Biophysical Journal Volume 73 September 1997 1135-1146 1135

A General Computational Framework for Modeling Cellular Structure

and Function

James Schaff, Charles C. Fink, Boris Slepohenko, John H. Carson, and Leslie M. Loew
Center for Biomedical Imaging Technology, Departments of Physiology and Biochemistry, University of Connecticut Health Center,
Farmington, Connecticut 06030 USA

ABSTRACT The "Virtual Cell" provides a general system for testing cell biological mechanisms and creates a framework for
encapsulating the burgeoning knowledge base comprising the distribution and dynamics of intracellular biochemical pro-
cesses. It approaches the problem by associating biochemical and electrophysiological data describing individual reactions
with experimental microscopic image data describing their subcellular localizations. Individual processes are collected within
a physical and computational infrastructure that accommodates any molecular mechanism expressible as rate equations or
membrane fluxes. An illustration of the method is provided by a dynamic simulation of IP3-mediated Ca2+ release from
endoplasmic reticulum in a neuronal cell. The results can be directly compared to experimental observations and provide
insight into the role of experimentally inaccessible components of the overall mechanism.

INTRODUCTION
Cellular structure and function are determined by an array Most approaches to modeling cellular processes describe
of interdependent biochemical reactions and processes, spa- their system as a single set of ordinary and partial differen-
tially distributed in a nonhomogeneous fashion within the tial equations. This limits the complexity of the model with
cell. There is a formidable amount of data concerning the respect to interacting features of the physiological system
biochemistry of cellular molecules on the one hand, and and forces a homogeneous representation of the cellular
microscopic image data on the structure and behavior of geometry. We have developed a system, the "Virtual Cell,"
cells on the other. Experimental methods such as patch for building complex cell models based on a hierarchical
clamp and site-directed mutagenesis have joined traditional assembly of molecules, reaction mechanisms, and their as-
biochemical and molecular biological approaches to gener- sociated intracellular structures. Once the cell model is
ate an ever-expanding catalog of biomolecular reaction created, it serves as the input to a numerical simulation of
mechanisms and kinetics. Electron microscopy, x-ray crys- intracellular dynamics that is calculated by using the finite-
tallography, and NMR spectroscopy are providing an volume formalism (Patankar, 1980). This permits a spatially
atomic-level understanding of the structure of biological heterogeneous collection of volume elements in which each
macromolecules and biomolecular assemblies. Tools such element is identified with the physiology of a single class of
as fluorescent probes and confocal microscopy have made cellular features. The geometrical organization of the ele-
possible the determination of three-dimensional intracellu- ments can correspond to the organization of their associated
lar spatial distributions of molecular species as well as structural features in an experimental micrograph. There-
subcellular structures in living cells. It should be possible to fore, complex behavior may be simulated with a collection
use basic physical-chemical principles to organize and un- of simple models. Each simple model maps directly onto a
derstand all of these data. This problem is close to being biological feature and encapsulates only the physiology
solved at the molecular level, where it is now possible to associated with that feature. A model will specify how the
calculate the structures and dynamics of large proteins and concentrations of molecules can change within an element
nucleic acids as well as their interactions with solvents, and how an element might exchange molecular species with
ligands, and substrates. At the level of the cell, however, its neighbors. Thus an assembly of different elements can
there is no general approach to modeling the spatially or- interact with complex behavior, even though their individ-
ganized and interdependent chemical processes that deter- ual behaviors may be simple. Indeed, this organization
mine cell function. This would be desirable, not only for the closely approximates the actual dynamics within the cell.
purpose of testing hypotheses, but also as a means of Comparison of simulation results and observed behavior is
collecting and organizing this massive quantity of experi- facilitated because the microscopy data and the simulation
mental data. model are mapped onto the same geometry.

Receivedfor publication 26 February 1997 and infinalform 29 May 1997. UNDERLYING PHYSICAL CONCEPTS
Address reprint requests to Dr. Leslie M. Loew, Department of Physiology,
University of Connecticut Health Center, MC3505, 263 Farmington Ave., A key feature of the Virtual Cell is that the physical laws
Farmington, CT 06030-1269. Tel.: 860-679-3568; Fax: 860-679-1269; controlling intracellular molecular dynamics are integral to
E-mail: les@volt.uchc.edu. the infrastructure and do not have to be reformulated for
© 1997 by the Biophysical Society every physiological process to be studied. Thus a physio-
0006-3495/97/09/1135/12 $2.00 logical model can be constructed from biochemical and
1136 Biophysical Journal Volume 73 September 1997

biophysical knowledge of a system with the assurance that diffusion flux, and the second term of Eq. 1 describes the
the underlying infrastructure will properly apply the appro- flux due to the electric field. The latter will only be impor-
priate physical laws. This simplifies the modeling process tant for very fast physiological processes (e.g., action po-
and makes it more accessible to the experimentalist. tentials) or very high spatial resolutions.
The problem consists of a set of continuous partial dif- The time rate of change of concentration defined over an
ferential equations relating the state variables of the system infmiitesimally small volume is the sum of the net flux into
and the input stimuli. The formulation accounts for diffu- the volume and the net reaction rate,
sion in the presence of concentration and/or electric gradi-
ents, chemical reactions, and membrane transactions due to acj -P

pumps and channels. The set of state variables includes the at -V'Ji + Si(Cl, C2, - - -
CN) (2)
concentrations of each molecular species, Ci, and the elec-
tric potential, V, defined over a subset of three-dimensional where the molecular flux, Ji, is defined by Eq. 1, and Si
space for times greater than zero. State variables are defined defines the net chemical reaction rate for species i (compo-
such that knowledge of their values at time t = 0 is suffi- nent 4). For a chemical species produced in an aqueous
cient to determine the behavior of the whole system for any compartment by a single chemical reaction, Si is given by a
known stimulus. In general, the rules governing the state simple rate equation, as exemplified by the common revers-
variables are a function of position (heterogeneous) and of ible reaction in Eqs. 3 and 4:
the state variables (nonlinear). k,
The physical infrastructure is divided into components A + B±
k-i
C (3)
describing
1. molecular flux, Ji, of diffusible molecules dependent SC = kl[A][B] -k-[C] (4)
on concentration and electrical gradients;
2. electric flux density, D, dependent on the distribution For a molecular species created or consumed in a number of
of charged molecules according to the laws of electrostatics; reactions, Si is simply the sum of the individual rate laws.
3. current density, I, defined by a cable theory formalism The physical infrastructure deals separately with molecular
(Jack et al., 1975b); fluxes through membranes (component 5, above), treating
4. source terms arising from reactions according to dif- them as boundary conditions between two features that can
ferential rate equations (Frost and Pearson, 1961); and also be functions of all concentrations and potentials. Thus
5. membrane transactions in which bulk concentrations of the dependence of Ci on the concentration of all of the other
molecular species and/or electrical potential regulate mem- molecular species is explicitly considered for both passage
brane surface enzyme kinetics or transmembrane flux. through membranes (e.g., ligand-gated channels) and pro-
Components 1-4 are defined throughout the physical duction or consumption of species i in chemical reactions
space, whereas membrane transactions are confined to the (e.g., enzyme-mediated transformations or buffer on/off
interfaces between distinct compartments (e.g., the plasma rates).
membrane separates extracellular and cytosolic compart- D and I (components 2 and 3) are actually redundant
ments). It is important to appreciate that all of the compo- alternatives for describing the electrical potential based on
nents are interdependent. Components 1, 4, and 5 are di- the instantaneous distribution of charge or the mobilities of
rectly involved in determining Ci, but the Ci are also charged molecules, respectively. The former is quite rigor-
dependent on V; on the other hand, V can be determined ous but requires small time steps to ensure convergence.
through 2 or 3, but these are dependent on the distribution The electric current density formulation permits an ap-
and motion of charged molecules. For a given physiological proach similar to the cable theory used in neuronal model-
process within a given time scale, however, some of these ing, which assumes that the physiological processes of
components may remain constant and therefore may not interest are slow compared to charge relaxation in the con-
have to be explicitly considered. ductive aqueous compartments. Details of both of these
Component 1 is the molecular flux, Ji, arising from alternative formulations for the electrical potential are pro-
diffusion and electric forces as described in the Nernst- vided in the Appendixes.
Planck equation (Eq. 1) (Hille, 1992). Note that the sub-
script, i, denotes values pertaining to the ith molecular CHEMICAL DYNAMICS COMPUTATION
species:
A key attribute of the Virtual Cell is the utilization of
empirical three-dimensional data and images to directly
i= -DiVCi - ZikCi,VV, =DiRT (1) build the simulation geometry. The anatomical features of
the cell are identified and extracted as geometrical descrip-
Ci is the concentration, Di is the diffusion constant, z; is the tions. These geometrical descriptions are sampled to create
ionic valence, ,ui is the mobility, V is the electric potential, a simulation volume. This simulation volume represents a
F is Faraday's constant, R is the gas constant, and T is the spatially heterogeneous collection of three-dimensional el-
absolute temperature. The first term of Eq. 1 describes the ements, in which each element pairs a subcellular anatom-
Schaff et al. Virtual Cell 1137

ical feature with its associated physiology. The spatial or- TABLE I Hierarchical organization of cellular components
ganization of cellular structure and function thus comprises Component Description
a collection of feature models. These cellular features can be Cell model A collection of feature models organized in space
identified by three-dimensional segmentation of digital con- to correspond to their organization in
focal or electron microscopy data. For cellular structures experimental or idealized images.
that are either too small or too convoluted to be extracted A list of initial conditions, reaction models,
Feature model
easily from actual data, feature geometry may be approxi- membrane reaction models, and species
mated by synthetic geometry. This synthetic geometry can parameters (ionic charges, diffusion
consist of any combination of geometric primitives (e.g., coefficients) associated with a particular
sphere, cylinder), mathematically defined solid models anatomical feature (e.g., cytosol feature
model).
(e.g., fractal solids), or arbitrary polygonal surfaces (e.g.,
CAD models). Reaction model Mechanisms for individual biochemical
The specification of the chemical dynamics is also very transformations with their associated rate
flexible. The simulation description specifies the molecular constants corresponding to reactions associated
species to be simulated, the number of elements in each with molecules "in solution" (e.g., calcium
buffering).
dimension, the size of the elements, and the simulation time
step. The dynamics are described as time-dependent con- Membrane reaction Equations that represent reactions and fluxes
centrations within elements and fluxes of molecular species model associated with membrane molecules,
across boundaries separating adjacent elements. These can including all relevant conductances and rate
constants (e.g., G-protein cascades; calcium-
be functions of diffusion, ion drift, potential, chemical re- induced calcium release through the ryanodine
action, or any other relationship involving the spatial dis- receptor)
tribution of molecular species. Membranes are represented
by element boundaries separating dissimilar elements. The Molecular species A molecule or a distinct conformation of a
output of the simulation provides the time course for the molecule that is typically present "in solution"
(e.g., cAMP, calcium ion). A state variable
concentrations of each molecular species within each ele- that represents a molecular species is defined
ment, the electric potential in each element, and the surface in each element and holds the concentration of
densities of the states of membrane-associated molecules. that species.
An object-oriented hierarchical scheme, summarized in Ta-
ble 1, is used to organize the components needed to describe Membrane molecular A molecule or a distinct conformation of a
species molecule that is present only within a
the physiological events under investigation. membrane (e.g., open confonnation of a
Membrane reaction models and local reaction models are ryanodine receptor). A state variable that
maintained as reaction descriptors set to define the reaction represents a membrane molecular species is
rates of molecular species as a function of the local envi- defined only for elements that contain
membranes and holds the surface density of
ronment. Each descriptor consists of a set of rate expres- that species.
sions that are functions of model variables and parameters
(e.g., rate constants, channel conductances, etc). These de-
scriptions are stored in a data base, and when needed, a
specific model object library is automatically generated for The fundamental step in the numerical formulation in-
input to a simulation. Thus new models may be specified volves integrating the equations in time over a single vol-
with little exposure to the underlying numerical methods. In ume element, using appropriate interpolation profiles and
addition, this object-oriented approach lends itself well to boundary conditions. The solution of each integration re-
the development of a data base of reaction and feature lates a small neighborhood of sample values over space and
models that can be readily merged with a data base of time. The resulting equations are solved iteratively by the
cellular and subcellular images. line-by-line method (Patankar, 1980). This is an implicit
The numerical solution to the set of partial differential method that guarantees stability, given the physically ap-
equations is achieved by the finite-volume method (Patan- propriate constraints associated with physiological models.
kar, 1980). This method is based on the subdomain formu- The ability of each iteration to converge to the solution of
lation of the method of weighted residuals, although the the nonlinear equations is made possible by an implicit
resulting discrete equations resemble those of the finite- underrelaxation scheme. The simulation geometry is com-
difference method. The finite-volume method allows for posed of unifornly sampled orthogonal elements (cubes).
very good control of boundary conditions and profile as- Piecewise linear interpolation functions are used to interpret
sumptions while preserving the conservative nature of many the values of molecular concentrations and electric poten-
of the equations of interest (e.g., Fick's law and Gauss' tials between element centers.
law). Most importantly, finite-volume formalism accommo- Within such uniformly sampled orthogonal elements, the
dates the heterogeneous spatial organization of cellular time dependence of Ci is simply the sum of fluxes entering
compartments. the volume element from its six adjacent neighbors plus the
1138 Biophysical Journal Volume 73 September 1997

