Fig 1.
Overview of digital twin tool.
(A) Illustration of the idea of a digital twin: use personalised predictions and visualisations to increase the metabolic insights and compliance of a patient to prescribed life-style changes. (B) The physiological and biochemical processes considered herein. (C) Workflow and steps taken in this work. (D) Overview of all reactions and regulations included in the model. (E) Comparison between new digital twin tool and previous physiologically based meal simulation models.
Fig 2.
Model features compared to previous model.
(A) Overview of the qualitative predictions and improvements in new model, compared to previous models. Previous models can only either i) ignore protein ingestion in glucose simulations, or ii) have a static formula for a phenomenological conversion between protein and plasma glucose, which incorrectly increases the hepatic glucose uptake. In contrast, iii) in the new model, we can have a protein-induced increase in gluconeogenesis and hepatic glucose production, which is much larger after fasting than in a fed state. The arrows and their sizes represent the qualitative relationships between the fluxes between plasma glucose and plasma liver during the second OPTT response, for the three model alternatives. (B) Experimental data (error bars) validating another key difference between the new (blue line) and old model simulations (orange line): the decrease of plasma glucose during fasting. (C) New model (blue line) can describe all of the mechanistic flux data (error bars) that led to the ‘Dalla Man model’ [21], equally as well as that model and its subsequent improvements (orange line). The data shows responses to a mixed meal at 0.5 h (black bar). (D) Examples of predictions of key mechanistic variables that new model can produce, which the original Dalla Man model and its subsequent improvements (including [23]) cannot produce. (E) Qualitative aspects of metabolism compared to simulations, where predictions (blue filled bars) are compared to data (grey filled bars) for organ-specific glucose uptake (i, [32]), organ-specific endogenous glucose production (EGP) (ii,[35]), and insulin clearance (iii, [33]). Model uncertainty is represented as error bars on blue filled bars, and error bars on grey filled bars are qualitative reasonable intervals of each respective metabolic flow or reaction.
Fig 3.
Model simulations of the four clinical studies used for parameter estimations.
In all plots, data points are represented as the centre of error bars (SEM) or crosses. Simulations with uncertainty are represented by areas, and the best agreeing simulation is the line, for both healthy populations (blue) and T2D population (red). Quantitative details regarding all studies are summarised in Table D in S1 Text. All of these data are mean responses, and simulations are made for a 175 cm, 75 kg male person. (A) Krssak data [26], which describes a mixed meal response (87 g carbohydrates and 23 g protein) happening at 1 h (black bar), whereafter plasma glucose, insulin, and EGP and hepatic glycogen are measured. (B) Magnusson data [27], which describes a fasting response, following a mixed meal (98.2 g carbohydrates and 26 g protein) which happens 4 h before t = 0. (C) Lerche data [25], which describes a 48 h fast, which starts at t = 0, which is preceded by an overnight fast, and which is followed by an OGTT of 1 g carbohydrate/total bodyweight (kg). In iii) and iv), the time has been shifted, so that the meal occurs at t = 3.5 h. (D) Individual meal from new clinical study, which in this training data describes the glucose response to a mixed meal (81 g carbohydrates and 41 g protein, in total 940 kcal) at t = 1 h (black bar).
Fig 4.
The first validation test done with the model, using the Rothman data [29].
In all plots, data values are depicted with a dot, SEM uncertainty by the error bar, simulation uncertainties are depicted by the areas, and a simulation using the model parameters that best agree with the data is depicted with the line. The first data point (also depicted with an X) is used for personalising the model, and the rest are used for validation. The study monitors the response of hepatic glycogen, and of the contribution of glycogenolysis to EGP, during a 68 h fast. (A) Prediction of hepatic glycogen levels and gluconeogenesis contribution to the endogenous glucose production on the full population. (B) Prediction of hepatic glycogen levels on an individual level.
Fig 5.
Second validation test with the model, using two clinical studies not used for training the model [30,31].
In all plots, data values are depicted with error bars (SEM), simulation uncertainties are depicted by the areas, and the simulation that best agrees with the data is depicted with the dash-dotted line. The first couple of data points (depicted with an X) are used for personalising the model, and the remaining data points are used for validation. (A) Prediction of mixed meal (black bar) from Taylor data [30], i) hepatic glycogen ii) plasma insulin iii) plasma glucose. (B) Prediction of OGTT (black bar) from Firth data [31], i) endogenous glucose production ii) plasma insulin iii) plasma glucose.
Fig 6.
Third validation test of new data testing the model’s ability in predicting fasting intervention and metabolism of proteins.
Model uncertainty is depicted with the area, the calibration data is depicted with an X, the calibrated CGM data are depicted with a dot, and OPTT event is depicted with a black line. (A) Study design. (B) OPTT response during the fed state, before the fast started. (C) Glucose levels during the 48 h fast. (D) OPTT response during the fasting state, after the 48 h fast. (E) Person-specific predictions of non-measured variables connected to protein metabolism.
Fig 7.
Simulations of five common diet schemes and varying different personalized variables.
(A) The model considers diets with different meal frequencies, ranging from Intermediate Fasting (IF) with two meals (depicted as black bars), to 5:2 with 2–3 meals, and even SFM (with three smaller meals). These three diets are isocaloric. (B) The model can also to some extent simulate different compositions, such as LCHF (low carbohydrate, high fat diet), and HCLF (high carbohydrate, low fat diet). (C) Mean values of key variables during the second week after the diet has started. (D) Impact of changing the rate of consumption. E) Impact of changing the body weight by 30 kg, while keeping all other variables constant. All simulations in A-D were done using a male who was 180 cm, 80 kg, and in E that male either weighed 110 or 80 kg.