Driver versus navigator causation in biology: the case of insulin and fasting glucose

Systematic review of experiments involving tissue specific insulin receptor knock-outs

The first step in insulin signalling is the binding of insulin to insulin receptor (Bevan, 2001). The downstream actions of this event finally lead to insulin-dependent glucose uptake in insulin dependent tissues of the body. Experimentally, disruption of insulin signalling is achieved by knocking out or inhibiting various players in the signalling cascade. We chose to look at the effects of knocking out the tissue specific insulin receptor on fasting and post-meal/feeding or post glucose load levels in rodent models. Studying tissue specific insulin receptor knockouts enables us to differentiate between the roles of insulin signalling in different tissues. A classical belief is that the post-meal glucose curve is mainly influenced by the rate of glucose uptake by tissues, mainly muscle, whereas the fasting glucose levels are mainly determined by the rate of liver glucose production (Bock et al., 2007). If this belief is true one expects that muscle specific knockout would mainly affect the GTT curve but may not affect fasting glucose level, whereas liver specific knockout would mainly affect the fasting glucose level. We searched the published literature for experiments in which the fasting and post-feeding levels of glucose are measured in tissue specific insulin receptor knock-outs.

Systematic review of experiments to chronically raise or lower insulin levels

The insulin receptor knockout experiments assume that the main action of insulin is through the specific receptors. It can be argued that insulin acts through other receptors or may have other mechanisms of action yet unknown and therefore receptor knockouts do not fully eliminate insulin action. Alternatively, we can alter insulin level itself to see how it affects glucose level in fasting state or post glucose load. Insulin is known to alter plasma glucose immediately on administration (Galloway et al., 1981), but this is not a steady state response. If insulin levels can be raised or lowered and sustained long enough to reach a steady state, the effect of insulin on glucose in a steady state can be studied. If insulin affects steady state glucose, a sustained rise in insulin will result into a sustained lower steady state glucose level. Conversely a sustained suppression of insulin would lead to higher steady state glucose. We studied published literature for experiments where a stable and sustained increase or decrease in insulin was achieved and then the effect on fasting glucose and GTT studied.

Fasting versus post feeding glucose in STZ rats

Streptozotocin induced diabetes is a popular model of rodent diabetes (Akbarzadeh et al., 2007; Gajdosík et al., 1999). STZ acts by specifically destroying the insulin producing β-cells of the pancreatic islets (Szkudelski, 2001). A low dose of STZ that destroys a substantial population of β cells but does not lead to total destruction of their population is often perceived as a good model for T2D, whereas a high dose of STZ that destroys the β-cell population almost entirely is perceived as a model of T1D (Akbarzadeh et al., 2007; Asrafuzzaman et al., 2017; Freitas et al., 2015; Gajdosík et al., 1999; Kagami et al., 2008; Kovacs et al., 2011; Szkudelski, 2001; Zhang et al., 2008). We searched literature to look for studies that carefully differentiated between steady state glucose from post load glucose in STZ models but did not find any studies that make this distinction clear. Therefore, we designed and conducted experiments to differentially study the steady state and perturbed state glucose levels in rats treated with STZ.

Theoretical and mathematical considerations

In this approach we elaborate on the theoretical underpinnings of insulin-glucose relationship. We also explore possible explanations for the unexpectedly consistent failure of experimental insulin signal impairment to alter steady state glucose level. Simultaneously we make differential predictions from alternative homeostasis models that can be tested in human epidemiological data. The two models which we use here answer the critical question in glucose homeostasis: whether the fasting/steady state glucose level is a consequential balance between glucose production and glucose utilisation rates (consequential steady state CSS) or whether there is a target glucose level that is maintained by sensing and correcting any changes in it (targeted steady state TSS).

Analysis of insulin-glucose relationship in steady and perturbed state in human data: Epidemiological inquiry

Here we use human epidemiological data to test the correlational predictions made by the CSS vs. TSS models of glucose homeostasis.

