensity of its interactions strongly correlates with the phenotype based on large numbers of interactions. Therefore, the gene is likely to regulate the trait. By integrating genetic, gene expression, and phenotypic trait information, the ranking algorithm identified biologically meaningful candidate insulin regulators.
A previous publication has shown that, compared to wild-type mice, whole-body App
knockout mice (App−/−
) have elevated insulin secretion in response to an intravenous glucose injection 
. A recent study of the cross of App
transgenic mice and T2D predisposition mice shows that increased Aβ production impairs insulin signaling and accelerates insulin resistance 
. To our knowledge, however, no other studies have demonstrated a direct effect of APP on islet function. Given that App
is highly expressed in pancreatic islets 
, we sought to determine if the changes observed in App−/−
mice reflect direct or indirect effects of App
on islet function. Our measurements of glucose stimulated insulin secretion in isolated islets from App
KO mice confirms our network analysis and is also consistent with the causality test 
which also indicates App
as a causal gene in pancreatic islet tissue (Figure S7
). The model demonstrates that App
is under the regulation of multiple genetic loci, and may function as an integrator for these perturbation signals, mediating interactions between two distinct gene sets that share a common genetic architecture with plasma insulin.
We have previously shown that Lepob/ob
mutation exposes a strain-dependent difference in diabetes susceptibility between BTBR and B6 mice 
. In the current study we exploited this difference and used it as a “sensitized screen” to genetically map genes and diabetes-related clinical traits that may underlie this difference. This approach allowed us to identify App
as a key negative regulator of insulin secretion from pancreatic islets. In this study, we compared wild-type and App−/−
mice to test for a direct role of App
in insulin secretion in mice not expressing the Lepob/ob
mutation. In these studies, the loss of App
resulted in enhanced insulin secretion, consistent with the strong negative relationship between islet App
and circulating insulin across the F2 samples. These results suggest that while leptin deficiency was critical in revealing the islet network involving App
and circulating insulin, it was not required to demonstrate the direct role of App
in insulin secretion.
Our results, which demonstrate a difference in insulin secretion between islets collected from wild-type and those collected from App−/−
mice at 17 weeks, but not 10 weeks of age, implies an age-dependence for the role that App
plays in the islet. However, studies in mouse 
and human islets 
have not reported an age-dependent change in App
expression. It is possible that proteolytic processing of App
mediated by the beta- and gamma-secretase enzymes, or other forms of post-translational modification, are necessary for App
to regulate insulin secretion.
Mouse and rat beta cells are more sensitive to oxidative stress than human beta cells 
, due to the relatively higher expression of antioxidant enzymes in human beta cells 
. We showed that the sub-network regulating plasma insulin level variation () is enriched for GO categories “neuron projection” (p
), “extracellular space” (p
), and “hormone activity” (p
). Genes involved in the stress response process are not enriched in the subnetwork. Recent RNAseq data 
suggests that APP
robustly expresses in human islet cells. In addition, it has been shown that aggregated amyloid-β peptide as well as other proteins have been detected at higher levels in pancreatic islets of T2D patients comparing to healthy control people 
. These suggest that the subnetwork and key regulators in mouse islet we identified in the F2 cross are expected to be relevant in human islets. Our findings support the hypothesis that APP contributes to the common pathogenesis of AD and T2D 
For the future development, (1) a generalized multi-way interaction model is needed to capture complex interaction networks underlying complex traits such as plasma insulin; (2) additional experiments are needed to systematically validate candidate genes (such as genes in Table S5
and genes connected to App
in ) for their roles in affecting β-cell function which in turn affect insulin production and insulin secretion; (3) the molecular mechanism of age-dependent App
regulating insulin secretion is warrant further study.
In conclusion, using an integrative analysis of gene expression, genotypes, and phenotypic traits of the B6xBTBR ob/ob F2 cross, we showed that plasma insulin is modulated by the variation of multiple genetic factors, presumably through expression changes of hundreds of genes in multiple tissues. Our approach focused on revealing the underlying disease network across loci and tissues. The model predicted that App acts in pancreatic islets to affect plasma insulin. This prediction was tested in isolated islets where the knockout of App was associated with increased insulin secretion. Considering App is known for Alzheimer's disease development and a strong association between T2D and AD, our findings point to a potential mechanism through which these two diseases are linked.