# SEM++: A particle model of cellular growth, signaling and migration

- 1.1k Downloads
- 11 Citations

## Abstract

We present a discrete particle method to model biological processes from the sub-cellular to the inter-cellular level. Particles interact through a parametrized force field to model cell mechanical properties, cytoskeleton remodeling, growth and proliferation as well as signaling between cells. We discuss the guiding design principles for the selection of the force field and the validation of the particle model using experimental data. The proposed method is integrated into a multiscale particle framework for the simulation of biological systems.

### Keywords

Cell mechanics Subcellular element model Cell growth Cell migration Cell signaling## 1 Introduction

Developmental processes such as embryogenesis, organ formation and tumor growth are orchestrated by multiscale bio-mechanical and chemical phenomena. In tumor-induced angiogenesis for instance, tip cell selection, sprout migration, proliferation of the stalk cells and the formation of lumen emerge as a result of such coordinated cell interactions [32, 39]. The phenotypic behavior of individual cells is determined by their response to a collection of stimuli from the surrounding environment, including mechanical and chemical signals provided by the extracellular matrix (ECM) and cells in their neighborhood. These processes can transform a homogeneous cell population to a heterogeneous one where individual cells interact in a coordinated way to give rise to a larger, hierarchical structure.

Continuum models are well suited to capture the large-scale dynamics in growing tissue at the millimeter to centimeter range. However, the approximation of cells by a continuum density prohibits the exploration of single cell dynamic that give rise to important biological processes such as polarization, growth and adhesion. When cells are modeled at a continuum level, a single cell does not exist. The emergence of phenomena initiated by single cells, a result of mutation or phenotypic activation, can therefore not be studied in these models [31]. We wish to study such phenomena in computer simulations and we employ a cell-based model that can represent individual cells subject to mechanical stresses, migration, proliferation, cell-cell and cell-substrate adhesion and cell-cell mediated signaling.

In the field of cell-based modeling (see Fig. 2), we discriminate between grid-based and lattice-free particle methods. Cellular Automata are structured grids where every grid point represents a single cell [4, 15]. In single particle models each cell is represented by a particle which can freely move in space [5, 35, 49]. Cell vertex models represent cells with polygonal (2D) or polyhedral (3D) shapes which are evolved by minimizing an energy potential [16, 43]. They are valuable tools to study adhesion, surface tension and pressure driven cell arrangement as they explicitly capture cell size and shape, however, they are limited to represent polygonal cell shapes in closely packed epithelial layers. Cellular potts models (CPM) are derived from the Ising model of ferromagnetism [26] and have been used to study adhesion driven sorting in cell layers and various other developmental and pathological systems in biology [12, 23]. In CPM, a cell is defined as a set of lattice sites which are updated according to a Monte Carlo Metropolis algorithm to minimize an energy potential. They are not not time-dependent and describe equilibrium situations [5]. In the Immersed Boundary Cell Model each cell is represented by an elastic body and the immersed boundary method [50] is used to capture the interactions between the cells and a surrounding, incompressible, viscous fluid [53, 54].

The methods presented in this paper belong to the lattice-free methods and extend the subcellular element model (SEM) [45, 57]. In the SEM, cells are represented by an agglomeration of computational particles whose interactions are defined via short-range potentials. The potential function [57] models the viscoelastic properties of the cytoskeleton of isolated cells. Model parameters are non-dimensionalized to scale with the number of particles per cell and the mechanical stiffness and homotypic and heterotypic cell-cell adhesion are tunable via the intra-cellular and inter-cellular potential parameters respectively.

We present extensions to this model to account for actin polymerization based migration. We introduce a model for cell growth and proliferation driven by the cell-cycle and we propose extensions to model cell-cell mediated signaling processes that can determine the cell phenotype. We refine the structure of the cell to account for the explicit inclusion of subcellular organelles such as the nucleus. The cell membrane is explicitly traced to model membrane bound receptor mediated signaling cascades and its mechanical distinct properties from the cytoskeleton.

The proposed methods have been implemented in a standalone, object-oriented C++ framework that features thread-based parallelism. The implementation allows for modular extensions and fast prototyping of the model and integrates visualization capabilities for 2D simulations. Furthermore, we present an integration of the model into the LAMMPS framework [33, 51], a highly optimized parallel C++ library for the computation of particle interactions.

The article is structured as follows: First, we introduce the SEM and specify our extensions of it. Section 3 describes our implementation of the model. Applications of the model are shown in Sect. 4, followed by a brief summary.

## 2 Model and methods

Model parameters of the extended SEM

Name | Description |
---|---|

\(d\) | dimensionality |

\(R_{cell}\) | cell radius |

\(N\) | number of particles per cell |

\(p_d\) | sphere close packing density |

\(\rho \) | scaling factor |

\(\kappa _0\) | elasticity scaling coefficient |

\(\eta _0\) | viscous damping scaling coefficient |

\(\lambda \) | tuning coefficient for varying \(N\) |

\(D\) | diffusion coefficient for particles |

\(\alpha \) | shifting parameter |

\(T_p\) | polymerization rate |

\(R_{2d},R_{3d}\) | radius for neighbor detection |

\(N_{int}^{th}\) | internal neighbor threshold |

\(N_{ext}^{th}\) | external neighbor threshold |

\(\beta ^{th}\) | neighbor sector threshold |

\(R_{nuc}\) | nucleus radius |

\(r_{min}\) | minimal shifted radius |

\(R_{dens}\) | radius for local density measure |

\(\pi _p\) | polarity threshold for polymerization |

\(\pi _{dp}\) | polarity threshold for depolymerization |

### 2.1 The subcellular element model (SEM)

### 2.2 SEM++

The extensions of the proposed model include a modification of the potential function to increase adhesion between SCE, the polarization of cells, the smooth insertion and removal of SCE to model polymerization and depolymerization, the detection of cell membrane elements and a special SCE for the cell nucleus. These extensions are detailed below.

#### 2.2.1 Potential functions

#### 2.2.2 Polarization

#### 2.2.3 Polymerization and depolymerization

Active, autonomous cell shape alteration and cell migration is the result of the assembly and disassembly of the cytoskeleton network via polymerization and depolymerization of actin filaments and microtubules. Although random polymerization events are accounted for in the noise term \(\varvec{\xi }_{\alpha _i}\) of Eq. (5), this model does not qualify to capture the active dynamics of cytoskeleton remodeling in response to external stimuli.

We propose to model the building up and breaking down of the cytoskeleton via the explicit insertion and deletion of SCEs. In order to achieve a seamless, smooth and location independent insertion and removal of the selected elements for polymerization and depolymerization, we propose the introduction of a polymerization factor \(\varphi _{\alpha _i}\) associated with each element. We define the time-interval for a polymerization event to occur as \(T_p\).

#### 2.2.4 Cell membrane

*Membrane detection in 2D models*

For every SCE, we consider the elements of the same cell in its neighborhood defined by a circle of size \(R_{2d}\). The neighborhood is then divided into angular sections defining the space between consecutive neighboring elements and the SCE under consideration. If the angle \(\beta _{max}\) defining the largest of these sections exceeds a predefined threshold value \(\beta ^{th}\), the element is considered part of the cell’s membrane. The membrane elements of cell \(i\) are hereafter denoted by the subset \(\alpha ^m_i\). We note that the algorithm works for both isolated and interacting cells. Furthermore, by taking the bisector of \(\beta _{max}\), we get an estimate for the surface normal \(\varvec{n}\) at the membrane element under consideration. A conceptual sketch of the membrane detection mechanism in 2D models is depicted in Fig. 6.

*Membrane detection in 3D models*

#### 2.2.5 Nucleus

## 3 Implementation

The proposed methods have been implemented inside a C++ framework developed at the CSE Lab. The code explores object-oriented programming and thread level based parallelism of Intel’s TBB library. The object-oriented implementation allows for modular development and supports the adaptation and extension of existing methods. New concepts can easily be integrated and existing modules can be combined to assemble complex models. The flexibility of this framework is of key importance in developing novel methods and algorithms for cell-based models where the method of choice to model a certain phenomena is highly ambiguous. Often, the choice of one specific model over another is the result of an extensive validation and optimization process during which a number of implementations need to coexist.The framework integrates a GLUT based visualization engine that allows for direct visual feedback during development.The thread-based parallelism of the framework allows for efficient computations on multi-core architectures. However, it does not explore multi-processor parallelism or GPU based acceleration. Furthermore, the methods have not been subjected to extensive performance optimization.

To address this issue, we started to implement the core features of the extended SEM inside the LAMMPS framework [33, 51], a highly optimized parallel C++ library to conduct particle simulations in the field of MD on multi-node, multi-GPU clusters. In this way, limitations associated with the computational costs of individual-based models are addressed inside a highly efficient and thoroughly tested modeling framework for particle simulations that enables a boost in the modeled system size at almost no developmental cost. We evaluated a number of existing libraries used mainly for MD simulations including NAMD [44], Amber [3] and DESMOND [13]. Although performance is the main reason for integrating the SEM within these frameworks, the software needs to offer a set of key features to allow for this integration. Most MD frameworks are designed to be used at an interface level where the user provides scripts that drive the simulations. LAMMPS can in addition be used as a library, is very well documented and encourages the integration of new models directly inside the source code. An extension to the source code, especially the memory management, is essential to the integration of the SEM. Classic MD simulations usually keep the number of atoms (particles) fixed throughout the simulation, as they model a closed system. Simulating growth and death, however, demands the seamless insertion and deletion of elements in an efficient way. Of the different MD software we evaluated using these criteria, our results supported the selection of the LAMMPS framework as our platform of choice.

## 4 Applications

### 4.1 Cell stretching

Cells are grown for \({200}\hbox { s}\) and relaxed for another \({100}\hbox { s}\) before the walls are removed. After removing the walls and assigning the wall layer particles, the system is relaxed for another \({4}\hbox { s}\). We perform creep response tests on the equilibrated system for applied stresses \(\sigma \) of \({6.59}\;\hbox {Pa},\,{8.66}\;\hbox {Pa},\; {11.40}\;\hbox {Pa},\,{15.00}\;\hbox {Pa},\; {19.75}\;\hbox {Pa},\,{25.00}\;\hbox {Pa}\) and \({30.00}\;\hbox {Pa}\) for \({7}\;\hbox {s}\). The first five values for \(\sigma \) correspond to simulation results reported in [57]. In order to capture recovery after the stress, we set \(\sigma ={0}\hbox {Pa}\) for a period of \({4}\hbox { s}\) after the stretching.

Simulations of cells without a nucleus (see Fig. 10) are compared to simulations of cells including a nucleus element (see Fig. 11). We investigate cells with a nucleus of a size equivalent to 5, 10 and 20 % of the cells volume. To initialize the simulations including a nucleus, we remove SCE elements at the center of mass corresponding to the desired nucleus volume from the equilibrated cell holding no nucleus. We then introduce a nucleus element at the center of the cell and ramp up its polymerization parameter \(\varphi \) over a time period of \(T_p = {100}\hbox { s}\), followed by an additional equilibration period of \({10}\hbox { s}\). The nucleus potential parameters are set to the same values as for the SCE potential. We compare the creep response for cells with \(N = 1,000\) and \(N = 8,000\) elements. The system was integrated with Brownian Dynamics and Forward Euler time integration. The time steps during growth were set to \(\varDelta t = {1.0 \times 10^{-3}}\hbox { s}\) and \(\varDelta t = {2.5 \times 10^{-4}}\hbox { s}\) for \(N = 1,000\) and \(N = 8,000\) respectively. During the creep response test, a 10 times smaller \(\varDelta t\) was used to ensure stability of the method. Additionally, we assessed two ways to control the random fluctuations during the creep response test: first with a constant diffusion coefficient \(D = {1.6 \times 10^{-13}}\;\hbox {m}\;\hbox {s}^{-2}\) as reported in [57] and then with \(D = k_B T / \eta \) computed with a fixed temperature of \(T = {298}\hbox {K}\). We observe that the simulation results for a fixed diffusion coefficient D show more fluctuations in the strain curve. Otherwise, the creep response curves for a fixed temperature agree well with the results obtained for a bigger diffusion coefficient (Fig. 10). For large stresses, cells can be observed to rupture or undergo plastic deformations and are very sensitive to the level of discretization. We therefore consider to ignore the results obtained for stresses above \({19.75}\;\hbox {Pa}\).

Considering the simulation results in the absence and in the presence of a nucleus of varying size, we observe a viscoelastic creep response in all scenarios. In the presence of a large nucleus element, we observe a reduction in the observed deformation response. However, we find qualitative differences in the simulations corresponding to \(N = 1,000\) SCEs and \(N = 8,000\) SCEs (Fig. 11). Whereas the deformation is monotonically decreasing in the case where \(N = 8,000\), for \(N = 1,000\), we observe an increase in the observed deformation for a nucleus of 5 % volume before the deformation decreases for larger nuclei with respect to the reference case without a nucleus.

In conclusion, these results suggest that inhomogeneities in the cytoplasm of the cell, as induced by a large organelle such as the nucleus, could have an important effect on the viscoelastic properties of the cell. We note, however, that the presence of a nucleus element does not change the response curve qualitatively. This effect could possibly be accounted for by adjusting the potential parameters. In the light of the presented results, the effect of a nucleus element on the creep response of cells under stress and the scaling behavior of the Subcellular Element Model for large numbers of elements per cell needs to be further investigated.

### 4.2 Cytoskeleton remodeling

Cytoskeleton remodeling in response to external mechanical and chemotactic stimuli can lead to alterations in the cell shape, structural reinforcement and migration, depending on the coordinating mechanisms that dictate the location of polymerization events. With the machinery for polymerization and depolymerization of distinct SCEs in place, we propose two different models for polymerization. We first introduce a polarity driven polymerization method that leads to directed cell migration. Secondly, we propose a density aware model for polymerization to account for the cell response to mechanical stimuli as a result of internal and external compression. We introduce the notion of a cytoskeleton density \(d_{\alpha _i}\) at each element to be defined as the number of SCEs \(\beta _j\) inside a defined neighborhood of radius \(R_{dens}\). The density is further refined into the internal density \(d^{int}_{\alpha _i}\), only considering neighborhood elements of the same cell \((i = j)\) and the external density \(d^{ext}_{\alpha _i}\), only considering the neighborhood elements of different cells \((i \ne j)\).

*Cell Migration (Polarity aware Polymerization)*Cells can be observed to hold a persistence of direction in their migrative behavior [19]. In mesenchymal cell migration, the persistence of direction is a result of the polarization of the cell regulating the assembly and destruction of the actin cytoskeleton, in combination with orchestrated binding and unbinding of adhesion molecules of the cell membrane to the ECM. In the current model, focal adhesion sites are not explicitly modeled. In the 3D model, substrate adhesion is accounted for by introducing a wall potential force. Here, we employ a (9,3) LJ potential

We present simulation results for single migrating cells in an in vitro environment in (Fig. 12, 2D) and (Fig. 13, 3D). Cell migration in the Subcellular Element Model has been considered before where cells are represented by single elements (\(N = 1\)) [46]. In this scenario, cell migration is realized by an advection velocity on the cell particle. Sandersius et al. [58, 59] have motivated coordinated polymerization as a driving mechanism in cell migration that resemble closely the model presented in this work.

*Cell Reinforcement (Density aware Polymerization)* Cells have been shown to react to external stimuli by mechanotaxis and cytoskeleton remodeling [29]. They do so in response to mechanical cues propagated by their environment, the extracellular matrix or migrating neighboring cells, in order to reinforce the stress fibers where needed or to evade highly populated tissue regions. We account for this behavior by including a mechanism to sense the cytoskeleton density at each element and coordinate the polymerization activity in response.