Increase in insulin

A model for sustained increase in insulin levels is a knock-out or inhibition of the insulin degrading enzyme (IDE). An interplay between insulin secretion and insulin degradation maintains the level of insulin in plasma (Authier, Posner & Bergeron, 1996; Duckworth, Bennett & Hamel, 1998; Hulse, Ralat & Wei-Jen, 2009; Shen et al., 2006). Plasma insulin has a half-life of 4–9 min (Hulse, Ralat & Wei-Jen, 2009; Tomasi et al., 1967) and it is degraded predominantly by the insulin degrading enzyme (IDE) (Hulse, Ralat & Wei-Jen, 2009; Shen et al., 2006). Inhibition of IDE has been considered as a therapeutic option for type 2 diabetes with limited and questionable success (Costes & Butler, 2014; Maianti et al., 2014). We performed a systematic literature review to find out experiments in which IDE was inhibited to obtain a sustained high plasma insulin level and, in such animals, GTT was performed (Table 2) (PROSPERO registration ID: CRD42019140619).

Decrease in insulin

We performed a systematic literature review for experiments in which there was sustained suppression of insulin production. Two insulin suppressing agents have been repeatedly used to lower insulin production in rodent models as well as in humans.

(i) Diazoxide (DZX): Diazoxide is a potassium channel activator which causes reduction in insulin secretion by the β-cells by keeping the cells in a hyperpolarized state by opening the channel (Panten et al., 1989). It has been used as a drug to modulate insulin secretion for research and therapeutic purposes (Doyle, 2003).

(ii) Octreotide (OCT): Octreotide is a somatostatin analogue which inhibits insulin and growth hormone. It has been used to reduce insulin secretion in vitro and in vivo (Lamberts et al., 1996).

We searched the literature systematically for studies where the insulin levels have been altered using either DZX or OCT and glucose tolerance has been examined using a GTT after DZX/OCT treatment (Table 3) (PROSPERO registration ID for DZX: CRD42020141688; PROSPERO registration ID for OCT: CRD42020141464). It should be noted that this literature includes a significant proportion of human trials. We also searched literature for studies in which insulin was suppressed by other methods.

Animal model and conditions

STZ treatment for insulin suppression

Male SD rats weighing 180–200 g were injected with STZ at 50 mg/kg body weight. The STZ was dissolved in Citrate Buffer (Citric Acid: 0.1M and Sodium Citrate: 0.1M). Injection of Citrate Buffer alone was used as control.

Fasting and post-feeding glucose in 12 day follow up

Three days after the STZ injection, the rats were fasted for 16 h and glucose was measured using the hand held Accu-Chek Glucometer. The rats were then given 40 g of Standard Chow for 8 h. Food was weighed and post-feeding glucose was measured after three hours. The protocol was repeated for 12 days and body weight, food weight and glucose readings were taken daily. A total of 12 animals per group were used for this experiment.

Duration of fasting

An experiment was also performed to see how much time was required to reach a steady state of glucose after removal of food. The food was removed from the STZ and Control animals after ad libitum availability and glucose readings were taken after 3, 6, 9, 12 and 16 h. After a recovery of three days, glucose levels were measured only at 16 hours after removing the food. 9 STZ treated animals and 10 Control animals (injected with Citrate Buffer) were used for this experiment.

Choice of models for glucose homeostasis

The fasting state has been generally accepted to be a steady state for glucose concentration for several reasons. In a given healthy individual the fasting glucose levels are stable in time (Halter et al., 1985; Lerner & Porte, 1972). The post-meal peak of glucose and insulin returns to the fasting level within a few hours and remains stable over a long time. The fasting state is considered and modelled as a steady state by the widely used HOMA model (Matthews et al., 1985; Turner et al., 1979). Classically the negative feedback loops are assumed to work through insulin and insulin is taken as a determinant of steady state glucose level. Most popular models of glucose homeostasis work on this assumption although non-steady state models of insulin resistance exist (Palumbo et al., 2013).

Models used for this analysis

A critical question in glucose homeostasis is whether the fasting steady state glucose level is a consequential balance between glucose production and glucose utilisation rates (consequential steady state CSS) or whether there is a target glucose level that is maintained by sensing and correcting any changes in it (targeted steady state TSS). The difference in the two can be visualised by the tank water level analogy (Fig. 2). If a tank has an input tap releasing water in it at a constant rate and has an outlet at the bottom through which water escapes proportionate to the pressure of the water column, a steady state is invariably reached (Fig. 2A). The steady state level is decided by the rate of intake and the size of the outlet. This is a CSS which will change with any change in the size/capacity of the input or the outlet tap. In contrast to CSS, in a TSS there is a desired water level and sensors are placed above and below the desired level such that when the level goes below the lower sensor the input is switched on or its rate increased and/or output switched off or its rate decreased (Fig. 2B). Refer to Supplementary Information 2 for the details of the CSS model.