production of molecular species i via reactions. In particu- TABLE 2 Discrete formulation for molecular flux
lar, as depicted in Fig. 1, the discretized form for the flux Boundary condition Discrete approximation for molecular flux
component of Eq. 2 between similar elements is simply
Interface between similar JXIX+ = -D,VCIx+ - zipC1VVb+
elements
Ji dS =
Areab.Ufdary C=(x) - CA(x
Ax
+ ax)

s (5)
V(x) - V(x + AX)
+ ZiI.i CiIx+ ax
* [JXIX+ - JXIX- + JYIY+ - JY1Y- + JZZ+ -JZIZ-];
Interface between dissimilar JX,I+ = flux due to membrane transport
(For boundaries between dissimilar elements, membrane elements (membrane) = AC,, c2 ... CN, V)
transport mechanisms are incorporated, as will be detailed Simulation boundary Flux need not be evaluated
below.) The x, y, and z components of the molecular flux Jx, (specified concentration)
JY, and Jz are evaluated at the six element boundaries by Simulation boundary Jxlx+ = Jboundary
assuming them to be constant across each boundary surface. (specified flux)
The molecular fluxes depend on the element boundary
conditions as summarized in Table 2. The simulation
boundary conditions of either fixed concentration, or of
fixed molecular flux, are easily applied. Similar discrete As in the case of the solution reaction models, all-mem-
formulations for electric flux density and electric current brane reactions are treated as coupled differential equations
density are detailed in Appendix II. describing their kinetics. However, the surface densities of
In general, the reaction rate for either elementary or the membrane molecular species can dictate the fluxes of
multistep reactions is a nonlinear relation that may be lin- solution molecular species through a "membrane," opera-
earized about a set of current conditions for sufficiently tionally defined within the Virtual Cell as a boundary be-
small time steps: tween dissimilar elements. Membrane reaction models re-
late the transactions controlled by the membrane to the
SA - fAlfinear([A], [B], ) [A] (6) accessible states of receptors, pumps, channels, or carriers.
+ gA-cons tan t(C],
[D], * * ) , This is illustrated for the open, closed, and inactive states of
a prototypical three-state channel selective for solute S, in
A particular simulation may not define one or more of the Fig. 2. The state transitions are defined in terms of reaction
reagents as state variables. The reaction model must then
rate equations that can be functions of membrane potential
either be disabled or simplified, depending on which re-
and species concentrations on either side of the membrane.
agents are missing. A simplified model uses a constant
concentration for each missing reagent as defined for that The equations are written such that the stoichiometry of the
biological feature type. This constant concentration can be whole model is consistent and flux is conserved. These
specified in the reaction model description, or the default
value for that feature type will be used.
o 0- ' .o
0 0
-AA1,,
o0
. 000
0
°O@O
0
~0
3
* 0 0
0
..

-.j,i A...,. i,:

o s1
____ )c 0,0 0 S F.
00.0 00 0
_.------ _- S, Flu.x
.1 0 000

t
: k
k - F V |S. ],I [S.,..
[A ]
k F VfSjj,,

[A / < [ '\
k = F IV [Sj] [S ,t, I..!

FIGURE 1 Finite volume formulation. Boundary conditions for individ- FIGURE 2 Transitions of membrane-bound molecules. This illustration
ual elements in the chemical simulation framework guarantee mass con- of membrane reactions is based on a three-state model for a gated channel
servation. The flux is calculated by the Nernst-Planck equation when the that controls the flux of solute S,. The rate constants for opening, kc0,
neighboring elements are of the same type. The flux is based on membrane inactivation, ki, and activation, kic, can each be functions of membrane
reactions (channels, pumps ) when the neighboring elements are differ-
... potential, V (voltage-gated channel), or the concentrations of ligands, Si
ent. The y and z components of the flux are omitted for clarity. (chemically gated channel).
Schaff et al. Virtual Cell 1139

