Summary of main findings
Herein, we have presented a first multi-level, multi-timescale, and mechanistic model of the progression of insulin resistance in humans. The model describes insulin resistance development on three different biological levels: whole-body composition (Fig. 2a, 3b, c, 4a, and 5c), plasma glucose and insulin (Fig. 2b, 3d, 4b, and 5d), and intracellular adipocyte insulin signaling (Fig. 2c, 3g, 4d, e and 5e). The model agrees with the multi-level dataset from the Fast-food study (18, 19), both describing estimation data (Fig. 3b) and correctly predicting independent validation data (Fig. 3c, d, g). For this weight-increase study, we predict traditional biomarkers which had not been measured, such as oral glucose tolerance test and other intracellular signaling intermediaries (Fig. 4). The model also agrees with whole-body weight-loss data from the Topiramate study (Fig. 5b, c) (16). Moreover, for this study, we use the model to predict changes in a glucose tolerance test and intracellular insulin signaling (Fig. 5d, e). Finally, we illustrate how this model potentially can be used to improve health in future eHealth technologies (Fig. 6).
New strengths and possibilities with our new modular and multi-scale model for insulin resistance
An important strength of this model is that it combines three well-determined and validated models into an interconnected multi-scale model. Having a multi-scale model is an important strength since the progression of diabetes in reality is multi-scale, as seen in the data (Fig. 3). Despite this importance, there existed no previously available multi-scale model that could describe such data. Nevertheless, there exists models that describe the different levels separately: the Hall model for whole-body weight describes changes over months and years (3); the Dalla Man model for the meal response describes the interplay between plasma glucose and insulin (4); and the Brännmark model that describes intracellular insulin signaling data in adipocytes (6, 7). However, these three models had previously not been connected into a single model, in part because the arguably most central connection between them – adiposity-driven insulin resistance – had not previously been modelled. Herein, we have for the first time connected these three well-established models and levels into a multi-scale model, by introducing a new model for the progression of insulin resistance. This is the first such human, multi-scale insulin resistance model. The connecting kit, the adiposity-driven insulin resistance model, has been adopted from a corresponding multi-scale model for mice (21), even though the three constituent models for the three levels and timescales, come from existing models that were specific to humans. Moreover, the cell level model has also been adjusted to allow a continuous development of insulin resistance, by allowing some of the model’s steady states to instead change over time. A final important aspect of this multi-scale model is that it is modular, meaning that the different subsystems and organs described in the model can be changed to other models with more or less details (22, 23).
Another strength with our new model is that it can describe not only estimation data, but also correctly predict independent validation data, and can thus also be used to predict non-measured variables. The model correctly describes estimation data of weight change from both the Fast-food study (Fig. 3b) and the Topiramate study (Fig. 5c). Furthermore, the model also correctly described independent validation data from both these studies. For the Fast-food study, the model describes independent data on changes in fat mass and fat free mass (Fig. 3c), fasting glucose and insulin concentrations (Fig. 3d), and the insulin response of the intracellular signaling metabolites (Fig. 3e). In all these predictions, the model only changed one parameter: the scale difference between mice and humans in the insulin resistance model. For the Topiramate study, the model describes independent data for weight change using a dosage of topiramate that was not used for fitting, 96 mg/day (Fig. 5d ). Because of the success of these validation tests, we then used the model to make predictions of the gradual changes of some of the things that were not measured in the original study. For example, we could predict how the glucose levels and fluxes change during the studies, as well as how intracellular signaling is changing. These kinds of predictions are something that the earlier model cannot do, since they require the interplay between the different layers. These predictions of additional non-measured variables can in principle be tested by doing new studies where these variables are measured, and this could either validate the current model even further, or reject the model, and both these outcomes would provide new mechanistic insights regarding the progression of insulin resistance.
Limitations with our model
The current version of the model has some limitations. One such limitation is that the implementation of the fat-dependent insulin resistance is a minimal model, using relatively simple expressions. Specifically, the model lacks relevant details and hypotheses assumed to be involved in insulin progression. One such mechanistic hypothesis is ectopic fat storage and inflammation in liver and pancreas (24, 25). Inflammation is also often believed to play a role in the adipose tissue itself, as is the varying cell size distributions of adipocytes (26, 27). These things could be included in future, more detailed versions of the model. However, all of these are processes that are not covered by the model’s current level of detail, and among processes currently included, the progression of insulin resistance is mechanistic, in the sense that it affects the right included mechanisms. For instance, the endogenous production of glucose from the liver is known to be impacted by insulin resistance, and this impact is included, even though the underlying mechanisms for this impact are not included. To include such underlying mechanisms would allow us to simulate a wider array of drugs, including e.g. anti-inflammatory drugs like cd44-inhibitors (28, 29), or drugs that influence the size of adipocytes like metformin (30). These potential additions could thus be useful for both drug development and individualized prevention. Apart from this lack of mechanistic detail, the current implementation of insulin resistance progression is given by a logarithmic expression (Eq. 153–155,160–162 in supplementary material). This expression thus excludes potential transient and/or adjustment processes in the body. Also, the current progression of insulin resistance has only been validated on a relatively small weight span and population, meaning that higher or lower weight changes and other time scales might not be accurately represented by the model.
Another potential limitation with the current model concerns how the interconnection was introduced. Specifically, the interconnection is top down only – the whole-body level only influences the organ/tissue level only goes in one direction, that is from the top-level (whole-body) to the lower level (organ/tissue/cell), and is not reversible. This implementation of the connection means that the meal response or meal response dynamics does not affect the whole-body composition changes, which, in reality, it does. A future implementation of the interconnection could describe how short-term changes in meal response dynamics would lead to short-term changes in ectopic fat storage, which over time would lead to long-term changes in fat mass, and therefore also overall body weight. To implement such a two-way interconnection between the levels, the model should represent fat tissue in greater detail, including e.g., proliferation and death of adipocytes, the effects of differently sized adipocytes, the amount of fat in each adipocyte, and ectopic fat storage (26, 27, 31, 32). Other more realistic interconnections include for example different hunger and fat-mass regulating hormones (such as leptin, adiponectin, various inflammation mechanisms, intracellular mechanisms on more organs than fat tissue), as well as the interplay between glucose, proteins, and fat (15, 33–36).
A third potential drawback is that the model is heavily focused on adipose tissue and the adipocytes, and their involvement in insulin resistance. This adipocentric explanation is one of the most popular ones for the progression of insulin resistance, but not the only one. Other explanations do exist, such as various genetic explanations, inflammation in other organs and/or due to ectopic fat storage (26, 26, 31, 32, 37–39). It is also possible that there are several mechanisms leading to insulin progression that are true at the same time or for different clusters of people with insulin resistance. There are also at least some evidence for the existence of a range of different diabetic subtypes (40, 41), and that there are also different possible pathways to diabetes and insulin resistance (9). Ideally, all different hypotheses should be implemented and compared.
Future applications of our multi-scale model: digital twins, eHealth, and drug development
The multi-scale model presented herein is a so called passive digital twin. A passive digital twin is, in contrast to an active digital twin, not personalized using individual data, even though it could be. Both active and passive digital twins can be useful in an eHealth scenario. Passive twins can for example be used to describe general dynamics of disease progression and be used as a medical pedagogics tool. For example, when looking at the progression of insulin resistance, the model can show how an increased energy intake can result in a weight increase, and eventually also to progression towards insulin resistance and type 2 diabetes. To simulate such illustrations of the effect of daily habits could both help to convey medical knowledge in a comprehensive way and increase motivation to making life-style changes. Active digital twins can also help with medical pedagogics and motivation, but with the additional benefit of being able to make personalized predictions. Such predictions could potentially also be used to help motivate patients to adhere to prescribed drugs or to more stringently follow their prescribed diet and exercise-schemes. Furthermore, mechanistically based, multi-scale models for the progression of insulin resistance and type 2 diabetes could potentially also be used to evaluate different care interventions. For example, when using weight loss as a prevention or treatment for diabetes, a digital twin can be used for comparison of different options - topiramate could be compared to other interventions, both by comparing the effects on weight loss and other relevant biomarkers. All of these potential applications of a digital twin could be further increased by connecting the digital twin with a machine learning risk model-based drug development, and systems pharmacology, creating a hybrid model. This hybrid model could then be used to calculate a personalized or general risk for different diseases, like diabetes or cardiovascular diseases, given a certain scenario simulated by the digital twin. Then, when comparing different weight-loss drugs, their relative effect on the risk of disease could also be compared (42, 43).