Epidemiological data

The three data sets used here come from two different studies: (i) Coronary Risk of Insulin Sensitivity in Indian Subjects (CRISIS) study, Pune, India (Yajnik et al., 2007) and (ii) Newcastle Heart Project (NHP), UK (Bhopal et al., 1999). Data from the latter is divided into two groups as the subjects belong to different ethnicities namely European white and south Asian and we will prefer to analyse the two groups separately since certain ethnic differences are likely to be present in the tendency to develop metabolic syndrome (Bhopal, 2013; Gujral et al., 2013). Hence all the comparison of predictions with the data has been done independently for the three data sets. All the studies are population surveys that include non-diabetic (fasting glucose values less than 110 mg/dl) and diabetic individuals (fasting glucose values above 110 mg/dl) and the clinical history, morphometric parameters, glucose and insulin during fasting and oral glucose tolerance test (OGTT) of the subjects were recorded. In the analysis, we included only the non-diabetic groups in which the homeostatic mechanism can be assumed to be intact and therefore any hypothesis about it can be tested. Most of the individuals in the diabetic group would be under different drug regime affecting glucose-insulin dynamics in different ways and therefore we exclude that group for the analysis.

Statistics

Linear regression and correlation were used to compare the glucose-insulin relationship in steady state (fasting) versus perturbed state (post-glucose load) in the three data sets along with the relationships between HOMA-IR and HOMA-β derived from the fasting data.

Pooled results of IRKO experiments

We shortlisted 16 articles describing 46 independent experiments in which glucose tolerance curves of insulin receptor knock-outs and controls were compared (Table 1 of Supplemental Information 1). The experiments could be segregated in four different tissue specific knockouts for the analysis: fat/adipose insulin receptor knock-out or FIRKO (Fig. 3), Muscle insulin receptor knock-out or MIRKO (Fig. 4), liver insulin receptor knock-out or LIRKO (Fig. 5) and β-cell insulin receptor knock-out or βIRKO (Fig. 6). A generalised trend in the total picture summed up over all four IRKOs was that along the GTT curve, significantly higher glucose levels are seen in the knockouts as compared to the controls, particularly and consistently at 30, 60 and 120 min. However, the fasting glucose level was not significantly different. In some studies, fasting glucose was significantly higher in the knock-outs than the controls, however in some other studies it was significantly lower as well. In 29 out of 46 experiments there was no significant difference (Table 4) in the fasting glucose levels of knock-outs and controls. This trend was consistently seen in MIRKO (Fig. 4), LIRKO (Fig. 5) and βIRKO (Fig. 6). Only in FIRKO (Fig. 3) there were greater number of studies showing fasting glucose significantly higher in the knock-outs than in the controls(Guerra et al., 2001; Sakaguchi et al., 2017; Softic et al., 2016), but in the non-parametric analysis the collective trend was not significant Table 4. These inconsistencies in the FIRKO fasting glucose levels compared to the MIRKO and LIRKO could be explained with the help of the duration of fasting used for the glucose tolerance tests. Although the glucose levels in the wild type are higher than the FIRKO/BATIRKO in the fasting conditions, there could be possible reasons for that. For example, in the study (Sakaguchi et al., 2017), the fasting for the glucose tolerance test was carried out only for 6 hours as compared to 16 hours/overnight in other studies. Secondly, one of the knock-outs used in this study is a double knock-out of insulin receptor as well as the insulin-like growth factor 1 receptor which could be a possible reason for the higher glucose levels in the fasting condition. In case of the Guerra et al., 2001; Softic et al., 2016, the animals have been fasted ON for the glucose tolerance test and the tests have been performed on FIRKO and WT of different ages. The impairment of glucose tolerance increases with age, though this also is not seen consistently across all the studies. In the case of Blüher et al. (2002), the fasting duration for the GTT is 16 h, highest in all the studies and in this case the treated glucose levels in the fasting condition are equal to or lower than that of the controls. It is important to note that in FRIKO, although there are maximum number of cases where in the treated is higher than the control in the fasting condition, this trend is not significant in the non-parametric analysis (Table 4). Also, only in FIRKO, the 30, 60- and 120-min glucose was not significantly different in the knock-outs than the controls (Fig. 3). It is notable that in none of LIRKO experiments the fasting sugar was significantly higher than the controls. This contradicts the classical belief that liver insulin resistance is mainly responsible for fasting hyperglycaemia in T2D (Bock et al., 2007; Johnson et al., 1972).