states are defined in terms of membrane surface densities essary for finding the proper initial conditions before stim-
and can be represented by state variables or by constants. ulation, the ability to test how a subset of physiological
There is only one value of the surface density of a mem- parameters relaxes to a steady state is also of great value in
brane variable that is defined for each element, although an evaluating hypotheses on the interrelations of complex cel-
element may have up to six separate membrane surfaces. lular processes.
Reaction rates for each membrane reaction model are aver-
aged over the membrane surfaces of an element. Missing
reagents are handled in a manner similar to that of the bulk EXAMPLE MODEL: CALCIUM WAVE IN A
reactions defined above. An important additional require- NEURONAL CELL
ment of the transmembrane reactions is that flux be
conserved. The origin, regulation, and role of calcium as a second
A primary application of the cell models is to investigate messenger in intracellular signal transduction is a subject of
the course of events after a homeostatic system is perturbed intense investigation (Berridge, 1993). The complexity of
by some stimulus. A cell model is constructed with a set of the interacting and varied mechanisms that have been dis-
initial conditions corresponding to a best guess for the covered at the biochemical and electrophysiological levels
prestimulus values of the state variables, i.e., all concentra- makes models for calcium dynamics a good initial test for
tions, surface densities, and potentials. Some of these rep- the usefulness of the Virtual Cell approach. We have con-
resent experimentally measured and estimated values, structed a limited cell model based on inositol-1,4,5-
whereas others may not be known a priori. However, this trisphosphate (1P3)-mediated calcium waves in differenti-
guess may not represent a real steady state once all of the ated NlE-1 15 neuroblastoma cells. In these cells, the
components of the model are permitted to interact. When an peptide hormone bradykinin triggers calcium waves that are
integrated cell model is built from an arbitrary collection of known to involve the IP3 receptor on the endoplasmic
individual reaction models, care must be taken to achieve a reticulum (ER) membrane, without involvement of the ry-
self-consistent initial condition. Starting at a steady state anodine receptor or calcium influx through plasma mem-
allows consideration of the system's response to the stim- brane channels (Monck et al., 1990; Wang and Thompson,
ulus without interference from the relaxation of initial 1995). Thus only calcium release through the IP3 receptor
conditions. channel need be considered in our model.
In general, there is not a unique steady state defined by The state variables and initial conditions associated with
the system and boundary conditions alone, but by fixing the the features of the cell model are given in Table 3. An
values of the set of known variables to their prestimulus effective diffusion constant for calcium of 40 ,tm2/s is
homeostatic values, it might be possible to allow the dy- chosen to implicitly account for calcium buffering in the
namic simulation to run until the desired steady state is cytosol (Jafri and Keizer, 1995). To initiate a calcium re-
reached. However, this could be wasteful of computer time. sponse, the bradykinin is stepped to the saturating level of
Moreover, the proposed cell model will not necessarily have 500 nM at t = 0 (although bradykinin is assigned a diffusion
a steady-state solution that is consistent with that fixed set coefficient, its concentration is uniform, so its flux is 0
of initial conditions. In this case either the initial conditions throughout the extracellular space). A two-state bradykinin
or the cell model must be modified. receptor model in the plasma membrane then responds with
A better and more direct approach involves the minimi- an initial burst of 5000 molecules IP3/4Mm2* s that exponen-
zation of the initial dynamics (i.e., finding a steady-state tially relaxes to a steady-state production of 100 molecules
equilibrium). The general method used to address this issue of IP3/pum2* s with a time constant of 50 ms. The ER
is nonlinear constrained optimization. The extent to which feature model contains a five-state membrane reaction
the system can be altered may be formalized by adding model for the type 1 1P3 receptor, for which the most
constraints or ranges for the undetermined or underdeter- detailed kinetic parameters are available in the literature
mined initial values and system parameters. The measure of (Bezprozvanny et al., 1991; Bezprozvanny, 1994; Bezproz-
initial dynamics is a function of the magnitude of net vanny and Ehrlich, 1994). This membrane reaction model is
reaction rates and fluxes for all species, where the steady fully detailed in Fig. 3. The sarcoplasmic/endoplasmic re-
state represents a system in which the time rate of change of ticulum calcium ATPase pump (SERCA pump) is a second
each species is zero. The minimum solution arrives at pa- membrane reaction model that is part of the ER feature
rameter values and initial conditions simultaneously. model. This is set to pump calcium back into the ER with a
The problem becomes manageable if the system is first calcium dependence given by Eq. 7 (Lytton et al., 1992;
simplified by assuming that each species is homogeneous Sneyd et al., 1995):
within the bounds of a feature. This reduces the problem to
the constrained minimization of a system of nonlinear al-
J = V[SER ] [Ca2 ]C2yt
gebraic equations. An initial implementation uses a conju- maxSERA]K2 + [Ca2+]2 (7)
gate gradient method for the minimization of the system
dynamics and a penalty function associated with the viola- where Vm. is taken as 2.52 X 10-2 (AM ,um3)/pump * s),
_

tion of constraints (Bryson and Ho, 1975). Although nec- [SERCA] is given in units of pumps/pIm2, and K, the
1140 Biophysical Journal Volume 73 September 1997

TABLE 3 Components of the neuroblastoma cell calcium wave model

Feature Species Initial condition Diffusion coefficient
ER [Ca2+] 2500 ,LM 40 gm2/s
R, IP3 Receptor (unbound) 9.43026 chan/4Lm2
RI, IP3 Receptor (with IP3) 0.14145 chan/Amm2
RIC, IP3 Receptor (with IP3 & 1 Ca) 0.02768 chan/lpm2
RICC, IP3 Receptor (with IP3 & 2 Ca) 0.39890 chan/.m2
RIC_Open, IP3 Receptor (with IP3 & 1 Ca, Open) 0.00171 chan/4Lm2
SERCA pump 5 X 103 pump/41m2
Cytosol [Ca2+] 0.05 AM 40 ,um2/s
[1P3] 0.01 uM 250 tm2/s
Bradykinin receptor (unbound) 200 receptor/4Lm2
Bradykinin receptor (bound) 0 receptor/4Lm2
Extracellular [Bradykinin] (stimulus) 0 ,uM, t = 0; 0.5 ,uM, t > 0 o .Lm2/s
100

