Category: 3D

Building a Cellular Automaton Model Using BioFVM

Note: This is part of a series of “how-to” blog posts to help new users and developers of BioFVM. See below for guides to setting up a C++ compiler in Windows or OSX.

What you’ll need

Matlab or Octave for visualization. Matlab might be available for free at your university. Octave is open source and available from a variety of sources.

Our modeling task

We will implement a basic 3-D cellular automaton model of tumor growth in a well-mixed fluid, containing oxygen pO2 (mmHg) and a drug c (e.g., doxorubicin, μM), inspired by modeling by Alexander Anderson, Heiko Enderling, Jan PoleszczukGibin Powathil, and others. (I highly suggest seeking out the sophisticated cellular automaton models at Moffitt’s Integrated Mathematical Oncology program!) This example shows you how to extend BioFVM into a new cellular automaton model. I’ll write a similar post on how to add BioFVM into an existing cellular automaton model, which you may already have available.

Tumor growth will be driven by oxygen availability. Tumor cells can be live, apoptotic (going through energy-dependent cell death, or necrotic (undergoing death from energy collapse). Drug exposure can both trigger apoptosis and inhibit cell cycling. We will model this as growth into a well-mixed fluid, with pO2 = 38 mmHg (about 5% oxygen: a physioxic value) and c = 5 μM.

Mathematical model

As a cellular automaton model, we will divide 3-D space into a regular lattice of voxels, with length, width, and height of 15 μm. (A typical breast cancer cell has radius around 9-10 μm, giving a typical volume around 3.6×103 μm3. If we make each lattice site have the volume of one cell, this gives an edge length around 15 μm.)

In voxels unoccupied by cells, we approximate a well-mixed fluid with Dirichlet nodes, setting pO2 = 38 mmHg, and initially setting c = 0. Whenever a cell dies, we replace it with an empty automaton, with no Dirichlet node. Oxygen and drug follow the typical diffusion-reaction equations:

\[ \frac{ \partial \textrm{pO}_2 }{\partial t} = D_\textrm{oxy} \nabla^2 \textrm{pO}_2 – \lambda_\textrm{oxy} \textrm{pO}_2 – \sum_{ \textrm{cells} i} U_{i,\textrm{oxy}} \textrm{pO}_2 \]

\[ \frac{ \partial c}{ \partial t }  = D_c \nabla^2 c – \lambda_c c – \sum_{\textrm{cells }i} U_{i,c} c \]

where each uptake rate is applied across the cell’s volume. We start the treatment by setting c = 5 μM on all Dirichlet nodes at t = 504 hours (21 days). For simplicity, we do not model drug degradation (pharmacokinetics), to approximate the in vitro conditions.

In any time interval [t,tt], each live tumor cell i has a probability pi,D of attempting division, probability pi,A of apoptotic death, and probability pi,N of necrotic death. (For simplicity, we ignore motility in this version.) We relate these to the birth rate bi, apoptotic death rate di,A, and necrotic death rate di,N by the linearized equations (from Macklin et al. 2012):

\[ \textrm{Prob} \Bigl( \textrm{cell } i \textrm{ becomes apoptotic in } [t,t+\Delta t] \Bigr)  = 1 – \textrm{exp}\Bigl( -d_{i,A}(t) \Delta t\Bigr) \approx d_{i,A}\Delta t \]

\[ \textrm{Prob} \Bigl( \textrm{cell } i \textrm{ attempts division in } [t,t+\Delta t] \Bigr)  = 1 – \textrm{exp}\Bigl( -b_i(t) \Delta t\Bigr) \approx b_{i}\Delta t \]

\[ \textrm{Prob} \Bigl( \textrm{cell } i \textrm{ becomes necrotic in } [t,t+\Delta t] \Bigr)  = 1 – \textrm{exp}\Bigl( -d_{i,N}(t) \Delta t\Bigr) \approx d_{i,N}\Delta t \]

Each dead cell has a mean duration Ti,D, which will vary by the type of cell death. Each dead cell automaton has a probability pi,L of lysis (rupture and removal) in any time span [t,t+Δt]. The duration TD is converted to a probability of cell lysis by

\[ \textrm{Prob} \Bigl( \textrm{dead cell } i \textrm{ lyses in } [t,t+\Delta t] \Bigr)  = 1 – \textrm{exp}\Bigl( -\frac{1}{T_{i,D}} \Delta t\Bigr) \approx \frac{ \Delta t}{T_{i,D}} \]

(Illustrative) parameter values

We use Doxy = 105 μm2/min (Ghaffarizadeh et al. 2016), and we set Ui,oxy = 20 min-1 (to give an oxygen diffusion length scale of about 70 μm, with steeper gradients than our typical 100 μm length scale). We set λoxy = 0.01 min-1 for a 1 mm diffusion length scale in fluid.

We set Dc = 300 μm2/min, and Uc = 7.2×10-3 min-1 (Dc from Weinberg et al. (2007), and Ui,c twice as large as the reference value in Weinberg et al. (2007) to get a smaller diffusion length scale of about 204 μm). We set λc = 3.6×10-5 min-1 to give a drug diffusion length scale of about 2.9 mm in fluid.

We use TD = 8.6 hours for apoptotic cells, and TD = 60 days for necrotic cells (Macklin et al., 2013). However, note that necrotic and apoptotic cells lose volume quickly, so one may want to revise those time scales to match the point where a cell loses 90% of its volume.

Functional forms for the birth and death rates

We model pharmacodynamics with an area-under-the-curve (AUC) type formulation. If c(t) is the drug concentration at any cell i‘s location at time t, then let its integrated exposure Ei(t) be

\[ E_i(t) = \int_0^t c(s) \: ds \]

and we model its response with a Hill function

\[ R_i(t) = \frac{ E_i^h(t) }{ \alpha_i^h + E_i^h(t) }, \]

where h is the drug’s Hill exponent for the cell line, and α is the exposure for a half-maximum effect.

We model the microenvironment-dependent birth rate by:

\[ b_i(t) = \left\{ \begin{array}{lr} b_{i,P} \left( 1 – \eta_i R_i(t) \right) & \textrm{ if } \textrm{pO}_{2,P} < \textrm{pO}_2 \\  \\ b_{i,P} \left( \frac{\textrm{pO}_{2}-\textrm{pO}_{2,N}}{\textrm{pO}_{2,P}-\textrm{pO}_{2,N}}\right) \Bigl( 1 – \eta_i R_i(t) \Bigr) & \textrm{ if } \textrm{pO}_{2,N} < \textrm{pO}_2 \le \textrm{pO}_{2,P} \\ \\ 0 & \textrm{ if } \textrm{pO}_2 \le \textrm{pO}_{2,N}\end{array} \right. \]

where pO2,P is the physioxic oxygen value (38 mmHg), and pO2,N is a necrotic threshold (we use 5 mmHg), and 0 < η < 1 the drug’s birth inhibition. (A fully cytostatic drug has η = 1.)

We model the microenvironment-dependent apoptosis rate by:

\[ d_{i,A}(t)  = d_{i,A}^* + \Bigl( d_{i,A}^\textrm{max} – d_{i,A}^* \Bigr) R_i(t)  \]

where di,Amax is the maximum apoptotic death rate. We model the microenvironment-dependent necrosis rate by:

\[ d_{i,N}(t) = \left\{ \begin{array}{lr} 0 & \textrm{ if } \textrm{pO}_{2,N} < \textrm{pO}_{2} \\ \\ d_{i,N}^* &  \textrm{ if } \textrm{pO}_{2} \le \textrm{pO}_{2,N} \end{array}\right. \]

for a constant value di,N*.
(Illustrative) parameter values

We use bi,P = 0.05 hour-1 (for a 20 hour cell cycle in physioxic conditions), di,A* = 0.01 bi,P, and di,N* = 0.04 hour-1 (so necrotic cells survive around 25 hours in low oxygen conditions).

We set α = 30 μM*hour (so that cells reach half max response after 6 hours’ exposure at a maximum concentration c = 5 μM), h = 2 (for a smooth effect), η = 0.25 (so that the drug is partly cytostatic), and di,Amax = 0.1 hour^-1 (so that cells survive about 10 hours after reaching maximum response).

Building the Cellular Automaton Model in BioFVM

BioFVM already includes Basic_Agents for cell-based substrate sources and sinks. We can extend these basic agents into full-fledged automata, and then arrange them in a lattice to create a full cellular automata model. Let’s sketch that out now.

Extending Basic_Agents to Automata

The main idea here is to define an Automaton class which extends (and therefore includes) the Basic_Agent class. This will give each Automaton full access to the microenvironment defined in BioFVM, including the ability to secrete and uptake substrates. We also make sure each Automaton “knows” which microenvironment it lives in (contains a pointer pMicroenvironment), and “knows” where it lives in the cellular automaton lattice. (More on that in the following paragraphs.)

So, as a schematic (just sketching out the most important members of the class):

class Standard_Data; // define per-cell biological data, such as phenotype, 
	// cell cycle status, etc..
class Custom_Data; // user-defined custom data, specific to a model.

class Automaton : public Basic_Agent
		Microenvironment* pMicroenvironment; 
		CA_Mesh* pCA_mesh; 
		int voxel_index; 
		// neighbor connectivity information
		std::vector<Automaton*> neighbors; 
		std::vector<double> neighbor_weights; 
		Standard_Data standard_data; 
		void (*current_state_rule)( Automaton& A , double ); 
		void copy_parameters( Standard_Data& SD  ); 
		void overwrite_from_automaton( Automaton& A );  

		void set_cellular_automaton_mesh( CA_Mesh* pMesh );  
		CA_Mesh* get_cellular_automaton_mesh( void ) const;  
		void set_voxel_index( int ); 
		int get_voxel_index( void ) const;  
		void set_microenvironment( Microenvironment* pME );  
		Microenvironment* get_microenvironment( void );  

		// standard state changes
		bool attempt_division( void ); 
		void become_apoptotic( void ); 
		void become_necrotic( void ); 
		void perform_lysis( void ); 
		// things the user needs to define
		Custom_Data custom_data;
		// use this rule to add custom logic 
		void (*custom_rule)( Automaton& A , double); 		

So, the Automaton class includes everything in the Basic_Agent class, some Standard_Data (things like the cell state and phenotype, and per-cell settings), (user-defined) Custom_Data, basic cell behaviors like attempting division into an empty neighbor lattice site, and user-defined custom logic that can be applied to any automaton. To avoid lots of switch/case and if/then logic, each Automaton has a function pointer for its current activity (current_state_rule), which can be overwritten any time.

Each Automaton also has a list of neighbor Automata (their memory addresses), and weights for each of these neighbors. Thus, you can distance-weight the neighbors (so that corner elements are farther away), and very generalized neighbor models are possible (e.g., all lattice sites within a certain distance).  When updating a cellular automaton model, such as to kill a cell, divide it, or move it, you leave the neighbor information alone, and copy/edit the information (standard_data, custom_data, current_state_rule, custom_rule). In many ways, an Automaton is just a bucket with a cell’s information in it.

Note that each Automaton also “knows” where it lives (pMicroenvironment and voxel_index), and knows what CA_Mesh it is attached to (more below).

Connecting Automata into a Lattice

An automaton by itself is lost in the world–it needs to link up into a lattice organization. Here’s where we define a CA_Mesh class, to hold the entire collection of Automata, setup functions (to match to the microenvironment), and two fundamental operations at the mesh level: copying automata (for cell division), and swapping them (for motility). We have provided two functions to accomplish these tasks, while automatically keeping the indexing and BioFVM functions correctly in sync. Here’s what it looks like:

class CA_Mesh{
		Microenvironment* pMicroenvironment; 
		Cartesian_Mesh* pMesh; 

		std::vector<Automaton> automata; 
		std::vector<int> iteration_order; 
		// setup to match a supplied microenvironment 
		void setup( Microenvironment& M ); 
		// setup to match the default microenvironment 
		void setup( void );  
		int number_of_automata( void ) const; 
		void randomize_iteration_order( void );

		void swap_automata( int i, int j ); 
		void overwrite_automaton( int source_i, int destination_i ); 
		// return the automaton sitting in the ith lattice site 
		Automaton& operator[]( int i ); 
		// go through all nodes according to random shuffled order 
		void update_automata( double dt ); 

So, the CA_Mesh has a vector of Automata (which are never themselves moved), pointers to the microenvironment and its mesh, and a vector of automata indices that gives the iteration order (so that we can sample the automata in a random order). You can easily access an automaton with operator[], and copy the data from one Automaton to another with overwrite_automaton() (e.g, for cell division), and swap two Automata’s data (e.g., for cell migration) with swap_automata().  Finally, calling update_automata(dt) iterates through all the automata according to iteration_order, calls their current_state_rules and custom_rules, and advances the automata by dt.

Interfacing Automata with the BioFVM Microenvironment

The setup function ensures that the CA_Mesh is the same size as the Microenvironment.mesh, with same indexing, and that all automata have the correct volume, and dimension of uptake/secretion rates and parameters. If you declare and set up the Microenvironment first, all this is take care of just by declaring a CA_Mesh, as it seeks out the default microenvironment and sizes itself accordingly:

// declare a microenvironment
Microenvironment M;
// do things to set it up -- see prior tutorials
// declare a Cellular_Automaton_Mesh
CA_Mesh CA_model;
// it's already good to go, initialized to empty automata:

If you for some reason declare the CA_Mesh fist, you can set it up against the microenvironment:

// declare a CA_Mesh
CA_Mesh CA_model;
// declare a microenvironment
Microenvironment M;
// do things to set it up -- see prior tutorials
// initialize the CA_Mesh to match the microenvironment
CA_model.setup( M );
// it's already good to go, initialized to empty automata:

Because each Automaton is in the microenvironment and inherits functions from Basic_Agent, it can secrete or uptake. For example, we can use functions like this one:

void set_uptake( Automaton&amp; A, std::vector<double>& uptake_rates )
	extern double BioFVM_CA_diffusion_dt; 
	// update the uptake_rates in the standard_data 
	A.standard_data.uptake_rates = uptake_rates; 
	// now, transfer them to the underlying Basic_Agent
	*(A.uptake_rates) = A.standard_data.uptake_rates; 
	// and make sure the internal constants are self-consistent 
	A.set_internal_uptake_constants( BioFVM_CA_diffusion_dt ); 

A function acting on an automaton can sample the microenvironment to change parameters and state. For example:

void do_nothing( Automaton& A, double dt )
{ return; }

void microenvironment_based_rule( Automaton& A, double dt )
	// sample the microenvironment 
	std::vector<double> MS = (*A.get_microenvironment())( A.get_voxel_index() );
	// if pO2 < 5 mmHg, set the cell to a necrotic state
	if( MS[0] < 5.0 ) { A.become_necrotic(); } // if drug > 5 uM, set the birth rate to zero
	if( MS[1] > 5 )
	{ A.standard_data.birth_rate = 0.0; }
	// set the custom rule to something else
	A.custom_rule = do_nothing;

Implementing the mathematical model in this framework

We give each tumor cell a tumor_cell_rule (using this for custom_rule):

void viable_tumor_rule( Automaton& A, double dt )
	// If there's no cell here, don't bother.
	if( A.standard_data.state_code == BioFVM_CA_empty )
	{ return; }
	// sample the microenvironment
	std::vector<double> MS = (*A.get_microenvironment())( A.get_voxel_index() );

	// integrate drug exposure 
	A.standard_data.integrated_drug_exposure += ( MS[1]*dt ); 
	A.standard_data.drug_response_function_value = pow( A.standard_data.integrated_drug_exposure,
		A.standard_data.drug_hill_exponent );  
	double temp = pow( A.standard_data.drug_half_max_drug_exposure,
		A.standard_data.drug_hill_exponent );  
	temp += A.standard_data.drug_response_function_value; 
	A.standard_data.drug_response_function_value /= temp; 
	// update birth rates (which themselves update probabilities) 
	update_birth_rate( A, MS, dt ); 
	update_apoptotic_death_rate( A, MS, dt ); 
	update_necrotic_death_rate( A, MS, dt );

The functional tumor birth and death rates are implemented as:

void update_birth_rate( Automaton& A, std::vector<double>& MS, double dt )  
	static double O2_denominator = BioFVM_CA_physioxic_O2 - BioFVM_CA_necrotic_O2;
	A.standard_data.birth_rate = 	A.standard_data.drug_response_function_value;
		// response
	A.standard_data.birth_rate *= A.standard_data.drug_max_birth_inhibition;
		// inhibition*response; 
	A.standard_data.birth_rate *= -1.0;
		// - inhibition*response
	A.standard_data.birth_rate += 1.0;
		// 1 - inhibition*response
	A.standard_data.birth_rate *= viable_tumor_cell.birth_rate; 
		// birth_rate0*(1 - inhibition*response)
	double temp1 = MS[0] ; // O2 
	temp1 -= BioFVM_CA_necrotic_O2; 
	temp1 /= O2_denominator; 
	A.standard_data.birth_rate *= temp1; 
	if( A.standard_data.birth_rate < 0 )
	{ A.standard_data.birth_rate = 0.0; }

	A.standard_data.probability_of_division = A.standard_data.birth_rate;
	A.standard_data.probability_of_division *= dt; 
		// dt*birth_rate*(1 - inhibition*repsonse) // linearized probability 
void update_apoptotic_death_rate( Automaton& A, std::vector<double>& MS, double dt )  
	A.standard_data.apoptotic_death_rate = A.standard_data.drug_max_death_rate; 
		// max_rate
	A.standard_data.apoptotic_death_rate -= viable_tumor_cell.apoptotic_death_rate; 
		// max_rate - background_rate
	A.standard_data.apoptotic_death_rate *= A.standard_data.drug_response_function_value; 
		// (max_rate-background_rate)*response 
	A.standard_data.apoptotic_death_rate += viable_tumor_cell.apoptotic_death_rate; 
		// background_rate + (max_rate-background_rate)*response 

	A.standard_data.probability_of_apoptotic_death = A.standard_data.apoptotic_death_rate;
	A.standard_data.probability_of_apoptotic_death *= dt; 
		// dt*( background_rate + (max_rate-background_rate)*response ) // linearized probability

void update_necrotic_death_rate( Automaton& A, std::vector<double>& MS, double dt )   
	A.standard_data.necrotic_death_rate = 0.0; 
	A.standard_data.probability_of_necrotic_death = 0.0; 
	if( MS[0] > BioFVM_CA_necrotic_O2 )
	{ return; }

	A.standard_data.necrotic_death_rate = perinecrotic_tumor_cell.necrotic_death_rate; 
	A.standard_data.probability_of_necrotic_death = A.standard_data.necrotic_death_rate; 
	A.standard_data.probability_of_necrotic_death *= dt; 
		// dt*necrotic_death_rate


And each fluid voxel (Dirichlet nodes) is implemented as the following (to turn on therapy at 21 days):

void fluid_rule( Automaton& A, double dt )
	static double activation_time = 504; 
	static double activation_dose = 5.0; 
	static std::vector<double> activated_dirichlet( 2 , BioFVM_CA_physioxic_O2 );
	static bool initialized = false; 
	if( !initialized )
		activated_dirichlet[1] = activation_dose; 
		initialized = true; 

	if( fabs( BioFVM_CA_elapsed_time - activation_time ) < 0.01 ) { int ind = A.get_voxel_index(); if( A.get_microenvironment()->mesh.voxels[ind].is_Dirichlet )
			A.get_microenvironment()->update_dirichlet_node( ind, activated_dirichlet );

At the start of the simulation, each non-cell automaton has its custom_rule set to fluid_rule, and each tumor cell Automaton has its custom_rule set to viable_tumor_rule. Here’s how:

void setup_cellular_automata_model( Microenvironment& M, CA_Mesh& CAM )
	// Fill in this environment 
	double tumor_radius = 150;  
	double tumor_radius_squared = tumor_radius * tumor_radius; 
	std::vector<double> tumor_center( 3, 0.0 ); 
	std::vector<double> dirichlet_value( 2 , 1.0 );
	dirichlet_value[0] = 38; //physioxia
	dirichlet_value[1] = 0; // drug 
	for( int i=0 ; i < M.number_of_voxels() ;i++ )
		std::vector<double> displacement( 3, 0.0 );
		displacement = M.mesh.voxels[i].center; 
		displacement -= tumor_center; 
		double r2 = norm_squared( displacement ); 
		if( r2 > tumor_radius_squared ) // well_mixed_fluid
			M.add_dirichlet_node( i, dirichlet_value ); 
			CAM[i].copy_parameters( well_mixed_fluid );
			CAM[i].custom_rule = fluid_rule; 
			CAM[i].current_state_rule = do_nothing; 	
		else // tumor 
			CAM[i].copy_parameters( viable_tumor_cell );
			CAM[i].custom_rule = viable_tumor_rule; 
			CAM[i].current_state_rule = advance_live_state; 		

Overall program loop

There are two inherent time scales in this problem: cell processes like division and death (happen on the scale of hours), and transport (happens on the order of minutes). We take advantage of this by defining two step sizes:

double BioFVM_CA_dt = 3;
std::string BioFVM_CA_time_units = "hr";
double BioFVM_CA_save_interval = 12;
double BioFVM_CA_max_time = 24*28;
double BioFVM_CA_elapsed_time = 0.0;

double BioFVM_CA_diffusion_dt = 0.05;

std::string BioFVM_CA_transport_time_units = "min";
double BioFVM_CA_diffusion_max_time = 5.0;

Every time the simulation advances by BioFVM_CA_dt (on the order of hours), we run diffusion to quasi-steady state (for BioFVM_CA_diffusion_max_time, on the order of minutes), using time steps of size BioFVM_CA_diffusion time. We performed numerical stability and convergence analyses to determine 0.05 min works pretty well for regular lattice arrangements of cells, but you should always perform your own testing!

Here’s how it all looks, in a main program loop:

BioFVM_CA_elapsed_time = 0.0;
double next_output_time = BioFVM_CA_elapsed_time; // next time you save data

while( BioFVM_CA_elapsed_time < BioFVM_CA_max_time + 1e-10 )
	// if it's time, save the simulation
	if( fabs( BioFVM_CA_elapsed_time - next_output_time ) < BioFVM_CA_dt/2.0 )
		std::cout << "simulation time: " << BioFVM_CA_elapsed_time << " " << BioFVM_CA_time_units
			<< " (" << BioFVM_CA_max_time << " " << BioFVM_CA_time_units << " max)" << std::endl;
		char* filename;
		filename = new char [1024];
		sprintf( filename, "output_%6f" , next_output_time );
		save_BioFVM_cellular_automata_to_MultiCellDS_xml_pugi( filename , M , CA_model ,
			BioFVM_CA_elapsed_time );

		cell_counts( CA_model );
		delete [] filename;
		next_output_time += BioFVM_CA_save_interval;

	// do the cellular automaton step
	CA_model.update_automata( BioFVM_CA_dt );
	BioFVM_CA_elapsed_time += BioFVM_CA_dt;

	// simulate biotransport to quasi-steady state

	double t_diffusion = 0.0;
	while( t_diffusion < BioFVM_CA_diffusion_max_time + 1e-10 )
		M.simulate_diffusion_decay( BioFVM_CA_diffusion_dt );
		M.simulate_cell_sources_and_sinks( BioFVM_CA_diffusion_dt );
		t_diffusion += BioFVM_CA_diffusion_dt;

Getting and Running the Code

  1. Start a project: Create a new directory for your project (I’d recommend “BioFVM_CA_tumor”), and enter the directory. Place a copy of BioFVM (the zip file) into your directory. Unzip BioFVM, and copy BioFVM*.h, BioFVM*.cpp, and pugixml* files into that directory.
  2. Download the demo source code: Download the source code for this tutorial: BioFVM_CA_Example_1, version 1.0.0 or later. Unzip its contents into your project directory. Go ahead and overwrite the Makefile.
  3. Edit the makefile (if needed): Note that if you are using OSX, you’ll probably need to change from “g++” to your installed compiler. See these tutorials.
  4. Test the code: Go to a command line (see previous tutorials), and test:

    (If you’re on windows, run BioFVM_CA_Example_1.exe.)

Simulation Result

If you run the code to completion, you will simulate 3 weeks of in vitro growth, followed by a bolus “injection” of drug. The code will simulate one one additional week under the drug. (This should take 5-10 minutes, including full simulation saves every 12 hours.)

In matlab, you can load a saved dataset and check the minimum oxygenation value like this:

MCDS = read_MultiCellDS_xml( 'output_504.000000.xml' );
min(min(min( MCDS.continuum_variables(1).data )))

And then you can start visualizing like this:

contourf( MCDS.mesh.X_coordinates , MCDS.mesh.Y_coordinates , ...
MCDS.continuum_variables(1).data(:,:,33)' ) ; 
axis image; 
xlabel('x (\mum)' , 'fontsize' , 12 );
ylabel( 'y (\mum)' , 'fontsize', 12 ); 
set(gca, 'fontsize', 12 );
title('Oxygenation (mmHg) at z = 0 \mum', 'fontsize', 14 ); 
print('-dpng', 'Tumor_o2_3_weeks.png' ); 
plot_cellular_automata( MCDS , 'Tumor spheroid at 3 weeks');

Simulation plots

Here are some plots, showing (left from right) pO2 concentration, a cross-section of the tumor (red = live cells, green = apoptotic, and blue = necrotic), and the drug concentration (after start of therapy):

1 week:

o2_1 weekTumor_spheroid_at_1_week__cross_section

Oxygen- and space-limited growth are restricted to the outer boundary of the tumor spheroid.

2 weeks:

o2_2 weeks Tumor_spheroid_at_2_weeks__cross_section

Oxygenation is dipped below 5 mmHg in the center, leading to necrosis.

3 weeks:

o2_3 weeks Tumor_spheroid_at_3_weeks__cross_section

As the tumor grows, the hypoxic gradient increases, and the necrotic core grows. The code turns on a constant 5 micromolar dose of doxorubicin at this point

Treatment + 12 hours:

o2_Treatment + 12 hours Tumor_spheroid_at_Treatment_+_12_hours__cross_section drug_Treatment + 12 hours

The drug has started to penetrate the tumor, triggering apoptotic death towards the outer periphery where exposure has been greatest.

Treatment + 24 hours:

o2_Treatment + 24 hoursTumor_spheroid_at_Treatment_+_24_hours__cross_section drug_Treatment + 24 hours
The drug profile hasn’t changed much, but the interior cells have now had greater exposure to drug, and hence greater response. Now apoptosis is observed throughout the non-necrotic tumor. The tumor has decreased in volume somewhat.

Treatment + 36 hours:

o2_Treatment + 36 hours Tumor_spheroid_at_Treatment_+_36_hours__cross_sectiondrug_Treatment + 36 hours
The non-necrotic tumor is now substantially apoptotic. We would require some pharamcokinetic effects (e.g., drug clearance, inactivation, or removal) to avoid the inevitable, presences of a pre-existing resistant strain, or emergence of resistance.

Treatment + 48 hours:

o2_Treatment + 48 hours Tumor_spheroid_at_Treatment_+_48_hours__cross_sectiondrug_Treatment + 48 hours
By now, almost all cells are apoptotic.

Treatment + 60 hours:

o2_Treatment + 60 hours Tumor_spheroid_at_Treatment_+_60_hours__cross_sectiondrug_Treatment + 60 hours
The non-necrotic tumor is nearly completed eliminated, leaving a leftover core of previously-necrotic cells (which did not change state in response to the drug–they were already dead!)

Source files

You can download completed source for this example here:

This file will include the following:

  1. BioFVM_cellular_automata.h
  2. BioFVM_cellular_automata.cpp
  3. BioFVM_CA_example_1.cpp
  4. read_MultiCellDS_xml.m (updated)
  5. plot_cellular_automata.m
  6. Makefile

What’s next

I plan to update this source code with extra cell motility, and potentially more realistic parameter values. Also, I plan to more formally separate out the example from the generic cell capabilities, so that this source code can work as a bona fide cellular automaton framework.

More immediately, my next tutorial will use the reverse strategy: start with an existing cellular automaton model, and integrate BioFVM capabilities.

Return to News • Return to MathCancer


Share this:

BioFVM: an efficient, parallelized diffusive transport solver for 3-D biological simulations

I’m very excited to announce that our 3-D diffusion solver has been accepted for publication and is now online at Bioinformatics. Click here to check out the open access preprint!

A. Ghaffarizadeh, S.H. Friedman, and P. Macklin. BioFVM: an efficient, parallelized diffusive transport solver for 3-D biological simulations. Bioinformatics, 2015.
DOI: 10.1093/bioinformatics/btv730 (free; open access)

BioFVM (stands for “Finite Volume Method for biological problems) is an open source package to solve for 3-D diffusion of several substrates with desktop workstations, single supercomputer nodes, or even laptops (for smaller problems). We built it from the ground up for biological problems, with optimizations in C++ and OpenMP to take advantage of all those cores on your CPU. The code is available at SourceForge and

The main idea here is to make it easier to simulate big, cool problems in 3-D multicellular biology. We’ll take care of secretion, diffusion, and uptake of things like oxygen, glucose, metabolic waste products, signaling factors, and drugs, so you can focus on the rest of your model.

Design philosophy and main capabilities

Solving diffusion equations efficiently and accurately is hard, especially in 3D. Almost all biological simulations deal with this, many by using explicit finite differences (easy to code and accurate, but very slow!) or implicit methods like ADI (accurate and relatively fast, but difficult to code with complex linking to libraries). While real biological systems often depend upon many diffusing things (lots of signaling factors for cell-cell communication, growth substrates, drugs, etc.), most solvers only scale well to simulating two or three. We solve a system of PDEs of the following form:

\[ \frac{\partial \vec{\rho}}{\partial t} = \overbrace{ \vec{D} \nabla^2 \vec{\rho} }^\textrm{diffusion}
– \overbrace{ \vec{\lambda} \vec{\rho} }^\textrm{decay} + \overbrace{ \vec{S} \left( \vec{\rho}^* – \vec{\rho} \right) }^{\textrm{bulk source}} – \overbrace{ \vec{U} \vec{\rho} }^{\textrm{bulk uptake}} + \overbrace{\sum_{\textrm{cells } k} 1_k(\vec{x}) \left[ \vec{S}_k \left( \vec{\rho}^*_k – \vec{\rho} \right) – \vec{U}_k \vec{\rho} \right] }^\textrm{sources and sinks by cells} \]
Above, all vector-vector products are term-by-term.

Solving for many diffusing substrates

We set out to write a package that could simulate many diffusing substrates using algorithms that were fast but simple enough to optimize. To do this, we wrote the entire solver to work on vectors of substrates, rather than on individual PDEs. In performance testing, we found that simulating 10 diffusing things only takes about 2.6 times longer than simulating one. (In traditional codes, simulating ten things takes ten times as long as simulating one.) We tried our hardest to break the code in our testing, but we failed. We simulated all the way from 1 diffusing substrate up to 128 without any problems. Adding new substrates increases the computational cost linearly.

Combining simple but tailored solvers

We used an approach called operator splitting: breaking a complicated PDE into a series of simpler PDEs and ODEs, which can be solved one at a time with implicit methods.  This allowed us to write a very fast diffusion/decay solver, a bulk supply/uptake solver, and a cell-based secretion/uptake solver. Each of these individual solvers was individually optimized. Theory tells us that if each individual solver is first-order accurate in time and stable, then the overall approach is first-order accurate in time and stable.

The beauty of the approach is that each solver can individually be improved over time. For example, in BioFVM 1.0.2, we doubled the performance of the cell-based secretion/uptake solver. The operator splitting approach also lets us add new terms to the “main” PDE by writing new solvers, rather than rewriting a large, monolithic solver. We will take advantage of this to add advective terms (critical for interstitial flow) in future releases.

Optimizing the diffusion solver for large 3-D domains

For the first main release of BioFVM, we restricted ourselves to Cartesian meshes, which allowed us to write very tailored mesh data structures and diffusion solvers. (Note: the finite volume method reduces to finite differences on Cartesian meshes with trivial Neumann boundary conditions.) We intend to work on more general Voronoi meshes in a future release. (This will be particularly helpful for sources/sinks along blood vessels.)

By using constant diffusion and decay coefficients, we were able to write very fast solvers for Cartesian meshes. We use the locally one-dimensional (LOD) method–a specialized form of operator splitting–to break the 3-D diffusion problem into a series of 1-D diffusion problems. For each (y,z) in our mesh, we have a 1-D diffusion problem along x. This yields a tridiagonal linear system which we can solve efficiently with the Thomas algorithm. Moreover, because the forward-sweep steps only depend upon the coefficient matrix (which is unchanging over time), we can pre-compute and store the results in memory for all the x-diffusion problems. In fact, the structure of the matrix allows us to pre-compute part of the back-substitution steps as well. Same for y- and z-diffusion. This gives a big speedup.

Next, we can use all those CPU cores to speed up our work. While the back-substitution steps of the Thomas algorithm can’t be easily parallelized (it’s a serial operation), we can solve many x-diffusion problems at the same time, using independent copies (instances) of the Thomas solver. So, we break up all the x-diffusion problems up across a big OpenMP loop, and repeat for y– and z-diffusion.

Lastly, we used overloaded +=, axpy and similar operations on the vector of substrates, to avoid unnecessary (and very expensive) memory allocation and copy operations wherever we could. This was a really fun code to write!

The work seems to have payed off: we have found that solving on 1 million voxel meshes (about 8 mm3 at 20 μm resolution) is easy even for laptops.

Simulating many cells

We tailored the solver to allow both lattice- and off-lattice cell sources and sinks. Desktop workstations should have no trouble with 1,000,000 cells secreting and uptaking a few substrates.

Simplifying the non-science

We worked to minimize external dependencies, because few things are more frustrating than tracking down a bunch of libraries that may not work together on your platform. The first release BioFVM only has one external dependency: pugixml (an XML parser). We didn’t link an entire linear algebra library just to get axpy and a Thomas solver–it wouldn’t have been optimized for our system anyway. We implemented what we needed of the freely available .mat file specification, rather than requiring a separate library for that. (We have used these matlab read/write routines in house for several years.)

Similarly, we stuck to a very simple mesh data structure so we wouldn’t have to maintain compatibility with general mesh libraries (which can tend to favor feature sets and generality over performance and simplicity).  Rather than use general-purpose ODE solvers (with yet more library dependencies, and more work for maintaining compatibility), we wrote simple solvers tailored specifically to our equations.

The upshot of this is that you don’t have to do anything fancy to replicate results with BioFVM. Just grab a copy of the source, drop it into your project directory, include it in your project (e.g., your makefile), and you’re good to go.

All the juicy details

The Bioinformatics paper is just 2 pages long, using the standard “Applications Note” format. It’s a fantastic format for announcing and disseminating a piece of code, and we’re grateful to be published there. But you should pop open the supplementary materials, because all the fun mathematics are there:

  • The full details of the numerical algorithm, including information on our optimizations.
  • Convergence tests: For several examples, we showed:
    • First-order convergence in time (with respect to Δt), and stability
    • Second-order convergence in space (with respect to Δx)
  • Accuracy tests: For each convergence test, we looked at how small Δt has to be to ensure 5% relative accuracy at Δx = 20 μm resolution. For oxygen-like problems with cell-based sources and sinks, Δt = 0.01 min will do the trick. This is about 15 times larger than the stability-restricted time step for explicit methods.
  • Performance tests:
    • Computational cost (wall time to simulate a fixed problem on a fixed domain size with fixed time/spatial resolution) increases linearly with the number of substrates. 5-10 substrates are very feasible on desktop workstations.
    • Computational cost increases linearly with the number of voxels
    • Computational cost increases linearly in the number of cell-based source/sinks

And of course because this code is open sourced, you can dig through the implementation details all you like! (And improvements are welcome!)

What’s next?

  • As MultiCellDS (multicellular data standard) matures, we will implement read/write support for  <microenvironment> data in digital snapshots.
  • We have a few ideas to improve the speed of the cell-based sources and sinks. In particular, switching to a higher-order accurate solver may allow larger time step sizes, so long as the method is still stable. For the specific form of the sources/sinks, the trapezoid rule could work well here.
  • I’d like to allow a spatially-varying diffusion coefficient. We could probably do this (at very great memory cost) by writing separate Thomas solvers for each strip in x, y, and z, or by giving up the pre-computation part of the optimization. I’m still mulling this one over.
  • I’d also like to implement non-Cartesian meshes. The data structure isn’t a big deal, but we lose the LOD optimization and Thomas solvers. In this case, we’d either use explicit methods (very slow!), use an iterative matrix solver (trickier to parallelize nicely, except in matrix-vector multiplication operations), or start with quasi-steady problems that let us use Gauss-Seidel iterative type methods, like this old paper.
  • Since advective flow (particularly interstitial flow) is so important for many problems, I’d like to add an advective solver. This will require some sort of upwinding to maintain stability.
  • At some point, we’d like to port this to GPUs. However, I don’t currently have time / resources to maintain a separate CUDA or OpenCL branch. (Perhaps this will be an excuse to learn Julia on GPUs.)

Well, we hope you find BioFVM useful. If you give it a shot, I’d love to hear back from you!

Very best — Paul

Share this:

Macklin Lab featured on March 2013 cover of Notices of the American Mathematical Society

I’m very excited to be featured on this month’s cover of the Notices of the American Mathematical Society. The cover shows a series of images from a multiscale simulation of a tumor growing in the brain, made with John Lowengrub while I was a Ph.D. student at UC Irvine. (See Frieboes et al. 2007, Macklin et al. 2009, and Macklin and Lowengrub 2008.) The “about the cover” write-up (Page 325) gives more detail.

The inside has a short interview on our more current work, particularly 3-D agent-based modeling. You should also read Rick Durrett‘s perspective piece on cancer modeling (Page 304)—it’s a great read! (And yup, Figure 3 is from our patient-calibrated breast cancer modeling in Macklin et al. 2012. 😉 )

The entire March 2013 issue can be accessed for free at the AMS Notices website:

I want to thank Bill Casselman and Rick Durrett for making this possible. I had a lot of fun in the process, and I’m grateful for the opportunity to trade ideas!

Return to News   Return to MathCancer   Follow @MathCancer
Share this: