Network signatures link hepatic effects of anti-diabetic interventions with systemic disease parameters

Background Multifactorial diseases such as type 2 diabetes mellitus (T2DM), are driven by a complex network of interconnected mechanisms that translate to a diverse range of complications at the physiological level. To optimally treat T2DM, pharmacological interventions should, ideally, target key nodes in this network that act as determinants of disease progression. Results We set out to discover key nodes in molecular networks based on the hepatic transcriptome dataset from a preclinical study in obese LDLR-/- mice recently published by Radonjic et al. Here, we focus on comparing efficacy of anti-diabetic dietary (DLI) and two drug treatments, namely PPARA agonist fenofibrate and LXR agonist T0901317. By combining knowledge-based and data-driven networks with a random walks based algorithm, we extracted network signatures that link the DLI and two drug interventions to dyslipidemia-related disease parameters. Conclusions This study identified specific and prioritized sets of key nodes in hepatic molecular networks underlying T2DM, uncovering pathways that are to be modulated by targeted T2DM drug interventions in order to modulate the complex disease phenotype.

circadian rhythm genes, in concordance with previous findings that hepatic circadian rhythm gene expression is strongly affected by HFD [1].

Correlation of co-expression to disease parameters
For the 14 modules with a valid eigengene, correlations were calculated with each of the 16 disease parameters measured in the ADT dataset. In total, 10 significant correlations between modules and disease parameters were found (|r| > 0.75, p < 1E-7, Table 1). These correlations include 4 coexpression modules (A, B, C, E) and 4 different parameters (liver weight, atherosclerosis, cholesterol, and triglycerides), with much overlap between these disease parameters (Supplementary figure 3-1, see below).
The disease parameters related to dyslipidemia, which were previously found to be deteriorated after both drug interventions (plasma and intrahepatic triglycerides, cholesterol, atherosclerotic lesion area, liver weight) correlated with multiple co-expression modules. In contrast, although glycemia / insulin sensitivity related disease parameters (glucose, insulin, QUICKI) and obesity (body weight, epididymal fat weight) are fully resolved by T0901317 and partly by fenofibrate [1], no co-expression module correlated with these parameters. This indicates that hepatic target activation determines changes in dyslipidemia rather than disglycemia.

Selection of co-expression modules driven by interventions
Since the co-expression network was generated based on data from all intervention groups (HFD, DLI, T0901317 and fenofibrate) there is a possibility that some interventions determine the correlations within the module more than others. To investigate which modules are driven by which treatments, we performed over-representation analysis of differentially expressed genes (DEGs) relative to the control (HFD) group (Supplementary table 3-1, see below). An enrichment with DEGs for a given intervention indicates that the gene expression in the module is affected by that intervention and thus that the module may be relevant for explaining the link between the intervention target and disease parameter.
Several modules showed a distinct enrichment for specific interventions, such as two modules that were enriched for DLI but not drug interventions (E, G), modules enriched for both drug interventions but not DLI (D, M), and modules enriched specifically for T0901317 (J, L), or fenofibrate (K). Interestingly, several modules were enriched specifically for DEGs of drug interventions and not for DEGs of HFD (D, J, L, M), which indicates the underlying mechanisms are not inherent to the disease induced by HFD but are nevertheless affected by drug intervention. For example, module M is enriched for genes for which the protein products participate in the electron transport chain, which are expected to increase especially after intervention with both fenofibrate and T0901317 [2], as result of increased hepatic fatty acid oxidation. Three modules (H, I, F) were not enriched for any of the diet or drug interventions. These modules also showed very poor correlations with any of the disease parameters, and their gene coexpression may be due to confounding factors not directly related to the disease process.
In addition to over-representation analysis, the DEGs for interventions were split into three groups indicating genes for which the expression is either reversed compared to HFD condition (changed in HFD vs chow and changed in opposite direction for intervention vs HFD), regulated in the same direction, i.e. deregulated compared to HFD condition (changed in HFD vs chow, and in same direction for intervention vs HFD), or specific for the intervention (not changed in HFD vs chow, but changed in intervention vs HFD). This provides insight in which intervention may revert, aggravate, or introduce additional deregulation of module genes.
The results for the three modules with significant GO annotation and correlation to a disease parameter (A, B, C) are shown in Table 2. All three modules are enriched for genes differentially expressed after HFD, which are largely reversed by DLI. This is in concordance with the observed improvement by DLI of the disease parameters correlating with these modules. In contrast, the drug interventions further deregulate nearly all genes in the module that were also regulated by HFD. These opposite effects match the observed deterioration of the corresponding disease parameters by both drug interventions.
In addition, the modules show a large part of additional genes regulated by drugs that were not deregulated by disease, indicating that the drug interventions target or result in different metabolic and immune-related mechanisms than those related to disease progression.
Module B, annotated to immune response and inflammation related processes, was most strongly enriched with DEGs for the T0901317 intervention. Notably, the majority of DEGs (141 out of 146) in the module were upregulated by T0901317 compared to HFD and show the opposite response for DLI where these are downregulated compared to HFD. These genes include genes encoding for several macrophage markers (CD14, CD68, LYZ), and immune cell specific proteins (CD86, CD74, CD83, CD52, CD53, Rac2).

Tables & Figures
Supplementary Figure 3