# Some ideas about how to efficently store simulation data

After my last post about visualization of 3D simulated tumors @jc_atlantis made an excellent point that another reason why 3D agent-based simulations are not performed frequently is the large amount of generated data. Of course, one approach to overcome that problem would be to store only information that is really essential and not the 1-1 simulation snapshot. However, in some cases we can’t know in advance what information will be important and thus, we need better ways to save our simulation outputs.

In this post I will show few tricks (in C++) how to efficiently store output of 2D agent-based model presented in one of the previous posts. As a result we will reduce the size of generated simulation output from about 1Gb to reasonable 35Mb. Of course, presented ideas can be utilized in 3D agent-based model and in other programing languages like Python, Java or MATLAB.

First of all let us define the task: we want to store information about spatial distribution of 1,000,000 cells on 2D lattice. Each cell is described by two variables: remaining proliferation potential (p, unsigned char variable) and if it is cancer stem cell (is_stem, boolean variable). All of the cells to be saved are stored in the STL vector. The above assumptions are defined by the following piece of the C++ code:

struct cell {//defining a cell
unsigned int place; //varible keeping linear index to cell's place on
unsigned char p; //remaining proliferation potential
bool is_stem; //is the cell a cancer stem cell?

cell(unsigned int a=0, unsigned char b=0, bool c=0):place(a), p(b), is_stem(c){}; //constructor
};

static const int N = 2000; //lattice size
vector<cell> cells; //vector containing all cells present in the system


Each cell is occupying  (usigned int = 4 bytes) + (unsigned char = 1 byte) + (bool  = 1 byte) = 6 bytes of memory. Hence, as we want to save 1 mln cells,  our output file shouldn’t exceed 6 megabytes (Mb) of memory.

1. Outputting to text file

Information about cells is typically written to a text file in which each line describes one cell and variables values are separated by special character (standard .csv format). That format allows simulation output to be easily loaded to other programs for further analysis. Moreover, the code generating and loading the output files is very simple:

void save_cells_ASCII(string fileName) {
ofstream file_op(fileName);

for(int i=0; i<cells.size();i++)
file_op << cells.at(i).place << " " << (int)cells.at(i).p << " " << cells.at(i).is_stem <<  endl;          file_op.close(); } void load_cells_ASCII(string fileName) {     ifstream myfile (fileName);     string line;          unsigned int place;     int p;     bool is_stem;          while ( getline (myfile,line) ) {             istringstream iss(line);             iss >> place >> p >> is_stem;
cells.push_back(cell(place, (unsigned char)p, is_stem));
}

myfile.close();
}


Code above generated a file of 11Mb of size – way more to be expected from the cell definition. Why is it so? Explanation is quite straightforward. In the generated text file each character occupies 1 byte of disk space – it is OK for the proliferation capacity (p) and boolean variable (is_stem), as they occupy the same amount of space in the memory. However, the cell position 2234521, which when stored as unsigned int occupies 4 bytes of space in the memory, occupies 7 bytes of space on the disk. In addition, each space between the written value in the text file occupies additional byte of space. All of the above together generates the wasteful 11Mb file on the disk.

2. Binary files

In order to to have a file on disk occupying exactly the same amount of space as the variables in the memory, we need to use binary files. It is a little bit more complicated to operate on binary files, but the idea is simple: we just write each variable byte by byte (1 byte = 8 bits). First, we write functions that 1) prepare array of pointers to char (byte) variables that will be saved; 2) read simulation snapshot from array of pointers to char variables.

void save_cells_binary(char** data, unsigned long* sizeData) {
unsigned long Ncells = cells.size(); //number of cells
int sizeOfCell = sizeof(int)+sizeof(char)+sizeof(bool);
*sizeData = Ncells*sizeOfCell;
*data = (char*)malloc( *sizeData );

memcpy( *data, (char*)&Ncells, sizeof(unsigned long));
memcpy((char*)&Ncells, *data, sizeof(unsigned long));

for(int i=0; i<Ncells; i++) {
memcpy(*data+i*sizeOfCell+sizeof(unsigned long), (char*)&cells.at(i).place, sizeof(int));
memcpy(*data+i*sizeOfCell+sizeof(int)+sizeof(unsigned long), (char*)&cells.at(i).p, sizeof(char));
memcpy(*data+i*sizeOfCell+sizeof(int)+sizeof(char)+sizeof(unsigned long), (char*)&cells.at(i).is_stem, sizeof(bool));
}
}

unsigned long Ncells;
memcpy((char*)&Ncells, data, sizeof(unsigned long));
int sizeOfCell = sizeof(int)+sizeof(char)+sizeof(bool);

unsigned int place;
unsigned char p;
bool is_stem;

for(int i=0; i<Ncells; i++) {         memcpy((char*)&place, data+sizeof(unsigned long) + i*sizeOfCell, sizeof(unsigned long));         memcpy((char*)&p, data+sizeof(unsigned long) + i*sizeOfCell+sizeof(int), sizeof(char));         memcpy((char*)&is_stem, data+sizeof(unsigned long) + i*sizeOfCell+sizeof(int)+sizeof(char), sizeof(bool));                  cells.push_back(cell(place, p, is_stem));     } } 

Now we need only functions that will operate on the binary files and 1) read the char array from the binary file; 2) save char array to binary file.

 void readWholeBinary(string fileName, char** data, unsigned long* sizeData) {     ifstream file_op(fileName, ios::binary | ios::ate);     *sizeData = file_op.tellg();     file_op.seekg(0, ios::beg);     *data = (char*)malloc( *sizeData );     file_op.read(*data, *sizeData);     file_op.close(); } void saveWholeBinary(string fileName, char* data, unsigned long sizeData, unsigned long originalSize) {     ofstream file_op(fileName, ios::binary | ios::out);     if (originalSize>sizeData) //there was compression, we ass original file size to the beginning of the file
file_op.write((char*)&originalSize, sizeof(unsigned long));
file_op.write(data, sizeData);
file_op.close();
}


Executing the above code generates 5.51 Mb file on the disk (we had about 950,000 cells to save). This is about half of the space that is occupied by the text file!

3. Using zip compression

All of us probably used zip compression to store or send files through e-mail. Why won’t we write files generated by our simulation already compressed with zip? There are quite a few C++ zip libraries, but I’ve chosen zlib library as it is quite easy to use (an excellent tutorial can be found here). It operates on the char arrays that we already generate on the way to save binary files, so compressing the file before writing takes only few lines of the code.

void compressFunction (char** dataCompressed, char* dataOriginal, unsigned long* sizeDataCompressed, unsigned long sizeDataOriginal) {
*sizeDataCompressed  = ((sizeDataOriginal) * 1.1) + 12;
*dataCompressed = (char*)malloc(*sizeDataCompressed);
compress((Bytef*)(*dataCompressed),sizeDataCompressed,(Bytef*)dataOriginal,sizeDataOriginal );// size of source data in bytes
}

void uncompressFunction (char** dataUncompressed, char* dataCompressed, unsigned long sizeDataCompressed, unsigned long* sizeDataUncompressed) {
memcpy((char*)sizeDataUncompressed, dataCompressed, sizeof(unsigned long));
*dataUncompressed = (char*)malloc( *sizeDataUncompressed );
uncompress((Bytef*)(*dataUncompressed), sizeDataUncompressed, (Bytef*)(dataCompressed+sizeof(unsigned long)), sizeDataCompressed-sizeof(unsigned long));
}


The binary file after compression takes 3.57 Mb of space on disk – way better.

4. Collapsing the variables

We can easily save a byte of memory per cell if we notice that the proliferation capacity variable in our simulations has the value smaller than 50. Unsigned char, however, can store the value up to 255. Thus, we can add the value 100 to the remaining proliferation capacity if the cell is cancer stem cell and forget about saving boolean variable.

 memcpy(*data+i*sizeOfCell+sizeof(unsigned long), (char*)&cells.at(i).place, sizeof(int));
val = cells.at(i).p+(unsigned char)cells.at(i).is_stem*100;
memcpy(*data+i*sizeOfCell+sizeof(int)+sizeof(unsigned long), (char*)&val, sizeof(char));


5. Using information about the space to reduce the file size

The most amount of disk space is used by the information about the cells location on the lattice (unsigned int = 4 bytes). Can we skip writing that information for most of the cells? Yes, we can. In the code below we just write the cells row by row, and store the information about the location of the first cell in the lattice row. If there is an empty space between the cells in the row we put the value 255 and it the remainder of the row is empty we put value 254.

void save_cells_usingSpace(char** data, unsigned long* sizeData) {

unsigned long Ncells = cells.size(); //number of cells
int sizeOfCell = sizeof(int)+sizeof(char)+sizeof(bool);
*sizeData = 0; //we will update that value
*data = (char*)malloc( Ncells*sizeOfCell ); //we alloc more than we need

unsigned char lat[N][N] = {0};
int x, y;
for (int i = 0; i<cells.size(); i++) {
x = cells.at(i).place % N;
y = floor((double)cells.at(i).place/(double)N);
lat[x][y] = (unsigned char)cells.at(i).is_stem*(unsigned char)100 + (unsigned char)cells.at(i).p + (unsigned char)1;
}

memcpy(*data, (char*)&N, sizeof(int));
*sizeData += sizeof(int);

//255 means empty space, 254 means end of line
unsigned char es = 255, el = 254;

for (unsigned short int i = 0; i<N; i++) {

int sep = 0;
bool st = false;

for (unsigned short int j = 0; j<N; j++) {             if(st == false && lat[i][j]>0) {
st = true;
memcpy(*data+*sizeData,(char*)&i, sizeof(unsigned short int));
memcpy(*data+*sizeData+sizeof(unsigned short int),(char*)&j, sizeof(unsigned short int));
memcpy(*data+*sizeData+2*sizeof(unsigned short int),(char*)&lat[i][j], sizeof(unsigned char));
*sizeData += 2*sizeof(unsigned short int)+sizeof(unsigned char);
} else if (st == true && lat[i][j] == 0 ) { //we have empty space
sep++;
} else if (st==true && lat[i][j]>0){
for (int k = 0; k<sep; k++) {
memcpy(*data+*sizeData,(char*)&es, sizeof(unsigned char));
*sizeData += sizeof(unsigned char);
}

sep = 0;
memcpy(*data+*sizeData,(char*)&lat[i][j], sizeof(unsigned char));
*sizeData += sizeof(unsigned char);
}
}

if (st == true) {
memcpy(*data+*sizeData,(char*)&el, sizeof(unsigned char));
*sizeData += sizeof(unsigned char);
}
}
}

void load_cells_usingSpace(char* data, unsigned long dataSize) {
unsigned short int x, y;

int Nm=0;
memcpy((char*)&Nm, data, sizeof(int));

unsigned long i = sizeof(int);

while(i < dataSize) {

memcpy((char*)&x,data+i, sizeof(unsigned short int));
memcpy((char*)&y,data+i+sizeof(unsigned short int), sizeof(unsigned short int));
i += 2*sizeof(unsigned short int);

while (true) {
i+=sizeof(unsigned char);
if (read == 254) { //end of line
break;
}
}
}

}


The above approach when combined with zip compression generated file that takes only 0.47 Mb of disk space. That is less then a byte per cell!

6. Overall comparison

Fig 1. shows the amount of used space by a file written using different approaches presented above. As it can be seen the last approach uses about 25x less disk space than the standard text files approach. When saving simulation snapshots every 10 days the total amount of generated data was reduced from 1 Gb when using standard text file approach to about 35 Mb!
Figure 1. Comparison of the amount of used disk space.

What is important binary files approach is way faster that text file approach, even when using zip compression algorithm, see Fig. 2.

Of course, the trade-off is that it won’t be that easy to read our simulation files to other programs. However, we can always write wrapper functions that will allow other programs to interpret our files (see e.g. this post).

# Some tips&tricks for 3D tumor visualization

There are several reasons why we rarely see 3D agent based models of tumor growth, with I think the two most important ones being: 1) 3D simulations are way more computationally expensive than 2D simulations; 2) it’s harder to visualize results of 3D simulations to present all necessary information. The other frequently used argument is that there is no need for 3D, because the crucial phenomena can be observed as well in 2D. However, when one considers e.g. stochastic models there might be a huge difference in the system behavior between 2D and 3D case (see very nice post Why is that a 2D random walk is recurrent, while a 3D random walk is transient?).

Possible improvements in the computational speed are highly dependent on the given model setting and typically require sophisticated techniques. Problem with nice visualizations, however, just requires proper tools. Today, I will show my attempts to make a good-looking 3D tumor visualizations, which I have taken over the last several years.

Several years ago when working with @heikman on the project about the impact of cellular senescence  I was using The Persistence of Vision Raytracer (POV-Ray) in order to obtain high quality 3D tumor visualization. After simulating the tumor up to a given cell number I was parsing the information about cells locations and their basic properties into a POV-Ray specific format. In my approach, each cell was separate actor in the scene with possible different surface/texture/color properties. As it can be expected renderings with raytracing software are very good (see Fig.1 below), but the whole procedure to prepare input files to POV-Ray, to set-up properly camera and light was really time-consuming. In addition, rendering takes a long time, so there is no chance for “live” rendering during the simulation, which would allow to observe simulation behavior.

Figure 1. Exemplary 3D tumor renderings obtained using POV-Ray software. Bottom-left picture shows tumor consisted of 9.4 million of cells.

I really wanted to have a tool that would allow me to seamlessly simulate and visualize 3D tumor at the same time. I quickly realized that rendering each cell as a separate object is not a good approach, when you simulate more than 1 million cells. Hence, I started looking around for an algorithm that would allow to extract tumor surface from the cloud of cells. I quickly found well known Marching Cubes algorithm  that does the trick very well.  I’ve implemented the algorithm and visualized resulting surface using Visualization Library in C++. I was quite happy with the results (see Fig. 2 below), however, at some point code became so complicated and so model specific that I assumed that it doesn’t have any future.

Figure 2. “Live” 3D tumor visualization during the simulation. Code developed in C++ using Visualization Library.

Quite recently I started using VTK library from within C++ (you can also use it form within Python) and I quickly realized that it might be the ultimate tool for 3D visualization. Now I will show how easy it is to produce great visualizations that can be used in “live” renderings during 3D simulations using VTK library.

In the following example I will use glyph mapper that basically takes my cloud of points (cells), creates cubes around them, and extracts the surface. First, let us include all necessary headers and define necessary variables.

#include "vtkCubeSource.h"
#include "vtkPolyDataMapper.h"
#include "vtkRenderer.h"
#include "vtkRenderWindow.h"
#include "vtkRenderWindowInteractor.h"
#include "vtkInteractorStyleTrackballCamera.h"
#include "vtkPolyData.h"
#include "vtkGlyph3DMapper.h"

vtkSmartPointer<vtkRenderer> renderer;
vtkSmartPointer<vtkRenderWindow> renderWindow;
vtkSmartPointer<vtkRenderWindowInteractor> renderWindowInteractor;
vtkSmartPointer<vtkPoints> points;


Now we can write the function that will initialize the renderer and rendering window together with its interactor object.

void initializeRendering() {
renderer = vtkSmartPointer<vtkRenderer>::New(); //creating new renderer
renderer->SetBackground(1., 1., 1.); //setting background color to white

renderWindow = vtkSmartPointer<vtkRenderWindow>::New(); //creating renderer window
renderWindow->SetSize(1000, 1000); //setting size of the window

renderWindowInteractor = vtkSmartPointer<vtkRenderWindowInteractor>::New(); //creating render
renderWindowInteractor->SetRenderWindow(renderWindow); //adding previously created renderer window

points = vtkSmartPointer<vtkPoints>::New();         //vector with points
}


I assume that we have our cell objects are stored in STL vector structure and that each cell contains information about its (x,y,z) position in 3D space. Having that we can write a really short function that adds our cells to the scene.

void addCells(vector<Cell>::iterator begin, vector<Cell>::iterator end) {
vector<Cell>::iterator cell = begin;
for (cell; cell<end; cell++)
points->InsertNextPoint(cell->x,cell->y,cell->z);
}


Finally, we can write the rendering function that is invoked in the main part of the code.

void visualizeLesion(vector<Cell>::iterator begin, vector<Cell>::iterator end) {

initializeRendering();

// Combine into a polydata
vtkSmartPointer<vtkPolyData> polydata = vtkSmartPointer<vtkPolyData>::New();
polydata->SetPoints(points);

vtkSmartPointer<vtkCubeSource> cubeSource = vtkSmartPointer<vtkCubeSource>::New();//you can use sphere if you want

vtkSmartPointer<vtkGlyph3D> glyph3D = vtkSmartPointer<vtkGlyph3D>::New();
glyph3D->SetSourceConnection(cubeSource->GetOutputPort());
glyph3D->ScalingOff();
glyph3D->SetInputData(polydata);
glyph3D->Update();

// Create a mapper and actor
vtkSmartPointer<vtkPolyDataMapper> mapper = vtkSmartPointer<vtkPolyDataMapper>::New();
mapper->SetInputConnection(glyph3D->GetOutputPort());

vtkSmartPointer<vtkActor> actor = vtkSmartPointer<vtkActor>::New();
actor->SetMapper(mapper);

// Add the actor to the scene

// Render
renderWindow->Render();

vtkSmartPointer<vtkInteractorStyleTrackballCamera> style = vtkSmartPointer<vtkInteractorStyleTrackballCamera>::New(); //like paraview

renderWindowInteractor->SetInteractorStyle( style );

renderWindowInteractor->Start();
}


Snapshot of resulting rendering window is presented in Fig. 3 – and yes, I agree, it’s not very impressive. However, we can easily tweak it and make it significantly better.

Figure 3. Exemplary 3D tumor visualized using 3D glyph mapper in VTK.

First of all to have better a better perspective we can slice the tumor in half and take resulting parts slightly apart. This can be achieved by modification of the addCells function (note that in my setting center of the tumor is at (x,y,z) = (250,250,250)).

void addCells(vector<Cell>::iterator begin, vector<Cell>::iterator end) {
vector<Cell>::iterator cell = begin;
for (cell; cell<end; cell++) {         if (cell->z < 250) {             points->InsertNextPoint(cell->x,cell->y,cell->z);
} else {
points->InsertNextPoint(cell->x,cell->y,cell->z+50);
}
}
}


Fig. 4 shows the snapshot of the rendering window after slicing the tumor – much better, but there is still place for easy improvement. Let’s add some color!

Figure 4. Exemplary 3D tumor visualized using 3D glyph mapper in VTK.

I assume that each cell has also information about the number of free spots in its direct neighborhood. We will color the tumor based on that information.

#include "vtkCubeSource.h"
#include "vtkPolyDataMapper.h"
#include "vtkRenderer.h"
#include "vtkRenderWindow.h"
#include "vtkRenderWindowInteractor.h"
#include "vtkInteractorStyleTrackballCamera.h"
#include "vtkPolyData.h"
#include "vtkGlyph3DMapper.h"
#include "vtkPointData.h"
#include "vtkFloatArray.h"
#include "vtkLookupTable.h"

vtkSmartPointer<vtkRenderer> renderer;
vtkSmartPointer<vtkRenderWindow> renderWindow;
vtkSmartPointer<vtkRenderWindowInteractor> renderWindowInteractor;

vtkSmartPointer<vtkPoints> points;

vtkSmartPointer<vtkFloatArray> scalar;


Then we need to modify the function that initializes rendering

void initializeRendering() {
renderer = vtkSmartPointer<vtkRenderer>::New(); //creating new renderer
renderer->SetBackground(1., 1., 1.); //setting background color to white

renderWindow = vtkSmartPointer<vtkRenderWindow>::New(); //creating renderer window
renderWindow->SetSize(1000, 1000); //setting size of the window

renderWindowInteractor = vtkSmartPointer<vtkRenderWindowInteractor>::New(); //creating render
renderWindowInteractor->SetRenderWindow(renderWindow); //adding previously created renderer window

points = vtkSmartPointer<vtkPoints>::New();         //vector with points
scalar = vtkSmartPointer<vtkFloatArray>::New(); //vector with scalar values of the points
scalar->SetNumberOfComponents(1); //one value per point
}


void addCells(vector<Cell>::iterator begin, vector<Cell>::iterator end) {
vector<Cell>::iterator cell = begin;
for (cell; cell<end; cell++) {         if (cell->z < 250) {             points->InsertNextPoint(cell->x,cell->y,cell->z);
} else {
points->InsertNextPoint(cell->x,cell->y,cell->z+50);
}

scalar->InsertNextValue((double)cell->nSpots/26.);
}
}


Finally, we can modify the visualization function.

void visualizeLesion(vector<Cell>::iterator begin, vector<Cell>::iterator end) {

initializeRendering();

//creating new color table
vtkSmartPointer<vtkLookupTable> colorLookupTable = vtkSmartPointer<vtkLookupTable>::New();
colorLookupTable->SetNumberOfColors(256);
colorLookupTable->SetHueRange( 0.667, 0.0);
colorLookupTable->Build();

// Combine into a polydata
vtkSmartPointer<vtkPolyData> polydata = vtkSmartPointer<vtkPolyData>::New();
polydata->SetPoints(points);
polydata->GetPointData()->SetScalars(scalar);

vtkSmartPointer<vtkCubeSource> cubeSource = vtkSmartPointer<vtkCubeSource>::New();//you can use sphere if you want

vtkSmartPointer<vtkGlyph3D> glyph3D = vtkSmartPointer<vtkGlyph3D>::New();
glyph3D->SetColorModeToColorByScalar();
glyph3D->SetSourceConnection(cubeSource->GetOutputPort());
glyph3D->ScalingOff();
glyph3D->SetInputData(polydata);
glyph3D->Update();

// Create a mapper and actor
vtkSmartPointer<vtkPolyDataMapper> mapper = vtkSmartPointer<vtkPolyDataMapper>::New();
mapper->SetInputConnection(glyph3D->GetOutputPort());
mapper->SetLookupTable(colorLookupTable);

vtkSmartPointer<vtkActor> actor = vtkSmartPointer<vtkActor>::New();
actor->SetMapper(mapper);

// Add the actor to the scene

// Render
renderWindow->Render();

vtkSmartPointer<vtkInteractorStyleTrackballCamera> style = vtkSmartPointer<vtkInteractorStyleTrackballCamera>::New(); //like paraview

renderWindowInteractor->SetInteractorStyle( style );

renderWindowInteractor->Start();
}


Snapshot of sliced and colored tumor is presented in Fig. 5 – fast, good-looking and easy.

Figure 5. Exemplary 3D tumor visualized using 3D glyph mapper in VTK. Color represents the number of free spots in the direct neighborhood of each cell (jet colormap).

# Avoiding boundary effects – dynamically expanding lattice for ABM

In a typical setting, agent-based model (ABM) simulations take place on a domain of fixed size, e.g. square lattice of size 500 x 500. Depending on the model, this can introduce so-called boundary effects, i.e. model predictions that are in part caused by the limited amount of space. Some time ago me and @theheikman published a paper in which we investigated the impact of evolution of cancer stem cells on the rate of tumor growth (link to the paper here). Because of the ABM formulation, we had to implement domain that can freely expand on demand. Otherwise, the tumor evolution would stop once the cells fill all the space. Of course, we could set the initial domain size to be so big that the space wouldn’t be filled in the simulated timeframe. However, it is hard to know a priori what domain size will be sufficient. In this post I will show how to implement in C++ an ABM lattice that dynamically expands if one of the cells touches its boundary. We will be working on pointers and memory allocating/freeing routines.

Our lattice will be a boolean array in which value of true will indicate that the spot is occupied by the cell. First, we need to create lattice of initial size NxN.

lattice=new bool *[N];
for (int i=0; i&lt;N; i++) {
lattice[i] = new bool [N] ();
fill_n(lattice[i], N, false);
}


An element of the lattice can be easily accessed using double indexing, i.e. lattice[i][j] = true.

Now we need to have a procedure that will expand our lattice, by fixed amount of rows (gN) from top, bottom and fixed amount of columns (also gN) to right and left.

void expandLattice() {
bool **aux;
aux=new bool *[N+2*gN];

for (int i=0; i&lt;gN; i++) { //adding empty columns to the left
aux[i] = new bool [N+2*gN] ();
fill_n(aux[i], N+2*gN, false);
}

for (int i=N+gN; i&lt;N+2*gN; i++) { //adding empty columns to the right
aux[i] = new bool [N+2*gN] ();
fill_n(aux[i], N+2*gN, false);
}

//copying the interior columns
for (int i=gN; i&lt;N+gN; i++) {
bool *aux2;
aux2 = new bool [N+2*gN] ();
fill_n(aux2, N+2*gN, false); //filling with false values
memcpy(aux2+gN, lattice[i-gN],N*sizeof(bool)); //copying existing values
free(lattice[i-gN]);
aux[i] = aux2;
}

free(lattice);
lattice=aux;
N += 2*gN;
}


Very important part of the above procedure is invoking free() function in order to deallocate the memory previously occupied by the lattice.

And that is about it: if an event of touching the boundary is detected, we just invoke the expandLattice() function. We also need to remember to free the allocated memory at the very end of the simulation, by putting

for (int i=0; i&lt;N; i++)
free(lattice[i]);

free(lattice);


at the end of main() function.

In CAexpandP file you can find the code for the cancer stem cell driven model of tumor growth considered in the previous posts that uses dynamically expanding lattice (change extension from .doc to .cpp). Plot below nicely shows how the domain size grows together with the tumor when using that code with initial N = 100 and gN = 100 (plot shows the average of 20 simulations).

# Julia – excellent opensource alternative for MATLAB

After my last post about ABM speed comparison @marcjwills_ suggested that Julia programming language could be a good alternative for slower, but easy to code programming languages – I’ve decided to give it a try. On the project webpage I found promising speed comparisons showing that Julia can indeed be my new tool of choice. After easy download and installation I saw the terminal window that brought a huge smile to my face.

Quick web search allowed me to find nice IDE for Julia called Juno – simple and not overloaded with features. What might be a great news to some people (especially @dbasanta) is that Julia after few easy tweaks can use matplotlib.

The obvious first test to do is to update speed comparison post with Julia. First, I’ve decided to rewrite the partially vectorized MATLAB code in Julia without making any essential changes. After browsing the web I found couple of useful webpages that helped me with doing that without much effort:

The differences between Julia and MATLAB are really minor and most of the functions work in exactly the same way. I only had to be careful with logical operations between matrices and scalars – in Julia you need to add “.” to the operator to make it work as in MATLAB. Other changes that I had to make were:

1. To create logical lattice: “false(N,N)” to “falses(N,N)”.
2. Array elements are accessed with “[i]” instead of “(i)”.
3. To create array of all permutations: “perms(uint8(1:8))'” to “hcat(collect(permutations(1:8))…)”.
6. Deleting elements by using “deleteat!()” function.

The working code in Julia is the following:

N = 1000; #square domain dimension

pprol = 1/24; #probability of proliferating
pmig = 15/24; #probability of migrating
pdeath = 5/100; #probability of dying
pmax = int8(10); #proliferation capacity
ps = 0.3; #probability of symmetric division

L = falses(N,N);
L[[1:N, 1:N:N*N, N:N:N*N, N*(N-1):N*N]] = true; #boundary

L[N*round(N/2)+round(N/2)] = true;
cells = int32([N*round(N/2)+round(N/2)]);
cellsIsStem = [true];
cellsPmax = int8([pmax]);

aux = int32([-N-1 -N -N+1 -1 1 N-1 N N+1])'; #indices to heighborhood
Pms = hcat(collect(permutations(1:8))...); #permutations
nP = size(Pms,2); #number of permutations
nSteps = 0;

while(length(cells)<10^4)
sh = randperm(length(cells));
cells = cells[sh];
cellsIsStem = cellsIsStem[sh];
cellsPmax = cellsPmax[sh];

S[L[S]] = 0; #setting occupied spots to 0
indxF = int32(find(any(S,1))); #selecting cells with at least one free spot
nC = length(indxF); #number of cells with free spot

P = rand(nC).<pprol; #proliferation
Ps = P & (rand(nC).<ps) & (cellsIsStem[indxF]);#symmetric division
De = P & (cellsPmax[indxF] .== 0);#proliferation exhaution
D = P & (rand(nC).<pdeath) & (~cellsIsStem[indxF]); #death at proliferation attempt
M = (~P) & (rand(nC).<pmig); #go when no grow del = D | De; #cells to delete act = find((P | M) & (~del)); #indices to the cells that will do something for ii = act #only for those that will do anything ngh = S[:,indxF[ii]]; ngh = ngh[ngh.>0];
indO = find(~L[ngh]); #selecting free spot
if ~isempty(indO) #if there is still a free spot
indO = indO[1];
L[ngh[indO]] = true;
if P[ii] #proliferation
push!(cells,int32(ngh[indO]));
if Ps[ii] #symmetric division
push!(cellsIsStem,true);
push!(cellsPmax,cellsPmax[indxF[ii]]);
else
push!(cellsIsStem,false);
push!(cellsPmax, cellsPmax[indxF[ii]]-1);
if ~cellsIsStem[indxF[ii]]
cellsPmax[indxF[ii]] = cellsPmax[indxF[ii]]-1;
end
end
else #migration
L[cells[indxF[ii]]] = false;
cells[indxF[ii]] = int32(ngh[indO]);
end
end
end

if any(del) #updating death
L[cells[indxF[del]]] = false;
deleteat!(cells,indxF[del]);
deleteat!(cellsIsStem,indxF[del]);
deleteat!(cellsPmax,indxF[del]);
end
end


To visualize the results of simulations we can use Gadfly plotting package and use simple spy() function (analog to MATLAB).

using Gadfly
spy(L)


Ok, so modifications were easy and straightforward, but what about speed? Amazingly Julia performed almost exactly the same as MATLAB (see plot below).

But that is not all… MATLAB code was using vectorization feature. Without that, when using only loops like in C++, it would be several times slower. Is that also the case in Julia?

Below is the C++ style code in Julia – only loops, no vectorization. One of the awesome feature of Julia, that MATLAB lacks, is definition of functions within the script (see “returnEmptySpace” function below).

N = 1000; #square domain dimension

aux = int32([-N-1 -N -N+1 -1 1 N-1 N N+1]); #indices to heighborhood

pprol = 1/24; #probability of proliferating
pmig = 15/24; #probability of migrating
pdeath = 5/100; #probability of dying
pmax = int8(10); #proliferation capacity
ps = 0.3;

L = falses(N,N);
L[[1:N, 1:N:(N*N), N:N:(N*N), (N*(N-1)):(N*N)]] = true; #boundary

L[N*round(N/2)+round(N/2)] = true;
cells = [int32(N*round(N/2)+round(N/2))];
cellsIsStem = [true];
cellsPmax = [int8(pmax)];

#defining function that returns random empty place around cell
function returnEmptyPlace(indx::Int32)
neigh = int32(zeros(1,8))
nF = 0
for i = 1:8
if ~L[indx+aux[i]]
neigh[nF+1] = indx+aux[i]
nF+=1
end
end
if nF>0
return neigh[rand(1:nF,1)];
else
return 0;
end
end

while length(cells)<10^5

sh = randperm(length(cells));
cells = cells[sh];
cellsIsStem = cellsIsStem[sh];
cellsPmax = cellsPmax[sh];

newCells = Int32[]
newCellsIsStem = Bool[]
newCellsPmax = Int8[]

i = 1;
while i<=length(cells) deleted = false; newSite = returnEmptyPlace(cells[i])[1]; if newSite > 0 #if there is a new spot
if rand()<pprol
if cellsIsStem[i] #is stem cell
push!(newCells, newSite);
L[newSite] = true;
push!(newCellsPmax,cellsPmax[i]);
if (rand()<ps) #symmetric division push!(newCellsIsStem, true); else push!(newCellsIsStem, false); end else if (cellsPmax[i]>0) && rand()>pdeath
cellsPmax[i]-=1
L[newSite]=true;
push!(newCells,newSite);
push!(newCellsPmax,cellsPmax[i]);
push!(newCellsIsStem, false);
else
L[cells[i]]=false;
splice!(cells,i)
splice!(cellsPmax,i)
splice!(cellsIsStem,i)
deleted = true;
end
end
elseif rand() < pmig
L[newSite]=true;
L[cells[i]]=false;
push!(newCells,newSite);
push!(newCellsPmax,cellsPmax[i]);
push!(newCellsIsStem, cellsIsStem[i]);
splice!(cells,i)
splice!(cellsPmax,i)
splice!(cellsIsStem,i)
deleted = true;
end
end

if ~deleted
i+=1;
end
end

cells = [cells, newCells];
cellsPmax = [cellsPmax, newCellsPmax];
cellsIsStem = [cellsIsStem, newCellsIsStem]
end


How about the speed of a new code? I was amazed with the results…

# Java, C++, MATLAB, Julia or Python for cellular automatons? Speed comparison.

Thanks to @spuri4096 and @chandlergatenbee, the basic cancer stem cell driven model of tumor growth that I implemented in the very first Compute Cancer blog is now programmed in 4 languages:

1. C++ (post with the code here)
2. MATLAB (post with the code here)
3. Java (post with the code here)
4. Python (post with the code here)
5. Octave – the same code as for MATLAB. Works without any modifications.
6. Julia (post with the code here)

Isn’t that the perfect opportunity to see which one is the fastest?

First of all, codes are not completely identical. There are small differences between C++ and MATLAB codes in how the death of a cell is updated, but that shouldn’t affect the comparison too much. The code in Java is much more sophisticated – it uses coded lattice concept and dynamically expanding lattice. Python code (as much as I can see…) does pretty much the same job as the C++ code, so that is the fairest comparison out of all.

I will compare the time needed to simulate cancer from a single cancer stem cell to 100, 250 and 500 thousand of cells. The domain size is set to 1000×1000 grid points (not for Java code, in which lattice expands on demand), so 1 million is the maximal cell number. Of course all simulation parameters are also the same for all codes: probability of symmetric division (ps) = 0.3; proliferation probability (pdiv) = 1/24; migration probability = 15/24; and probability of spontaneous death (alpha) = 0.05. Because of the stochastic nature of the model I will run 20 simulations for each code and report the average time (+- standard deviation).

The results are summarized on the plot below.

As probably everyone expected C++ outcompeted other codes.
Java also did great.
What might surprise some people, MATLAB didn’t perform that bad (it doesn’t suck that bad as people usually think).
The speed of Octave was a surprise for me – I had patience only to run simulations for 100K cell limit…

Of course in each case there is definitely a room for improvement using even more sophisticated language specific programming tricks.

However, from my perspective – if you don’t want to spend too much time on learning programing language, MATLAB (Octave) might be not such a bad idea.

# Cancer Stem Cell CA in Python

Python doesn’t seem to be the first programming language people go to when developing cellular automata models. However, given that Python is an object-orientated language that is easy to read and write, it might actually be ideal for such models, especially if you prefer to think from the perspective of the agent (if you’d rather model using matrices you can do that too by using Python’s NumPy package). As we’ll see, it can also be pretty fast. Below I’ll cover one way to implement the Cancer Stem cell CA described in the Cancer Stem Cell CA in Matlab post.

GETTING PYTHON

If you don’t have Python, the easiest way to get it and nearly all of its scientific packages is to download the Anaconda distribution from https://store.continuum.io/cshop/anaconda/. This installation and all modules will be kept in it’s own folder, so it won’t interfere with the Python that comes bundled in your OS. It also means that uninstalling is just a matter of dropping the folder in the trash. You may also want to download an IDE, a list of which can be found on the Anaconda website (http://docs.continuum.io/anaconda/ide_integration ;my personal favorite is PyCharm).

As far as versions are concerned, the “go-to” version is Python 2.7. However, 2.7 is no longer being developed so you will eventually have to migrate to Python 3. For a long time many avoided this upgrade because several of the most important packages hadn’t been ported over, but that has since changed and most of the critical packages like Numpy, SciPy, Matplotlib, Pandas, etc… are now available in Python 3. If you choose to go with Python 2, it’s important to note that when dividing you need to use a float in the denominator, as division by an integer returns a rounded down integer. This isn’t the case with Python 3, however.

OVERVIEW

Before we begin coding, I’d like to provide an overview of how the model will be implemented. Instead of using a matrix to track cells, we’ll create a dictionary of cells. The keys to this dictionary will be the cell’s position, while the value will be the cell itself. Each cell we create will know the positions of it’s neighbors, and the cell will use the dictionary of cells to look up its neighbors (a dictionary in Python is really just a hash table, so this look-up is fairly fast). A nice advantage of using a dictionary is that we can easily add and remove entries from it, giving us the ability to iterate only across cells, not empty positions in a matrix. This effectively allows the domain to grow and shrink according to how many cells are present. During each round of this model, we’ll shuffle the values in this dictionary and iterate through them, having each cell execute it’s actions.

One final note. I’ve taken screenshots of the code to make it easier to read, but the raw code can be found in this .doc file: CSC_CA_in_python. It’s not possible to upload .py files, so if you download the code, just change the extension from .doc to .py and then run the CA. Or, just copy and paste into your own file and run it.

Let’s start coding by importing the packages we will be using.The first is NumPy, Python’s package for scientific computing (http://www.numpy.org), a package you’ll almost always need. Second, we’ll import the package we’ll use to visualize our results, matplotlib (http://matplotlib.org). Next, we’ll import Numba, a ‘just in time’ compiler than can dramatically speed up sections of code in which numerical calculations are called for (http://numba.pydata.org). Finally, importing the random module will allow us to randomly select cells.

We’ll begin writing up our CA by defining a few methods that we’ll use to construct the world of our CA. First, lets construct that dictionary which will contain the list of positions in a central position’s Moore neighborhood. To do this, we can use both list comprehension and dictionary comprehension. Basically, these approaches allow us to build lists or dictionaries without creating an empty list and messy for loop to populate that list/dictionary. Not only is list comprehension more convenient and cleaner, but it is also faster than the alternative.

SPEEDUP TRICKS

Here we’ll setup a few methods to speed up our CA. We’ll frequently request a random binomial number and shuffle cells, so we’ll call those methods once, here at the top, so they don’t have to be reevaluated each time we call them, something that would slow down our model. Next, we’ll create a few methods to speed up the generation of random binomial numbers (something we’ll do frequently) by creating methods to call binomial using numba’s just in time compiler (jit). To do this, all we have to do is create a wrapper method, and then add the @jit decorator. This small modification provides a dramatic speed-up; without it the CA and plotting takes 16 min, but adding the simple @jit decorator reduces the time it takes down to around 5 minutes! Who says Python is always slow 🙂

CANCER CELL CLASS

OK, now that everything is set up, we’ll start creating our Cancer Cell class, which defines all of the cancer cell’s attributes and behaviors. First, we define the mandatory __init__ method, which defines the attributes each newly created cancer cell will have. In this case, each new cancer cell will be assigned a position (pos), the number of divisions it has remaining (divisions_remaining), and the list containing the positions in its neighborhood (neighbor_pos_list). This list is assigned simply by using the cell’s position to look up that list in the dictionary_of_neighbor_pos_lists that we created earlier.

Now that we’ve defined our cell’s attributes, we can start defining what exactly our cells will do. Because this is a fairly simple CA, we’ll just define all of the behaviors in one method (sort of…), which we’ll call act. First we’ll write up the bit of code for cell division. Recall that a cancer cell can only divide if that the cell is lucky and has an empty space in its neighborhood. In this case, if divide_q is 1, the cell is lucky and has a chance to divide. If the cell is indeed lucky, it then uses the locate_empty_neighbor_position method to find all of the free spaces in its neighborhood (the second requirement for division). Jumping to that locate_empty_neighbor_position method, the first thing we’ll do is create a list of positions that are NOT in the cell dictionary, that is, spaces which are currently unoccupied. If there are actually empty positions, a randomly selected empty position in the neighborhood is returned. Moving back to the act method, a new Cancer Cell is created, assigned that random position, made to look up its neighbor list in the dictionary_of_neighbor_pos_lists, and then added to the cell dictionary. Finally, the dividing cell decreases the number of divisions that it can undergo by one.

After a cell has tried to divide, it will determine if it needs to die. Recall that a cell will die for one of two reasons: 1) it dies because it has exceed its maximum proliferative potential, that is,  divisions_remaining <= 0; or 2) it dies spontaneously, that is, if die_q = 1 . If either of these conditions are true, the cell and it’s position are deleted from the cell dictionary, thus removing it from the CA and freeing up space.

CANCER STEM CELL CLASS

Now that our Cancer Cell class is created, we can start defining the behaviors of Cancer Stem Cells. We could simply clog up our code by adding a bunch of if/else statements to our Cancer Cell’s act method, but instead we’ll create a Cancer Stem Cell class, which will be a subclass of the Cancer Cell class. Classes in Python have inheritance, which means that our new Cancer Stem Cell class has all of attributes and methods of the Cancer Cell class, and so we don’t need to redefine all attributes or re-write the locate_empty_neighbor_position method. The only attribute we’ll redefine is the PLOT_ID, which we’ll use when visualizing the results of our CA. Next, we’ll redefine the act method. The process of division is the same as before, except that the stem cell can either divide symmetrically or asymmetrically. Which type of division occurs is determined by calling the divide_symmetrically_q method. If divide_symmetrically = 1, a new Cancer Stem Cell is created and added to the cell dictionary; if divide_symmetrically = 0, a normal Cancer Cell is created and added to the dictionary instead. Now we have our Cancer Cells and Cancer Stem Cells setup and ready for action!

RUN THE MODEL

The last thing we have to do is setup the initial conditions of our CA, which we’ll do under the if __name__ == “__main__” line. First, we’ll define the constants at the top. Next, we’ll create the initial Cancer Stem Cell, which will be positioned in the middle of world. We do this by initializing this first stem cell, using the position in the center of the world and DICTIONARY_OF_NEIGHBOR_POS_LISTS, and then add it to the cell dictionary. Next, we’ll copy the cells in the cell dictionary to a list (line 189 for Python 2, or line 192 for Python 3). While copying the values to a list isn’t ideal, it does provide two advantages. First, we can shuffle this list, allowing us to move through our cells randomly. Second, this list is what allows us to change the cell dictionary on the fly, as the length of a dictionary cannot be changed while iterating through it. Finally, we loop through our list of randomly ordered cells, having each one conduct their actions by calling their act method, a process that is repeated until maximum number of reps has been met.

VISUALIZE RESULTS

After we’ve completed our simulation, we would like to visualize our results, so we create a visualization_matrix, which is just matrix of zeros. Because we don’t care about the order of our dictionary, and won’t be modifying it’s length, we just iterate through the cells in our cell dictionary. Each cell then adds it’s PLOT_ID number to the matrix, so that stem cells will be one color, and normal cells another color. Now we use matplotlib’s imshow and show methods to visualize the results:

CONCLUSION

Using Python has its pros and cons. The pros are that it’s easy to read and write, reducing development time and making it easier to share code. The con is that, in its native form, Python is not the fastest. However, as we’ve seen, there are tools to increase performance. Numba is only one of those tools, but there are others, with the most popular probably being Numba Pro, CUDA, Cython, and PyPy. There are also ways to take advantage of multithreading and mulitprocessing to speed up your models even more. So, if its possible to use these tools, you can have both fast development and execution speed! I hope you found this post helpful, and that maybe you’ll even consider using Python for your next project.

# Quick implementation of hexagonal lattice for ABM

In the agent based modeling we are typically more interested in the rules governing the cell fate rather than the basic setting of the lattice (if we don’t look at the off lattice model). However, the particular setting of the computational domain can have an effect on the model dynamics. In 2D we typically consider the following lattices and neighborhoods:

with the rectangular grid with Moore neighborhood being probably most frequently utilized. With von Neumann neighborhood we have the fewest number of neighbors and cells are spatially saturated earlier. The problem with Moore neighborhood is that the distance to all of the sites is not the same. Hexagonal grid has good properties (same distance, 6 neighbors), but is less frequently utilized, because implementation is more involved. In todays post I will show that essentially there is no difference in implementation between all of those lattices.

We will use the codes for basic ABM model posted before as a template (MATLAB version here and C++ version here). In both cases we considered rectangular lattice with Moore neighborhood. If we set migration probability to zero, set large value of proliferation capacity and set the spontaneous death rate to zero, i.e. we simulate essentially only division events, we will see in visualizations of simulated tumors what kind of neighborhood we assumed:

Let us see how the visualization looks like for different types of neighborhoods/lattices.

In both implementations we had a separate array defining the neighborhood of the cell. Thus, in order to modify the code to von Neumann neighborhood we need to change only two consecutive lines in the MATLAB code in which we define the neighborhood and the permutations table:

aux = int32([-N -1 1 N])'; %indices to heighborhood
Pms = perms(uint8(1:4))'; %permutations


In C++ implementation we need to modify the definition of the neighborhood:

static const int indcNeigh[] = {-N, -1, 1, N};//neighborhood


and the for loop in the returnEmptyPlace function:

for(int j=0;j<4;j++) {//searching through neighborhood
if (!lattice[indx+indcNeigh[j]]) {
neigh[nF] = indx+indcNeigh[j];
nF++;
}
}


As it could be expected, switching to von Neumann neighborhood made the simulated tumor diamond shaped.

Let us now consider hexagonal lattice. The very first observation that we need to make is that hexagonal lattice is essentially an rectangular lattice with shifted odd (or even) rows and two definitions of neighborhood (one for cells in odd rows, second for cells in even rows), see picture below

Thus, in the modified implementation we again only need to change the definition of the neighborhood and add the condition for choosing the proper one. In MATLAB it can be achieved by introducing two definitions of neighborhoods

auxR = int32([-N -1 1 N-1 N N+1])'; %indices to heighborhood
auxL = int32([-N -1 1 -N-1 N -N+1])'; %indices to heighborhood

Pms = perms(uint8(1:6))'; %permutations


and modify the lines in the main loop in which we create the neighborhood for all viable cells (create variable S)

    odd = mod(cells,2) ~= 0; %selecting cells in the odd rows
S = zeros(6,length(cells));
if any(odd) %creating neighborhood for the cells in the odd rows
SR = bsxfun(@plus,cells(odd),auxR(Pms(:,randi(nP,1,sum(odd)))));
S(:,odd) = SR;
end
even = ~odd;
if any(even) %creating neighborhood for the cells in the even rows
SL = bsxfun(@plus,cells(even),auxL(Pms(:,randi(nP,1,sum(even)))));
S(:,even) = SL;
end


In C++ the changes are even easier. We again introduce two neighborhoods:

static const int indcNeighR[] = {-N,-1,1,N-1,N,N+1};//neighborhood
static const int indcNeighL[] = {-N,-1,1,-N-1,N,-N+1};//neighborhood


and modify the returnEmptyPlace function

for(int j=0;j<6;j++) {//searching through neighborhood
if (indx % 2) {//if odd row
if (!lattice[indx+indcNeighR[j]]) {
neigh[nF] = indx+indcNeighR[j];
nF++;
}
} else {//if even row
if (!lattice[indx+indcNeighL[j]]) {
neigh[nF] = indx+indcNeighL[j];
nF++;
}
}
}


We might also want to change the code for visualization. In the previous settings it was enough to represent each cell as a pixel in the image. In case of hexagonal lattice in order to be correct we need to somehow shift the pixels in odd (or even rows). The easiest way to achieve that is to represent each cell as 4 pixels and adjust position according to the row number:

The modified visualization code in MATLAB is the following

function visualizeHex( N, cells, cellsIsStem, cellsPmax, pmax )

%select cells in the odd rows
odd = mod(cells,2) ~= 0;

M = ones(2*N,2*N,3); %matrix for image, we expand it

%disperse cell by one spot
i = mod(double(cells)-1,N)+1; %row, REMEMBER THAT CELLS ARE UINTs!
j = ceil(double(cells)/N); %column
cells = (i*2-1)+(2*j-1)*2*N;

%add cells to the top right in odd rows
cells = [cells reshape(bsxfun(@plus,cells(odd),[-1; 2*N-1; 2*N]),1,[])];
cellsIsStem = [cellsIsStem reshape(repmat(cellsIsStem(odd),3,1),1,[])];
cellsPmax = [cellsPmax reshape(repmat(cellsPmax(odd),3,1),1,[])];

%add cells to top left in even rows
even = [~odd false(1,3*sum(odd))];
cells = [cells reshape(bsxfun(@plus,cells(even),[-1; -2*N-1; -2*N]),1,[])];
cellsIsStem = [cellsIsStem reshape(repmat(cellsIsStem(even),3,1),1,[])];
cellsPmax = [cellsPmax reshape(repmat(cellsPmax(even),3,1),1,[])];

%plotting
color = hot(3*pmax);
M(cells(~cellsIsStem)) = color(cellsPmax(~cellsIsStem)+1,1);
M(cells(~cellsIsStem)+4*N*N) = color(cellsPmax(~cellsIsStem)+1,2);
M(cells(~cellsIsStem)+8*N*N) = color(cellsPmax(~cellsIsStem)+1,3);

CSCs = cells(cellsIsStem);
M(CSCs) = color(2*pmax,1);
M(CSCs+4*N*N) = color(2*pmax,2);
M(CSCs+8*N*N) = color(2*pmax,3);

figure(1)
clf
imshow(M);

end


Finally, simulating the tumor with exactly the same parameters settings on the hexagonal lattice results in way more circular tumor.

# Agent-Based Modeling (ABM) of Tumors in Java

Before I begin I would like to quickly introduce myself — my name is Sameer Puri and I am a high school student working in an internship under Dr. Heiko Enderling through the HIP IMO program at Moffitt Cancer Center.

Over the past 5 weeks, I have been developing an ABM in Java for my research. My principal goal in developing this ABM was to quickly generate tumors with 1 million cells. (And prove that Java can perform better in some cases…!)

When I started, I was working with an ABM written by @jpoleszczuk and Heiko in C++ (available here). Having a strong background in Java and little past experience with C++, it was difficult for me to understand. Structs, vectors, memcpy, and other features of C++ are not present in Java. Instead of trying to learn C++, I converted the code to Java, my language of expertise.

Optimization

I decided to optimize the code before making any extensions for my research. In Java, finding slow-down points is simple, since an array of profiling tools are available to find which part of the code is slowing down the ABM. I utilized the web-based profiler WarmRoast, developed by sk89q (on GitHub here).

• One of the biggest time hogs was the Random class which was implemented in the 1990s, and after 20 years, much more efficient pseudo-random number generators have emerged. I replaced it with the XorShift generator, developed by George Marsaglia. The XorShift is theoretically 3 times faster than the default generator. However, the speed-up was far more significant — the call for a random double was so fast that it disappeared from the WarmRoast profiler’s results.
• The returnEmptyPlace method was optimized with a strategy known as pregeneration, or exchanging memory for lower CPU time. The returnEmptyPlace method looks for an empty spot around a cell in the ABM, shuffling a list of 8 directions each time. Instead, I pregenerated all 40,320 permutations of the 8 direction vectors surrounding a lattice square. Now, the returnEmptyPlace method simply uses the speedy XorShift generator to select a random permutation of the directions list to use.
• I also optimized the shuffling of the list of cells in the main simulation loop. I created a new class called SPCollections which uses the XorRandom generator for shuffling. There was a large speed-up here as well.
• Using NetBeans’ built-in hashcode generator, I added hashcodes for the Cell and Indx classes, which will speed up the usage of HashMaps or other hash-based structures.
• Since memory was not a serious concern for me, I made one optimization. The original model in C++ used the char primitive type for its lattice. I turned the lattice into a byte lattice, which has half the memory footprint (chars are 16 bits, bytes are 8 bits).

• For my research, I had to grow 10 tumors in silico with one million cells each. To speed up simulation I utilized parallel computing with an ExecutorService. With my ABM, you can specify the number of “trials” you want to run. For each trial run, a simulator instance is created and run on a separate thread. Most CPUs today have at least 4 cores. Code in most programming languages tend to make use of only one core unless programmers specify methods that allow for parallel execution of code. By creating a separate thread for each simulation, the threads can operate on separate cores and maximize CPU usage!
• To universalize my data, I saved it utilizing Java serialization. With serialization, Java can output objects to files, which can be loaded later on without needing to code a loader. Others will be able to easily load data generated using my ABM and run their own operations on it.
• I coded an image generator to automatically generate tumor images. Once a simulation is finished, the image generator takes the serialized output of the simulation and generates three types of PNG images for each dumped instance: proliferation, quiescence, and nutrition. It also creates an animated GIF to show the growth of the tumor over time. Each image has the time step (1 step = 1 hour), number of stem cells and number of regular cancer cells at the upper left corner.

Proliferation during Development

Nutrition during Development (on rim)

Quiescence during Development

# Parameter dependent implementations

When dealing with ordinary or partial differential equations we know that existing numerical solvers have different performance depending on the specific mathematical problem. In particular, the performance of any given solver can dramatically change if we change model parameters. Let us take the well known logistic equation as an example. Let C(t) be the current number of cancer cells, r the growth rate, and K the maximal number of cancer cells that can be sustained by the host. Then the growth dynamics can be described by the following equation:

$\frac{dC(t)}{dt}=rC(t)\left(1-\frac{C(t)}{K}\right).$

The above equation has of course analytical solution, but let us use a MATLAB ordinary differential equations solver to find a numerical solution. Generally it is a good idea to start with the solver that has the highest order – in our case it will be ode45 based on the Runge-Kutta method. We will also consider another solver, ode15s, which is designed for solving stiff systems. We fix the value of carrying capacity K, initial condition C(0), and time interval in which we solve the equation. We will investigate the performance of both solvers for different values of growth rate r. The plots below show that for lower values of growth rate ode45 outperforms the ode15s, but situation changes when the growth rate is large (while keeping small relative difference between the calculated solutions).

That transition in solvers performance can be explained when one looks how solution changes for different values of growth rates (see figure below) – the larger the growth rate the steeper the solution. Stiff solvers, like ode15s, are designed to effectively deal with such phase transitions.

So, if we need to consider parameter dependent performance of the numerical solvers for differential equations, should we do the same for agent based models of cancer growth? Should we create parameter dependent implementations and switch between them on the fly? I think that yes, and I’ll try to convince you with the simple example below.

Let us consider a very simple CA: cell can either divide or migrate if there is a free spot – nothing else. We can adapt the very basic C++ implementation from the previous post as a baseline: we have a boolean matrix (in which value true indicates that the spot is occupied by the cell) and additional vector of cells keeping information about locations of all viable cells present in the system. At each simulation step we go through all of the cells in a vector in random order and decide about faith of each cell.

Let us tweak the code and at each iteration drop from memory (from vector) cells that are quiescent (completely surrounded). If the cell stays quiescent for prolonged time we save some computational time by skipping its neighborhood search. However, after each migration event we need to add some cells back to the vector – that is additional cost generated be searching the lattice again. The essential part of the tweaked code in C++ is:

   for (int i=0; i<nSteps; i++) {
random_shuffle(cells.begin(), cells.end()); //shuffling cells
while (!cells.empty()) {
currCell=cells.back(); //pick the cell
cells.pop_back();
newSite = returnEmptyPlace(currCell);
if (newSite) {//if there is a free spot in neigh
if ((double)rand()/(double)RAND_MAX < pDiv) {//cell decides to proliferate
lattice[newSite] = true;
cellsTmp.push_back(currCell);
cellsTmp.push_back(newSite);
nC++;
} else if ((double)rand()/(double)RAND_MAX < pmig) {//cell decides to migrate
toDelete.push_back(currCell);
lattice[newSite] = true;
} else {//cell is doing nothing
cellsTmp.push_back(currCell);
}
}
}
cells.swap(cellsTmp);

//freeing spots after cell migration and adding cells to vector
for(int j =0; j<toDelete.size(); j++)
lattice[toDelete.at(j)] = false;

c = !toDelete.empty();
while(!toDelete.empty()) {
newSite = toDelete.back();
toDelete.pop_back();
//adding cells that can have space after migration event
for(int j=0;j<8;j++) {//searching through neighborhood
if (newSite+indcNeigh[j]>=0) {
cells.push_back(newSite+indcNeigh[j]);
}
}
}

//removing duplicates in cells vector
if (c) {
sort( cells.begin(), cells.end() );
cells.erase( unique( cells.begin(), cells.end() ), cells.end() );
}
}


Depending on the proliferation/migration rates we can imagine that the tweaked code will have different performance. For low  probability of migration event only cells on the tumor boundary will have some space (green cells on the left figure below) and we save a lot of time by dropping quiescent cells (blue cells in figures below). For higher migration rates, however, less than 30% of the cells are quiescent at each proliferation step (compare right figure below) and the benefit of dropping cells can be outweighed by the after migration update steps.

The plot below shows the computational time needed to reach 500,000 cells when using the baseline code (all cells, blue line) and tweaked one (only proliferation, dashed green line) for a fixed proliferation probability. For lower proliferation rates change in the performance would occur even earlier.

I think that this simple example show that we always need to think how our model will behave before choosing the way to implement it.

# Quick parallel implementation of local sensitivity analysis procedure for agent-based tumor growth model

In the last couple of posts I’ve shown how to implement agent-based model of cancer stem cell driven tumor growth, both in MATLAB and C++. Having the implementations we can go one step further and perform some analysis of the tumor growth characteristics predicted by the model. We will start with performing local sensitivity analysis, i.e. we will try to answer the question how the perturbation of parameter values impacts the growth dynamics. Typically it is being done by perturbing one of the parameters by a fixed amount and analyzing the response of the model to that change. In our case the response will be the percentage change in average tumor size after 3 months of growth. Sounds fairly simple, but…

We have 5 different parameters in the model: proliferation capacity, probability of division, probability of spontaneous death, probability of symmetric division, and probability of migration. Moreover, let us assume that we would like to investigate 3 different values of parameter perturbation magnitude (5%, 10% and 20%). In order to be able to analyze the change in the average size we need to have its decent estimator, i.e. we need sufficient number of simulated stochastic trajectories – let us assume that 100 simulations is enough to have a good estimation of “true” average. So in order to perform the sensitivity analysis we will need to perform 100 + 3*5*100 = 1600 simulations (remember about the growth for nominal parameters values). Even if single simulation takes typically 30 seconds, then we will wait more than 13 hours to obtain the result using single CPU – that is a lot!

After looking at the above numbers we can make a straightforward decision right now – we will use C++ instead of MATLAB, because the model implementation in C++ is a several times faster. However 1) we will need to write a lot of a code in order to perform sensitivity analysis, 2) using multiple CPU is not that straightforward as in MATLAB. Is there a better way to proceed?

Few weeks ago I’ve shown here how to wrap your C++ in order to use it from within MATLAB as a function without loosing the C++ performance. Why not to use it to make our lives easier by making sensitivity code short and harness easily the power of multiple CPUs?

We will start the coding (in MATLAB) with setting all simulation parameters

nSim = 100; %number of simulations to perform for a given set of parameters
tmax = 30*3; %number of days to simulate
N = 1000; %size of the simulation domain

%nominal parameter values [rhomax, pdiv, alpha, ps, pmig]
nominal = [10, 1/24, 0.05, 0.3,10/24];

perturb = [0.05 0.1 0.2]; %perturbation magnitudes, percent of initial value up


Now we just need to construct a loops that will iterate through all possible perturbations and simulations. If we don’t use Parallel Toolbox, i.e. don’t use multiple CPUs, it doesn’t really matter how we will do that – performance will be similar. Otherwise, implementation is not that straightforward even though from MATLABs documentation it seems that it is enough to change for to parfor. The most important thing is how will we divide the work between the CPUs. The simples idea is to spread the considered perturbations values among the CPUs – that will allow to use 15 CPUs in our setting. However, I’ve got machine with 24 CPUs, so that would be a waste of resources – bad idea. The other idea would be to use parfor loop to spread all 100 simulations for a given perturbation value on all 24 CPUs and go through all perturbation values in a simple loop – now we are using all available CPUs. But are we doing that efficiently? No. The thing is that CPUs need to be synchronized before proceeding to the next iteration of the loop going through perturbation values. So some of the CPUs will be idle while waiting for other ones to finish with parfor loop. In order to make the code even more efficient we will just use one parfor loop and throw all 1600 simulation on 24 CPUs at the same time. Let us first prepare the output variable.

HTCG = zeros(1,nSim + length(nominal)*length(perturb)*nSim);


Before writing the final piece of the code we need to solve one more issue. Namely, in the C++ implementation we used srand(time(NULL)) to initiate the seed for random number generator. It is perfectly fine when we use single CPU – each simulation will take some time and we don’t need to worry about uniqueness of the seed. The problem is when we want to use multiple CPUs – all initial parallel simulations will start with exactly the same seed. One way to solve that is to pass the current loop iteration number (i) to C++ and use srand(time(NULL)+i) – that is what I have done. After solving that issue we can write the final piece of the code.

parfor i = 1:length(HTCG)
%%PREPARING PARAMETERS VALUES
params = nominal; %setting parameters to nominal values
if i>nSim %simulation is for perturbed parameters
%translating linear index to considered parameter and perturbation value
j = ceil((i-nSim)/(nSim*length(perturb)));
k = ceil((mod((i-nSim)-1,nSim*length(perturb))+1)/nSim);
%updating parameters
params(j) = params(j)*(1+perturb(k));
if k == 1 %if proliferation capacity parameter we need to round it
params(j) = round(params(j));
end
end

%%RUNNING SIMULATION AND SAVING OUTPUT
[~, cells] = CA(params,[tmax*24,N,i]);
HTCG(i) = length(cells)/3;
clear mex %important! without that the internal
%variables in CA functions won't be cleared and next simulation
%won't begin with single initial cell
end


Then in the command line we start the parallel pool with 24 workers (CPUs), by typing parpool(24) command, and run the code. Screenshot below shows nicely how all of the 24 CPUs are being used – no resource wasted!

We can then add few additional lines of the code to plot the results.


nom = mean(HTCG(1:100)); %average for nominal parameters value
%calculating averages for perturbed sets values
av = squeeze(mean(reshape(HTCG(101:end),nSim,length(perturb),length(nominal))));

%plotting results of sensitivity analysis
set(0,'DefaultAxesFontSize',18)
figure(1)
clf
bar(perturb*100,(av-nom)/nom*100)
legend({'\rho_{max}', 'p_{div}', '\alpha', 'p_s', 'p_{mig}'})
xlabel('% perturbation')
ylabel('% change')



And “voila!” – the resulting figure shows that the perturbation in the proliferation capacity has the highest impact on the tumor growth dynamics.