half-saturation concentration for calcium binding to growth cone; the ER was excluded from the region defined
SERCA, is taken as 270 nM. by the cell nucleus, but a nucleus feature model was not
The model was completed by using an experimental separately defined. Although the elements were defined as
confocal micrograph of a differentiated neuroblastoma cell three-dimensional cubes 200 nm on a side, only one plane of
as the basis for the geometric distribution of cytosolic elements was employed for the computation, with dimen-
elements and ER elements. The ER in the cell was visual- sions of 54 ,um X 20 ,u m X 0.2 ,um; thus the cell model was
ized by microinjecting a soybean oil solution of the fluo- effectively two-dimensional.
rescent dye di-18:2-1-(3-sulfanatopropyl)-4-[,B-[2-dilinoleyl- A 2-s simulation was generated with a time step of 100
amino)-6-naphthyl]vinyl]pyridinium betaine) (ANEPPS) ,us. The first 1.6 s of the simulation was sampled at 200-ms
(synthesized by Joseph Wuskell according to methods de- intervals (Fig. 5). The ER elements correspond to the white
veloped in our laboratory (Hassner et al., 1984; Tsau et al., pixels in the t = 0 calcium image, as the pseudocolor
1996) and available from Molecular Probes, Eugene, OR) scheme maps the luminal ER [Ca2+] of 2.5 mM to white. A
according to the procedure developed by Terasaki et al. most useful feature of the Virtual Cell is the ability to
(Terasaki and Jaffe, 1991). A confocal image of such a cell visualize the dynamics of intracellular components, such as
is displayed in Fig. 4, along with the derived distribution of [IP3] and IP3R in Fig. 5, that are not accessible experimen-
elements associated with intracellular features. The distri- tally. The 1P3 concentration is seen to build up rapidly in the
bution of ER elements was set in proportion to the density neurite, resulting in a nucleation point for a calcium wave at
of ER visualized in the confocal image, with a high density t = 200-400 ms. Ca2+ fills the entire neurite within 100 ms
in the cell soma and a lower density in the neurite and and rapidly declines, except for a wavefront that gradually
migrates into the soma. The pattern and speed of these
calcium dynamics are similar to those found experimentally
IP3 Receptor Rate Constants: (Wang and Thompson, 1995; Fink et al., unpublished re-
(Membrane Reaction k, 12.0 (.LM sec)'
Model) k., 8.0 (sec)-' sults). A key conclusion from this simulation is that the
[RIC Open] k.2 4.6 (sec)' calcium wave can originate in the neurite, even when the
density of plasma membrane hormone receptors is uniform
k4 t \k5 k3 4.0 (9M sec)-'
k-3 0.8 (sec)-' and when the density of ER is actually higher in the soma.
k4 25.2 (sec)-' This is the pattern consistently observed experimentally
k-4 224.0 (sec)-'
k5 183.0 (sec)-' (Fink et al., unpublished results; cf. http://www2.uchc.edu/
k3[Ca]
htbit). As revealed in Fig. 5, it can be explained by the more
k, [1P3] k2[Ca]
[R] " [RI] > [RIC] )> [RICC] rapid buildup of [IP3] in the neurite, where the surface-to-
volume ratio, and therefore the rate of production of IP3, is
-
-(
k-, k-2k-
higher than in the soma. This factor overrides the higher
density of 1P3 receptor in the soma dictated by the higher
Calcium flux through RIC_Open (based on 0.47pA/channel and [Ca]ER = density of ER. It is also clear that the calcium signal decays
25009M): rapidly, despite the continuous buildup of 1P3 in our rudi-
Dca= 0.974.[RIC_Open].Ax (P.m2/s) Ca2' Diffusion Coefficient mentary model. This can be understood by the combined
(Ax = width of a simulation element = 200nm) action of SERCA pumps and the negative feedback on
channel open probability associated with the RICC state in
the five-state model of Fig. 3. Indeed, the drop in [Ca2+]
FIGURE 3 IP3 receptor membrane model based on the model of Bez-
allows the IP3R to partially recover and permit a secondary,
prozvanny et al. (Bezprozvanny et al., 1991; Bezprozvanny and Ehrlich, highly damped calcium rise in the neurite to appear at t =
1994). 800 ms. Although a more realistic reaction model for 1P3
Schaff et al. Virtual Cell 1141

Original Confocal Image

a
FIGURE 4 Derivation of cell model for
neuronal calcium signaling from a confocal
micrograph of an ER labeled cell. (a) A
differentiated NIE-115 neuroblastoma cell
was microinjected with di-18:2-ANEPPS
dissolved in soybean oil, and the fluorescent
dye was allowed to diffuse from the oil drop-
lets into the-ER for 1 h. The displayed image
was chosen to represent an optical slice fully
within both the neurite and soma. The ER
has its highest density in the soma, with a
more diffuse distribution in the neurite and
growth cone. The kidney-shaped dark region Corresponding Simulation Geometry
in the soma is the nucleus, and the two small
dark circles are the injected oil droplets (now
free of dye). (b) The image in a was seg-
b
mented and classified by simple thresholding
techniques to define the indicated cellular
and extracellular regions. The segmented im-
age was then sampled at a lower pixel den-
sity to limit the mesh resolution for the sim-
ulation geometry. To eliminate the obvious
experimental artifact, a few ER elements
were added manually to the region obscured
by the oil droplets. The small neurites pro-
jecting from the top and left of the soma
were omitted to reduce the computational
size of the simulation. Extracellular
Nucleus
Endoplasmic Reticulum
Cytosol
Perinuclear ER

production and decay would have to be incorporated to try approaches as exemplified by detailed simulations of intra-
to fully simulate experimental behavior and verify such cellular calcium dynamics (Atri et al., 1993; Jafri, 1995;
subtle effects as damped oscillations, it is revealing that Jafri and Keizer, 1995; Sneyd et al., 1995; Dupont and
only the cell geometry, the properties of the IP3 receptor, Swillens, 1996; Keizer and Levine, 1996; Smith, 1996;
and the presence of the SERCA pump can suffice to pro- Smith et al., 1996; Wu et al., 1996). However, such formu-
duce calcium waves with the essential spatial and temporal lations preclude the expression of an observed physiological
features observed experimentally. Furthermore, the example phenomenon as it maps to the actual geometry of an intact
model is uses too coarse a mesh to resolve microdomains cell. Thus features of the physiology that depend on the
around channels; however, higher spatial resolution, perhaps spatial organization of molecules within and between cel-
while using only a small fragment of a cell, could help eluci- lular organelles cannot be accommodated by current methods.
date the dynamics of calcium close to the ER membrane. There have also been numerous simulations in neuro-
physiology involving formulations based on cable theory
DISCUSSION (Hodgkin and Huxley, 1952; Jack et al., 1975a; Hines,
1989; Smith, 1992). These decompose a single neuronal cell
Most current approaches to the modeling of cellular phys- into branching cylindrical computational elements repre-
iology are based on highly restricted sets of biochemical senting axons and dendrites. Although these simulations
reactions and idealized geometries, and include many other reproduce neuronal transmission with good fidelity, the
simplifying assumptions. Cells are represented as simple inherent geometric simplicity of the formulation restricts
geometrical shapes consisting of spatially homogeneous their application to the class of problems in which the cell
behavior. Significant insights can be attained with such interior is considered a homogeneous conductive medium.
1142 Biophysical Journal Volume 73 September 1997

FIGURE 5 Dynamic simulation of intracellular signaling in a neuronal cell. The cell model specified by Figs. 3 and 4 and Table 3 was used as the input
for a dynamic simulation with time steps of 100 ,us. The results are shown at 200-ms intervals for three of the state variables: [Ca21] (left), [JP3] (center),
and the density of open IP3 receptor channels (right). The latter are only associated with the ER and are displayed with a white outline of the cell to aid
visualization. The [Ca2+] and [IP3] pseudocolor scales are logarithmic, and the RIC_Open scale is linear.
Schaff et al. Virtual Cell 1143

It is difficult to extend this approach to the simulation of defined membrane feature, the mesh size would have to be
intracellular ion dynamics, for example. Furthermore, spe- set at a prohibitively fine resolution. A novel aspect of our
cialized structures such as boutons or spines are necessarily formulation for the Virtual Cell is the treatment of mem-
beyond the scope of methods based on linear cable theory. branes as discontinuities at boundaries between elements
We have developed a comprehensive, extensible, and associated with different features. This permits us to deal
flexible framework ("Virtual Cell") for organizing, model- explicitly with the properties of membranes without having
ing, simulating, and visualizing cell structure and physiol- to define them as separate features.
ogy. The spatial organization of cellular components is The current implementation of the Virtual Cell uses struc-
captured by incorporating geometries from volume data tured input files, each representing separate reaction models
sets. These data may be derived from experimental micro- and feature models. The input includes initial conditions for
graphs or may be synthesized. Each geometrical feature is each of the molecular species, rate constants for each of the
identified as a specific cellular compartment and associated reaction steps, and, for the case of membrane reaction
with a mathematical model that represents its function in a models, permeabilities or conductances in addition to the
cell. A chemical dynamics simulation is created by sam- relevant reaction rates. The model objects are generated and
pling the geometry into a three-dimensional mesh of simu- compiled automatically from these input files. A cell model
lation elements. The approach uses a finite-volume formal- is then created based on the spatial organization of the
ism that assigns appropriate properties (starting features derived from appropriately segmented experimen-
concentrations of chemical species, diffusion constants, tal images. During the course of a computation, values of
ionic charges, reaction rates, etc.) to elements identified the state variables are periodically stored to files that
with the different compartments involved in the process. uniquely represent the simulated physiology at a discrete
Boundaries between compartments are given the properties point in time. Any state variable can then be visualized over
of the corresponding membranes, including channels, space and time by reading these files, with their associated
pumps, receptors, etc. The partial differential equations of geometrical models, into a 3D viewer. Fig. 5 is an example
electrodiffusion, plus source terms corresponding to reac- of the display of the time dependence of three state variables
tions, are solved for each chemical species and electric as a collage, but the 3D viewer can also be used to generate
potential in each volume element. The formulation is suf- movies. In addition, the state variables can be used in
ficiently general that any mechanism for a dynamic physi- postprocessing of the data to determine derived states. Such
ological process can be modeled over the domain repre- derived states can be reaction rates, fluxes, or other func-
sented by a three-dimensional image. Simulation elements, tions of state variables that are explicitly or implicitly rep-
concentrations, electric potential, receptor and channel resented in the constituent models.
states, and original 3D images can be rendered into the same It is important to realistically assess the magnitude of the
3D scene, mapped to the same coordinate system for direct computational power required for the Virtual Cell. The
comparison, and are viewed as a time-dependent series of computation in our example simulation was of a relatively
images displaying the intracellular chemical dynamics un- modest 2D cell model containing a limited number of reac-
derlying the physiological response. The same physiologi- tion models and molecular species. The computations be-
cal models can be verified in multiple experiments (differ- hind Fig. 5 comprised 27,000 elements, three nonlinear
ent geometry and different stimuli). As related models are partial differential equations, and seven nonlinear ordinary
examined under various conditions and with different cell differential equations. This cell model required only the
types, general themes will emerge and can be evaluated in solution of the diffusion and source equations; it did not
a systematic way. Thus the Virtual Cell creates a framework include the electric potential formulation. The 2-s simula-
for testing mechanistic hypotheses and encapsulating tion required 2 days of computation on a Silicon Graphics
knowledge about the interactions and dynamics of intracel- workstation using a R8000 MIPS cpu rated at 300 Mflops.
lular components. Although this system is only 1 year old, faster desktop
The finite-volume formalism (Patankar, 1980) is a special workstations are already available that will improve the
case of the method of finite difference (Smith, 1985; Mas- speed by at least a factor of 3, and supercomputers can be
cagni, 1989) that has been successful in simulations of large expected to speed up the calculation still further. Further-
complex systems such as heat transfer, fluid mechanics, more, the numerical methods used in the first implementa-
astrophysics, geology, or global climate patterns (Sellers et tion are conservative, and have not been adapted to run on
al., 1997). These problems are similar to the Virtual Cell in multiprocessor systems. These can be replaced with more
that they require the solution of a series of ordinary and efficient methods that can be made parallel in anticipation
partial differential equations with structures equivalent to of improvements in processing speeds of several orders
those describing electrodiffusion with source terms. The of magnitude. Thus much more complex cell models may
Virtual Cell may be unique, however, in the sheer number be realistically accommodated in the next generation of
and variety of such equations that are potentially required to implementation.
describe a cell physiological system. Furthermore, the treat- The modular object-oriented structure of the framework
ment of membranes within the finite-volume formalism was will permit the next implementation to include a graphical
a particular challenge. To capture the relative scale of a fully user interface and will support the compilation of a data
1144 Biophysical Joumal Volume 73 September 1997

TABLE 4 Discrete formulation for electric flux density

Boundary condition Discrete approximation for electric flux density, D

.nterface between similar elements = ~VV~,~+S - V(x) - V(x + Ax)

Interface between dissimilar elements V(x)-V(x + Ax)

(capacitive membrane) = VjX+ membrane Ax 6membran

Simulation boundary (specified voltage) D need not be evaluated

Simulation boundary (specified field) DX1X+ = ox+ Ex = e Ex

base of reaction models and feature models. Together with of charged molecules and structures. The electric potential can be defined
a library of images, combinations of feature models can according to the laws of electrostatics, using Gauss's law (Al), which
states that the net charge, Q, within a closed surface is equal to the net
generate cell models for different cell types and physiolog- electric flux density, D, leaving that surface (Hayt, 1981):
ical problems. A data base of specific and customizable
reaction and membrane mechanisms can be made available
for building cellular models through a central repository. Q= D.dS (Al)
Such a repository of models could be maintained at a World
Wide Web site. A naming standard for both models and
molecular species must be created to allow models to be D is related to the permittivity (or dielectric constant) (e), the electric field
interoperable. Run-time binding of a model object library to (E), and the electric potential (V) according to
the simulation allows a single version of the simulation
framework to service any simulation with recompilation of D = sE= -eVV (A2)
only the newly specified models. The total charge within a volume can be defined as the concentration of
We will also need to focus on several issues to increase molecular species, C1, the ionic valence, zi, and Faraday's constant, F,
the range of problems to which the Virtual Cell may be integrated over that volume:
applied. The current treatment of membrane channel and
pump surface densities assumes a single value for each
element, representing the average over all of that element's (A3)
membrane boundaries; this will have to be modified to Q [ ziFCidv
account for heterogeneously distributed two-dimensional
membrane patches. The approach may be easily extended to Combining these three equations gives a form of Gauss's Law (Eq. A4)
phenomena involving intercellular interactions with very that can be evaluated for an incremental volume and relates electric
potential to ionic concentrations:
little change in the basic formulation. A greater challenge is
posed by processes involving changes in cell shape or
dynamic changes in the location or number of intracellular
organelles. Such processes will not necessarily be fully ( ziFCi] dv =
(-sV V) dS (A4)
covered by the physical laws of electrodiffusion. Additional ivLi Js

physical formulations will be required and coupled with

algorithms that incorporate deformable volume elements
and/or methods by which elements may be variably associ- An approach based on electric currents
ated with feature models. Thus the Virtual Cell should be The electric potential can also be defined as a function of electric currents
able to accommodate cell models describing any cellular through a conductive medium and across capacitive boundaries (Hodgkin
process. But the availability of such models will ultimately and Huxley, 1952; Jack et al., 1975b). Conservation of charge states that in
depend on the continued generation of experimental data a conductor, the net current density, I, leaving any closed surface, dS, is
describing the elementary components of cellular biochem- equal to 0:
istry, biophysics, and physiology.
dQ |j_
dt
:-
I*dS = O (A5)
APPENDIX I s

Electric potential equations This formulation assumes that the interior of all connected regions within
An approach based on electrostatics (Gauss's Law) a component (cytosol, organelle lumen, extracellular milieu, etc.) can be
modeled as conductors or, equivalently, that the relaxation of a free charge
For charged molecules, Eqs. 1 and 2 show that the time dependence of C; from the interior to the surface (membrane) is much faster than the cellular
is also dependent on V. However, V is itself dependent on the distribution dynamics of interest. I can be related to the electric potential V by
Schaff et al. Virtual Cell 1145

TABLE 5 Discrete formulation for electric current density

Boundary condition Discrete approximation for electric current density, J
Interface between similar elements IXIX+ = -AVx+ - F E ziDiVC

A(x) + A(x + Ax) V(x)-V(x + Ax)

2 Ax + F:zj( )

Interface between dissimilar elements ,IXI+ = C. at +FEzj F,O

(capacitance membrane)
= Cm( Vn -V+AX)( VO -V 1d-!+Axt + F>zif ( )

Simulation boundary (specified voltage) Ixlx+ does not need to be evaluated

Simulation boundary (specified field)
_ A(x) + A(x + Ax)
Ixjx+ = +AElx+ - F E ziDiVCi = + 2

CE(x) - CA(x + Ax)

*Ebondl:-F>: zjDj Ax

expansion of Eq. 1:
D * dS = Area
I = F zi1Ji = -AVV- F E ziDiVCi
(A6)
* h[Dxlx+ -Dxx- Dyly+-Dyly- Dzlz+
+ + - Dz|zIi (A)
where A = F E Zi IkiCi Area and Volume refer to the size of a single element. The x, y, and z
components of the electric flux density D, Dy, and Dz are evaluated at the
six element boundaries by assuming them to be constant across each
The first term represents Ohm's law, and the second is the current density boundary surface. The electric potential of neighboring elements are then
due to the concentration gradients. For a conductor (i.e., extracellular related by one of the forms of Table 4. The simulation boundary conditions
milieu, cytosol, or organelle lumen), the component of current density of either fixed potential or fixed electric field (or electric flux density) are
normal to a membrane, In, is related to the membrane-specific capacitance, easily applied.
Cm, the potential drop across the membrane, AVm, and the net charge flux
density passing through the membrane, Im (via pumps, channels, etc.)
according to Electric current density
This formulation requires a representation for current density within the
a(Avm) interior of a conductor as well as across membranes. Referring to Eq. A5,
In =-Cm at + Im (A7) the sum of the electric current densities entering an element is due to each
molecular species (Eq. A10). Assuming a linear profile of conductance
between adjacent elements, the average conductance A over an element is
Equations A5 and A6, and the membrane boundary condition (Eq. A7) give
computed as a linear combination of the sample concentrations of the
a form of conservation of current that can be applied to both membranes
current element and its nearest neighbors.
and interior spaces without the geometrical constraints of cable theory.

I * dS = Area
APPENDIX 11 (A10)
Discretization of electrical potential formulations * [IX|X+ - + -Iy|y_ + Izjz+ - IZZ] = 0
Electric flux density Area refers to the side of a single cubic element. The x, y, and z components
of the electric flux density Ix,, IY, and Iz are evaluated at the six element
Within an element, the left side of Eq. A4 becomes the sum of the average boundaries by assuming them to be constant across each boundary surface.
charges due to each molecular species. Assuming a linear profile of The electric potential of neighboring elements are then related by the
concentration between adjacent elements, the average concentration C, appropriate entry in Table 5. The simulation boundary conditions of either
over an element is computed as a linear combination of the sample fixed potential, or of fixed electric field (or electric flux density) are easily
concentration of the current element and its nearest neighbors: applied.

