in room "Poster 2" on gather.town

1 - A computational model of the competition at the cell-extracellular matrix interface
Zeynep Karagöz, Thomas Geuens, Vanessa LaPointe, Martijn van Griensven and Aurélie Carlier

3 - Understanding overflow metabolism in E. coli through the inclusion of enzymatic constraints
Javiera Cortés-Ríos, Ricardo Leiva-Guerra, Cain Acuña-Benavides, Maria Rodriguez-Fernandez and Pedro A. Saa

5 - A Monod model of competitive PCR
John Goertz, Ruby Sedgwick, Ruth Misener, Mark van der Wilk and Molly Stevens

7 - Data-driven mechanistic modelling of N-acetylaspartate metabolism
Polina Lakrisenko, Marcel Kretschmer, Andre Wegner and Daniel Weindl

9 - A mechanistic model for endocrine profiles of female puberty maturation
Sophie Fischer and Susanna Röblitz

11 - Spatial modelling of the nuclear exclusion of oncogenic transcription co-activators YAP and TAZ in liver cancer cells
Lilija Wehling, Liam Keegan, Jennifer Schmitt, Ursula Kummer, Kai Breuhahn and Sven Sahle

13 - Efficient parameter inference framework for dynamic flux balance models
Erika Dudkin, Yannik Schaelte, Moritz E. Beber, David S. Tourigny and Jan Hasenauer

15 - Kinetic modelling of toxin transport in a bio-artificial kidney
Jasia King, Sangita Swapnasrita, Stefan Giselbrecht, Roman Truckenmüller and Aurélie Carlier

17 - Towards understanding vascular (dys)regulation integrating endothelial morphology and blood flow response signaling: Development of geometric blood vessel model
Daniel Seeler, Nastasja Grdseloff, Claudia Jasmin Rödel, Charlotte Kloft, Salim Abdelilah-Seyfried and Wilhelm Huisinga

19 - UNCOVer - UNcertainty quantification of COVID-19 epidemiological models
Elba Raimundez, Paul Jost, Iva Ewert, Vanessa Nakonecnij, Jan Hasenauer and Dilan Pathirana

21 - yaml2sbml: Human readable and writable specification of ODE models in SBML
Jakob Vanhoefer, Marta R.A. Matos and Jan Hasenauer

23 - Repairing dynamic models: a method to obtain identifiable and observable reparameterizations with mechanistic insights
Gemma Massonis, Julio Banga and Alejandro Villaverde

25 - FitMultiCell: simulating and parameterization of computational models of multi-cellular processes
Emad Alamoodi, Jan Hasenauer, Jörn Starruß and The Fitmulticell Consortium

27 - Efficient robust adaptive distance functions in Approximate Bayesian Computation
Yannik Schälte and Jan Hasenauer

29 - Approximating Fisher Information for efficient MCMC
Federica Milinanni, Olivia Eriksson, Pierre Nyquist and Andrei Kramer

---

Zeynep Karagöz, Thomas Geuens, Vanessa LaPointe, Martijn van Griensven and Aurélie Carlier

Integrin transmembrane proteins play a central role in mechanotransduction at the cell–extracellular matrix interface. This process is central to cellular homeostasis and therefore particularly important when designing instructive biomaterials and organoid culture systems. Previous studies suggest that fine-tuning the extracellular matrix (ECM) composition and mechanical properties can improve organoid development [1-3]. Although experimentally very complex, computational models provide novel avenues to testing the effect of numerous different ECM ligands and mechanical properties on cell decision-making mechanisms.

We developed an ordinary differential equation–based model that enabled us to simulate three main interactions, namely integrin activation, competitive ligand binding and integrin clustering. Our results clearly indicate that this competition between ligands defines the fate of the system. By simulating the model with different initial conditions for competing ligands and testing different sets of binding rates, we demonstrated that the ligand with the higher binding rate (L1) occupies more integrins at the steady state than does the competing ligand (L2). We have also demonstrated that an increase in the initial concentration of ligands does not ensure an increase in the steady state concentration of ligand-bound integrins. Local parameter sensitivity analysis was in accordance with these observations. The L2-bound integrin concentration (IL2) was most sensitive to changes in binding and unbinding rates of the two competing ligands. Furthermore, the IL2 concentration was sensitive to the decrease in L1 initial concentration but not to changes in L2 initial concentration. In summary, with cell type specific, quantitative input, this computational model can be used to develop instructive cell culture systems.

[1]T. Geuens et al., npj Regenerative Medicine, 2020.

[2]N. Gjorevski et al., Nature, 2016. [3]E. Garreta et al., Nat. Mater., 2019.

Javiera Cortés-Ríos, Ricardo Leiva-Guerra, Cain Acuña-Benavides, Maria Rodriguez-Fernandez and Pedro A. Saa

Overflow metabolism refers to the phenomenon in which cells use fermentative pathways instead of respiration to generate energy in the presence of oxygen. The latter is highly unintuitive as cellular respiration is a more energy-efficient process. Interestingly, this process has been observed in many bacteria, yeasts and even mammalian cells since many years, and it was not until recently that a fundamental modelling framework based on resource (protein) allocation was established to describe it. Protein allocation models explain overflow metabolism based on the higher proteome cost for energy biogenesis of respiration compared to fermentation at high growth rates. To enable predictions, coarse-grained allocation models incorporate cross-sectioned proteome constraints such as respiratory, fermentation, and biomass fractions. While the latter have been effective for reproducing the observed data, they lack mechanistic detail. In this work, we developed a large metabolic model of E. coli (ECC2) provided with enzymatic constraints for representing overflow metabolism in greater detail. To this goal, turnover numbers from BRENDA and the typical total amount of enzyme in the cell (0.26 g/g) were mapped onto the model and used to constrain reaction fluxes. A sensitivity analysis was performed to identify the turnover constants with the highest contribution to overflow metabolism, which were then tuned using global optimization to reproduce the observations. The calibrated model accurately predicted acetate production (fermentation product) in the presence of oxygen from 0.52 1/h up to a maximum growth rate of 0.73 1/h. Notably, the maximum occupancy of the enzyme pool was reached just before acetate production began, highlighting the presence of an enzymatic bottleneck. Overall, this work sheds light on the metabolic mechanisms governing overflow metabolism.

John Goertz, Ruby Sedgwick, Ruth Misener, Mark van der Wilk and Molly Stevens

Despite its ubiquity in both research and clinical diagnostics, mechanistic models for polymerase chain reaction (PCR) remain incomplete or inflexible. Existing models are either too fine-grained for practical use and efficient parameter estimation or lack modular structure. This hampers our ability to adequately engineer PCR, in particular more complex systems such as asymmetric and competitive designs. Competitive PCR, where multiple distinct amplicon sequences compete for the same primers, is of particular interest for development of novel medical diagnostics. Taking inspiration from advances in systems biology, we propose a Monod model for PCR, treating each amplicon strand as a reproducing organism that consumes resources (primers, reagents) to generate its complement. We employ a Bayesian parameter estimation strategy along with Gaussian Process regression to capture the rich landscape of PCR behaviour across dozens of amplicon sequences. Our results challenge classical amplicon design strategies and suggest more nuanced refinements. We also believe our approach could inform the analysis and engineering of synthetic gene regulatory pathways. Finally, this model allows simulation of complex competitive dynamics, enabling exploration of novel PCR architectures.

Polina Lakrisenko, Marcel Kretschmer, Andre Wegner and Daniel Weindl

N-acetylaspartate (NAA) is the second most abundant metabolite in the brain, which is linked to Canavan disease, gestational diabetes and cancer. Various hypotheses have been proposed regarding its functions, but none of them has been conclusively confirmed. In the PeriNAA project we apply dynamical mechanistic modelling to better understand the role of NAA in cellular metabolism.

In order to learn more about NAA metabolism in health and disease, we’ve created a dynamical mechanistic model describing central carbon metabolism, TCA cycle and NAA metabolism. This large-scale compartmental model allows us to combine time-resolved stable isotope labelling data with metabolite pool size measurements and to analyse dynamics of cellular metabolism, centred around NAA. We use rule-based modelling approach and have developed a computational pipeline for model simulation and parameter estimation.

The calibrated model provides information on latent variables, such as metabolic fluxes and pools of metabolites that are not accessible experimentally. It will be used for generating and testing hypotheses regarding the role of NAA in cellular metabolism.

Sophie Fischer and Susanna Röblitz

The hypothalamus-pituitary-gonadal axis (HPG axis) has a central role for female reproduction. During pubertal development, the reactivation of the HPG axis causes characteristic physical and psychological changes. Breast development is the initial clinical sign of female pubertal onset. The 5-stage Tanner scale is commonly used to estimate the stage of female pubertal maturation by breast development. However, a trend of earlier breast development in girls since 1977 is reported. Hence, the necessity to establish new references to monitor pubertal development.

Our aim is to set up a mechanistic model that can be used to make predictions about pubertal maturation given the current hormone state and the age. The model consists of ordinary differential equations (ODE) describing the time courses of the main hormones GnRH, LH, FSH, and E2. We assume that all key mechanisms exist from early childhood on and that the main pacemaker for pubertal maturation is the development of a regular GnRH release pattern. Parameters for the GnRH release as well as the FSH synthesis and release are sampled from normal distributions in order to study individual puberty progression.

Our deterministic simulation results are comparable to the moving average of data from the Bergen Growth Study 2. Simulations with sampled parameters reflect inter-individual variabilities. Our results suggest that the development of a regular GnRH release pattern is of importance to describe the pubertal hormonal changes and go along with experimental results suggesting that pubertal onset in girls is associated with individual FSH dynamics. In future work, we plan to set up an “average model” as a prior using longitudinal hormone data and take individual hormone time series to estimate individual model parameters.

Lilija Wehling, Liam Keegan, Jennifer Schmitt, Ursula Kummer, Kai Breuhahn and Sven Sahle

The Hippo pathway inhibits nuclear translocation of two oncoproteins YAP (yes-associated protein) and TAZ (WW domain containing transcription regulator 1). The nuclear localization of the oncoproteins YAP and TAZ is regulated by various factors, including, cell substrate stiffness, and surrounding cell density. The nuclear localization of YAP/TAZ is an indicator of poor prognosis in many cancers, including hepatocellular carcinoma (HCC) [1]. Therefore the regulation of the nuclear exclusion of YAP and TAZ could be a promising approach of inhibiting cancer progression.

First, we established an HCC cell line expressing fluorescent tags for YAP, TAZ and H2B (nuclear protein) that allowed us to study nuclear-cytoplasmic shuttling of YAP/TAZ via time-lapse high-throughput confocal microscopy. Next, we created an image analysis pipeline, with which we quantified the relative nuclear abundance of YAP/TAZ. Finally, using the live-cell imaging data, we modelled the spatial distribution of YAP and TAZ under different cell density conditions. The mathematical models were based on the simulation of 2D PDEs. Model simulation and the parameter fitting to the experimental data was performed with the Spatial Model Editor (https://github.com/spatial-model-editor).

Our experiments showed that YAP localized predominantly in the nucleus, whereas TAZ was excluded from it. Using the PDE models we could attribute the characteristic distribution patterns of YAP and TAZ to a few key reactions, which govern observed localization patterns of YAP and TAZ. Based on our model, we propose parameter combinations whose perturbation might precisely target the oncogenes YAP/TAZ for the nuclear exclusion.

Being able to influence the nuclear exclusion of YAP and TAZ and testing medical substances (e.g. kinase inhibitors and other cancer drugs) with respect to their effect on the Hippo singling pathway could be of clinical relevance.

[1] Weiler, S et al. Gastroenterology, 2017, 152, 2037-2051.e22

Erika Dudkin, Yannik Schaelte, Moritz E. Beber, David S. Tourigny and Jan Hasenauer

Dynamic flux balance (DFBA) models are used to model genome-scale metabolic networks and investigate the dynamics of metabolic fluxes within time dependent environmental changes. DFBA models are formulated as linear programming (LP) problems embedded within systems of ordinary differential equations (ODEs). Due to the hybrid nature of DFBA models, these systems are particularly challenging to solve numerically.

Furthermore, because of these numerical and computational challenges associated with simulating DFBA models, parameter inference required to fit models to experimental observations has also not been performed systematically. An efficient and general framework for simulation and parameter inference is therefore required by researchers working with DFBA models. In this contribution, we present a simulation and parameter inference framework for a wide range of DFBA models, taking advantage of the speed and robustness of the simulation package dynamic-fba (Tourigny et al. (2020), doi.org/10.21105/joss.02342) and the parameter inference package pyPESTO (Schälte et al. (2020), doi.org/10.5281/zenodo.3928322) for parameter estimation and parameter uncertainty analysis. pyPESTO offers various optimization and uncertainty quantification methods and allows to employ high-performance computing infrastructure. We tested the framework on a toy DFBA model with simulated data and showed that the model describes the data well with the estimated parameters. Currently, we are testing the framework on actual application problems where systematic parameter analysis has not been performed so far. Goals of this project include to find and give guidelines to the most suitable optimization methods to infer the best parameter sets, as well as investigating the parameters uncertainties.

Jasia King, Sangita Swapnasrita, Stefan Giselbrecht, Roman Truckenmüller and Aurélie Carlier

Organic anion transporters (OATs) are mainly responsible for transepithelial movement of many small molecules including toxins in the kidney and other major organs. This movement includes several steps such as the uptake of uremic toxins, formation of protein-toxin complexes, toxin binding and unbinding to OATs. For simplification, the transport steps are often lumped together to model the overall uptake of the toxin-protein-OAT complex. We have decoupled the individual toxin transport steps to shed more light on the independent contributions of the activity and density of the transporters. We have shown that the final toxin concentration saturates after a certain increase in both the transporter uptake activity and density but increases steeper with the latter. Detailed mechanistic models such as this offer interesting avenues to accurately model the competitive binding between multiple toxins as well as direct organ-on-chip development with more accurate portrayal of the in vitro data.

Daniel Seeler, Nastasja Grdseloff, Claudia Jasmin Rödel, Charlotte Kloft, Salim Abdelilah-Seyfried and Wilhelm Huisinga

Cerebral cavernous malformations (CCMs) are lesions mainly found in the brain vasculature that may result in hemorrhagic stroke. Currently, no drug therapy is approved for their treatment. CCM pathology is caused by aberrant blood flow sensing and leads to changes of endothelial cell (EC) morphology. The objective is to develop a mathematical model capturing the bidirectional feedback between blood flow and EC morphology to understand vascular (dys)regulation. As a first step, we aimed to develop a geometric model linking EC morphology to blood vessel geometry. We used 3d positional data of EC junctions in the dorsal aorta (DA) of zebrafish embryos to construct vessel cross-sections (CSs). First, we used data-rich angiogram slices to identify a suitable family of shapes for the DA CS. To finally fit CSs to sparse junctional data, we enriched single data points by their neighbors with Gaussian weights decaying with distance. To ensure robust estimation, we chose the width of the weight function adaptively in space. We investigated the uncertainty caused by manual data annotation by comparing projection errors, CS areas and diameters between two annotations.

Finally, we compared vessel geometry changes over time with published results. To account for horizontal and vertical asymmetries, we split CSs into four quarter superellipses. Fitting to angiogram slices, we found low projection errors and similar CS areas. Errors when fitting to junctional data were mostly low but data was partially too sparse to identify shapes that change smoothly along the vessel axis. The two annotations resulted in errors, CS areas and diameters of similar magnitude. Our model predicted decreases in vessel diameter between 48 and 72 hours post fertilization matching experimental data.

Our next steps are to (1) employ a nonlinear mixed-effects modeling approach to improve estimation with sparse junction data; and (2) use the reconstructed vessel surface to characterize EC morphology in 3d.

Elba Raimundez, Paul Jost, Iva Ewert, Vanessa Nakonecnij, Jan Hasenauer and Dilan Pathirana

Epidemiological models are widely used to analyze the spread of diseases such as the global COVID-19 pandemic. However, all models are based on simplifying assumptions and often on sparse data. This can limit the reliability of parameter estimates, which often translates to uncertain predictions. The latter is of great importance and should be considered when designing intervention policies.

UNCOVer aims to identify aspects of compartmental models used for COVID-19 that improve the reliability of predictions. To do this, we will first create a curated collection of already published compartmental models that are converted into standard formats, SBML (for model definition) and PEtab (for parameter estimation problems), to ensure reusability and reproducibility. This model collection will be used to perform parameter estimation and uncertainty analysis with a comprehensive dataset. The analysis will provide insight into the parameters in the models that are critical to the reliability of predictions. Moreover, we will gain information about optimal model topology to describe the given dataset by creating a superset model and using model selection techniques. Lastly, we will provide an intervention specification format within the PEtab framework to enable and facilitate the testing of several types of interventions in the model collection. The outcomes of UNCOVer are not restricted to COVID-19 but can be used to assess the average population dynamics of other infectious diseases.

Jakob Vanhoefer, Marta R.A. Matos and Jan Hasenauer

Ordinary Differential Equations are a common model type throughout systems biology. The Systems Biological Markup Language (SBML) on the other hand is a widely adopted community standard for specifying dynamic models in the field of systems biology, that was recently extended to the PEtab format for describing parameter estimation problems. There exist a large number of software pipelines for simulating SBML models and performing parameter estimation for PEtab problems. Specifying an ODE model in SBML would grant direct access to these software pipelines. Since SBML is considered neither human-readable nor human-writable, an easy to use approach to construct SBML/PEtab models for a given ODE will facilitate model generation. In this contribution we present `yaml2sbml`, a python tool for converting ODE models specified in an easy to read and write YAML file into SBML/PEtab. Further `yaml2sbml` comes with a validator for the input YAML, a Command Line Interface and a Model Editor to build an input YAML within python code. Several examples highlight the usage of `yaml2sbml` on realistic problems.

Gemma Massonis, Julio Banga and Alejandro Villaverde

Mechanistic dynamic models allow for a quantitative and systematic interpretation of data and the generation of testable hypotheses. However, these models are often over-parameterized, leading to non-identifiability and non-observability, i.e. the impossibility of inferring their parameters and state variables. In many cases, the source of over-parameterization is structural, that is, it stems from the model equations. The lack of structural identifiability and observability (SIO) compromises a model’s ability to make predictions and provide insight. Here we present a methodology, AutoRepar, that corrects SIO deficiencies automatically, yielding reparameterized models that are structurally identifiable and observable [1]. The reparameterization preserves the mechanistic meaning of selected variables, and has the exact same dynamics and input-output mapping as the original model. We implement AutoRepar as an extension of the STRIKE-GOLDD software toolbox for SIO analysis [2], applying it to several models from the literature to demonstrate its ability to repair their structural deficiencies. AutoRepar increases the applicability of mechanistic models, enabling them to provide reliable information about their parameters and dynamics.

 

[1] Massonis, G., Banga, J. R., & Villaverde, A. F. (2020). Repairing dynamic models: a method to obtain identifiable and observable reparameterizations with mechanistic insights. arXiv preprint arXiv:2012.09826.
[2] Villaverde, A. F., Tsiantis, N., & Banga, J. R. (2019) Full observability and estimation of unknown inputs, states, and parameters of nonlinear biological models. Journal of the Royal Society Interface, 16:20190043 doi:10.1098/rsif.2019.0043

Emad Alamoodi, Jan Hasenauer, Jörn Starruß and The Fitmulticell Consortium

Biological tissues tend to be dynamic and highly organized. Multi-cellular models are getting more attention as a way to explain and understand this organization. The parameterization of these models is essential to understand the multi-cellular systems, to predict perturbation experiments, and to compare competing hypotheses. However, multi-cellular and multiscale models have been proven to be highly difficult to parameterize. A method that has been proven to be applicable to multi-cellular models is Approximate Bayesian Computation (ABC). Unfortunately, ABC is a computationally expensive approach, as it requires a large number of simulations. Thus, there is an increased need for a fast and general-purpose pipeline for modeling and simulating multi-cellular systems that can exploit HPC systems for faster computations. To this end, we started the development of a user-friendly, open-source, and scalable platform, called FitMultiCell, that can handle modeling, simulating, and parameterizing multicellular systems. To achieve the goal of FitMultiCell, we combine the modeling and simulation tool Morpheus with the advanced statistical inference tool pyABC. In this contribution, we present an overview of the FitMultiCell project and demonstrate an application example. An HCV model that was described by Kumberger et al. (Viruses, 10(4), 2018) was used to illustrate the flow of the platform.

Yannik Schälte and Jan Hasenauer

Approximate Bayesian Computation (ABC) is a likelihood-free inference method for complex stochastic models. Its general applicability and independence of model type have in the past years led to a growing popularity in many research areas, ranging from systems biology over population genetics, ecology, economics, to astronomy and epidemiology. A typical application are agent-based modes. In a nutshell, in ABC likelihood evaluations are circumvented by assessing the similarity of simulated and observed data via some distance metric and threshold. In particular, ABC is often combined with a sequential Monte-Carlo scheme (ABC-SMC) where the acceptance threshold is iteratively reduced while improving the posterior approximation, which allows to employ high-performance infrastructure with hundreds of cores, and thus scale to computationally expensive models.

It can be shown that under certain assumptions, if the acceptance threshold converges to zero, the sampling distribution converges to the true Bayesian posterior. However, the simplicity of ABC has in the past also seduced to wrong applications. In this contribution, we highlight potential problems induced by the distance metric and introduce novel adaptive metrics.

While, theoretically, neglecting model error, any distance metric can be employed, practically, due to limited resources, different metrics can lead to considerably different results, e.g. of the inferred parameter uncertainties and predictions. In the field, self-learning metrics have been developed which try to adjust to the problem structure. We demonstrate how such metrics can improve efficiency considerably, but also show pitfalls if underlying implicit assumptions are not met. We present a set of novel adaptive distance metrics in an ABC-SMC context that are more robust to model error and directly regress for the information different data sets yield on parameters, and demonstrate their applicability on a set of application problems.

Federica Milinanni, Olivia Eriksson, Pierre Nyquist and Andrei Kramer

Parameter estimation can be a difficult task in systems biology. In the modeling class of ordinary differential equations (ODEs), the parameters can be reaction rate coefficients, dissociation constants, stable concentrations of environmental compounds, Hill kinetic exponents and many other entities that are very different in scale and unit. The available experimental data may contain little information about some parameters and much information about others. This creates objective functions and/or Bayesian posterior probability densities that are difficult to explore. Some Markov Chain Monte Carlo (MCMC) methods aim to reduce auto-correlation within the chain by devising advanced transition rules. Among these methods, the Riemann Manifold versions of the Metropolis Adjusted Langevin Algorithm (SMMALA/RMMALA) and Hamiltonian Monte Carlo method (RMHMC) have update rules that require the gradients of the log-likelihood function and Fisher Information as a metric tensor. This choice of metric makes the moves in parameters space much more efficient than an isotropic transition rule.

Within the ODE framework, which is very common in system biology, likelihood functions arise from noisy measurements and require sensitivity analysis to compute the Fisher Information. However, forward sensitivity analysis is numerically expensive and large models (more than ~30 parameters) require a different approach.

In our project we developed an approximation to the sensitivity matrix. We started from the exact solution, that makes use of the related fundamental system, which is given by the the Peano Baker series (PBS). Our approach is the approximation of the PBS by truncating the series and approximating the integrals in it. We implemented the algorithm in the Julia language and tested it on a model of intracellular pathways related to synaptic plasticity. We compared our results (from the PBS approximation) with the sensitivity computed with Julia's forward sensitivity solver.