A.3.3. Aim
1(iii): Mathematical Modeling of the Nonlinear Dynamics of Cellular Networks
Computational
tools for modeling biological processes, at the cellular
scale using a series of differential equations: XPPAUT
Numerous
approaches have been used in the mathematical modeling of cellular and
biochemical processes starting from the work of Jacob & Monod (100;101). On one end is to
include as many details as possible. Some of the packages now available for
these types of simulations are listed in the first three rows of Table A.2.
This massive modeling approach has the advantage that every known reaction is
included in the system thus giving some hope for comparisons to experiments.
However, in many cases, the rates of reactions and the original concentrations
are not known, nor are all the intermediate states and connectivities (state/block
diagram); and the complexity of the system makes it difficult to study
sensitivity to parameters and initial conditions. On the other extreme is the
abstract approach taken by Glass and Kaufmann (102); the individual components
are taken to be Boolean variables (either on or off) and the behavior is completely
determined by the topology of the interactions and the switching rules. While
this greatly simplifies the models, the graded nature of responses is often
important.
We aim for an intermediate regime of modeling between these two extreme
cases, which will take into account the known biology but then simplifies it
through a series of mathematical steps. A description of the methods for
simplifying and reducing complex models is presented by Ermentrout (103). These methods
essentially exploit differences in time and spatial scales, when these scales
are separable. Many of the systems
we intend to model are inherently stochastic and we will utilize a master equation formalism (104) where transition (or jump) probabilities between states
control the probabilistic evolution of states (see for example the
analysis of voltage-gated or ligand-gated ion channels (105)). Closely associated
are the hybrid models that involve stochastically
forced differential equations (Langevin dynamics), or the time evolution of probability density functions
(Fokker-Planck formalism). These approaches have been successfully used in
other disciplines (see for example the review by Bahar and coworkers (106)), but have not yet been
exploited by theoretical biologists (107). Deterministic differential equations for mean concentrations
can be derived directly from the stochastic formulations. Associated approaches for obtaining
differential equation descriptions include the mean-field approximation, which is appropriate for interactions
that are sufficiently fast and pre-equilibrated within the time scale of an
examined slower process, and/or spatial
averaging for movements or paths that are again fast. The differential
equation description can be further reduced to produce simpler models that
still capture the essential properties of the system. An example is the pseudo-steady state approximation for
the reactants and products of the fast steps in serial reactions (e.g.
Michealis-Menten mechanism). The solution of averaged differential equations in
terms of the slow variables, as used in neural networks or weakly coupled
oscillators, is another mathematical tool for model reduction.
Therefore,
we will consider a hierarchy of models from discrete representation all the way
up to continuous levels for proper reduction of our models. In addition to these reductions, equally
important is to develop efficient methods for output analysis and
interpretation. Mathematical tools such as PCA for decomposing the dynamics
into its different modes, filtering out the noise or reconstructing the
dominant pathways will be utilized to this aim. For example, the eigenvector
corresponding to the zero eigenvalue of the transition matrix in the master
equation formalism yields the steady state probabilities of the individual
components of the system, while the eigenvector associated with the smallest
eigenvalue extracts the slowest (or least probable) passage.
These approaches should help in developing models that are
(i) more amenable to understanding, (ii) connected to experimental parameters,
and (iii) have far fewer free parameters. Methods of dynamical systems or
control theory can then be used to study how these systems change with changes
of parameters. In such a way, the different qualitative regimes of behavior can
be mapped out explicitly.
Some of the
underlying ideas of using simplified models and methods of dynamical systems
have been applied in our previous work. We modeled and analyzed the
spatio-temporal distribution of F actin in the cell in the presence of choppers
and initiators (9;10). Other
investigators used a simplified model to determine the point in the cell cycle
when apoptosis is triggered, also showing the utility of such simple models (108). Even though the latter model was
lacking in many details, there was a clear distinction between six possible
mechanisms for cell death, which could be compared to experimental data. Simple
systems of differential equations employing mass-action kinetics were the basis
for the design of a bistable genetic switch (109) and a genetic oscillator (110), lending hope
that reduced descriptions can capture natural cellular mechanisms. Finally, in the
absence of sufficient data on kinetic and transport parameters, or initial
concentrations, we will exploit cellular automata approaches (11;111).
We have also been
involved in the last few years in other modern techniques for simplifying
complex models (112;113) in the presence of multiple scale effects. Its
relevance for biological systems is apparent when considering that even a small
subset of the metabolic reactions may evolve into a chaotic behavior. Simple averaging
techniques may not be applicable in these cases, and more advanced mathematical
concepts, such as “measures” are needed, which we will use when appropriate.
Two challenges are, however, to
ensure that the important details are not being neglected, and to be able to
construct simplified models that are quantitative,
rather than just qualitative. For
example, the mean-field approximation is appropriate only if the system has
intrinsic randomness and separable fast processes; and it is usually a major
task to optimize the parameters representative of the mean-field effects in
different environments. Likewise the network connectivity and pseudo-steady
state conditions for individual steps must be validated by supporting data,
either from experiments or simulations with more detailed models. More
importantly, the continuous, deterministic approaches may not be adequate in
most cases, and in particular in cell signaling and regulatory processes, due
to the intrinsic discrete and stochastic nature of molecular interactions.
Specific Aim 2 indeed focuses on the integration of the results from
simulations and analytical methods at different levels for defining the reduced
models and parameters that contain the ‘important details’.