- 1
Only the internal density is considered and either all or only the membrane elements are subject to polymerization. Depolymerization happens at regions of low density and polymerization at regions of high density. This leads to a reinforcement of the cell in regions where it is compressed. This behavior is consistent with observations of cells reinforcing their actin network or basal lamina outside the cell in response to mechanical stress.

- 2
Only the external density is considered and only the cell membrane elements are subject to polymerization. This leads to a short migrative behavior of the cells to evade cell contact. This mechanism could initiate the polarization of the cell to migrate away from others.

In Fig. 14, we report simulation results of cell migration subject to cell-cell adhesion. We observe one cell migrating along its major elongation axis subject to cell-cell adhesion forces exerted by a neighboring cell that is not migrating. For the non-migrating cell, we investigate three different scenarios: no active remodeling of the cytoskeleton (Fig. 14 a) and the two scenarios outlined above (Fig. 14 b,c). We find that with no active remodeling of the cytoskeleton in the non-migrating cell, the two cells loose contact very quickly. In contrast, if polymerization is active from low density to high density regions as defined by the metric \(d_{\alpha _i}\), the cells remain in close contact and a circling pattern of the migrating cell around its neighboring cell can be observed. Such circling patterns are commonly observed in real cells. In Fig. 14 d, we show how two keratinocytes migrate on a collagen-coated substrate. The growth factor FGF7 was added to some of the cells to stimulate migration and the cells start circling around each other due to cell-cell adhesion [37]. The experimental images were provided by Michael Meyer from the Werner lab at ETH Zürich (HPL, Otto-Stern-Weg 7, 8093 Zürich, Switzerland).

From the simulations we observe that this behavior is very sensitive to the density sensing mechanism. Removing the external elements from the density estimation \((d^{int}_{\alpha _i})\) does not lead to the same behavior. These observations suggest that cytoskeleton remodeling plays a major role in cell-cell interactions observed during migration and pattern formation. Not only the processes that drive polymerization in the migrating cells but also the adhesion interactions and polymerization dynamics in neighboring cells can strongly influence and direct cell migration. It is therefore important for future works to carefully calibrate such migration models with experimental data. Data from wound healing assays [20, 40] or the growth of blood vessels in the mouse retina [41] could be used to fit parameters for the models and to quantify their uncertainties [6].

We note that the active reinforcement of the cell’s cytoskeleton has previously been addressed in the context of tissue stretching and during migration in epithelial like tissue sheets [59].

### 4.3 Growth and proliferation

The study of cellular growth and proliferation within the SEM framework has previously been proposed by Alt et al. [2]. The method suggests that at every time point, a subset of SCEs located in the center of the cell is allowed to duplicate with a small probability. For each duplicating element, a location at a distance \(r_e\) from the element is selected. If the point is sufficiently far away from all neighboring elements, the new element is introduced at that location. Iterative equilibration of the potential forces will integrate the new element into the structure. Once a cell reaches an appropriate size, the cell is divided by a plane through the center of mass, perpendicular to the longest axis of the cell.

- G0
After a cell division has occurred, the two daughter cells enter the quiescent state. The cells stay in the quiescent state to equilibrate for the time period \(T_{G0}\) before the cells transit to the growth phase.

- I
Cell growth is modeled via the duplication of SCEs at regular intervals over the time period \(T_{I}\). The duplicating element is chosen at random and a new element is inserted at the exact location of the duplicated element. The polymerization factor is set to \(\varphi _{\alpha _i}=0\) and ramped up to \(\varphi _{\alpha _i}=1\) over the time interval \(T_p={T_{i}}/{N}\). In this way we ensure that the location of insertion does not lead to any numerical instabilities and cell growth proceeds in a smooth manner. Variations to the proposed method to confine the region allowed for duplication to the center of the cell can easily be implemented.

- M1
When the cell has doubled in size, the cell cycle advances to the M1 phase where the nucleus breaks up into chromosomes which are divided to create two new nuclei. The nucleus is broken up by setting its polymerization factor \(\varphi _{\alpha _i}=0\) and a second nucleus is introduced at the location of the original one. During the M1 and M2 phase, the nuclei elements take on the function of the two poles of the mitotic spindle. Over the time period \(T_{M1}\), the two nuclei polymerization factors are ramped up. At the same time, the polymerization factor of the SCEs in the dividing cell is slightly increased to model swelling in size of the dividing cell [10]. Together, this leads to the separation of the two nuclei and the aggregation of SCEs around them. This way, the separation axis of the two nuclei, and accordingly the elongation of the cell is driven by the mechanical properties of the surrounding tissue.

- M2
After the two new nucleus elements have polymerized to attain their final size, the cell is split by assigning every SCE to the closest nucleus element inside the cell. The method does not depend on the introduction of a splitting plane and the orientation of cell division is driven by the mechanical properties of the surrounding tissue. Over the time period \(T_{M2}\), the two new cells are equilibrated and the polymerization factor of the individual SCEs is smoothly decreased back to unity, reverting the swelling of the cell.

We would like to stress that in the context of cell proliferation, the polymerization factor is used to model phenomena different than actin polymerization. The observed swelling of cells during proliferation is not a result of increased polymerization, but a result of an increase in the intracellular osmotic pressure.

In the proposed simulation framework, the polymerization factor acts as a scaling coefficient to the potential function and thus can be used to model the swelling of the SCEs as a result of osmotic swelling. Likewise, as mentioned above, the nucleus elements take on the role of the mitotic spindle in phases M1 and M2. A 3D simulation result of the proliferation algorithm is shown in Fig. 15.

### 4.4 Juxtacrine signaling

Multicellular organisms depend on effective and robust communication channels in order to coordinate collective behavior during development, growth and pathology (e.g. tip cell selection during angiogenesis). In time, a number of different signaling channels have evolved to integrate mechanical, chemical, even electrical information form close and distant locations. Here, we focus on the chemical signaling of membrane bound ligand-receptor complexes acting between neighboring, physically touching cells, referred to as juxtacrine signaling [36].

The mechanism has been studied in the context of spatially heterogeneous pattern formation and gene expression. Stochastic simulations of pattern forming reaction–diffusion systems can be used to study the effect of noise at low molecule numbers [7, 9, 21, 22]. Here, we propose an implementation of juxtacrine signaling inside the extended SEM framework and study the influence of added noise and intrinsic noise as a result of the number of SCEs per cell.

*Model of Juxtacrine Signaling*The model studied here was presented for homogeneous systems in [47] and serves as a simple, generic model of juxtacrine signaling. The model consists of three membrane bound molecular entities per cell \(i\): ligand \(a_i\), receptor \(f_i\) and the bound ligand-receptor complex \(b_i\). It describes the basic dynamics of reversible ligand-receptor binding, internalization and production and decay of ligand and receptor molecules as

*Pattern quantification*In order to measure progression during pattern formation, we define a metric that quantifies patterning at a certain point in time during simulation. Here, a pattern is characterized as a spatial inhomogeneity in some chemical species. Hence, we define the patterning intensity \(I\) that sums up the concentration differences over each cell and its neighbors:

We focus on the concentration of receptor-ligand complex \(b\), as the complex induces the activation of a signaling pathway that could lead to migration, changes in morphology and gene expression. Note that the intensity \(I\) is not normalized by the number of cells in the simulation and does not qualify to compare systems of different size. We do not consider the absolute magnitude of this function but are interested in its temporal evolution to determine when a pattern has stabilized (reached a constant value of \(I\)). A constant patterning intensity does not necessarily indicate a static pattern. Traveling or alternating patterns could also reflect a constant value of \(I\).

*Implementation details*We determine the fraction of membrane protein \(p\) of cell \(j\) that is presented to cell \(i\) as follows. For each membrane element \(\alpha _i^m\) of cell \(i\), we count the membrane elements \(\beta _j^m\) of neighboring cells inside an interaction radius \(r_m\) and store their number \(N_{\alpha _i^m}\) on the membrane particle \(\alpha _i^m\). We define the neighborhood of interacting elements around \(\alpha _i^m\) as

*Results*

All presented results are conducted in a periodic simulation domain containing 64 cells arranged in an eight by eight grid (see Fig. 16). The parameter values are set as proposed in [56]. All simulations are subject to the following initial conditions: \(a={477.5}\;\hbox {mol}\), \(f={1924.9}\;\hbox {mol}\) and \(b={198.7}\;\hbox {mol}\), the steady state solution in the absence of any noise.

In a first set of simulations, each cell is discretized with only one SCE per cell. This removes any intrinsic noise in the system that might result from spatial inhomogeneities in cell contact areas of cells discretized with many elements. The cell conformation corresponds to the situation of a 2D regular, hexagonal grid. To initiate the pattern forming process, we add a defined amount of uniform multiplicative noise (0.01, 5, 10 and 50 %) and study the pattern forming dynamics. For each noise level, we ran 19 simulations and present the patterning intensity \(I\) (see Fig. 17).

For all five investigated noise levels, we observe an initial steep increase of the intensity until a maximum level is reached, followed by a drop to about \(80-85\%\) of the maximum. Thereafter, it rises again by \(\sim 1\%\) and stabilizes. The steady state intensity represents the formation of a stable pattern and is located around the same level for all simulated noise levels, whereas the maximally attained intensity is slightly lower in the presence of \(50\%\) noise as compared to the other noise levels. However, a striking difference can be seen in the transient phase during the beginning of the simulations. Whereas the slope of the intensity is similar for all simulated noise levels, they are shifted along the time axis with the maximum appearing earlier for higher simulated noise levels. This indicates that pattern formation in the studied system is faster as the initial noise level is increased, which is consistent with previous observations [56].

The observations above suggest that pattern formation in juxtacrine signaling networks that feature a feedback mechanism are not only dependent on noise but converge faster as the noise level increases. In multicellular tissues, intrinsic noise is present in the form of local variations in membrane bound protein expression and cell shape irregularity.

We report the evolution of the patterning intensity \(I\) for a varying number of SCEs per cell in Fig. 18. We note that the intrinsic noise in the system is sufficient to initiate pattern formation in all five configurations and an intensity plateau is reached after 800 min. Again, all intensity curves show a steep initial increase until they reach a maximum intensity. For 25 and more SCEs per cell, the maximum is reached at around 400 min and the attained intensity maximum increases with the number of SCEs. After reaching the maximum in \(I\), the intensity level decreases to stabilize at around 800 min. In contrast to the situation of one SCE per cell and different noise levels, the intensity functions for different numbers of SCEs do not collapse to the same value.From the presented results we deduce that higher numbers of SCEs lead to a higher intensity plateau. We also note that these steady state intensity levels are higher than what was observed for the single SCE case. The situation for 10 SCEs is markedly different from the other configurations. The characteristic peak in \(I\) is not present and the intensity converges to a much higher value than what is observed for the other cases.

For higher numbers of SCEs (>25) we conclude that the intrinsic noise of the SEM and its variations due to cell resolution are different from what was observed for added initial noise simulations of single SCE cells. Cell resolution does not affect the speed of pattern formation but influences the observed intensity level. Comparing the time it takes to reach the maximal intensity, the simulations with many SCEs roughly correspond to the ones with a single SCE with 5 % added noise. The differences in the intensity plateau level could be a result of the intrinsic noise (a result of spatial contact inhomogeneities) that is continuously influencing the system as opposed to noise that is added at the beginning of the simulation.

## 5 Summary and outlook

In this work, we presented SEM++, an extension to the Subcellular Element Method [45, 57] that explicitly considers sub-cellular components such as the cell nucleus and the cell membrane. We introduce modifications to the potential functions of the Subcellular Elements that allow for polymerization and depolymerization as observed during migration and active cytoskeleton remodeling. SEM++ introduces a refined, unsupervised proliferation mechanism, actin polymerization driven cell migration and membrane protein mediated juxtacrine signaling.

SEM++ is implemented in an object-oriented, standalone C++ framework that allows for fast prototyping and model development. Furthermore, the methods have been integrated into the LAMMPS framework [33, 51] to benefit from its highly efficient and parallel implementation for particle simulations.

The model is well suited to resolve shape, structure and spatial interactions between cells. The computational elements inside the cell could further be used to model intra-cellular transport and signaling pathways operating in different locations of the cell. To accurately account for the extracellular milieu, the model needs to be extended or coupled to a model that captures the interaction of the cells with the surrounding ECM and can simulate signaling and transport processes inside the ECM. Since the method relies on discrete computational elements, it can readily be coupled to multiscale particle methods which can be used resolve the processes within the ECM [30].

Currently, the SEM does not consider directionality or heterogeneity inside a cell. The cytoskeleton of a cell is by no means a homogeneous network, as it is composed of oriented and elongated fiber bundles that are cross-linked to give the cell structure and shape. This could limit the application of the SEM in scenarios where cells at equilibrium assume an elongated shape, such as epithelial cells of the connective tissue or neurons. The introduction of spatially varying SCE potentials, elliptically shaped potential functions or fixed bonds in-between SCEs could address these limitations. Ongoing work focuses on the coupling of particle model with experimental data [20, 40, 41] and the quantification of uncertainties in the force field following related work of our group on Molecular Dynamics [6].

## Notes

### Acknowledgments

We wish to thank the Swiss National Supercomputing Centre (CSCS) for providing high performance computing platforms for this work.

### References

- 1.Alberts B, Johnson A, Lewis J, Raff M, Roberts K, Walter P (2008) Molecular biology of the cell, 5th edn. Garland Science, New YorkGoogle Scholar
- 2.Alt W, Adler F, Chaplain M, Deutsch A, Dress A, Krakauer D, Tranquillo RT, Anderson ARA, Chaplain MAJ, Rejniak KA, Newman T (2007) Modeling multicellular structures using the subcellular element model, Birkhäuser, BaselGoogle Scholar
- 3.Amber (2012) http://ambermd.org. Accessed 21 April 2012
- 4.Anderson ARA (2005) A hybrid mathematical model of solid tumour invasion: the importance of cell adhesion. Math Med Biol 22(2):163–186MATHCrossRefGoogle Scholar
- 5.Anderson ARA, Chaplain MAJ, Rejniak KA (eds) (2007) Single-cell-based models in biology and medicine. Birkhäuser, BaselGoogle Scholar
- 6.Angelikopoulos P, Papadimitriou C, Koumoutsakos P (2012) Bayesian uncertainty quantification and propagation in molecular dynamics simulations: a high performance computing framework. J Chem Phys 137(14):144,103CrossRefGoogle Scholar
- 7.Auger A, Chatelain P, Koumoutsakos P (2006) R-leaping: accelerating the stochastic simulation algorithm by reaction leaps. J Chem Phys 125(8):084,103CrossRefGoogle Scholar
- 8.Bausch AR, Moller W, Sackmann E (1999) Measurement of local viscoelasticity and forces in living cells by magnetic tweezers. Biophys J 76(1):573–579CrossRefGoogle Scholar
- 9.Bayati B, Chatelain P, Koumoutsakos P (2011) Adaptive mesh refinement for stochastic reaction–diffusion processes. J Comput Phys 230(1):13–26MATHMathSciNetCrossRefGoogle Scholar
- 10.Burg MB (2002) Response of renal inner medullary epithelial cells to osmotic stress. Comp Biochem Phys A 133(3):661–666CrossRefGoogle Scholar
- 11.Cascales JJL, de la Torre JG (1991) Simulation of polymer chains in elongational flow. steady-state properties and chain fracture. J Chem Phys 95(12):9384–9392CrossRefGoogle Scholar
- 12.Cickovski T, Aras K, Swat M, Merks RMH, Glimm T, Hentschel HGE, Alber MS, Glazier JA, Newman SA, Izaguirre JA (2007) From genes to organisms via the cell: a problem-solving environment for multicellular development. Comput Sci Eng 9(4):50–60CrossRefGoogle Scholar
- 13.Shaw DE (2012) Desmond. http://www.deshawresearch.com/resources.html. Accessed 21 April 2012
- 14.Desprat N, Richert A, Simeon J, Asnacios A (2005) Creep function of a single living cell. Biophys J 88(3):2224–2233CrossRefGoogle Scholar
- 15.Düchting W, Vogelsaenger T (1985) Recent progress in modelling and simulation of three-dimensional tumor growth and treatment. Biosystems 18(1):79–91CrossRefGoogle Scholar
- 16.Farhadifar R, Roper JC, Algouy B, Eaton S, Julicher F (2007) The influence of cell mechanics, cell-cell interactions, and proliferation on epithelial packing. Curr Biol 17(24):2095–2104CrossRefGoogle Scholar
- 17.Frieboes HB, Cristini V, Lowengrub J (2010) Continuum tumor modeling: single phase. Cambridge University Press, CambridgeGoogle Scholar
- 18.Frieboes HB, Jin F, Cristini V, Lowengrub J (2010) Continuum tumor modeling: multi phase. Cambridge University Press, CambridgeGoogle Scholar
- 19.Friedl P, Bröcker EB (2000) The biology of cell locomotion within three-dimensional extracellular matrix. Cell Mol Life Sci 57(1):41–64CrossRefGoogle Scholar
- 20.Gebaeck T, Schulz MMP, Koumoutsakos P, Detmar M (2009) Tscratch: a novel and simple software tool for automated analysis of monolayer wound healing assays. Biotechniques 46(4):265–274Google Scholar
- 21.Gillespie DT (1976) A general method for numerically simulating the stochastic time evolution of coupled chemical reactions. J Comput Phys 22(4):403–434MathSciNetCrossRefGoogle Scholar
- 22.Gillespie DT (2001) Approximate accelerated stochastic simulation of chemically reacting systems. J Chem Phys 115(4):1716–1733CrossRefGoogle Scholar
- 23.Glazier JA, Graner F (1993) Simulation of the differential adhesion driven rearrangement of biological cells. Phys Rev E 47(3):2128–2154CrossRefGoogle Scholar
- 24.Guilak F, Tedrow JR, Burgkart R (2000) Viscoelastic properties of the cell nucleus. Biochem Biophys Res Commun 269(3):781–786CrossRefGoogle Scholar
- 25.Hamant O, Heisler MG, Jonsson H, Krupinski P, Uyttewaal M, Bokov P, Corson F, Sahlin P, Boudaoud A, Meyerowitz EM, Couder Y, Traas J (2008) Developmental patterning by mechanical signals in arabidopsis. Science 322(5908):1650–1655CrossRefGoogle Scholar
- 26.Ising E (1925) Beitrag zur theorie des ferromagnetismus. Z Phys A-Hadron Nucl 31(1):253–258Google Scholar
- 27.Khain E, Sander LM (2006) Dynamics and pattern formation in invasive tumor growth. Phys Rev Lett 96(18):188,103CrossRefGoogle Scholar
- 28.Kierzkowski D, Nakayama N, Routier-Kierzkowska AL, Weber A, Bayer E, Schorderet M, Reinhardt D, Kuhlemeier C, Smith RS (2012) Elastic domains regulate growth and organogenesis in the plant shoot apical meristem. Science 335(6072):1096–1099CrossRefGoogle Scholar
- 29.Kim DH, Han K, Gupta K, Kwon KW, Suh KY, Levchenko A (2009) Mechanosensitivity of fibroblast cell shape and movement to anisotropic substratum topography gradients. Biomaterials 30(29):5433–5444CrossRefGoogle Scholar
- 30.Koumoutsakos P (2005) Multiscale flow simulations using particles. Annu Rev Fluid Mech 37:457–487MathSciNetCrossRefGoogle Scholar
- 31.Koumoutsakos P, Bayati B, Milde F, Tauriello G (2011) Particle simulations of morphogenesis. Math Models Methods Appl Sci 21:955–1006MATHMathSciNetCrossRefGoogle Scholar
- 32.Koumoutsakos P, Pivkin I, Milde F (2013) The fluid mechanics of cancer and its therapy. Annu Rev Fluid Mech 45(1):325–355MathSciNetCrossRefGoogle Scholar
- 33.LAMMPS (2012) Molecular dynamics simulator. http://lammps.sandia.gov. Accessed 21 April 2012
- 34.Lemon G, King JR, Byrne HM, Jensen OE, Shakesheff KM (2006) Mathematical modelling of engineered tissue growth using a multiphase porous flow mixture theory. J Math Biol 52(5):571–594MATHMathSciNetCrossRefGoogle Scholar
- 35.Macklin P, Edgerton ME, Lowengrub JS, Cristini V, Lowengrub JS (2010) Discrete cell modeling. Cambridge University Press, CambridgeCrossRefGoogle Scholar
- 36.Massagué J (1990) Transforming growth factor-\(\alpha \). A model for membrane-anchored growth factors. J Biol Chem 21(35):21,393–21,396Google Scholar
- 37.Meyer M, Müller AK, Yang J, Moik D, Ponzio G, Ornitz DM, Grose R, Werner S (2012) FGF receptors 1 and 2 are key regulators of keratinocyte migration in vitro and in wounded skin. J Cell Sci 125(23):5690–5701CrossRefGoogle Scholar
- 38.Micoulet A, Spatz JP, Ott A (2005) Mechanical response analysis and power generation by single-cell stretching. Chem Phys Chem 6(4):663–670Google Scholar
- 39.Milde F, Bergdorf M, Koumoutsakos P (2008) A hybrid model for three-dimensional simulations of sprouting angiogenesis. Biophys J 95(7):3146–3160CrossRefGoogle Scholar
- 40.Milde F, Franco D, Ferrari A, Kurtcuoglu V, Poulikakos D, Koumoutsakos P (2012) Cell image velocimetry (CIV): boosting the automated quantification of cell migration in wound healing assays. Integr Biol 4(11):1437–1447CrossRefGoogle Scholar
- 41.Milde F, Lauw S, Koumoutsakos P, Iruela-Arispe ML (2013) The mouse retina in 3d: Quantification of vascular growth and remodeling. Integr Biol 5(12):1426–1438CrossRefGoogle Scholar
- 42.Muñoz JJ, Barrett K, Miodownik M (2007) A deformation gradient decomposition method for the analysis of the mechanics of morphogenesis. J Biomech 40(6):1372–1380CrossRefGoogle Scholar
- 43.Nagai T, Honda H (2009) Computer simulation of wound closure in epithelial tissues: Cell-basal-lamina adhesion. Phys Rev E 80(6):061,903CrossRefGoogle Scholar
- 44.NAMD (2012) Scalable molecular dynamics. http://www.ks.uiuc.edu/Research/namd. Accessed 21 April 2012
- 45.Newman TJ (2005) Modeling multicellular systems using subcellular elements. Math Biosci Eng 2(3):613–624MATHMathSciNetCrossRefGoogle Scholar
- 46.Newman TJ (2008) Grid-free models of multicellular systems, with an application to large-scale vortices accompanying primitive streak formation. In: Schnell S, Maini PK, Newman SA, Newman TJ (eds) Multiscale modeling of developmental systems, current topics in developmental biology, vol 81. Academic Press, New York, pp 157–182CrossRefGoogle Scholar
- 47.Owen MR, Sherratt JA (1998) Mathematical modelling of juxtacrine cell signalling. Math Biosci 153(2):125–150MATHCrossRefGoogle Scholar
- 48.Owen MR, Sherratt JA, Wearing HJ (2000) Lateral induction by juxtacrine signaling is a new mechanism for pattern formation. Dev Biol 217(1):54–61CrossRefGoogle Scholar
- 49.Pathmanathan P, Cooper J, Fletcher A, Mirams G, Murray P, Osborne J, Pitt-Francis J, Walter A, Chapman SJ (2009) A computational study of discrete mechanical tissue models. Phys Biol 6(3):036,001CrossRefGoogle Scholar
- 50.Peskin CS (1972) Flow patterns around heart valves: a numerical method. J Comput Phys 10(2):252–271MATHCrossRefGoogle Scholar
- 51.Plimpton S (1995) Fast parallel algorithms for short-range molecular dynamics. J Comput Phys 117(1):1–19MATHCrossRefGoogle Scholar
- 52.Puthur R, Sebastian KL (2002) Theory of polymer breaking under tension. Phys Rev B 66(2):024,304CrossRefGoogle Scholar
- 53.Rejniak K, Kliman H, Fauci L (2004) A computational model of the mechanics of growth of the villous trophoblast bilayer. Bull Math Biol 66(2):199–232MathSciNetCrossRefGoogle Scholar
- 54.Rejniak KA (2007) An immersed boundary framework for modelling the growth of individual cells: an application to the early tumour development. J Theor Biol 247(1):186–204MathSciNetCrossRefGoogle Scholar
- 55.Rubin MB, Bodner SR (2002) A three-dimensional nonlinear model for dissipative response of soft tissue. Int J Solids Struct 39(19):5081–5099MATHCrossRefGoogle Scholar
- 56.Rudge T, Burrage K (2008) Effects of intrinsic and extrinsic noise can accelerate juxtacrine pattern formation. Bull Math Biol 70(4):971–991MATHMathSciNetCrossRefGoogle Scholar
- 57.Sandersius SA, Newman TJ (2008) Modeling cell rheology with the subcellular element model. Phys Biol 5(1):015,002 Google Scholar
- 58.Sandersius SA, Chuai M, Weijer CJ, Newman TJ (2011a) A ‘chemotactic dipole’ mechanism for large-scale vortex motion during primitive streak formation in the chick embryo. Phys Biol 8(4):045008CrossRefGoogle Scholar
- 59.Sandersius SA, Weijer CJ, Newman TJ (2011b) Emergent cell and tissue dynamics from subcellular modeling of active biomechanical processes. Phys Biol 8(4):045007CrossRefGoogle Scholar
- 60.Swanson KR, Bridge C, Murray JD, Alvord EC Jr (2003) Virtual and real brain tumors: using mathematical modeling to quantify glioma growth and invasion. J Neurol Sci 216(1):1–10CrossRefGoogle Scholar
- 61.Wearing HJ, Owen MR, Sherratt JA (2000) Mathematical modelling of juxtacrine patterning. Bull Math Biol 62(2):293–320CrossRefGoogle Scholar
- 62.Wottawah F, Schinkinger S, Lincoln B, Ananthakrishnan R, Romeyke M, Guck J, Käs J (2005) Optical rheology of biological cells. Phys Rev Lett 9(94):098,103CrossRefGoogle Scholar