A possible problem in comparing fasting glucose across different studies was that different fasting intervals have been used ranging from 4 to 16 h. No study clearly reported how much time is required to reach a steady state in a knock-out. In 10 of the experiments in which fasting time was reported as 16 h, none had fasting sugar significantly different from controls. In the 13 experiments in which it was high, the fasting duration was between 4 and 12 h or not precisely reported. Therefore, it is likely that in at least some of the experiments, glucose steady state was not yet achieved at the time point defined as fasting. This bias increases the probability that higher fasting glucose is reported for the knock-outs. However, since we do not see a significant difference in the collective analysis, the inference that IRKO does not alter fasting glucose is unlikely to be a result of the bias. In fact, any possible correction to the bias might further reduce the apparent residual difference. Therefore, in spite of some inconsistency across studies, a robust generalisation is that IRKOs have significantly increased plasma glucose over controls at 30–120 min post-glucose load but they do not appear to affect steady state fasting glucose. The time required to reach the steady state is evidently increased.

Increase in insulin by suppression of IDE

We found six publications that described 18 experiments that allowed comparison of GTT between raised insulin groups and control group (Table 2 of Supplemental Information 1). Analysis revealed no significant difference in the fasting glucose. In only one out of 18 experiments the treatment group had lower fasting glucose than the control. During the GTT curve, at 90 and 120 min the difference between treatment and control were significant but in the opposite direction of the expectation (Table 5; Fig. 7). While rise in insulin level should reduce plasma glucose, it increased in 15 out of 18 studies, two of which were individually significant and the difference was significant in non-parametric collective analysis (Table 5; Fig. 7). Across all time points along the GTT, the plasma glucose in the treated group was greater than the control group in majority of the experiments. Thus, in this class of experiments increasing insulin failed to reduce glucose at the steady state as well as post glucose load.

Decrease in insulin: suppression by diazoxide or octreotide

We found eight articles describing 14 experiments for diazoxide treatment and 10 articles with 15 experiments for octreotide treatment (Tables 3 and 4 from Supplemental Information 1 respectively). It can be seen from Table 6 and Fig. 8 that for both insulin suppressing agents, suppression of insulin did not result into increased fasting glucose. Further at 120 min post glucose load there was a marginally significant rise in glucose in the insulin suppressed group as compared to control group. This demonstrates that pharmacological suppression of insulin was unable to raise plasma glucose level in a fasting steady state. There was somewhat inconsistent but significant rise post glucose load. Diazoxide is the only agent in which both rodent and human data are available. The results of the two are very similar and separating them does not alter the significance.

We found more means of insulin suppression in which GTT after suppression was reported, but there were not many published replications of the experiments coming independently from different research groups. Therefore, any systematic review was not warranted. We briefly review their results here.

Suppression by protein restriction

Dietary protein deprivation is another method of insulin suppression. This also led to a decrease in plasma insulin levels; however fasting glucose levels did not increase (Schteingart et al., 1979).

Suppression by insulin siRNA

Transgenic mice for insulin-siRNA along with IDE overexpression, showed decreased levels of insulin. Again the fasting glucose levels remained normal while there was a change in glucose tolerance curve (Fig. 9) (Hwang et al., 2007). The curves in Fig. 9 are typical of insulin receptor knockout or insulin suppression experiments where, in individuals with impaired insulin signalling the glucose peak is higher which returns to steady state much later than the controls, but the fasting steady state level is not different.

Suppression of insulin by partial gene ablation

In rodents, there are two insulin genes Ins1 and Ins2 (Duvillie et al., 1997). A double knock-out of both the genes results in death, but ablation of either of the genes does not alter the glucose tolerance significantly suggesting redundancy (Mehran et al., 2012). There are studies in which one gene is completely knocked out and the other one is a heterozygote (Dionne et al., 2016; Mehran et al., 2012; Page et al., 2019; Templeman, Clee & Johnson, 2015; Templeman et al., 2017). Reduced insulin gene dosage did not consistently result into fasting hyperglycaemia in these studies although it offered protection against some of the effects of hyperinsulinemia.

Testable predictions from the models

It is possible to make other testable predictions of TSS and CSS models. In the normal healthy individual, increased glucose utilisation is expected to decrease fasting glucose levels by the CSS model but not by the TSS model. Human experiments have shown that sustained exercise does not reduce plasma glucose, in fact it might increase (Coggan, 1991). In order to match with experimental data, CSS based models of glucose dynamics during exercise need to include additional terms which involve neuronal mechanisms such as direct stimulation of liver glucose production in response to exercise through sympathetic route (Roy & Parker, 2006). This brings the model close to a TSS model. If TSS model describes glucose homeostasis more appropriately, reduced insulin signalling is not expected to change steady state glucose but only alter the time course to reach a steady state.

The mechanism of attaining a hyperinsulinemic normoglycemic prediabetic state is different by the CSS and TSS models. By the classical CSS based pathway, obesity induced insulin resistance is believed to be primary. The insulin resistance reduces glucose uptake and the excess glucose triggers a compensatory insulin response. The resultant hyperinsulinemia compensates for insulin resistance keeping the fasting glucose levels normal. Detailed analysis of the model and matching its prediction with empirical data has refuted this model (Chawla et al., 2018). One of the intuitively appealing reasons for this refutation is that after the heightened insulin levels normalise glucose, there is no reason why insulin levels remain high. Therefore, a steady state with hyperinsulinemia and normoglycemia is impossible by the CSS model but it exists in a prediabetic state. If a “compensatory” insulin response is mediated by glucose, one would expect a positive correlation between fasting glucose (FG) and fasting insulin (FI) and no correlation between insulin resistance and β-cell responsiveness.

By the TSS model, on the other hand, compensatory response is possible in either way. Primary insulin resistance may increase the glucose levels transiently, but when glucose sensing mechanisms detect the change a compensatory response can be operational. By this mechanism a hyperinsulinemic normoglycemic state is possible. Alternatively, primary hyperinsulinemia (Corkey, 2012; Garvey, Olefsky & Marshall, 1986; Shanik et al., 2008; Weyer et al., 2000) can also be compensated by increased insulin resistance by hitting the lower level of sensing which would trigger compensatory insulin resistance. Even in this case a hyperinsulinemic normoglycemic state is possible. Both glucose sensing neurones and neuronal regulation of insulin release and liver glucose production are well known. In the compensatory response mediated by TSS pathways there need not be a correlation between fasting insulin and fasting glucose, but insulin resistance and β cell response would be correlated.

Also using a simple CSS model (see Supplemental Information 2 for details), simulations show that the correlation coefficient and regression slope in the insulin-glucose relationship would remain the same in the fasting as well as post-meal state although the range of glucose and insulin levels will be different. On the other hand, in a TSS model the post-meal glucose and insulin levels are expected to be correlated but the steady state levels may not. We test these predictions by the two alternative models using human epidemiological data below.

We argued above that since on impairment of insulin signalling, the time required to reach a steady state can be substantially longer, overnight fasting may not ensure a steady state in all individuals. Fasting hyperglycaemia in T2D can have two alternative (but not mutually exclusive) causes. Either it represents the failure to reach a steady state in the specified fasting period, or it is because of mechanisms other than reduced insulin action. The TSS model can make differential predictions from the two alternative causes since it predicts a positive correlation between plasma glucose and plasma insulin in the post-meal state but loss of this correlation on reaching a steady state. In population data, if some individuals have reached a steady state but a few others haven’t we would expect a correlation significantly weaker than the post-meal correlation. These predictions can be tested in epidemiological data.