That’s the second blog post of three reflecting on the modeling.

  1. Performance - read this first !
  2. Future plans - the future is already here
  3. Visualization - even more future !


I started optimizing, then refactoring the leg-joint code to gain performance. As I was doing this, I realized it would be cleverer to go all the way back to the drawing board (actually a paper notebook), as maybe some parts of the simulation could be re-used with different geometries. After all, many epithelia share common traits with the leg imaginal disk. At the same time, Magali and I started talking with two physicists, François Molino and Cyprien Gay, who are applying bleeding edge soft-matter physics to biological tissues. It turns out Cyprien and his colleges recently published a milestone paper on continuous models; both were keen on trying cell based models, and the leg disk is of course very rich on this regard. So while I’m generalizing geometry, why not physics?

The road ahead

So how do we get there?

  1. Classify existing biophysical tissue models.

  2. Specify the subset of said models we want to address.

  3. Try to craft an API.

  4. Choose the proper libraries.

  5. Code simple test cases (simple geometry, simple physics).

  6. Repeat from 3 until the API and tech are +/- stable…

  7. Implement a real world problem, same physics as in 5.

  8. Try new physics at constant geometry.

The route through those points will start a bit far from code.

A very brief tour of the biophysical models of living tissues

As they imply dramatic changes in a tissue shape and organization, morphogenetic events have long been the focus of mechanical modeling efforts. As early as 1917, D’Arcy Thomson exposed the underlying mathematical and physical relationships underpinning cells and multi-cellular organism shapes and dynamical behaviors (in the seminal work On Growth and Form). In recent years, technical progresses have provided very detailed geometrical, dynamical and biochemical data on morphogenetic processes, allowing for ever finer mathematical modeling and computer simulations. Meanwhile, many models and simulations have been developed. The interested reader will find a very clear, useful and complete discussion of the topic in C.V.T Tamulonis PhD dissertation. Seriously, I can’t stress enough how this work is nice and complete!

For now, I just drafted a tree of the different models, as a basis for the discussion:

Tissue models family tree

The underlying dragonfly wing is from Figure 162 of On Growth and Form (p.476 of the above linked edition). On the tree tips are examples (not even close to exhaustive) of implementation for each models. Bold names refer to softwares, slanted to a researcher and it’s publications.

Continuous models deal with phenomena that span multiple cells in size, such that the changes in shape can be smoothed out, the detailed cell-cell interactions are not necessary to understand the tissue’s shape. Star among this huge family are organ simulations, among which the heart. To grasp the state of the art on the matter, you should rush to see the video of a heart simulation by Shu Takagi’s group at Riken. Cell agregates, such as multi-cellular tumor spheroids, are studied and modeled as continuous tissues. The paper cited above (Sham Tilli et al, 2015) precisely sets down a formalism that includes cell biology specific components (i.e. cell population dynamics and re-arangements, more on that later) in a continuous dynamical (rheological, more precisely) framework. See Cyprien’s site for more details on that. Finally, I must mention the incredible OpenWorm project, which tackles the neurobiology of C. elegans and integrates it to a mechanical (particle based) framework.

For now let’s concentrate on Cell Based models, in which leg-joint fits.

By definition cell based models are described by a multi-agent design pattern. This is where biological tissues radically differ from any other material: cells can act by themselves and exchange information between each others, an individual’s behavior influences the overall shape of the tissue. We’ll see consequences of this for the API design. For now, let’s go further down the tree.

The next branching is between lattice based and lattice free models.

Lattice based were developed first, inheriting directly from early cybernetics research (Norman Wiener, John Von Neumann) on cellular automata. Conway’s game of life was described in 1970, and you can find a nice python implementation by Jake VanderPlas himself here. It’s not a biological tissue model, really, but it captures the essence of lattice based modeling: cells are pixels or collections of such on a fixed grid. The evolution of the system is solved by looking at interactions between each pixel and it’s neighbors. That’s what goes on in a more detailed manner in cellular Potts models and their descendant the Glazier-Graner-Hogweg (GGH) model. These models are well established, and CompuCell3D provides a very handy, optimized and scriptable software for defining and running GGH simulation in 2D and 3D geometries. Lattice based models are well suited to study phenomena such as collective migration (e.g. tumor invasion) or cell population dynamics. As partial differential equations can be solved across the grid, they can also deal with reaction-diffusion mechanisms, and thus signaling. That’s all nice and well, but the grid is also a constrain (if precise shapes are of interest), and the physics governing pixel state transition is very phenomenological, it does not really capture the mechanical aspects of the tissue.

In lattice free models, the system’s space (usually 2 or 3 dimensional) is continuous and the objects are described by their metric in that space. An early split is between models made of descrete spherical elements and the ones relying on a vector based description. In the former class, cells are described as spheres (like in Chaste) or smaller particle clouds, as in the work by P.E van Liedekerke et al. cited by Tamulonis.

The later branch divides into finite elements models (see CVT Tamulonis again) and vertex models, where leg-joint and tyssue fit (ouf! as we say in French). Thanks to the close correspondence between those model architecture and the cell boundaries, they are well adapted to the description of contiguous, one cell thick tissues as the epithelium we’re interested in.

Now that we know where we are in the grand scheme of things, let’s dig on the library’s structure. We’re at step 3 already!

Library architecture and API design.

Prolégomène: libraries vs standalone software.

Software can come in various forms. Traditionally, and as is the case for most of the works presented above, simulations would be developed in a compiled language such as C++ (the uncontested giant in the field) and distributed as standalone executables. I’m a python user, and prefer to use libraries, and develop by importing what I need when I need it. I feel that doing development in the Jupyter notebook gives me a lot of freedom to explore and hack. The user/developer frontier is blurry in the scipy community for a good reason: that’s a very efficient way of developping software. So, contrary to e.g. Chaste, tyssue is designed as a modular, hackable library.

Object architecture and design patterns

Objects to consider

To fix the ideas, let’s work on a minimal 2D example of what we try to model. Generalizing to more complex geometries is deferred to further headaches :-p.

Here is our 2D three cells epithelium:

A minimal 3 cells epithelium

The epithelium contains cells indexed by Greek letters, junction vertices, indexed by Latin letters, and junction edges. The blue links denote a neighborhoud relation between two cells. In 2D, the edges can be described as pairs of Halfedges (see the documentation on Halfedge Data Structures in CGAL). In 3D, this concept is generalized by the Linear cell complex structure, made of connected Darts. Both Halfedges and Darts hold information on their source, target and the cell they are associated to. So the pair of Halfedges between vertices \(i\) and \(j\) should be indexed as \(i j, \alpha\) and \(j i, \beta\). In 3D, there will be a fourth index for a given Dart, giving the associated face of the cell, see the discussion on Darts orbit and \(\beta_i\) operators on CGAL’s doc.

The ensemble of those objects and their relations is called the topology of the system. Perhaps abusively, said topology also includes the set of geometrical points associated with the vertices (but nothing more, see below).

Data Structures

To the topology is associated data: geometric characteristics, parameter values and so on. We want to be able to get and set these data by single elements or through fancy indexing. Ideally, without copying it, and in a transparent way to the python user. But the above mentioned concepts are well defined and optimized in CGAL, and it would be a waste not to rely on all this good work. Yet, most of this data is irrelevant to CGAL: we could for example associate a color to a cell for representation purpose, that needs to be dynamically allocated, etc., all that in an interactive python session. graph-tool does a very good job at managing that with PropertyMaps, but we saw it was not fitting exactly our needs. Wrapping C++ and Numpy arrays is not trivial.

So here is how I see this: let the C++ side of things completely ignore the data — to the exception of the positions of the vertices in space, which CGAL need, and will receive special treatment —, and just let it manage topology. This way the only CellAttribute associated with a CGAL object is its index (plus the Point for vertices).

The CGAL/python interface is then just a matter of passing the indices (as std::vectors<int>) in read only mode to python, and updating the points back and forth. Python side, the core data structure is then comprised of 3 DataFrames (4 in 3D actually):

  1. cell_df, indexed by the Index cell_idx

  2. jv_df, indexed by the Index jv_idx

  3. je_df, indexed by the MultiIndex je_idx, itself comprised of a (srce, trgt, cell) triple (e.g. \({i, j, \alpha}\)). In 3D, it would be a quadruple (srce, trgt, face, cell).

Here is a sketch summarizing the above, along with the behavior and visualization aspects I’ll discuss next.

Data flows and management

You can find a toy implementation (CGAL independent) in this notebook, along with examples of simple computation combining data from cells and junctions.

A note on time

I haven’t spoken of the time dimension yet. The above description gives a static view of the tissue. Of course, the whole goal is to look at evolutions. Time will be a global attribute of the system. For all the DataFrames, we can construct a Panel where the fourth dimension is the time component, stacking up static views of the tissue (this can also be achieved via a supplementary ‘t’ index for each dataframe). Whether this is feasible, or it’s better to record the data at each time step is to be determined. For small systems, the former will be easier, but might not scale, and some kind of buffering might be needed (to be continued, data management is not my strong suit).

External constrains and supra cellular components

Further down the road, it might be necessary to include other elements, for example the extra-cellular matrix, which by definition is not included in this framework. If its shape is simple and static enough, this could be described in terms of a force field in the space surrounding the tissue. One could also envision a mixed continuous - cell based model, where the ECM is described as a finite elements triangulated volume. It is not clear to me how to manage contact points here, I’m sure that will be fun. Apart from the ECM, one can think of trans-cellular actin cables or an egg shell constraining the epithelium.

On that matter, management of contact points and mesh collision is not trivial, but it looks like the great folk at CGAL have this sorted out for us.


The architecture described above describes the state space of the epithelium, and it’s associated parameters. We can then add physical data: forces or gradients, for example. If the specific columns to consider depend on the physics engine, the resolution of the dynamical equations (wether through gradient descent, ODEs, etc.) should be independent of the topology at one point during the simulation.

Agent-like behavior

As I said earlier, cells are not passive chunks of material, but individuals displaying different behaviors, either individually or collectively. In this sense, cells are agents. This must be reflected in the library architecture as to make it easy to translate in the simulation the biologist’s insight of the modeled biological scenario. Here are some examples:

  • Cell growth
  • Cell division
  • Cell intercalation (aka Type 1 transition)

  • Apoptosis

For each of those behaviors, one cell or a group of cells will be implicated (the actors), and some specification on the physics involved that might look like a list of actuators (I’m thinking for example at the actin apico-basal cable in the fold formation scenario) specifying the interactions and their application points. Every behavior can trigger a change in topology, requiring a re-indexing from CGAL, and sets the system off-equilibrium, which is resolved by the physics engine. Those concept are still in early development, and the API is still sketchy here.

Events, signals and asynchronicity

Associated with the multi-agent pattern comes the idea that those agents, the cells, could act asynchronously, each behavior sending signals to neighboring cells and the hole epithelium. At each time step, we can imagine to gather all the ongoing behaviors (cell 123 might divide, while cells 234, 235, 236, and 244 undergo a type 1 transition) and solve the physics system only once for the whole tissue. Once again, this is still a bit sketchy, and any comments are welcome.

Next step: data viz!

This vast subject (3D! 3D+time!, vispy!, webGL!) will wait next post.