The authors are pleased to acknowledge valuable comments on the manu-

[ziFCidv = Volume E ziFC' (A8) script from James Watras, Mark Terasaki, Ann Cowan, Barbara Ehrlich,
and Laurinda Jaffe.
v LiC]d
1146 Biophysical Journal Volume 73 September 1997

Supported by grants from the U.S. Public Health Service (GM35063) and Keizer, J., and L. Levine. 1996. Ryanodine receptor adaptation and Ca21-
the State of Connecticut (Critical Technologies Program). induced Ca2' release-dependent Ca21 oscillations. Biophys. J. 71:
3477-3487.
Lytton, J., M. Westlin, S. E. Burk, G. E. Shull, and D. H. MacLennan.
REFERENCES 1992. Functional comparisons between isoforms of the sarcoplasmic or
endoplasmic reticulum family of calcium pumps. J. Biol. Chem. 267:
Atri, A., J. Amundson, D. Clapham, and J. Sneyd. 1993. A single-pool 14483-14489.
model for calcium oscillations and waves in the Xenopus laevis oocyte. Mascagni, M. V. 1989. Numerical methods for neuronal modeling. In
Biophys. J. 65:1727-1739. Methods in Neuronal Modeling: From Synapses to Networks. C. Koch
Berridge, M. J. 1993. Inositol trisphosphate and calcium signalling. Nature. and I. Segev, editors. MIT Press, Cambridge, MA. 439-484.
361:315-325. Monck, J. R., R. E. Williamson, I. Rogulja, S. J. Fluharty, and J. R.
Bezprozvanny, I. 1994. Theoretical analysis of calcium wave propagation Williamson. 1990. Angiotensin II effects on the cytosolic free Ca2+
based on inositol(1,4,5)trisphosphate (InsP3) receptor functional proper- concentration in N1E-1 15 neuroblastoma cells: kinetic properties of the
ties. Cell Calcium. 16:151-166. Ca21 transient measured in single fura-2 loaded cells. J. Neurochem.
Bezprozvanny, I., and B. E. Ehrlich. 1994. Inositol(1,4,5,)-trisphosphate 54:278-287.
(InsP3)-gated Ca channels from cerebellum: conduction properties for Patankar, S. V. 1980. Numerical Heat Transfer and Fluid Flow. Taylor and
divalent cations and regulation by intraluminal calcium. J. Gen. Physiol. Francis, Washington, DC.
104:821-856.
Sellers, P. J., R. E. Dickinson, D. A. Randall, A. K. Betts, F. G. Hall, J. A.
Bezprozvanny, I., J. Watras, and B. E. Ehrlich. 1991. Bell-shaped calcium- Berry, G. J. Collatz, A. S. Denning, H. A. Mooney, C. A. Nobre, N. Sato,
response curves of Ins(1,4,5)P3- and calcium-gated channels from en- C. B. Field, and A. Henderson-Sellers. 1997. Modeling the exchange of
doplasmic reticulum of cerebellum. Nature. 351:751-754. energy, water, and carbon between continents and the atmosphere.
Bryson, J. A. H., and Y.-C. Ho. 1975. Applied Optimal Control. Hemi- Science. 275:502-509.
sphere Publishing, New York. 39-41. Smith, G. D. 1985. Numerical Solution of Partial Differential Equations:
Dupont, G., and S. Swillens. 1996. Quantal release, incremental detection, Finite Difference Methods. Clarendon Press, Oxford.
and long-period Ca21 oscillations in a model based on regulatory Ca2+-
binding sites along the permeation pathway. Biophys. J. 71:1714-1722. Smith, G. D. 1996. Analytical steady-state solution to the rapid buffering
approximation near an open Ca21 channel. Biophys. J. 71:3064-3072.
Frost, A. A., and R. G. Pearson. 1961. Kinetics and Mechanism. John
Wiley and Sons, New York. Smith, G. D., J. Wagner, and J. Keizer. 1996. Validity of the rapid
Hassner, A., D. Birnbaum, and L. M. Loew. 1984. Charge shift probes of buffering approximation near a point source of calcium ions. Biophys. J.
memrane potential. Synthesis J. Org. Chem. 49:2546-2551. 70:2527-2539.
Hayt, W. H., Jr. 1981. Engineering Electromagnetics. McGraw-Hill, New Smith, R. G. 1992. NeuronC: a computational language for investigating
York. 57-176. functional architecture of neural circuits. J. Neurosci. Methods. 43:
83-108.
Hille, B. 1992. Ionic Channels of Excitable Membranes. Sinauer Associ-
ates, Sunderland, MA. Sneyd, J., J. Keizer, and M. J. Sanderson. 1995. Mechanisms of calcium
Hines, M. 1989. A program for simulation of nerve equations with branch- oscillations and waves: a quantitative analysis. FASEB J. 9:1463-1472.
ing connections. Int. J. Biomed. Comput. 24:55-68. Terasaki, M., and L. A. Jaffe. 1991. Organization of the sea urchin egg
Hodgkin, A. L., and A. F. Huxley. 1952. A quantitative description of endoplasmic reticulum and its reorganization at fertilization. J. Cell Biol.
membrane current and its application to conduction and excitation in 114:929-940.
nerve. J. Physiol. (Lond.). 117:500-544. Tsau, Y., P. Wenner, M. J. O'Donovan, L. B. Cohen, L. M. Loew, and J. P.
Jack, J. J. B., D. Noble, and R. W. Tsien. 1975. Electric Current Flow in Wuskell. 1996. Dye screening and signal-to-noise ratio for retrogradely
Excitable Cells. Clarendon Press, Oxford. transported voltage-sensitive dyes. J. Neurosci. Methods. 170:121-129.
Jafri, M. S. 1995. A theoretical study of cytosolic calcium waves in Wang, S. S.-H., and S. H. Thompson. 1995. Local positive feedback by
Xenopus oocytes. J. Theor. Biol. 172:209-216. calcium in the propagation of intracellular calcium waves. Biophys. J.
Jafri, M. S., and J. Keizer. 1995. On the roles of Ca2' diffusion, Ca2+ 69:1683-1697.
buffers, and the endoplasmic reticulum in IP3-induced Ca21 waves. Wu, Y.-C., T. Tucker, and R. Fettiplace. 1996. A theoretical study of
Biophys. J. 69:2139-2153. microdomains in turtle hair cells. Biophys. J. 71:2256-2275.