CENIIT PROJECT - Multi-level modelling of diabetes for improved drug development

The project at a glance

CENIIT - the CENter for Industrial Informatics Technologies - is a Linköping based research centre, intended to support junior researchers at Linköping university, who do industrially relevant research in topics related to informatics. In this project, we will take our already world-unique model for type 2 diabetes (T2D) to a new level A) by including crucial non-modelled features, B) by establishing a method to use a new type of large-scale data, C) by improving upon my new methods for identification of unique predictions to also work in large multi-level unidentifiable models. Our model is used by several drug development companies, who co-fund the project.


T2D occurs when the body no longer is able to regulate the blood sugar level. T2D is one of our most costly and rapidly spreading diseases: around 300 million people are afflicted world wide, and around 300 000 of them live in Sweden. For this reason, many of the major drug development companies are devoting large resources to the development of new drugs, which could improve or potentially even cure the disease. Such drugs act on particular proteins inside cells, by binding to them, and thereby modifying their properties. The development of such drugs is complicated, because the proteins inside living cells are part of large inter-connected networks of reactions, and because these networks are not fully understood. In a series of papers, we have undertaken an integrated experimental/theoretical approach to start the process of unravelling this network.

Figure 1: A) Example of our data (dots), where healthy (blue) and T2D (red) signalling can be compared. Overview of our multi-level model. The signal goes from the insulin receptor, IR, through a series of proteins, leading to GLUT4 translocation to the membrane, which leads to increased glucose uptake. The key malfunction that happens in diabetes is depicted with the red inhibition.

Our latest model (Brännmark, 2013) is unique because our data allow us to work in a new way. Our data are unique, because they have been collected in a way where values can be compared between normal responses, and those seen in T2D (Fig 1A). This internal consistency of the data means that we can require that a single model should explain all the data, and that models that can not describe the data should be discarded. In other words, our unique data have allowed us to systematically unravel the network, and test and refine hypothesis regarding the structure of the network (Fig 1B). In particular, we have identified a single-feedback (green arrow) that if malfunctional (red inhibition, T2D), will spread this malfunction, in a way that explains all the observed differences seen between normal and T2D signalling. We have also unravelled how to translate this signalling inside individual fat cells, to the dynamics on the whole-body glucose control level. This new model is is already used by several of the major drug development companies. Nevertheless, the current model still only describes a small fraction of the total number of proteins involved in the insulin signalling network: 20 of 3000.


Modelling of crucial new features: The first of the three sub-projects is relatively straightforward: expansion of the already developed model to include some new previously non-modelled parts. These new parts are important, because they describe the most important functionality in the studied fat cells (adipocytes): the cells' ability to take up and release fatty acids from and to the blood (called lipolysis). This uptake and release is the most important communication between the fat cells and the rest of the body, and the addition will mean that the whole-body predictions of the model will become more realistic. This part of the project is straightforward, because it will not involve the development of any new methods, but simply model-building based on new data, which will be collected by my collaborators at LiU (Peter Strålfors' group) and at AstraZeneca.

Figure 2: Overiview of our new multi-resolution approach.

Methods for a multi-resolution model incorporating large-scale data The second sub-project concerns method developments. More specifically, a new hybrid approach will be developed, which incorporates the strengths of statistically oriented large-scale methods such as LASSO with my more detailed mechanistic small-scale modelling methods. Historically, these approaches have seldom been combined, and there does today not exist a nonlinear ordinary differential equation model inferred from large-scale data, such as proteomics data simultaneously measuring the response of thousands of proteins. Such proteomics data have recently been collected by another experimental group (Humphrey, 2013) and in order to incorporate these kind of data in our models, we need to develop new methods. The main idea behind this new approach is to combine three types of sources, and to create a multi-resolution model (Figure 2). The innermost and most well-determined layer is our already developed model for the 20 already modelled key players. The second layer combines the information in databases of protein-protein interactions, which are used to describe the data. The final layer is for those proteins for which there is no previous understanding: for those proteins the interactions are only inferred from the data.

Methods for prediction uncertainty in large-scale multi-level models My new approach (Cedersund, 2012) which allows me to identify the true prediction uncertainty also for unidentifiable models, is centered around a modified optimization problem. This modified optimization problem put new requirements on optimization algorithms, and the existing methods so far still only works for small-scale problems, with up to around 25 parameters. To be able to obtain truly useful prediction uncertainties, larger multi-level models need to be handled. Our multi-level models have a modular structure, and this structure will be exploited to obtain the uncertainty in each module separately. The transference of the uncertainties between the models will be incorporated using the method in (Schelker, 2012). This idea will be tested, refined, and applied to our multi-level models for T2D.


Brännmark C, Nyman E, Fagerholm S, Bergenholm L, Ekstrand EM, Cedersund G, Strålfors P (2013) J Biol Chem, 288, 9867-80
Cedersund G (2012) FEBS J, 279, 3513-27
Schelker M, Raue A, Timmer J, Kreutz C (2012) Bioinformatics, 28, i529-i534
Humphrey SJ, et al, (2013) Cell Metab., 17, 1009-20

News and updates: