BIOLOGO

1
BIOLOGO

• A domain-specific language for morphogenesis.
• Trevor M. Cickovski
• December 9, 2004

2
What is Morphogenesis?

• A stage in embryonic development
• Mesenchymal cells begin to cluster and form
  patterns
• Chemical secretion/resorption/diffusion
• Cell differentiation
• Cell growth
• Cell division
• Cell migration

3
Chemical secretion/resorption/diffusion

• Single Dictyostelium cell undergoing chemotaxis
  (Source DictyBase)

4
Cell Growth and Division

• Division of a Dictyostelium cell (Source
  DictyBase)

5
Cell migration

• Migration of a Dictyostelium slug (Source
  DictyBase)

6
Example Avian Limb Bud

• Chick limb bud, after 3.5 days
• Source (Hinchliffe and Johnson, 1980)

7
Avian Limb Stages

• Schematic Representation
• Forelimb Pattern Formation Order
• Humerus
• Radius/Ulna
• Carpals/Metacarpals
• Digits

8
Example Dictyostelium Discoideum
Dictyostelium life cycle from 12h to 24h.
Multicellular slug body migrates towards a
chemical gradient to the right. Eventually the
slug culminates into a fruiting body, with a
spore, stalk and tip. (Source,
Dictyostelium-homepage zl-munich).
9
Software Modeling of Morphogenesis

• CompuCell3D, a C framework for
  three-dimensional simulation of morphogenesis
• BIOLOGO, a domain-specific language for
  morphogenesis, used to extend CompuCell3D

10
Motivation For BIOLOGO

• The Problem That Arises
• Morphogenesis researchers may not have C
  experience
• CompuCell3D is extended through plugins, a C
  design pattern
• This poses a serious challenge for extensibility
• Namely, the real users wont be able to extend
  it!
• Thats because the plugin method for extending
  CompuCell3D is too low-level for them if they
  dont know C

11
Solution

• Extend CompuCell3D through a domain-specific
  language which we call BIOLOGO
• Make syntax of BIOLOGO understandable to
  morphogenesis researchers
• Create a higher level of abstraction
• Allow representation of biological phenomena in a
  more structured and straightforward manner
• Generate plugin extensions for CompuCell3D,
  readable and with no performance penalty

12
Validation

• Generality
• We apply BIOLOGO to three biologically relevant
  validation simulations.
• Accuracy
• We compare the visualization of results using
  BIOLOGO-extended CompuCell3D with visualization
  using handcoded extensions, and verify they are
  the same.

13
Related Work Similar Domains, Different Design
Patterns

• Cell Programming Language (CPL) (Agarwal, 1994)
• Simulates morphogenetic cellular behavior
• Each cell is a program, different cell types run
  different programs, represented as pixels
• CPL instructions are mapped to Objective C
  functions, we allocate objects dynamically
• Nonscalable move, divide and grow algorithms

14
Related Work Similar Domains, Different Design
Patterns

• CellML (Nelson, 2004)
• Also based on XML
• Very high flexibility by allowing
  embedding/inclusion of other supported languages
• For simulation of cellular/subcellular processes
• Cell cycle, cell electrophysiology, muscle
  methods
• Operates at a different level of modeling than
  BIOLOGO

15
Related Work Similar Design Patterns, Different
Domains

• SPIRAL (Puschel et al., 2004)
• Domain Signal Processing
• Implements a compiler for the DSL SPL
• Translates SPL source to C/Fortran source with
  optimizations
• TCE (Baurngartner et al., 2002)
• Domain Tensor Contractions (Chem/Phys)
• Compiler for a Mathematica-style DSL
• Runs generalized matrix multiplications
• Translates DSL source to Fortran source with
  optimizations

16
The CPM A Mathematical Model for Morphogenesis

• Cells represented in a 3D lattice
• Each unique cell given a different integer index,
  indices stored in pixels
• Extracellular matrix has 0 index
• Neighbors and levels (1-4) given for a pixel S

17
Metropolis Algorithm in the CPM

• Choose a pixel at random
• Propose to change the pixel index to that of one
  of its neighbors (index flip)
• Execute the flip with Monte Carlo probability
  based on the resulting energy from the flip

18
CPM Energy Calculation

• Three terms
• results from adhesion between adjacent
  pixels
• results from cell deviation from their
  target volume and surface area
• results from cell chemotaxis or
  haptotaxis to a secreted or diffusing chemical.

19
CPM Energy Equations
20
Computational Modeling Issues

• Software must be extensible, flexible and easy to
  use, allowing
• Extensible CPM Hamiltonians
• Cell Type Maps for various organisms
• Arbitrary number of superimposed chemical fields
• Large 3D CPM lattices
• Speed and memory usage concerns

21
Miscellaneous Data

• Class Count
• CompuCell3D 134
• BIOLOGO 85
• Lines of code
• CompuCell3D 18347
• BIOLOGO 9197
• Plus external Xerces libraries (Apache, 2003)

22
CompuCell3D Overview
23
Customizing CompuCell3D
CompuCell3D defines a set of classes that
can be extended to add features to a simulation.
Steppables execute once per Monte Carlo step and
once before and after the main loop. They are
the main hooks for initialization and rendering.
Plugins load at runtime. They are the main way
of adding new features to CompuCell. They can be
Steppables, Steppers, CellChangeWatchers, or
Automatons.
CellChangeWatchers execute once per each
successful spin flip. They are useful for
adjusting values that depend on the number of
lattice points in a cell.
Steppers execute once per spin flip attempt.
They are the main hooks for energy functions.
Some simulation features, such as Renders are so
common that they are built into the system.
Automatons enable cell state to change their
state as the simulation evolves.
24
What BIOLOGO Saves

• Requirements to add a new Cell Type Map to
  CompuCell3D
• Example, simulating a brand new organism with
  different cell types and transitions
• 5 C classes plus 1 for each transition, and
  appropriate methods and implementation of rules

25
What BIOLOGO Saves

• Requirements to add a new Hamiltonian to
  CompuCell3D
• Example, we want to add an energy due to cell
  chemotaxis
• 2 C classes including full implementation of
  the energy calculations, plus any superimposed
  fields

26
Generated Code

• In the form of CompuCell3D plugins
• Plugins are the standard method for extending
  CompuCell3D
• Plugins are dynamically loaded libraries
• They are loaded upon reference in the
  configuration file

27
Three Validation Simulations

• Cell Sorting (Beysens et al., 2000)
• Basic CPM adhesion and volume
• No chemical energy or superimposed chemicals
• Avian Limb (Cickovski et al., 2004)
• Cells undergo haptotaxis with chemical
  fibronectin
• Domain grows in time
• Dictyostelium discoideum (Maree, 2000)
• Polarity within the slug
• Dynamic activator field

28
Predefined BioLogo variables

• Reference Cellular Potts Model algorithm
• potts.cellfield 3D field of type cell. The
  central Potts lattice.
• pt The selected Potts pixel
• oldcell The selected Potts cell
• newcell The candidate Potts cell

29
Validation Simulation Avian Limb (Cickovski et
al., 2004)

• Cells haptotax to secreted chemical fibronectin
• Secretion stimulated by exterior activator
• Cells can be one of two types Condensing and
  NonCondensing
• Condensing location of high activator
• Condensing more adhesive

30
Validation Simulation Avian Limb Growth
(Cickovski et al., 2004)
31
Validation Simulation Avian Limb Growth
(Cickovski et al., 2004)
ltcellmodel nameChickGrowthgt ltuseplugin
nameLimbChemical /gt ltcelltype
nameNonCondensinggt ltupdatecelltypesgt
ltchangeif currenttypeCondensing
conditionLimbChemical.
ReactionDiffusionpt.xpt.ypt.z less

LimbChemical.ActivatorThreshold /gt
lt/updatecelltypesgt lt/celltypegt


lt/cellmodelgt
32
Validation Simulation Avian Limb Growth
(Cickovski et al., 2004)
ltcellmodel nameChickGrowthgt ltuseplugin
nameLimbChemical /gt ltcelltype
nameNonCondensinggt ltupdatecelltypesgt
ltchangeif currenttypeCondensing
conditionLimbChemical.
ReactionDiffusionpt.xpt.ypt.z less

LimbChemical.ActivatorThreshold /gt
lt/updatecelltypesgt lt/celltypegt ltcelltype
nameCondensinggt ltupdatecelltypesgt
ltchangeif currenttypeNonCondensing
conditionLimbChemical.R
eactionDiffusionpt.xpt.ypt.z greater

LimbChemical.ActivatorThreshold /gt
lt/updatecelltypesgt lt/celltypegt lt/cellmodelgt
33
Validation Simulation Avian Limb Growth
(Cickovski et al., 2004)
Generated Code
34
Validation Simulation Avian Limb Growth
(Cickovski et al., 2004)
Generated Code
35
Validation Simulation Avian Limb Growth
(Cickovski et al., 2004)
Generated Code
36
Validation Simulation Avian Limb Growth
(Cickovski et al., 2004)
Generated Code
37
Validation Simulation Avian Limb Growth
(Cickovski et al., 2004)
Generated Code
38
Validation Simulation Avian Limb Growth
(Cickovski et al., 2004)
39
Validation Simulation Avian Limb Growth
(Cickovski et al., 2004)
ltHamiltonian nameLimbChemicalgt ltInput
nameActivatorThreshold typedouble /gt
ltInput nameLambda typeint /gt ltInput
nameFibroRate typedouble /gt



40
Validation Simulation Avian Limb Growth
(Cickovski et al., 2004)
ltHamiltonian nameLimbChemicalgt ltInput
nameActivatorThreshold typedouble /gt
ltInput nameLambda typeint /gt ltInput
nameFibroRate typedouble /gt ltInput
nameConcentrationFile typefile
fieldnameReactionDiffusion fieldtypefloat
/gt ltField nameFibroField typefloat /gt



41
Validation Simulation Avian Limb Growth
(Cickovski et al., 2004)
ltHamiltonian nameLimbChemicalgt ltInput
nameActivatorThreshold typedouble /gt
ltInput nameLambda typeint /gt ltInput
nameFibroRate typedouble /gt ltInput
nameConcentrationFile typefile
fieldnameReactionDiffusion fieldtypefloat
/gt ltField nameFibroField typefloat /gt
ltStepgt ltif conditionoldcell.type
notequal Mediumgt ltdeclaregtltfloat
namefibro valueFibroFieldpt.xpt.ypt.z
/gtlt/declaregt ltif conditionReactionDi
ffusionpt.xpt.ypt.z greaterequal
ActivatorThresholdgt ltcopy
toFibroFieldpt.xpt.ypt.z
fromfibroFibroRate /gt ltcopy
tofibro fromFibroFieldpt.xpt.ypt.z
/gt ltreturn valuefibroLambda
/gt lt/ifgt lt/ifgt
ltreturn value0 /gt lt/Stepgt lt/Hamiltoniangt
42
Validation Simulation Avian Limb Growth
(Cickovski et al., 2004)
Instantiation in CompuCell3D Configuration File
ltPlugin NameLimbChemical
ltActivatorThresholdgt0.7lt/ActivatorThresholdgt
ltLambdagt10lt/Lambdagt ltFibroRategt0.01lt/FibroRate
gt ltConcentrationFilegtbnewSys123_71x31x281.datlt
/ConcentrationFilegt lt/Plugingt
43
Validation Simulation Dictyostelium discoideum
(Maree, 2000)

• Dictyostelium discoideum is a type of slime mould
• Begins as unicellular
• Develops into a multicellular fruiting body
• Follows a Cyclic AMP (cAMP) gradient that leads
  the slug to the soil surface
• Marees simulation starts from this point

44
Validation Simulation Dictyostelium discoideum
(Maree, 2000)

• Piecewise Puschino kinetics determine cAMP
  gradient
• This field is input from a file, but is dynamic
• Means We need a file read frequency!!!!
• Cell type map has three cell types
• Autocycling (tip)
• Prestalk (form an upper stalk, initially 40 of
  body)
• Prespore (form a lower spore, initially 60 of
  body)
• Simple map with no type transitions

45
Validation Simulation Dictyostelium discoideum
(Maree, 2000)
ltHamiltonian nameDictyChemicalgt ltInput
nameLambda typeint /gt

lt/Hamiltoniangt
46
Validation Simulation Dictyostelium discoideum
(Maree, 2000)
ltHamiltonian nameDictyChemicalgt ltInput
nameLambda typeint /gt ltInput
nameConcentrationFile typefile
fieldnameReactionDiffusion fieldtypefloat
frequencysteps /gt

lt/Hamiltoniangt
47
Validation Simulation Dictyostelium discoideum
(Maree, 2000)
ltHamiltonian nameDictyChemicalgt ltInput
nameLambda typeint /gt ltInput
nameConcentrationFile typefile
fieldnameReactionDiffusion fieldtypefloat
frequencysteps /gt ltStepgt
ltif conditionoldcell.type equal
Autocyclinggt ltreturn
valueReactionDiffusionpt.xpt.ypt.z
Lambda /gt lt/ifgt lt/return
value0.0 /gt lt/Stepgt lt/Hamiltoniangt
48
Validation Simulation Dictyostelium discoideum
(Maree, 2000)
Reference in CompuCell3D Configuration File
ltPlugin NameDictyChemicalgt
ltLambdagt200lt/Lambdagt ltConcentrationFilegtbincac
tivatorlt/ConcentrationFilegt
ltConcentrationFileFrequencygt100lt/ConcentrationFile
Frequencygt lt/Plugingt
49
Validation Simulation Dictyostelium discoideum
(Maree, 2000)
50
Implications

• Extension Core of CompuCell3D is untouched by
  BIOLOGO
• Modeling BIOLOGO is a modeling language it does
  not need to be run multiple times for the same
  features
• Instantiation BIOLOGO generates extensions to
  CompuCell3D that can be instantiated through the
  CompuCell3D configuration file

51
Implementation Details

• Framework consists of two modules
• Compiler
• Generates Intermediate File
• Code Generator
• Generates CompuCell3D Source

52
Known DSL Design Patterns Used

• Source-To-Source Transformation
• Translation is done from BIOLOGO source to
  CompuCell3D source
• Data Structure
• Complex data structures (i.e. Cell Type Maps)
  represented as simpler BIOLOGO cell models
• Language Extension
• BIOLOGO extends XML

53
Video Demo
54
Future Work

• 1) Higher abstractions
• Have Hamiltonians use mathematical notation
• Make BIOLOGO a visual language
• Interface to a GUI (i.e., in Java)
• Potentially follow the example of Modeler
  (Michel et al., 2001)

55
Future Work

• 2) Web service
• Help for those who do not have built-in compilers
• Use compile farm
• Compile and create plugins on the server, client
  downloads them and runs CompuCell3D

56
Future Work

• 3) Management of Changes
• If a user runs BIOLOGO and wants to permanently
  commit their new feature to CompuCell3D, this
  could affect other users
• Need a change manager
• Distributed plugin repository

57
Future Work

• 4) Implement a hybrid XML and scripting DSL
• XML is not as ideal for computation, better for
  structural modules
• Computation can be done through scripting
• For example, embedding Lua or Python
• Pros More flexibility, less setup time
• Cons Interpreted

58
Future Work

• 5) Intermediate-level optimizations (cf. Aho et
  al., 1986)
• Right now no performance penalty for generated
  code, but this could ensure high efficiency
• Examples
• Optimal algorithm selection
• Optimize code for CompuCell3D compiler on target
  machine

59
Future Work

• 6) Improve DSL performance by making BIOLOGO a
  telescoping language (Kennedy et al., 2000)
• Generate a custom compiler for intermediate
  language with extensive knowledge of underlying
  libraries on the target machine
• Generates code with maximum efficiency
• Downside Creating custom compiler takes a long
  time

60
Publications

• 1) R. Chaturvedi, J. A. Izaguirre, C. Huang, T.
  Cickovski, P. Virtue, G. Thomas, G. Forgacs, M.
  Alber, S. A. Newman, and J. A. Glazier,
  Multi-Model Simulations of Chicken Limb
  Morphogenesis. Springer Verlag LNCS 2659,
  Computational Science - ICCS 2003, International
  conference Melbourne, Australia and St.
  Petersburg, Russia, Part II, pp. 39-49, June
  2003.
• 2) J. A. Izaguirre, R. Chaturvedi, C. Huang, T.
  Cickovski, J. Coffland, G. Thomas, G. Forgacs, M.
  Alber, S. A. Newman, and J. A. Glazier,
  CompuCell, A Multi-Model Framework for Simulation
  of Morphogenesis. Bioinformatics,
  20(7)1129-1137, 2004.
• 3) T. Cickovski, T. Matthey and J. A. Izaguirre,
  Design Patterns for Generic Object-Oriented
  Software. University of Notre Dame Technical
  Report 2004-29, November 2004.

61
Submitted/In Progress

• 1) T. Cickovski, C. Huang, R. Chaturvedi, T.
  Glimm, H. G. E. Hentschel, M. Alber, J. A.
  Glazier, S. A. Newman, J. A. Izaguirre, A
  Framework for Three-Dimensional Simulation of
  Morphogenesis. Submitted to IEEE/ACM
  Transactions on Computational Biology and
  Biocomplexity, 2004.
• 2) T. Cickovski and J. A. Izaguirre, BIOLOGO, A
  Domain-Specific Language for Morphogenesis. In
  progress, to be submitted to ACM Transactions on
  Programming Languages and Systems.

62
Conferences

• Already Attended
• Midwest Numerical Analysis Day (Northern Illinois
  University, May 2003)
• Biocomplexity IV (Indiana University, May 2003)
• Biocomplexity VI (Indiana University, May 2004)
• To Attend
• SIAM Computational Science and Engineering
  (CSE05) (Orlando, February 2005)

63
Acknowledgements

• Special thanks to the following
• Mark Alber, Kedar Aras, Brian Bien, Rajiv
  Chaturvedi, David Cieslak, Joseph Coffland, James
  A. Glazier, Tilmann Glimm, George Hentschel,
  Chengbang Huang, Thierry Matthey, Chris Mueller,
  Stuart A. Newman, Troy Raeder, Matthew Rissler
  and Todd Schneider.
• Last and most
• To my advisor, Jesus A. Izaguirre
• Funding
• My funding was provided by a Schmitt Fellowship
  and NSF grants IBN-0083653 and IBN-0313730

64
Supplementary Material Starts Here

• Slides Available Upon Request
• Cell Sorting Example
• More Syntax
• CompuCell3D Applied Patterns
• Generated C For Avian Limb Chemical Hamiltonian
• Miscellaneous Data

65
Example Basic Cell Sorting

• Two different types of cells
• One type is very adhesive to other cells of the
  same type
• All cells are repelled by the medium

Source Beysens et al., 2000, from the
experiments of Steinberg et al., 1963 and 1998.
66
Validation Simulation Cell Sorting (Beysens et
al., 2000)

• A basic model of cells of two different types
  (Light and Dark), clustering.
• Dark cells are highly adhesive
• Light cells are less adhesive
• All cells are repelled by the medium

67
Validation Simulation Cell Sorting (Beysens et
al., 2000)
ltcellmodel nameCellSortgt ltdeclaregtltint
nameflag /gtlt/declaregt ltcelltype
nameLightgt ltcreationgt
ltcopy toflag from0 /gt lt/creationgt
ltupdatevariablesgt ltcopy
toflag from1 /gt lt/updatevariablesgt
lt/celltypegt ltcelltype nameDarkgt
ltupdatecelltypesgt ltchangeif
currenttypeLight
condition((flag equal 0) and (drand48()
greaterequal .5)) /gt lt/updatecelltypesgt
lt/celltypegt lt/cellmodelgt
68
Validation Simulation I Cell Sorting (Beysens et
al., 2000)

• Ran on CompuCell3D, Visualized with Ogle

69
Syntax Basic Statements
70
Syntax Declare Stmt

• ltdeclaregtlttype name valuegtlt/declaregt
• Types (data members use . operator)
• boolean, int, float, double, char, string
• pixel
• Data members x, y, z
• cell
• Data members are cell state variables
• Defined by the cell model

71
Syntax Expressions

• Infix notation
• Symbols
• , -, , /, , (, )
• greater, less, greaterequal, lessequal, equal,
  notequal
• - encompasses characters and strings
• Can enscript C functions as well
• A note Do minimally and with care right now,
  BIOLOGO does not have complete built-in C
  parsing. Future goal is to improve this.

72
Syntax ForNeighbors

• ltforneighbors variable point grid
  distance depth checkboundsgt
• .. BIOLOGO Statements
• lt/forneighborsgt
• Function
• At each iteration select a neighbor to
  ltpointgt within ltgridgt, store in ltvariablegt
• Store distance between ltvariablegt and ltpointgt
  in ltdistancegt
• Loop until ltdistancegt hits ltdepthgt
• Implement neighbor bounds checking if
  ltcheckboundsgt is true

73
CompuCell3D Features/Patterns

• Allows different boundary conditions per axis
  through the Strategy and Factory design patterns
• Dynamic class nodes contiguously allocate all
  attributes of a particular cell, reducing cache
  misses and page faults
• Singleton object for medium pixels
• Lazy pixel neighbor evaluation
• Factory pattern for cell object creation

74
C Code Generation For Avian Limb Chemical
Hamiltonian

• BIOLOGO code
• ltHamiltonian nameLimbChemicalgt
• ltInput nameActivatorThreshold
  typedouble /gt
• ltInput nameLambda typeint /gt
• ltInput nameFibroRate typedouble /gt
• ltInput nameConcentrationFile typefile
  fieldnameReactionDiffusion fieldtypefloat
  /gt
• ltField nameFibroField typefloat /gt
• ltStepgt
• ltif conditionoldcell.type notequal
  Mediumgt
• ltdeclaregtltfloat namefibro
  valueFibroFieldpt.xpt.ypt.z /gtlt/declaregt
• ltif conditionReactionDiffusionpt.x
  pt.ypt.z greaterequal ActivatorThresholdgt
• ltcopy toFibroFieldpt.xpt.yp
  t.z fromfibroFibroRate /gt
• ltcopy tofibro
  fromFibroFieldpt.xpt.ypt.z /gt
• ltreturn valuefibroLambda /gt
• lt/ifgt
• lt/ifgt
• ltreturn value0 /gt

75
C Code Generation For Avian Limb Chemical
Hamiltonian (1)
76
C Code Generation For Avian Limb Chemical
Hamiltonian (2)
77
C Code Generation For Avian Limb Chemical
Hamiltonian (3)
78
C Code Generation For Avian Limb Chemical
Hamiltonian (4a)
79
C Code Generation For Avian Limb Chemical
Hamiltonian (4b)
80
C Code Generation For Avian Limb Chemical
Hamiltonian (5a)
81
C Code Generation For Avian Limb Chemical
Hamiltonian (5b)
82
Miscellaneous Data