1: \documentclass[11pt]{article}
2:
3:
4: %To compile with bibtex:
5: % > pdflatex iee-unravelling %%no .tex extension here!!
6: % > bibtex iee-unravelling
7: % > pdflatex iee-unravelling
8: % > pdflatex iee-unravelling
9:
10:
11:
12: % The following packages can be found on http:\\www.ctan.org
13: \usepackage{graphics} % for pdf, bitmapped graphics files
14: \usepackage{psfig} % for postscript graphics files
15: %\usepackage{mathptmx} % assumes new font selection scheme installed
16: \usepackage{times} % assumes new font selection scheme installed
17: \usepackage{amsmath} % assumes amsmath package installed
18: \usepackage{amssymb} % assumes amsmath package installed
19:
20: \setlength{\topmargin}{-0.9in} %%US letter: 8.5x11in
21: \setlength{\textheight}{9.5in}
22: \setlength{\oddsidemargin}{-.0 in}
23: \setlength{\textwidth}{6.5in}
24: \setlength{\parskip}{0pt plus 1pt minus 1pt}
25: \setlength{\parsep}{1pt plus 1pt minus 1pt}
26: \setlength{\itemsep}{-1pt plus 1pt minus 1pt}
27: \setlength{\floatsep}{2pt plus 1pt minus 1pt}
28: \setlength{\textfloatsep}{5pt plus 1pt minus 1pt}
29: \renewcommand{\baselinestretch}{1.1}
30:
31:
32:
33: \newtheorem{theorem}{Theorem}
34: \newtheorem{lemma}{Lemma}[section]
35: \newtheorem{proposition}[lemma]{Proposition}
36: \newtheorem{corollary}[lemma]{Corollary}
37: \newtheorem{fact_}{Fact}
38:
39: \newenvironment{theo}[1]{\begin{theorem}\label{#1}\rm}{\end{theorem}}
40: \newenvironment{lem}[1]{\begin{lemma}\label{#1} \rm}{\end{lemma}}
41: \newenvironment{prop}[1]{\begin{proposition}\label{#1}\rm}{\end{proposition}}
42: \newenvironment{cor}[1]{\begin{corollary}\label{#1}\rm}{\end{corollary}}
43: \newenvironment{fact}[1]{\begin{fact_}\label{#1}\rm}{\end{fact_}}
44:
45:
46: \newcommand{\beqn}[1]{\begin{eqnarray}\label{#1}}
47: \newcommand{\eeqn}{\end{eqnarray}}
48: \newcommand{\beq}{\begin{eqnarray*}}
49: \newcommand{\eeq}{\end{eqnarray*}}
50: \newcommand{\bit}{\begin{itemize}}
51: \newcommand{\eit}{\end{itemize}}
52: \newcommand{\rf}[1]{~(\ref{#1})}
53: \newcommand{\comment}[1]{}
54: \newcommand{\bprf}{{\it Proof: }}
55: \newcommand{\bbox}{\hfill\rule{1ex}{1.4ex}}
56: \newcommand{\qed}{\hfill $\Box$}
57: \newcommand{\abs}[1]{|{#1}|}
58: \newcommand{\dnorm}[1]{\|{#1}\|}
59: \newcommand{\eps}{\varepsilon}
60: \newcommand{\N}{{\mathbb N}}
61: \newcommand{\tL}{\mbox{\tiny $L$}}
62: \newcommand{\Prot}{\mbox{\it Prot}}
63: \newcommand{\mRNA}{\mbox{\it mRNA}}
64: \newcommand{\tProt}{\mbox{{\tiny\it Prot}}}
65: \newcommand{\tmRNA}{\mbox{{\tiny\it mRNA}}}
66: \newcommand{\cX}{\widehat{\mbox{\it X}}}
67: \newcommand{\cwg}{\widehat{\mbox{\it wg}}}
68: \newcommand{\cWG}{\widehat{\mbox{\it WG}}}
69: \newcommand{\cen}{\widehat{\mbox{\it en}}}
70: \newcommand{\cEN}{\widehat{\mbox{\it EN}}}
71: \newcommand{\chh}{\widehat{\mbox{\it hh}}}
72: \newcommand{\cHH}{\widehat{\mbox{\it HH}}}
73: \newcommand{\cci}{\widehat{\mbox{\it ci}}}
74: \newcommand{\cCI}{\widehat{\mbox{\it CI}}}
75: \newcommand{\cCIA}{\widehat{\mbox{\it CIA}}}
76: \newcommand{\cCIR}{\widehat{\mbox{\it CIR}}}
77: \newcommand{\cPTC}{\widehat{\mbox{\it PTC}}}
78: \newcommand{\RX}{\mbox{\it RX}}
79: \newcommand{\slp}{\mbox{\it slp}}
80: \newcommand{\SLP}{\mbox{\it SLP}}
81: \newcommand{\wg}{\mbox{\it wg}}
82: \newcommand{\WG}{\mbox{\it WG}}
83: \newcommand{\en}{\mbox{\it en}}
84: \newcommand{\EN}{\mbox{\it EN}}
85: \newcommand{\hh}{\mbox{\it hh}}
86: \newcommand{\HH}{\mbox{\it HH}}
87: \newcommand{\ci}{\mbox{\it ci}}
88: \newcommand{\CI}{\mbox{\it CI}}
89: \newcommand{\CIA}{\mbox{\it CIA}}
90: \newcommand{\CIR}{\mbox{\it CIR}}
91: \newcommand{\PTC}{\mbox{\it PTC}}
92: \newcommand{\ptc}{\mbox{\it ptc}}
93: \newcommand{\twg}{\mbox{\tiny\it wg}}
94: \newcommand{\tWG}{\mbox{\tiny\it WG}}
95: \newcommand{\ten}{\mbox{\tiny\it en}}
96: \newcommand{\tEN}{\mbox{\tiny\it EN}}
97: \newcommand{\thh}{\mbox{\tiny\it hh}}
98: \newcommand{\tHH}{\mbox{\tiny\it HH}}
99: \newcommand{\tci}{\mbox{\tiny\it ci}}
100: \newcommand{\tCI}{\mbox{\tiny\it CI}}
101: \newcommand{\tCIA}{\mbox{\tiny\it CIA}}
102: \newcommand{\tCIR}{\mbox{\tiny\it CIR}}
103: \newcommand{\tPTC}{\mbox{\tiny\it PTC}}
104: \newcommand{\tptc}{\mbox{\tiny\it ptc}}
105:
106: \newcommand{\lthw}[1]{\ln\, \frac{{#1}}{\theta}} %1/\theta
107: \newcommand{\lthow}[1]{\ln\, \frac{{#1}}{1-\theta}} %1/(1-\theta)
108: \newcommand{\lth}{\ln \frac{1}{\theta}}
109: \newcommand{\ltho}{\ln \frac{1}{1-\theta}}
110:
111: \newcommand{\ea}{et. al.}
112:
113:
114: %\bibliographystyle{pnas}
115: %\bibliographystyle{plain}
116: \bibliographystyle{unsrt} %% Same as plain except the entries appear in the order of
117: %% their first citation.
118: %\bibliographystyle{abbrv} %%Only initials for first names
119: %\bibliographystyle{acm} %%A lot like abbrv.bst, but names come out
120: %%"Last, initials", with last names capitalized (like SICON)
121:
122:
123: \title{
124: Methods of robustness analysis for Boolean models \\of gene control networks
125: \footnote{M. Chaves is with the Institute for Systems Theory and Automatic Control,
126: University of Stuttgart, Pfaffenwaldring 9, 70550 Stuttgart, Germany
127: {\tt\small chaves@ist.uni-stuttgart.de}.
128: E.D. Sontag is with the Department of Mathematics,
129: Rutgers University, Piscataway, NJ 08854 USA
130: {\tt\small sontag@math.rutgers.edu}.
131: R. Albert is with the Department of Physics and Huck Institutes for the Life Sciences,
132: Pennsylvania State University, University Park, PA 16802 USA
133: {\tt\small ralbert@phys.psu.edu}.
134: }
135: }
136:
137: \author{Madalena Chaves, Eduardo D. Sontag and R\'eka Albert\thanks{Corresponding author.}
138: }
139:
140: \date{}
141:
142: \begin{document}
143:
144:
145: \maketitle
146:
147: \begin{abstract}
148: As a discrete approach to genetic regulatory networks, Boolean models
149: provide an essential qualitative description of the structure of interactions
150: among genes and proteins. Boolean models generally assume only two possible states (expressed
151: or not expressed) for each gene or protein in the network as well as a
152: high level of synchronization among the various regulatory processes.
153:
154: In this paper, we discuss and compare two possible methods of
155: adapting qualitative models to incorporate
156: the continuous-time character of regulatory networks.
157: The first method consists of introducing asynchronous updates in the Boolean
158: model. In the second method, we adopt the approach introduced by L. Glass
159: to obtain a set of piecewise linear differential equations which continuously
160: describe the states of each gene or protein in the network.
161:
162: We apply both methods to a particular example: a Boolean model of the
163: segment polarity gene network of {\it Drosophila melanogaster}. We analyze the
164: dynamics of the model, and provide a theoretical characterization of
165: the model's gene pattern prediction as a function of the timescales
166: of the various processes.
167: \end{abstract}
168:
169:
170:
171: \section{Introduction}
172:
173: Genes and gene products interact on several levels. At the genomic level, transcription
174: factors can activate or inhibit the transcription of genes to give mRNAs. Since these
175: transcription factors are themselves products of genes, the ultimate effect is that genes
176: regulate each other's expression as part of gene regulatory networks. Similarly, proteins
177: can participate in diverse post-translational interactions that lead to modified protein
178: functions or to formation of protein complexes that have new roles; the totality of these
179: processes is called a protein-protein interaction network. In many cases different
180: levels of interactions are integrated - for example, when the presence of an external
181: signal triggers a cascade of interactions that involves biochemical
182: reactions, protein-protein interactions and transcriptional regulation.
183:
184: During the last decade, genomics, transcriptomics and proteomics have produced an
185: incredible quantity of molecular interaction data, contributing to genome-scale maps of
186: protein interaction networks~\cite{giot03,han04,li04} and transcriptional regulatory
187: networks~\cite{lee02}. Network analysis of these maps revealed intriguing topological
188: similarities~\cite{a05}. At the same time, it has been increasingly realized
189: that cellular interaction maps only represent a network of possibilities, and
190: not all edges are present and active {\it in vivo} in a given condition or
191: in a given cellular location~\cite{han04,lbyst04}.
192: Therefore, only an integration of (time-dependent) interaction and activity information
193: will be able to give the correct dynamical picture of a cellular network.
194:
195: For many biological networks, and in particular genetic control or regulatory networks,
196: detailed information on the kinetic rates of protein-protein or protein-DNA
197: interactions is rarely available. However, for many biological systems,
198: evidence shows that regulatory relationships can be sigmoidal and be
199: well approximated by step functions.
200: In this case, Boolean models, where every variable has only two states (ON/OFF),
201: and the dynamics is given by a set of logical rules, are frequently appropriate
202: descriptions of the network of interactions among genes and proteins.
203: Examples include models of genetic networks in the fruit fly
204: {\it Drosophila melanogaster}~\cite{st01,ao03} and the flowering plant
205: {\it Arabidopsis thaliana}~\cite{mta99,epa04}.
206:
207: While Boolean models introduce biologically unrealistic time constraints
208: (typically, such models use synchronous updates, which inherently assume
209: that the various biological processes have the same duration),
210: they still provide significant qualitative information on the
211: underlying structure of the network.
212: On the other hand, continuous models certainly have a more realistic time description of
213: a biological system. But, in the absence of information on the kinetic rates, continuous
214: models include many unknown parameters, and analysis of the system involves exploring
215: the (often large) state space of parameters.
216: An important (continuous) model for {\it Drosophila melanogaster} segment polarity genes
217: was first developed in~\cite{dmmo00}, where a thorough investigation of the parameter
218: space showed that the system is very robust with respect to variations in the kinetic
219: constants.
220: Both this continuous model and the discrete model~\cite{ao03} agree in their overall
221: conclusions regarding the robustness of the segment polarity gene network.
222:
223: In this paper, we propose two new approaches to the analysis of Boolean models,
224: which combine discrete logical rules and structure with more realistic
225: assumptions regarding the relative timescales of the genetic processes.
226: The first method introduces asynchronous updates in the Boolean model; since
227: update times are randomly chosen, the model is now stochastic
228: %thereby also building an element of stochasticity into the system
229: (see also~\cite{cas05}). The second method associates to the discrete variables a set of continuous variables,
230: whose dynamics is given by a piecewise linear system of differential equations,
231: thus introducing a simple ``hybrid'' model in the manner first suggested by L. Glass and
232: collaborators in~\cite{g75,eg00,gk73}.
233: These methods allow us to naturally probe the system with respect to perturbations in
234: the time dynamics to analyze its performance.
235: Both methods uncover the robustness of the segment polarity gene
236: model~\cite{ao03}, and its ability to correctly predict the final gene
237: expression pattern.
238: %
239:
240: Section~\ref{sec-drosophila} summarizes the Boolean model~\cite{ao03}, to be analyzed
241: with our methods. The asynchronous updates and the Glass-type methods are described
242: and applied in Sections~\ref{sec-asynch} and~\ref{sec-glass}, respectively.
243: In Section~\ref{sec-mutants}, we discuss the effects of perturbations in initial gene expression,
244: and the development of mutant phenotypes.
245: Finally, in Section~\ref{sec-timesep} we show that under a timescale separation
246: between posttranslational and translational processes in the model~\cite{ao03},
247: we are able to analytically and exactly compute how frequently the model will predict the
248: correct gene pattern.
249:
250:
251:
252:
253: \section{The segment polarity gene network in {\it Drosophila}}
254: \label{sec-drosophila}
255: We will apply our analysis to a Boolean model of the interactions among the
256: {\it Drosophila melanogaster} segment polarity genes.
257: This gene network represents the last step in the
258: hierarchical cascade of gene families initiating the segmented body of the fruit fly.
259: While genes in preceding stages of development act transiently, the segment polarity genes
260: are expressed throughout the life of the fly.
261: The best characterized segment polarity genes include {\it engrailed} ($\en$),
262: {\it wingless} ($\wg$), {\it hedgehog} ($\hh$), {\it patched} ($\ptc$),
263: {\it cubitus interruptus} ($\ci$) and {\it sloppy paired} ($slp$), coding for
264: the corresponding proteins, which we will represent by capital letters
265: ($\EN$, $\WG$, $\HH$, $\PTC$, $\CI$ and $\SLP$).
266: Two additional proteins, resulting from transformations of the protein $\CI$,
267: also play important roles: $\CI$ may be converted into a transcriptional activator, $\CIA$,
268: or may be cleaved to form a transcriptional repressor $\CIR$.
269:
270: The expression pattern of the {\it Drosophila} segment polarity genes
271: (see Table~\ref{table_steady_states})
272: is maintained almost unmodified for three hours,
273: during which time the embryo is divided into $14$ parasegments.
274: Each of these parasegments is composed of about 4 cells, delimited by furrows positioned
275: between the $\wg$ and $\en$ -expressing cells~\cite{hs92}.
276:
277: The Boolean model that we will study was introduced and developed by one of us in~\cite{ao03}.
278: (Further robustness analysis was also developed in~\cite{cas05}.)
279: In this model, a parasegment of four cells is considered: the variables are the expression levels
280: of the segment polarity genes and proteins (listed above) in each of the four cells
281: (the total number of nodes in the network is thus $4\times 13=52$.).
282: The expression level of each gene or protein is assumed to be either 0 (OFF) or 1 (ON).
283: The model successfully describes the transition from the initial expression
284: pattern\rf{eq-initial-wt} to a final pattern two or three developmental stages later,
285: when the embryo has been clearly divided into parasegments
286: (see first entry of Table~\ref{table_steady_states}).
287: As discussed in~\cite{ao03}, the evolution of these gene expression patterns is well described
288: by a set of logical rules, which are depicted in Table~\ref{table_rules}.
289:
290: We adopt the notation ``$\wg_1^k$'' or ``$\wg_1(k)$'' to represent the state of
291: {\it wingless} mRNA in the first cell of the parasegment at time $k$.
292: Similar notations apply for other mRNAs and proteins.
293: Periodic boundary conditions are assumed, meaning that: $node_{4+1}=node_1$ and
294: $node_{1-1}=node_4$. The wild type initial pattern corresponds to:
295: \beqn{eq-initial-wt}
296: \wg_4^0=1,\ \en_1^0=1,\ \hh_1^0=1, \ \ptc_{2,3,4}^0=1, \ \ci_{2,3,4}^0=1,
297: \eeqn
298: with the remaining nodes zero.
299:
300: \begin{table}
301: \caption{Regulatory functions governing the states of segment polarity gene products in the model.
302: Each node is labeled by its biochemical symbol and subscripts signify cell number.}
303: \label{table_rules}
304: \centering{
305: \begin{tabular}{ll}
306: \hline %\toprule
307: Node & Boolean updating function (synchronous algorithm) \\
308: \hline\hline%\midrule
309: $\SLP_i$ & $\SLP_i(k+1)=\left\{\begin{array}{lllll}
310: 0 &\mbox{if}& i\in\{1,2\}\\
311: 1 &\mbox{if}& i\in \{3,4\}\\
312: \end{array}\right.$ \\
313: $\wg_i$ & $\wg_i(k+1)=(\CIA_i(k)$ and $\SLP_i(k)$
314: and not $\CIR_i(k))$ \\
315: & or $[\wg_i(k) $ and $(\CIA_i(k)$ or $\SLP_i(k) )$
316: and not $\CIR_i(k)]$ \\
317: $\WG_i$ & $\WG_i(k+1)=\wg_i(k)$ \\
318: $\en_i$ & $\en_i(k+1)=(\WG_{i-1}(k)$ or $\WG_{i+1}(k))$
319: and not $\SLP_i(k)$ \\
320: $\EN_i$ & $\EN_i(k+1)=\en_i(k)$ \\
321: $\hh_i$ & $\hh_i(k+1)=\EN_i(k)$ and not $\CIR_i(k)$ \\
322: $\HH_i$ & $\HH_i(k+1)=\hh_i(k)$ \\
323: $\ptc_i$ & $\ptc_i(k+1)=\CIA_i(k)$ and not $\EN_i(k)$ and not $\CIR_i(k)$ \\
324: $\PTC_i$ & $\PTC_i(k+1)=\ptc_i(k)$ or $(\PTC_i(k)$ and not $\HH_{i-1}(k)$ \\
325: & and not $\HH_{i+1}(k))$ \\
326: $\ci_i$ & $\ci_i(k+1)=$ not $\EN_i(k)$ \\
327: $\CI_i$ & $\CI_i(k+1)=ci_i(k)$ \\
328: $\CIA_i$ & $\CIA_i(k+1)=CI_i(k)$ and [not $\PTC_i(k)$ or $\HH_{i-1}(k)$\\
329: & or $\HH_{i+1}(k)$ or $\hh_{i-1}(k)$ or $\hh_{i+1}(k)$]\\
330: $\CIR_i$ & $\CIR_i(k+1)=\CI_i(k)$ and $\PTC_i(k)$ and not $\HH_{i-1}(k)$ \\
331: & and not $\HH_{i+1}(k)$ and not $\hh_{i-1}(k)$ and not $\hh_{i+1}(k)$\\
332: \hline%\bottomrule
333: \end{tabular}}
334: \end{table}
335:
336:
337: \subsection{Steady states of the Boolean model}
338: A complete analysis of the steady states is found in~\cite{ao03}.
339: Table~\ref{table_steady_states} summarizes these results, indicating the
340: expressed nodes in each of the six steady-states.
341: We note that three of the four main steady states agree perfectly with
342: experimentally observed states corresponding to wild type, to $\ptc$ knockout mutant (broad striped)
343: and to $\en$, $\wg$ or $\hh$ knockout mutant (non-segmented)
344: embryonic patterns (\cite{gakt00,tek92}; see~\cite{ao03} for more references).
345: \begin{table}[h]
346: \caption{Complete characterization of the model's steady states.}
347: \label{table_steady_states}
348: \begin{center}
349: \begin{tabular}{ll}
350: \hline
351: Steady state & Expressed nodes \\
352: \hline\hline
353: wild type & $\wg_{4}$, $\WG_{4}$, $\en_{1}$, $\EN_{1}$, $\hh_{1}$, $\HH_{1}$, \\
354: & $\ptc_{2,4}$, $\PTC_{2,3,4}$, $\ci_{2,3,4}$, \\
355: & $\CI_{2,3,4}$, $\CIA_{2,4}$, $\CIR_{3}$ \\
356: \hline
357: broad stripes & $\wg_{3,4}$, $\WG_{3,4}$, $\en_{1,2}$, $\EN_{1,2}$,
358: $\hh_{1,2}$, $\HH_{1,2}$, \\
359: & $\ptc_{3,4}$, $\PTC_{3,4}$, $\ci_{3,4}$, $\CI_{3,4}$, $\CIA_{3,4}$ \\
360: \hline
361: no segmentation & $\ci_{1,2,3,4}$, $\CI_{1,2,3,4}$,
362: $\PTC_{1,2,3,4}$, $\CIR_{1,2,3,4}$ \\
363: \hline
364: wild type variant & $\wg_{4}$, $\WG_{4}$, $\en_{1}$, $\EN_{1}$, $\hh_{1}$, $\HH_{1}$, \\
365: & $\ptc_{2,4}$, $\PTC_{1,2,3,4}$, $\ci_{2,3,4}$, \\
366: & $\CI_{2,3,4}$, $\CIA_{2,4}$, $\CIR_{3}$ \\
367: \hline
368: ectopic & $\wg_{3}$, $\WG_{3}$, $\en_{2}$, $\EN_{2}$, $\hh_{2}$, $\HH_{2}$, \\
369: & $\ptc_{1,3}$, $\PTC_{1,3,4}$, $\ci_{1,3,4}$, \\
370: & $\CI_{1,3,4}$, $\CIA_{1,3}$, $\CIR_{4}$ \\
371: \hline
372: ectopic variant & $\wg_{3}$, $\WG_{3}$, $\en_{2}$, $\EN_{2}$, $\hh_{2}$, $\HH_{2}$, \\
373: & $\ptc_{1,3}$, $\PTC_{1,2,3,4}$, $\ci_{1,3,4}$, \\
374: & $\CI_{1,3,4}$, $\CIA_{1,3}$, $\CIR_{4}$ \\
375: \hline
376: \end{tabular}
377: \end{center}
378: \end{table}
379:
380:
381: \subsection{The regulatory function of the {\it sloppy paired} gene}
382:
383: The rule for $\SLP$ protein in Table~\ref{table_rules} summarizes in a simple way the experimental observations
384: on the expression and regulatory activity of the {\it sloppy paired} gene in the segment polarity network~\cite{cgg94}.
385: A more detailed rule for the {\it sloppy paired} expression pattern can be created to incorporate recent evidence of engrailed protein inhibiting $\slp$ transcription~\cite{av03}.
386: However, inhibition by engrailed accounts only partially for the experimentally observed restriction of
387: $\slp$ to the posterior half of the parasegment. Thus we need to invoke an additional regulatory
388: effect, which we denote by $\RX$. $\RX$ probably represents a combination of regulation by the
389: pair-rules responsible for the establishment of $\slp$, namely {\it runt}, {\it opa} and
390: Factor X~\cite{sg04} and of $\slp$ autoregulation.
391:
392: Therefore, $\SLP$ expression in Table~\ref{table_rules} can be replaced by
393: the following set of equations:
394: \beqn{eq-RX}
395: \RX_i(k+1)&=&\left\{\begin{array}{lllll}
396: 0, &\mbox{ if }& i\in\{1,2\}\\
397: 1, &\mbox{ if }& i\in \{3,4\}\\
398: \end{array}\right. \nonumber \\ & & \\
399: \slp_i(k+1)&=&\RX_i(k) \mbox{ and not }\EN_i(k) \nonumber\\
400: \SLP_i(k+1)&=&\slp_i(k) \ . \nonumber
401: \eeqn
402: The {\it sloppy paired} initial conditions would then be:
403: \beqn{eq-initial-slp}
404: \slp_{3,4}^0=1,\ \SLP_{3,4}^0=1.
405: \eeqn
406:
407: This generalization of the segment polarity network model introduces additional steady states,
408: such as a two-stripe $\en$ pattern characterized by $\slp_4=\SLP_4=1$, $\wg_4=\WG_4=1$,
409: $\en_{1,3}=\EN_{1,3}=1$, $\hh_{1,3}=\HH_{1,3}=1$, $\ptc_{2,4}=\PTC_{2,4}=1$,
410: and $\ci_{2,4}=\CI_{2,4}=\CIA_{2,4}=1$
411: (this pattern was also found in~\cite{ao03} as a result of $\slp$ mutation). This expression pattern is non-viable since it has two $\it en- wg$ borders and would lead to an ectopic parasegment structure. On the other hand, it is not difficult to see that, starting from conditions\rf{eq-initial-wt}
412: and\rf{eq-initial-slp}, none of the ``new'' steady states are reachable, since these
413: initial conditions imply that neither $\slp$ nor $\SLP$ can change their expression at any time.
414: Indeed we have the following result:
415:
416: \begin{lem}{lm-slp}
417: Consider the extended model of Table~\ref{table_rules} together with\rf{eq-RX}.
418: Assume that initial conditions are given by\rf{eq-initial-wt} and\rf{eq-initial-slp}.
419: Then $\slp_i(t)=\slp_i^0$ and $\SLP_i(t)=\SLP_i^0$ for all times .
420: \end{lem}
421:
422: A sketch of proof is as follows. Note first that $\slp_{1,2}(t)=\SLP_{1,2}(t)\equiv0$ follows
423: from $\RX_{1,2}(t)=0$ for all $t$.
424: Next, observe that:
425: \beq
426: \slp_4(t_1)=0 \Rightarrow \EN_4(t_2)=1 \Rightarrow \en_4(t_3)=1
427: \Rightarrow \SLP_4(t_4)=0 \Rightarrow \slp_4(t_5)=0,
428: \eeq
429: where $t_1\geq t_2\geq t_3\geq t_4\geq t_5\geq0$.
430: Thus, in order for $\slp_4$ to become zero at any time, it had to be so at
431: some previous instant ($t_5\leq t_1$). If $\slp_4(0)=1$, and if $T>0$ is the first
432: instant such that $\slp_4(T)=0 $ then we have a contradiction. Hence $\slp_4(t)=1$
433: for all times. (Note that this argument is independent of the order in
434: which the nodes are updated.)
435: Similar arguments show that $\slp_3(t)=\slp_3(0)=1$ and $\SLP_{3,4}(t)=1$,
436: for all $t$.
437:
438: Thus the extended model leads to the same results as assuming a constant $SLP$ pattern. For this reason, and lacking more specific biological evidence on the regulation of {\it sloppy paired}, our present analysis is focused on the simpler biologically relevant
439: model of Table~\ref{table_rules}.
440:
441:
442: \section{Asynchronous algorithms}
443: \label{sec-asynch}
444: In general, for a network of $N$ gene products (denoted $x_1$, $\ldots$, $x_N$),
445: the dynamics of a Boolean model is typically studied by simultaneously updating
446: the state of all the nodes in the network, according to
447: \beq
448: X_i^{k+1}=F_i(X_1^k,X_2^k,\ldots,X_N^k),\ \ \ i=1,\ldots,N
449: \eeq
450: where $F_i$ is the regulating function for mRNA or protein $X_i$.
451: An underlying hypothesis is the existence of perfect synchronization among
452: the various regulatory processes. However, it is well known that
453: the timescales of transcription, translation, and degradation processes
454: can vary widely from gene to gene and can be anywhere from minutes to hours.
455:
456: In analogy with task coordination and data communication procedures
457: in the context of parallel computation systems~\cite{tsi}, we have previously
458: developed several methods that introduce different timescales for the different
459: regulatory processes within the network~\cite{cas05}.
460: %A first algorithm consists of allowing the (constant) time unit of the
461: %synchronous model to be randomly perturbed for each node.
462: These include algorithms that randomly choose the order in which the nodes are updated
463: (this random order algorithm is summarized below in~\ref{sec-rand0},
464: and in Section~\ref{sec-timesep}),
465: and a totally asynchronous algorithm, where the next updating times for each node are
466: randomly chosen at each instant.
467:
468: We now introduce a more intuitive asynchronous algorithm, where each node is updated
469: according to its own specific time unit.
470: The time units for the nodes are randomly chosen from a uniform distribution in an
471: interval $[1-\eps,1+\eps]$, where $\eps\in(0,1)$. The updating times of $i$-th node are
472: then pre-specified as:
473: \beq
474: T_i^1, T_i^2,\ldots, T_i^k,\ldots \ \ \ k\in\N,
475: \eeq
476: with
477: \beqn{eq-itimes}
478: T_i^{k+1} = T_i^k +\gamma_i = k\gamma_i,\ \ \ k\in\N.
479: \eeqn
480: For instance, $\gamma_{\tWG_4}<\gamma_{\twg_4}$ means that wingless protein in the fourth cell
481: is translated at a faster rate (shorter time intervals) then {\it wingless} mRNA is produced.
482: At any given time $t$, the next node(s) to be updated is(are) $i$
483: such that $T_i^k=\min_{j,\ell}\{ T_j^\ell\geq t\}$, for some $k$.
484: The variables $X_i$ are updated according to:
485: \beqn{eq-inode}
486: X_i(T_i^k) = F_i(\, X_1(\tau^k_{1i}),\ldots,X_N(\tau^k_{Ni})\,),
487: \eeqn
488: where $\tau^k_{ji}$ defines the most recent instant when node $j$ was updated, that is
489: \beqn{eq-ijtau}
490: \tau^k_{ji}= \max_\ell\ \{ \ T_j^\ell: \ T_j^\ell < T_i^k \ \}.
491: \eeqn
492: By ordering all the time sequences $\{T^k_{i}:\ i=1\ldots N,\ k=1,2,\ldots\}$, into a single
493: nondecreasing sequence, say $\{t_1,t_2,\ldots\}$, the asynchronous model can also
494: be written in the form
495: \beq
496: X_i(t_{k+1})=\left\{
497: \begin{array}{ll}
498: F_i(X_1(t_k),X_2(t_k),\ldots,X_N(t_k)),
499: & \mbox{if } t_k=\ell\gamma_i \mbox{ for some $i$, $\ell$} \\
500: X_i(t_k), & \mbox{otherwise.}
501: \end{array}
502: \right.
503: \eeq
504: It is clear that the steady states of this model must satisfy
505: $X_i(t_{k+1})=X_i(t_k)=F_i(X(t_k))$,
506: and therefore are the same as those of the synchronous boolean model in
507: Table~\ref{table_steady_states}.
508:
509: Note that the case $\eps=0$ reduces to the synchronous model, where every node is updated
510: simultaneously ($\gamma_i=1$), at the same time instants: $T_i^k=k$, for all $i=1,\ldots,N$.
511: This algorithm allows great variability in each process' duration, exploring the gene
512: expression patterns due to all possible combinations of individual timescales.
513: {Implementation of this asynchronous algorithm shows that, if started from the initial wild
514: type state\rf{eq-initial-wt}, any of the steady states of the model
515: (Table~\ref{table_steady_states}) may occur with a certain probability.}
516: {The probability of occurrence of each pattern depends on the range over which
517: the individual time units $\gamma_i$ are allowed to vary (see Fig.~\ref{fig-interval-length}).
518: For $\eps=0$, the wild type steady state is attained with probability 100\%
519: (corresponding to the synchronous
520: Boolean model). As $\eps$ increases to 0.01 (resp. 0.1) this value decreases to 60\% (resp. 44\%).
521: However, further increase in $\eps$ (hence larger time intervals) unexpectedly leads to an
522: increase in the occurrence of the wild type state, up to 51\% for $\eps=0.9$.}
523: Other final states observed are the
524: broad-striped pattern ($25\%-38\%$) observed in heat-shock experiments and $\ptc$
525: mutants~\cite{gakt00} and the pattern with no segmentation ($12\%-15\%$) observed
526: in $\en$, $\hh$ or $\wg$ mutants~\cite{tek92}, the latter two corresponding to
527: embryonic lethal phenotypes~\cite{gakt00}.
528: Each of the other three steady states occurs with frequencies less than $5\%$.
529: (These values were obtained from 10000 numerical experiments.)
530:
531: A possible extension of this algorithm would be to consider a discrete model with a finite
532: number of logical levels describing ON, OFF as well as other intermediate steps of the
533: system~\cite{st01}. This would involve decisions about the number of intermediate steps,
534: their values, and development of new transition rules.
535: Instead, in this paper we will focus on a ``hybrid'' model, that takes into account
536: the continuous nature of the biological processes, while still using Boolean rules to
537: describe ON/OFF transitions (Section~\ref{sec-glass}).
538:
539: \subsection{Random order algorithm}
540: \label{sec-rand0}
541: For comparison purposes, we briefly summarize an alternative asynchronous algorithm,
542: which guarantees that every node is updated exactly once during each unit time
543: interval~\cite{cas05}.
544: A random order of updates for the $N$ nodes is generated as a permutation
545: $\phi^k$ of $\{1,\ldots N\}$. This permutation is randomly chosen out of a uniform
546: distribution over the set of all $N!$ possible permutations, at the beginning of
547: the time unit $k$. The updating times for each node are now written as
548: \beq
549: T_i^k = N(k-1) + \phi^k(i),\ \ \ k\in\N,
550: \eeq
551: so that $\phi^k(j)<\phi^k(i)$ implies $T_j^k<T_i^k$, and node $j$ is updated before node $i$
552: at the $k$-th iteration.
553: The results of this algorithm are qualitatively similar to those of the asynchronous
554: algorithm (Table~\ref{table_steady_states}).
555:
556: \begin{table}
557: \caption{The frequencies of the six steady states observed with the three different methods when
558: starting from the wild type initial condition.}
559: \label{tab-all-permutations}
560: \begin{center}
561: \begin{tabular}{lccc}
562: \hline
563: Steady state & Asynchronous & Random order & Glass-type\\
564: pattern & algorithm & algorithm & model\\
565: \hline\hline
566: wild type & 44-51\% & 56\% & 89-100\% \\
567: %\hline
568: broad stripes & 25-38\% & 24\% & 0-6\% \\
569: %\hline
570: no segmentation & 12-15\% & 15\% & 0-3\% \\
571: %\hline
572: wild type variant & 4-5.6\% & 4.2\% & 0-1\% \\
573: %\hline
574: ectopic & 0.4-1\% & 0.98\% & 0\% \\
575: %\hline
576: ectopic variant & 0.1-0.5\% & 0.68\% & 0\% \\
577: \hline
578: \end{tabular}
579: \end{center}
580: \end{table}
581:
582: \section{Glass-type networks}
583: \label{sec-glass}
584: The asynchronous algorithm defined by\rf{eq-itimes},\rf{eq-inode} and\rf{eq-ijtau}
585: allows the introduction of distinct timescales for each regulation process in a
586: Boolean model.
587: We next propose an alternative method which provides a bridge between discrete and continuous
588: approaches, resulting in a more realistic model, but without the necessity of specifying
589: any kinetic or binding parameters (which are typically unknown).
590: In this method, the gene and protein levels are represented
591: as continuous variables, and their time evolution is described by differential equations,
592: but the interactions among nodes are still modeled by Boolean functions~\cite{g75,eg00,gk73,jghpsg04}.
593: Glass~\cite{g75} introduced a class of piecewise linear differential equations that
594: combine logical rules for the synthesis of gene products with linear (free) decay by describing
595: each node with two variables, one discrete and one continuous.
596: For simplicity of notation, in what follows we will let $\cX_i$ denote the continuous
597: variable associated with node $i$, its discrete variable $X_i$, and the discrete variable's
598: Boolean rule by $F_i$. The Glass-type model is then
599: \beqn{eq-glass}
600: \frac{d\, \cX_i}{dt} = -\cX_i + F_i(X_1,X_2,\ldots,X_N),\ \ \ i=1,\ldots,N.
601: \eeqn
602: At each instant $t$, the discrete variable $X_i$ is defined as a function of
603: the continuous variable according to a threshold value:
604: \beqn{eq-threshold}
605: X_i(t)=\left\{ \begin{array}{ll}
606: 0, & \cX_i(t)\leq\theta \\
607: 1, & \cX_i(t)>\theta\ ,
608: \end{array}
609: \right.
610: \eeqn
611: where $\theta\in(0,1)$.
612: The discrete variables $X_i$ represent the ON and OFF levels of the nodes
613: in the Boolean model.
614: The underlying assumption in this Glass model is that the decay rate and activation
615: threshold of each gene product is identical.
616: Since the initial condition for the piecewise linear system\rf{eq-glass} is also\rf{eq-initial-wt}
617: (i.e., $X(0)=\cX(0)$) and $F_i\in\{0,1\}$, it is easy to see that solutions of\rf{eq-glass}
618: evolve in the hypercube $[0,1]^N$.
619: Under these conditions, the limiting values ``0'' and ``1'' of the continuous variable $\cX_i$
620: represent, respectively, ``absence of species $i$'' and ``maximal concentration of species $i$''
621: -- thus we can view the $\cX_i$ as dimensionless variables, scaled to attain their maximal values at 1.
622: %
623: The continuous dynamics is translated into a Boolean ON/OFF response, according to $\theta$:
624: as soon as $\cX_i$ increases above $\theta$, species $i$ is considered to be in the ON state; otherwise
625: it remains in the OFF state (see also~\cite{jghpsg04}).
626: Thus the parameter $\theta$ defines the fraction of ``maximal concentration''
627: necessary for a protein or mRNA to effectively perform its biological function.
628: This method allows us to study the continuous evolution of the genetic network
629: simply by specifying $\theta$, the fraction of maximal concentration that is effective as ON level,
630: avoiding the need to specify any kinetic parameters.
631: Below and in Section~\ref{sec-convergence}, we will see that system\rf{eq-glass} exhibits
632: distinct dynamics in the two regions $\theta\leq1/2$ and $\theta>1/2$.
633: %
634: It is easy to see that the steady states of the piecewise linear equations\rf{eq-glass}
635: are still those of the Boolean model, since:
636: \beq
637: \frac{d\, \cX_i}{dt} = 0 \ \Leftrightarrow \ \cX_i=F_i(X_1,X_2,\ldots,X_N),\ i=1,\ldots,N,
638: \eeq
639: independently of $\theta$.
640: Applying this method to the {\it Drosophila} segment polarity gene network, we find an
641: exact convergence to the wild type steady state when started from
642: the wild type initial condition (Fig.~\ref{fig-glass-wt}),
643: independently of the ON/OFF threshold value $\theta$.
644: This result supports the Boolean model as a suitable description of the underlying network of
645: gene interactions.
646:
647: \subsection{Introducing distinct timescales}
648: The assumption of equal decay rates and activation thresholds for all nodes is
649: an oversimplification similar to that made in synchronous Boolean models.
650: However, for further robustness analysis, one may introduce different timescales for
651: the different processes, by scaling the time units in each differential equation according to:
652: \beqn{eq-glass-a}
653: \frac{d\, \cX_i}{dt} = \alpha_i(\ -\cX_i + F_i(X_1,X_2,\ldots,X_N) \ ),
654: \eeqn
655: with $\alpha_i\geq\eps$ for some fixed $\eps>0$, $i=1,\ldots,N$.
656: Each $X_i$ is a discrete variable defined as before
657: (note that steady states of this new system are still those of the
658: Boolean model).\footnote{We will analyze the behaviors of trajectories of systems of the
659: form\rf{eq-glass}, assuming that trajectories are well-defined. Since the
660: right-hand sides of equations of these type are discontinuous, it is very
661: difficult to give general existence and uniqueness theorems for solutions of
662: inital-value problems. One must impose additional assumptions, insuring that
663: only a finite number of switches can take place on any finite time interval,
664: and often tools from the theory of differential inclusions must be applied,
665: see for instance~\cite{gedeon} for more discussion. See also~\cite{jghpsg04}.}
666:
667: This method represents a continuous equivalent to the asynchronous algorithm described in
668: Section~\ref{sec-asynch}. Here, the (inverse) scaling factors $\alpha_i^{-1}$ may be viewed
669: as half-lives of mRNA or proteins. These may be directly compared to the individual time
670: units $\gamma_i$ as follows. Using Euler's method to discretize system\rf{eq-glass-a} obtains:
671: \beq
672: \cX_i(t+\Delta t)= \cX_i(t)+\alpha_i\Delta t (- \cX_i(t)+F_i(X(t))).
673: \eeq
674: Now notice that choosing the integrating time interval to be such that $\alpha_i\Delta t=1$
675: recovers the discrete asynchronous algorithm with specific time units
676: \beq
677: \gamma_i=\Delta t=\alpha_i^{-1}.
678: \eeq
679: For comparison to the discrete algorithm, we choose both the scale factors $\alpha_i^{-1}$
680: and the time units $\gamma_i$
681: randomly from a uniform distribution in intervals of the form $[1-\eps,1+\eps]$, $\eps\in(0,1)$.
682: The numerical experiments are reported in Fig.~\ref{fig-interval-length}, where the threshold
683: value $\theta$\rf{eq-threshold} was set to $1/2$.
684: We again observed that, starting from wild type initial
685: conditions\rf{eq-initial-wt}, all steady states may occur with a certain frequency
686: (see Table~\ref{tab-all-permutations}). But, in contrast to the asynchronous Boolean model,
687: the wild type pattern occurs with frequencies that decrease monotonically with $\eps$,
688: down to 89\% for $\eps=0.9$ (Fig.~\ref{fig-interval-length}).
689: The next more frequently achieved patterns are the broad stripes
690: (Fig.~\ref{fig-glass-bs}), with probability 6\% for $\eps=0.9$, the no segmentation
691: (Fig.~\ref{fig-glass-ns}), with probability 3\%, and the wild type
692: variant, with probability 1\%.
693:
694: The three methods we have described (Table~\ref{tab-all-permutations})
695: produce qualitatively compatible results, in the sense that the wild
696: type pattern is always the most frequently occurring steady state,
697: followed by the broad stripes, no segmentation, and wild type variant patterns.
698:
699: \subsection{Fraction of maximal concentration that defines an ON state}
700: The piecewise linear system\rf{eq-glass-a} follows the threshold\rf{eq-threshold} to decide
701: whether a given node is ON or OFF. While this value $\theta$ did not affect the
702: dynamics of the system in the case $\alpha_1=\alpha_2=\cdots=\alpha_N$, it plays a significant role
703: in the general case.
704: In Figure~\ref{fig-threshold}, it is immediate to see that the effect of $\theta$ depends on the
705: length of the interval allowed for the timescales. Indeed, there is a marked difference between
706: narrow ($\eps\leq0.5$) and wide ($\eps\geq0.6$) intervals.
707: For narrow intervals, numerical experiments indicate that, starting from initial condition\rf{eq-initial-wt},
708: system\rf{eq-glass-a} converges to wild type steady state with probability around 90\% or more;
709: whereas, for wider intervals, this probability may decrease down to 68\% ($\eps=0.9$, $\theta=0.9$).
710:
711: On the other hand, the threshold value also divides the dynamics into two regions.
712: For $\theta\leq0.5$, the probability that initial condition\rf{eq-initial-wt} leads to
713: the wild type steady state is above 90\%, independently of $\eps$. For $\theta\leq0.5$
714: it is less probable to reach the non segmented (around 1\%) than the broad striped pattern
715: (around 10\%).
716: For higher $\theta\geq0.6$, the probability of reaching the wild type steady state (from
717: initial condition\rf{eq-initial-wt}) decreases very significantly, and in addition, it becomes
718: more problable to reach the non segmented (around 20\%) than the broad striped pattern (around 7\%).
719:
720: In our Glass-type model, $\theta$ represents the fraction of maximal concentration above which
721: an mRNA or protein is considered ON, or biologically effective.
722: Our results indicate that quite small fractions of the maximal concentration can (and should) be
723: interpreted as sufficient amounts for an mRNA or protein to be in the ON state.
724: Namely, when the concentration of mRNAs or proteins has increased to a fraction up to half
725: its maximum possible value, it is already present in a sufficient amount to perform its function.
726: This follows from observation of Figure~\ref{fig-threshold}: $0<\theta\leq0.5$ leads to
727: a good (realistic) performance of the model, with 90\% convergence to wild type pattern.
728: Our results also indicate that a higher threshold is, on the contrary, not very realistic.
729: For instance, setting $\theta\geq0.6$ leads to a fairly high incidence on the mutant patterns.
730: But by letting $\theta\geq0.6$ one is assuming that an mRNA or protein is not present in a sufficient
731: amount to perform its function until it reaches at least sixty per cent of its maximal value.
732: Typically such high thresholds are not observed: indeed, in~\cite{dmmo00} (supplementary
733: material) where continuous dynamics are also transformed into an ON/OFF response,
734: a threshold of 10\% is used, and considered very reasonable.
735: In our experiments, unless otherwise indicated, we have used $\theta=0.5$.
736:
737: To further generalize system\rf{eq-glass-a}, it is natural to allow each variable to respond to its
738: own threshold and consider distinct $\theta_i$ values.
739: We will address this problem in Section~\ref{sec-theta-vary}, and see that the system's dynamics is
740: preserved in each $\theta$ region.
741: In fact, our simulations with distinct $\theta_i$ recover many of the theoretical results we obtained
742: for a species-independent $\theta$ (Section~\ref{sec-convergence}).
743:
744:
745: \section{Pre-patterning errors and knockout mutant situations}
746: \label{sec-mutants}
747:
748: In~\cite{ao03,cas05} we identified some sufficient or necessary initial conditions for obtaining the wild
749: type steady state (minimal pre-patterns).
750: We now analyze how mutant patterns arise from gene ``knockout'' experiments or delay in establishing the
751: initial pre-pattern.
752:
753: Gene knockout experiments consist of completely supressing the expression of a given gene in all cells.
754: In our models this is equivalent to setting the corresponding mRNA permanently zero in all equations.
755: Thus a {\it wingless} knockout can be analyzed by setting $\wg_i=0$ for all $i=1,2,3,4$, and
756: in every equation of the Boolean rules in Table~\ref{table_rules}.
757: The steady states of the resulting system can now be computed from:
758: \beqn{eq-knockoutF}
759: X^* = F^*(X^*),
760: \eeqn
761: where the vector $X^*$ and functions $F^*$ include all but the knockout variables.
762: For example, for the {\it wingless} knockout, we drop the functions $F_{\twg_i}$ and for
763: the other mRNAs and proteins we have
764: \beq
765: && \WG_i=F^*_{\tWG_i} = 0, \ \ i=1,2,3,4 \\
766: && \mbox{node }=F^*_{\mbox{\tiny node}} = F_{\mbox{\tiny node}}, \
767: \mbox{ for nodes } \en,\hh,\ci,\ptc,\EN,\HH,\PTC,\CI,\CIA,\CIR,
768: \eeq
769: because $\wg$ appears explicitly only in the rules $F_{\tWG_i}$.
770: So, it follows immediately that $\WG_i=0$, and the remaining nodes are then
771: also easy to compute.
772: %
773: It is easy to check (see also~\cite{ao03}) that knockouts of $\wg$, $\en$, $\hh$ or $\ptc$
774: exhibit only one steady state, while knockouts of $\ci$ exhibit three steady states (summarized
775: in Table~\ref{tab-knockouts}) in both asynchronous and Glass-type models.
776:
777: \begin{table}
778: \caption{The steady states corresponding to gene knockouts
779: in the segment polarity network model, calculated according to\rf{eq-knockoutF}.
780: Here, for the $\ci$ knockout, the ``wild type'' state is interpreted
781: as the wild type pattern in all but the $\ci$, $\CI$, $\CIA$,$\CIR$ mRNA/proteins.}
782: \label{tab-knockouts}
783: \begin{center}
784: \begin{tabular}{cc}
785: \hline
786: Knockouts & Mutant steady states \\
787: \hline\hline
788: $\wg$, $\en$, $hh$ & no segmentation \\
789: \hline
790: $\ptc$ & broad stripes \\
791: \hline
792: & wild type, \\
793: $\ci$ & broad stripes , \\
794: & ectopic \\
795: \hline
796: \end{tabular}
797: \end{center}
798: \end{table}
799:
800: From another point of view, one may consider a delay in the establishment of the pre-pattern
801: (that is, the full initial condition\rf{eq-initial-wt}). If expression of a given gene is delayed,
802: does the system recover, and how soon?
803: To answer this question, we simulated a delay in expression of gene $X$, by
804: setting the corresponding discrete variable $X(t)=0$ in all cells,
805: for all $t\leq T_{\mbox{\tiny delay}}$.
806: We then varied $T_{\mbox{\tiny delay}}$ between 0 and 7 time units, and measured the frequency of
807: occurrence of each steady state, both for the asynchronous and Glass-type models.
808: (In the latter we set the concentration threshold $\theta$ equal to $1/2$ for all nodes.)
809:
810: The results are shown in Figures~\ref{fig-delay-wg-en-hh} and~\ref{fig-delay-ptc-ci}.
811: We can see that, for short $T_{\mbox{\tiny delay}}$, both models recover their original
812: frequencies of occurrence of each steady state, while for long $T_{\mbox{\tiny delay}}$
813: both models converge to the corresponding mutant steady state.
814: A curious exception is the case of $\ci$, where long delays in its initial expression do not significantly
815: change the probability that the wild type steady state is achieved (and even slightly increase it
816: in the asynchronous model). This agrees with the conclusion of~\cite{ao03} that {\it cubitus interruptus} knockout
817: coupled with an otherwise wild-type initial condition converges to a state close
818: to the wild type steady state. Remarkably, delays in both cubitus interruptus transcription factors
819: (CIA, CIR) have a lesser effect than an imbalance in their expression (see~\cite{cas05}). This leads us to
820: predict that, during the pre-segmentation stage of embryo development, the cubitus interruptus proteins'
821: expression is the last to be established.
822:
823: Another noteworthy observation is the fact that small delays in $\wg$ expression have much more drastic
824: effects on the system than the same delay in $\en$ or $\hh$.
825: This phenomenon reflects the one-way signaling cascade starting with expression of $\wg$,
826: which induces $\en$, which in turn induces $\hh$.
827: We see that a total disruption in the system is caused by a delay of only 3 time units in $\wg$
828: or $\en$ expression, which cause the system to fail to reproduce the wild type pattern, and
829: settle into a non-segmented pattern. However, if only $\hh$ is delayed, the system
830: is disrupted only after a delay of 5 time units.
831: In other words, recovery of the system back to the ``good'' developmental process is more probable in
832: the event of a $\hh$ expression delay, than a $\wg$ or $\en$ expression delay.
833:
834:
835: \section{Robustness of the model under timescale separation}
836: \label{sec-timesep}
837:
838: In the previous algorithms, the space of all possible timescales for protein/mRNA regulatory
839: processes was explored, with no assumptions on the characteristic duration of translational
840: or post-translational processes.
841: As a consequence, the robustness analysis shows that the model diverges from the wild type
842: pattern very often, with the biologically inviable states occurring with a noticeable
843: frequency.
844: However, it is also well known that post-translational processes such as protein
845: conformational changes or complex formation, usually have shorter durations than transcription,
846: translation or mRNA decay.
847: This fact justifies the introduction of a distinct timescale separation among processes,
848: by choosing to update proteins first and mRNAs later.
849:
850: \subsection{Timescale separation in the random order algorithm}
851: Timescale separation is straightforwardly implemented in the random order algorithm
852: presented in Section~\ref{sec-asynch}; at the $k$-th updating step we generate
853: two random permutations, $\phi^k_{\tProt}$ and $\phi^k_{\tmRNA}$,
854: within the set of proteins and mRNAs, respectively. Then the $N$ nodes are updated in the order
855: given by
856: \beq
857: \phi^k=(\phi^k_{\tProt},\;\phi^k_{\tmRNA}).
858: \eeq
859: {This method again shows that the Boolean model is very robust, in the sense that
860: when started from the wild type initial condition, the wild type
861: pattern occurs with a frequency of $87.5\%$ and only one other steady state is observed, the
862: broad striped pattern, with a frequency of $12.5\%$.}
863: Furthermore, these frequencies are exact as we show in~\cite{cas05}, where we also completely
864: characterize the model resulting from incorporation of a protein/mRNA timescale separation
865: into the random order algorithm.
866: We show that the wild type state is in fact an attractor for the system,
867: while the pathway to the broad stripes state may exhibit oscillatory cycles.
868: We summarize this and other results in the next theorem, stated without proof, and refer
869: to~\cite{cas05} for more details.
870:
871: \begin{theo}{th-random}
872: In the random order algorithm with timescale separation,
873: let $\wg_3^0=0$, $\ptc_3^0=1$, $\hh_{2,4}^0=0$ and $\ci_3^0=1$
874: (as satisfied by the initial pattern\rf{eq-initial-wt}).
875: Then system diverges from the wild type pattern if and only if the
876: permutation $\phi^1$ satisfies the following sequence among the proteins
877: $\CI$, $\CIA$, $\CIR$ and $\PTC$:
878: \beqn{eq-perm-wg3=1}
879: \begin{array}{cccccc}
880: \CIR_3 & \CI_3 & & \CIA_3 & & \PTC_3, \\
881: % & & & & & \\
882: & \CI_3 & \CIR_3 & \CIA_3 & & \PTC_3, \\
883: % & & & & & \\
884: & \CI_3 & & \CIA_3 & \CIR_3 & \PTC_3,
885: \end{array}
886: \eeqn
887: while the other proteins may appear in any of the remaining slots.
888: \bbox
889: \end{theo}
890:
891: Thus we can compute the exact probability with which the random order algorithm (with
892: timescale separation) leads to either the wild type or broad stripes pattern:
893: the latter is simply the fraction of sequences of the form\rf{eq-perm-wg3=1},
894: and equals $12.5\%$~\cite{cas05}.
895:
896: \subsection{Timescale separation in the Glass-type and asynchronous algorithms}
897: As shown in Section~\ref{sec-glass}, the (discrete) asynchronous algorithm and the (piecewiese continuous)
898: Glass-type system provide equivalent representations of a gene expression network. Indeed,
899: the ``specific time units'' $\gamma_i$ and the inverse ``scaling factors'' $\alpha_i^{-1}$,
900: both represent the rate of dynamical evolution of each individual node.
901: For these two models, we implement time separation among processes by using two
902: non-overlapping intervals for the scaling factors:
903: \beq
904: && \gamma_i^{-1},\alpha_i \in A_{\tmRNA},\ \ \ \mbox{if }\ X_i\in\{\wg,\en,\hh,\ptc,\ci \} \\
905: && \gamma_i^{-1},\alpha_i \in A_{\tProt},\ \ \ \mbox{if }\ X_i\in\{\WG,\EN,\HH,\PTC,\CI,\CIA,\CIR\},
906: \eeq
907: with, for instance, $A_{\tmRNA}=[0.2,0.6]$ and $A_{\tProt}=[1.4,1.8]$.
908: Under these conditions, choosing the factors $\alpha_i$ from a uniform distribution in these
909: intervals, numerical experiments indicate that the two methods respond in mostly similar ways,
910: with only two patterns occurring at steady state when the systems
911: start from (wild type) initial condition\rf{eq-initial-wt}.
912: The two possible steady states are the wild type and broad stripes patterns.
913:
914: For the asynchronous algorithm, the probabilities of convergence to
915: each of the steady states clearly depend on the {\it distance} between the two intervals:
916: convergence to wild type is $93\%$ in the case where intervals $A_{\tmRNA}$, $A_{\tProt}$
917: are consecutive, and up to $100\%$ for the case $2a<b$, $a\in A_{\tmRNA}$, $b\in A_{\tProt}$.
918:
919: For the Glass-type model, two cases can be distinguished: $\theta\leq0.5$ and $\theta>0.5$.
920: For $\theta\leq0.5$, numerical simulations show that the
921: model reaches wild type pattern with probability near 100\%, even when there
922: is some overlap between $A_{\tmRNA}$ and $A_{\tProt}$ (see Fig.~\ref{fig-time-separation}).
923: In fact, we next theoretically prove that {\em the wild type pattern is indeed
924: the unique possible steady state} of the hybrid system\rf{eq-glass-a}
925: and initial condition\rf{eq-initial-wt}, as indicated by
926: the simulations, when there is a suitable distance between the intervals,
927: and a lower bound on $\theta$ (Theorem~\ref{th-glass1}, Section~\ref{sec-convergence}).
928: For $\theta>0.5$, we have found no condition that guarantees convergence to the
929: wild type steady state, and indeed numerical simulations show that, even for large interval
930: separation, the system may converge to one of the mutant patterns.
931:
932: A comparison of Theorems~\ref{th-random} and~\ref{th-glass1}
933: emphasizes differences and similarities between discrete and continuous models:
934: intuitively, the single discrete event described by Theorem~\ref{th-random} cannot take
935: place in a continuous model. Therefore $\wg_3$ remains ``0'' (OFF) for all times, ruling out
936: the possibility that the broad stripes pattern is reached.
937: Indeed, Theorem~\ref{th-random} establishes that (in the discrete
938: case, with the random order algorithm)
939: divergence from wild type pattern occurs if and only if $\wg_3^1=1$.
940: This fact involves a jump in $\wg_3$ from ``0'' to ``1'' at precisely
941: the first iteration.
942: On the other hand, in the Glass-type model, the continuous variable $\cwg_3$ cannot
943: instantaneously jump from ``0'' to ``1''.
944: Since the discrete ON/OFF levels are defined by a threshold on $\cwg_3$, there will
945: necessarily be a smoothing effect on any transition between ``0'' and ``1''.
946: This is what happens in case (a) of Theorem~\ref{th-glass1}.
947:
948: The second (sufficient) condition of Theorem~\ref{th-glass1} guarantees convergence to
949: the wild type steady state for all $0<\theta\leq0.5$ (while condition (a) is only for
950: $0.382\leq\theta\leq0.5$), but assumes that $\alpha_{\tPTC_3}>\alpha_{\tCI_3}$.
951: This is an analog to Theorem~\ref{th-random}: if $\alpha_{\tPTC_3}>\alpha_{\tCI_3}$,
952: then (starting from $\PTC_3(0)=\CI_3(0)=0$ and assuming $F_{\tPTC_3}=F_{\tCI_3}=1$)
953: $\cPTC_3$ increases faster than $\cCI_3$, implying that $\PTC_3$ becomes ON faster
954: than $\cCI_3$. Such response prevents the events listed in Theorem~\ref{th-random},
955: which would lead to a mutant state.
956: Thus, both discrete and piecewise linear model predict that the sequence of $\PTC$,
957: $\CI$ expression in the third cell is one of the fundamental pieces in establishing
958: the correct development of embryo segmentation.
959:
960: It is also worth pointing out that the three methods provide qualitatively similar
961: results under timescales separation, all predicting that only wild type and broad stripes
962: patterns to occur, the latter with considerably smaller frequency.
963: \begin{table}
964: \caption{Probabilities of convergence to a given steady state,
965: under the separation of timescales assumption.
966: These values are theoretically exact for both the random order algorithm and Glass-type model.}
967: \label{tab-two-timescales}
968: \begin{center}
969: \begin{tabular}{lccc}
970: \hline
971: Steady state & Random order & Asynchronous & Glass-type\\
972: pattern & algorithm & algorithm & model\\
973: \hline\hline
974: wild type & 87.5\% & 93.7-100\% & 100\% \\
975: %\hline
976: broad stripes & 12.5\% & 0-6.3\% & 0\% \\
977: %\hline
978: other & 0\% & 0\% & 0\% \\
979: \hline
980: \end{tabular}
981: \end{center}
982: \end{table}
983:
984: \subsection{Distinct concentration thresholds for ON state}
985: \label{sec-theta-vary}
986: A natural question arising in the analysis of equation\rf{eq-glass-a} concerns
987: the dynamics of the system under more general concentration thresholds.
988: We have seen that $\theta$ (even when equal for all species) plays an important role
989: in establishing basins of attraction to each of the steady states of the model.
990:
991: We have, in particular, identified three distinct regions of behavior:
992: \beq
993: \begin{array}{lc}
994: \mbox{Region 1}: & 0<\theta<\frac{3-\sqrt{5}}{2}\\
995: & \\
996: \mbox{Region 2}: & \frac{3-\sqrt{5}}{2}\leq\theta\leq\frac{1}{2}\\
997: & \\
998: \mbox{Region 3}: & \frac{1}{2}<\theta\leq1
999: \end{array}
1000: \eeq
1001: When $\theta$ is equal for all nodes, the probability that the system evolves into
1002: the wild type pattern is above 90\% in regions 1 and 2, but may be as low as 68\%
1003: in region 3.
1004: Furthermore, for region 2, we theoretically prove that that probability is exactly 100\%
1005: under the timescale separation assumption.
1006:
1007: We now associate to each node a specific $\theta_i$, so that\rf{eq-threshold} is modified to:
1008: \beqn{eq-threshold2}
1009: X_i(t)=\left\{ \begin{array}{ll}
1010: 0, & \cX_i(t)\leq\theta_i \\
1011: 1, & \cX_i(t)>\theta_i\ .
1012: \end{array}
1013: \right.
1014: \eeqn
1015: To test the performance of the system and compare it to previous results,
1016: we considered two timescale situations: $\alpha_i\in[0.5,1.5]$ for all $i$,
1017: or the timescale separation $A_{\tmRNA}=[0.2,0.6]$, $A_{\tProt}=[1.4,1.8]$.
1018: In each case, we randomly assigned values to $\theta_i$ from
1019: uniform distributions in the intervals $(0,1)$, $(0,0.5)$ and $(0.4,0.5)$.
1020: (Note that $0.4$ is close to $(3-\sqrt{5})/2$.)
1021: The numerical results with varying $\theta_i$ extend and confirm our previous observations
1022: for $\theta_i=\theta$, $i=1,\ldots,N$.
1023:
1024: Table~\ref{tab-theta-vary} summarizes the results for all combinations of $\theta_i$ and
1025: $\alpha_i$ regions.
1026: The most general case, allowing a large degree of freedom in both timescales and concentration
1027: thresholds, indicates the vulnerability of the network, with a very high incidence on mutant
1028: patterns ($\theta_i\in(0,1)$, $\alpha_i\in[0.5,1.5]$).
1029: Again we see that there is a marked difference in the $\theta_i$ regions below or above $0.5$.
1030: Comparing the reasonable results obtained for $\theta_i\leq0.5$ with the bad performance for
1031: $0.5<\theta_i<1$, we conclude that the optimal ON concentratin for proteins or mRNA is below 50\%
1032: of maximal concentration.
1033:
1034: Restricting $\theta_i$ even further to one of the conditions given in Theorem~\ref{th-glass1}
1035: (Section~\ref{sec-convergence}; conditions developed for the case $\theta_i=\theta$, $i=1,\ldots,N$)
1036: dramatically increases the probability that the system develops in the correct way.
1037:
1038: These simulations suggest, moreover, that some of the theoretical results obtained for each of the
1039: three regions may extend to the case of distinct $\theta_i$.
1040: Indeed, note that, under timescale separation,
1041: when $\theta_i$ are chosen from region 2, convergence to wild type steady state is 100\%
1042: -- compare to part (a) of Theorem~\ref{th-glass1}.
1043:
1044: These simulations further confirm the role of $\alpha_{\tPTC_3}$ and $\alpha_{\tCI_3}$
1045: in the segmentation network.
1046: Requiring $\alpha_{\tPTC_3}>\alpha_{\tCI_3}$ decreases the probability of formation of the broad stripes
1047: pattern (any timescales), but doesn't influence the probability of a non segmented embryo. The latter
1048: mutant is prevented only by a complete separation of timescales of the regulatory processes.
1049:
1050:
1051: \begin{table}
1052: \caption{Probabilities of convergence to a given steady state,
1053: with distinct concentration thresholds $\theta_i$ and distinct timescales $\alpha_i$,
1054: in the Glass-type model. (Probabilities computed out of 1000 simulations for each case.)}
1055: \label{tab-theta-vary}
1056: \begin{center}
1057: \begin{tabular}{lcccccc|r}
1058: \hline
1059: Steady state & $(0,1)$ & $(0.5,1)$ & $(0,0.5]$ & $(0,0.5]$ & $[0.4,0.5]$ & $0.5$ & \multicolumn{1}{l}{$\theta_i$} \\
1060: pattern & & & & $\alpha_{\tPTC_3}>\alpha_{\tCI_3}$ & & & $\alpha_i$ \\
1061: \hline\hline
1062: wild type & 45.6\% & 57.1\% & 84.1\% & 90\% & 92.6\% & 94.2\% & $A_{\tmRNA}=$ \\
1063: %\hline
1064: broad stripes & 27.8\% & 15.1\% & 12\% & 6.2\% & 7.3\% & 4.5\% & $A_{\tProt}=$ \\
1065: %\hline
1066: no segmentation & 24.4\% & 25.8\% & 0.9\% & 0.9\% & 0.05\% & 1.3\% & $[0.5,1.5]$\\
1067: %\hline
1068: wild type, variant & 2.1\% & 1.9\% & 2.9\% & 2.9\% & 0\% & 0\% & \\
1069: \hline\hline
1070: wild type & 74.1\% & 52.7\% & 96.6\% & 97.1\% & 100\% & 100\% & $A_{\tmRNA}=$ \\
1071: %\hline
1072: broad stripes & 10.8\% & 3.3\% & 3.3\% & 2.8\% & 0\% & 0\% & $[0.2,0.6]$\\
1073: %\hline
1074: no segmentation & 14.1\% & 43.9\% & 0\% & 0\% & 0\% & 0\% & $A_{\tProt}=$ \\
1075: %\hline
1076: wild type, variant & 1.0\% & 0\% & 0.1\% & 0.1\% & 0\% & 0\% & $[1.4,1.8]$\\
1077: \hline
1078: \end{tabular}
1079: \end{center}
1080: \end{table}
1081:
1082:
1083: \subsection{Glass-type model provides exact convergence to wild type pattern}
1084: \label{sec-convergence}
1085: In this section we will require that the intervals $A_{\tmRNA}$ and $A_{\tProt}$
1086: do not overlap, by satisfying the following assumption:
1087: \beqn{eq-separation}
1088: \mbox{For all }\ \ a\in A_{\tmRNA}\ \ \mbox{and}\ \ b\in A_{\tProt}:\ \ \ 0<2a<b.
1089: \eeqn
1090: A second assumption is that the effective maximal concentration is equal for all nodes and
1091: satisfies $\theta\leq1/2$, which is equivalent to:
1092: \beqn{eq-lowthres}
1093: \ltho\leq\lth
1094: \eeqn
1095:
1096: \begin{theo}{th-glass1}
1097: Consider system\rf{eq-glass-a} with initial condition\rf{eq-initial-wt}.
1098: Assume that the scaling factors $\alpha_i$ satisfy\rf{eq-separation}.
1099: Assume also that one of the following conditions holds:
1100: \bit
1101: \item[(a)]
1102: $\theta\leq1/2$ and $(1-\theta)^2\leq\theta$
1103: or equivalently $0.382\approx(3-\sqrt{5})/2\leq\theta\leq1/2$;
1104:
1105: \item[(b)]
1106: $\theta\leq1/2$ and $\alpha_{\tPTC_3}>\alpha_{\tCI_3}$;
1107: \eit
1108: then $\wg_3(t)=0$ for all $t$.
1109: \end{theo}
1110: This shows that the steady state representing the broad stripes pattern
1111: cannot ever be reached in system\rf{eq-glass-a} from the initial
1112: condition\rf{eq-initial-wt}, when $\theta\leq1/2$ and either of the extra
1113: conditions holds.
1114:
1115:
1116: \begin{theo}{th-glass2}
1117: Consider system\rf{eq-glass-a} with initial condition\rf{eq-initial-wt}.
1118: Assume that the scaling factors $\alpha_i$ satisfy\rf{eq-separation}, and
1119: that $\theta\leq1/2$.
1120: Then $\wg_4(t)=1$ and $\PTC_1(t)=0$ for all $t$.
1121: \end{theo}
1122: This shows that the steady states represented by the no segmentation, wild type variant or the two ectopic patterns
1123: also cannot ever be reached in system\rf{eq-glass-a} from the initial condition\rf{eq-initial-wt}.
1124: From Theorems~\ref{th-glass1} and~\ref{th-glass2}
1125: we conclude that, under the timescale separation assumption, the
1126: Glass-type model\rf{eq-glass-a} can only converge to the wild type pattern,
1127: when starting from the initial condition\rf{eq-initial-wt}, and for appropriate $\theta$ values
1128: (Table~\ref{tab-two-timescales}).
1129:
1130:
1131:
1132: We first summarize some useful observations.
1133: Let $X$ denote any of the nodes in the network, and $\alpha$ its time rate.
1134: Since equations\rf{eq-glass-a} are either of the form $d\cX/dt=\alpha(-\cX+1)$
1135: or $d\cX/dt=-\alpha\cX$, their solutions are continous functions,
1136: piecewise combinations of:
1137: \beqn{eq-hybrid-sol1}
1138: \cX^1(t) &=& 1 - (1-\cX^1(t_0))\,e^{-\alpha(t-t_0)} \\
1139: \cX^0(t) &=& \cX^0(t_0)\,e^{-\alpha(t-t_0)} \label{eq-hybrid-sol0}
1140: \eeqn
1141: $\cX^1(t)$ (resp. $\cX^0(t)$) is monotonically increasing (resp. decreasing).
1142: In addition, note that discrete variables $X$ can only switch between 0 and 1
1143: at those instants when $\cX(t_{\mbox{\tiny switch}})=\theta$, that is:
1144: \beqn{eq-hybrid-switch1}
1145: t_{\mbox{\tiny switch}}^{1} &=& t_0 +\frac{1}{\alpha} \,\lthow{(1-\cX(t_0))}\\
1146: t_{\mbox{\tiny switch}}^{0} &=& t_0 +\frac{1}{\alpha} \,\lthw{\cX(t_0)}
1147: \label{eq-hybrid-switch0}
1148: \eeqn
1149:
1150: From the initial conditions, together with the constant values of $\SLP_i$ ($i=1,2,3,4$), we
1151: can immediately conclude:
1152: \beqn{eq-wgen}
1153: && \cwg_{1,2}(t)=\WG_{1,2}(t)=0,\\
1154: && \cen_{3,4}(t)=\cEN_{3,4}(t)=0, \nonumber \\
1155: && \chh_{3,4}(t)=\cHH_{3,4}(t)=0, \label{eq-wgen2}
1156: \eeqn
1157: for all $t\geq0$.
1158: Then, because $\ci_{3,4}(0)=1$ and $F_{\tci_{3,4}}=\mbox{not }\EN_{3,4}$,
1159: \beqn{eq-ciCI}
1160: \cci_{3,4}(t)=1\ \mbox{ and }\ \cCI_{3,4}(t)=1-e^{-\alpha_{\tCI_{3,4}}\, t}.
1161: \eeqn
1162:
1163:
1164: \begin{lem}{lm-wg3}
1165: Let $0\leq t_0<t_3\leq t_1$ and $0\leq t_2<t_3$.
1166: Define $\delta=\ltho/\max_{1,\ldots,N}{\alpha_i}$.
1167: Assume $\CIA_3(t)=0$ for $t\in(t_2,t_3)$, and $\wg_3(t)=0$ for $t\in[0,t_3)$.
1168: Then
1169: \bit
1170: \item[(a)] $\wg_3(t)=0$ for $t\in[0,t_3+\delta)$;
1171: \item[(b)] $\WG_3(t)=0$ for $t\in[0,t_3+\delta)$;
1172: \item[(c)] $\en_{2}(t)=\EN_{2}(t)=0$ for $t\in[0,t_3+\delta)$;
1173: \item[(d)] $\hh_{2}(t)=\HH_{2}(t)=0$ for $t\in[0,t_3+\delta)$.
1174: \eit
1175: Assume further that $\PTC_3(t)=1$ for $t\in(t_0,t_1)$. Then
1176: \bit
1177: \item[(e)] $\PTC_3(t)=1$ for all $t\in(t_0,t_3+\delta)$.
1178: \item[(f)] $\CIA_3(t)=0$ for all $t\in(t_2,t_3+\delta)$.
1179: \eit
1180: \end{lem}
1181:
1182: \bprf
1183: Part (a) follows directly from the fact that $F_{\twg_3}(t)=0$ on $[0,t_3)$, and
1184: from\rf{eq-hybrid-switch1}.
1185:
1186: To prove parts (b), (c), and (d), first note that initial conditions together with
1187: $\wg_3(t)=0$ for $t\in[0,t_3)$ imply
1188: \beq
1189: && \cWG_3(t)=0,\ \cen_{2}(t)=\cEN_{2}(t)=0,\\
1190: && \chh_{2}(t)=\cHH_{2}(t)=0,
1191: \eeq
1192: for $t\in[0,t_3]$. Then, from equations\rf{eq-hybrid-sol1} to\rf{eq-hybrid-switch0}
1193: we conclude that the corresponding discrete variables cannot switch from 0 to 1 during
1194: an interval of the form $[0,t_3+\frac{1}{\alpha_j}\ltho)$. Taking the largest common interval
1195: yields the desired results.
1196:
1197: To prove parts (e) and (f), assume also that $\PTC_3(t)=1$ for $t\in(t_0,t_1)$.
1198: From\rf{eq-wgen2} and part (d), it follows that function $F_{\tPTC_3}$ does not switch
1199: in the interval $(t_0,t_3+\delta)$ and in fact $\PTC_3(t)=1$ for all $t$ in this interval.
1200: This, together with\rf{eq-wgen2} and part (d) yield $F_{\tCIA_3}(t)=0$ for $(t_0,t_3+\delta)$,
1201: so that $\cCIA_3$ cannot increase in this interval and the discrete level satisfies
1202: $\CIA_3(t)=0$ for all $t\in(t_2,t_3+\delta)$, as we wanted to show.
1203: \bbox
1204:
1205:
1206: \begin{cor}{cr-wg3}
1207: Let $0\leq t_0<t_3\leq t_1$ and $0\leq t_2<t_3$.
1208: If $\PTC_3(t)=1$ for $t\in(t_0,t_1)$,
1209: $\CIA_3(t)=0$ for $t\in(t_2,t_3)$, and $\wg_3(t)=0$ for $t\in[0,t_3)$,
1210: then $\wg_3(t)=0$ for all $t$.
1211: \end{cor}
1212:
1213: \bprf
1214: Applying Lemma~\ref{lm-wg3} we conclude that, given any $k\geq0$:
1215: \beq
1216: && \CIA_3(t)=0,\ \mbox{ for }\ t\in(t_2,t_3+k\delta)\\
1217: && \wg_3(t)=0,\ \mbox{ for }\ t\in[0,t_3+k\delta)\\
1218: && \PTC_3(t)=1\ \mbox{ for }\ t\in(t_0,t_3+k\delta)
1219: \eeq
1220: imply
1221: \beq
1222: && \CIA_3(t)=0,\ \mbox{ for }\ t\in(t_2,t_3+(k+1)\delta)\\
1223: && \wg_3(t)=0,\ \mbox{ for }\ t\in[0,t_3+(k+1)\delta)\\
1224: && \PTC_3(t)=1\ \mbox{ for }\ t\in(t_0,t_3+(k+1)\delta).
1225: \eeq
1226: Since $\delta$ is finite, we conclude by induction on $k$
1227: that $\wg_3(t)=0$ for all $t$.
1228: \bbox
1229:
1230:
1231: {\it Proof of Theorem~\ref{th-glass1}:}
1232: The rule for $\CIA_3$ may be simplified to (by\rf{eq-wgen2})
1233: \beq
1234: F_{\tCIA_3}=\CI_3\mbox{ and }[\mbox{not}\PTC_3\mbox{ or }\hh_2 \mbox{ or }\HH_2].
1235: \eeq
1236: From equation\rf{eq-ciCI}, we have that
1237: \beqn{eq-CI3}
1238: \CI_3(t)=1,\ \mbox{ for all } t>\frac{1}{\alpha_{\tCI_3}}\ltho.
1239: \eeqn
1240: On the other hand, since $\ptc_3(0)=1$, by continuity of solutions
1241: $\ptc_3(t)=1$ for all $t<\frac{1}{\alpha_{\tptc_3}}\lth$.
1242: This implies that the Patched protein satisfies
1243: \beq
1244: \cPTC_3(t)=1-e^{-\alpha_{\tPTC_{3}}\, t},\ \ 0\leq t\leq\frac{1}{\alpha_{\tptc_3}}\lth
1245: \eeq
1246: and therefore
1247: \beqn{eq-PTC3}
1248: \PTC_3(t)=\left\{
1249: \begin{array}{ll}
1250: 0, & 0\leq t\leq\frac{1}{\alpha_{\tPTC_3}}\ltho\\
1251: 1, & \frac{1}{\alpha_{\tPTC_3}}\ltho<t<\frac{1}{\alpha_{\tptc_3}}\lth.
1252: \end{array}
1253: \right.
1254: \eeqn
1255: By assumption, $\alpha_{\tPTC_3}>\alpha_{\tptc_3}$ and also $\ltho\leq\lth$,
1256: defining a nonempty interval where $\PTC_3$ is expressed.
1257: Now let $t_c=\frac{1}{\alpha_{\tCI_3}}\ltho$ and $t_p=\frac{1}{\alpha_{\tPTC_3}}\ltho$.
1258: %
1259: $\cCIA_3(t)$ starts at zero and must remain so while $\CI_3=0$, so that
1260: \beq
1261: \CIA_3(t)=0\ \mbox{ for } 0<t<t_c.
1262: \eeq
1263: In the case $t_c>t_p$,
1264: letting $t_0=t_p$, $t_1=\frac{1}{\alpha_{\tptc_3}}\lth$, $t_2=0$, and $t_3=t_c$
1265: in Corollary~\ref{cr-wg3}, obtains $\wg_3(t)=0$ for all $t$.
1266: %
1267: This proves item (b) of the theorem, and part of (a).
1268:
1269: To finish the proof of item (a), we assume that $(1-\theta)^2<\theta$ and must now consider the
1270: case $t_c\leq t_p$.
1271: Then
1272: \beq
1273: \cCIA_3(t)=\left\{
1274: \begin{array}{ll}
1275: 0, & 0\leq t\leq t_c \\
1276: 1-e^{-\alpha_{\tCIA_3}\,(t-t_c)}, & t_c < t \leq t_p \\
1277: \cCIA_3(t_p)\,e^{-\alpha_{\tCIA_3}\,(t-t_p)}, &
1278: t_p < t\leq \frac{1}{\alpha_{\tptc_3}}\lth,
1279: \end{array}\right.
1280: \eeq
1281: Following equation\rf{eq-hybrid-switch1} with $t_0=t_c$ and $\cCIA_3(t_0)=0$,
1282: $\CIA_3$ might become expressed at time $t_c<t_a<t_p$:
1283: \beq
1284: t_a=t_c+\frac{1}{\alpha_{\tCIA_3}}\ltho,
1285: \eeq
1286: but it would then become zero again at (equation\rf{eq-hybrid-switch0} with
1287: $t_0=t_p$)
1288: \beq
1289: t_b=t_p + \frac{1}{\alpha_{\tCIA_3}}\lthw{\cCIA_3(t_p)}.
1290: \eeq
1291: Finally, we show that, even if $\CIA_3(t)=1$ for $t\in(t_a,t_b)$, $\wg_3$ cannot become
1292: expressed in this interval. In this interval, $\cwg_3$ evolves according to
1293: $
1294: \cwg_3(t)=1-e^{-\alpha_{\twg_3}\,(t-t_a)},
1295: $
1296: and $\wg_3$ can switch to 1 at time
1297: \beq
1298: t_w=t_a+\frac{1}{\alpha_{\twg_3}}\ltho.
1299: \eeq
1300: We will show that $t_w>t_b$, so $\wg_3(t)=0$ in the interval $[0,t_b)$.
1301: Writing
1302: \beq
1303: \lthw{\cCIA_3(t_p)}&=&\lthow{\cCIA_3(t_p)}\frac{1-\theta}{\theta} \\
1304: &=&\lthow{\cCIA_3(t_p)} +\ln\, \frac{1-\theta}{\theta} \\
1305: &\leq&\ltho +\ltho
1306: \eeq
1307: where we have used $\cCIA_3(t_p)\leq1$ and the assumption on
1308: $\theta$: $\frac{1-\theta}{\theta}\leq\frac{1}{1-\theta}$.
1309: Therefore
1310: \beq
1311: t_b & \leq & t_p +\frac{2}{\alpha_{\tCIA_3}}\ltho \\
1312: & < & \frac{1}{\alpha_{\twg_3}}\ltho+\frac{1}{\alpha_{\tCIA_3}}\ltho<t_w
1313: \eeq
1314: where we have used the timescale separation assumption\rf{eq-separation}.
1315: Letting $t_0=t_p$, $t_2=0$, and $t_1=t_3=\min\{t_b,\alpha_{\tptc_3}^{-1}\lth\}$ in the Corollary,
1316: obtains $\wg_3(t)=0$ for all $t$.
1317: \bbox
1318:
1319: We will next show that if $\wg_4(t)=1$ in a given interval $[0,T)$,
1320: then in fact $\wg_4(t)$ remains expressed for a longer time, up to $T+\delta$,
1321: with $\delta>0$.
1322: This is mainly due to assumption\rf{eq-separation}, which says that mRNAs take longer
1323: than proteins to update their discrete values, because they have longer half-lives:
1324: $\alpha_{\tmRNA}^{-1}>\alpha_{\tProt}^{-1}$.
1325: This allows the initial signal ``$\wg_{4}=1$'' to travel down the network, sequencially
1326: affecting the wingless protein, {\it engrailed, hedgehog} and CIA, and feed back into
1327: {\it wingless}
1328: allowing $\wg_4$ to remain expressed for a further time interval.
1329:
1330: \begin{lem}{lm-wg4}
1331: Let $T\geq \frac{1}{\alpha_{\twg_4}}\lth$ and define
1332: \beqn{eq-delta}
1333: \delta=\frac{1}{\alpha_{\tWG_4}}\,\lthw{ (1-e^{-\frac{\alpha_{\tWG_4}}{\alpha_{\twg_4}}\lth})}.
1334: \eeqn
1335: If $\wg_4(t)=1$ for $0\leq t< T$, then
1336: \bit
1337: \item[(a)] $\WG_4(t)=1$ for $t\in(\frac{1}{\alpha_{\tWG_4}}\ltho,T+\delta)$;
1338: \item[(b)] $\en_1(t)=1$ for $t\in[0,T+\delta)$;
1339: \item[(c)] $\cEN_1(t)=1-e^{-\alpha_{\tEN_1}t}$ for $t\in[0,T+\delta)$,
1340: and $\EN_1(t)=1$ for $(\frac{1}{\alpha_{\tEN_1}}\ltho,T+\delta)$;
1341: \item[(d)] $\ci_1(t)=0$, $\CI_1(t)=0$, $\CIA_1(t)=0$, and $\CIR_1(t)=0$ for $t\in[0,T+\delta)$;
1342: \item[(e)] $\hh_1(t)=1$, for $t\in[0,T+\delta)$;
1343: \item[(f)] $\CIA_4(t)=1$, for $t\in(\frac{1}{\alpha_{\tCI_4}}\ltho
1344: +\frac{1}{\alpha_{\tCIA_4}}\ltho,T+\delta)$,
1345: and $\CIR_4(t)=0$, for $t\in[0,T+\delta)$;
1346: \item[(g)] $\wg_4(t)=1$ for $t\in[0,T+\delta)$.
1347: \eit
1348: \end{lem}
1349:
1350: \bprf
1351: Let $T\geq\frac{1}{\alpha_{\twg_4}}\lth$, and assume that $\wg_4(t)=1$ for $0\leq t< T$.
1352: To prove part (a), note that $\cWG_4(t)$ is of the form\rf{eq-hybrid-sol1}
1353: (with $t_0=0$, and $\cWG_4(0)=0$) and the corresponding discrete variable is
1354: $\WG_4(t)=1$, for $t\in(\frac{1}{\alpha_{\tWG_4}}\ltho,T)$.
1355: Moreover, suppose that $\wg_4(t)=0$ for $t>T$, then
1356: \beq
1357: \cWG_4(t)=(1-e^{-\alpha_{\tWG_4}T})e^{-\alpha_{\tWG_4}(t-T)},\ \ \ t>T.
1358: \eeq
1359: But $\WG_4$ remains 1 until the switching threshold is attained, that is up to time
1360: \beq
1361: & & T+\frac{1}{\alpha_{\tWG_4}}\lthw{ (1-e^{-\alpha_{\tWG_4}T})}\\
1362: &\geq &
1363: T+\frac{1}{\alpha_{\tWG_4}}\lthw{ (1-e^{-\alpha_{\tWG_4}\frac{1}{\alpha_{\twg_4}}\lth})} \\
1364: &\equiv & T+\delta.
1365: \eeq
1366: Thus we conclude that $\WG_4(t)=1$ in the desired interval.
1367:
1368: To prove part (b), observe that $F_{\ten_1}(t)=\WG_4(t)$ for all $t$, from\rf{eq-wgen},
1369: and recall that $\en_1(0)=1$. From part (a), $F_{\ten_1}(t)=1$ for
1370: $t\in(\frac{1}{\alpha_{\tWG_4}}\ltho,T+\delta)$.
1371: On the other hand, $\en_1$ can only switch from 1 to 0 at
1372: $t=\alpha_{\ten_1}^{-1}\lth$ which is larger than $\alpha_{\tWG_4}^{-1}\ltho$. So, in fact,
1373: $\en_1(t)=1$ for all $0\leq t< T+\delta$.
1374:
1375: Part (c) follows immediately by integration of the $\cEN_1$ equation.
1376:
1377: To prove part (d), first recall $F_{\tci_1}=$ not $\EN_1$ and the initial conditions
1378: $\ci_1(0)=0=\CI_1(0)=\CIA_1(0)=\CIR_1(0)$.
1379: Therefore $\cci_1(t)$ increases up to $t=\frac{1}{\alpha_{\tEN_1}}\ltho$ and then decreases in
1380: $\alpha_{\tEN_1}^{-1}\ltho<t<T+\delta$. Now note that the discrete variable $\ci_1(t)$ remains 0
1381: in the whole interval $[0,T+\delta)$. This is because $\cci_1$ never reaches the $\theta$ threshold:
1382: this would be attained at some $t\geq\alpha_{\tci_1}^{-1}\ltho$ but, since
1383: $\alpha_{\tci_1}^{-1}\ltho>\alpha_{\tEN_1}^{-1}\ltho$, the function $\cci_1$ starts decreasing
1384: before it could reach the value $\theta$.
1385: Finally, from the rules of the Cubitus proteins it is immediate to see that
1386: $\CI_1(t)=\CIA_1(t)=\CIR_1(t)=0$ for $t\in[0,T+\delta)$.
1387:
1388: To prove part (e), recall that $F_{\thh_1}=\EN_1$ and not $\CIR_1$. From part (a), it follows
1389: that $F_{\thh_1}(t)=0$ in the interval $[0,\alpha_{\tEN_1}^{-1}\ltho)$ and
1390: $F_{\thh_1}(t)=1$ in the interval $(\alpha_{\tEN_1}^{-1}\ltho,T+\delta)$. Since $\hh_1(0)=1$, $\chh_1(t)$
1391: decreases in the interval $[0,\alpha_{\tEN_1}^{-1}\ltho)$ but increases in
1392: $(\alpha_{\tEN_1}^{-1}\ltho,T+\delta)$.
1393: The discrete value is $\hh_1(t)=1$ in the whole interval, since $\chh_1(t)$ remains above the
1394: $\theta$ threshold. (The justification is similar to the case of $\ci_1(t)$ in part (d).)
1395:
1396: To prove part (f), note that part (e) and then the use of\rf{eq-ciCI}, allows us to simplify
1397: $F_{\tCIA_4}$:
1398: \beq
1399: F_{\tCIA_4}(t)=\CI_4(t)\ \mbox{ and }\ \hh_1(t) = 1,\ \ \ t\in(\frac{1}{\alpha_{\tCI_4}}\ltho,T+\delta).
1400: \eeq
1401: Thus
1402: \beq
1403: \cCIA_4(t)=\left\{
1404: \begin{array}{ll}
1405: 0, & 0\leq t\leq \frac{1}{\alpha_{\tCI_4}}\ltho \\
1406: 1-e^{-\alpha_{\tCIA_4}\left(t-\frac{1}{\alpha_{\tCI_4}}\ltho\right)}, &
1407: \frac{1}{\alpha_{\tCI_4}}\ltho<t\leq T+\delta,
1408: \end{array}
1409: \right.
1410: \eeq
1411: and $\CIA_4(t)=1$ for $t\in[\frac{1}{\alpha_{\tCI_4}}\ltho+\frac{1}{\alpha_{\tCIA_4}}\ltho,T+\delta)$.
1412: Observe that this interval is indeed nonempty, by assumption\rf{eq-separation}.
1413: Finally, $F_{\tCIR_4}(t)=\CI_4(t)\mbox{ and not}\hh_1(t) = 0$, and hence $\CIR_4(t)=0$ for
1414: $t\in[0,T+\delta)$.
1415:
1416: To prove part (g), we note that (from part (f))
1417: \beq
1418: F_{\twg_4}(t)=1,\ \ \ t\in(\frac{1}{\alpha_{\tCI_4}}\ltho
1419: +\frac{1}{\alpha_{\tCIA_4}}\ltho,T+\delta),
1420: \eeq
1421: implying that $\cwg_4(t)$ increases in this interval. On the other hand, we know that
1422: $\cwg_4(t)\geq\theta$ and $\wg_4(t)=1$ up to at least
1423: $t=\frac{1}{\alpha_{\twg_4}}\lth>\frac{1}{\alpha_{\tCI_4}}\ltho+\frac{1}{\alpha_{\tCIA_4}}\ltho$.
1424: This shows that in fact $\wg_4(t)=1$ for all $t\in[0,T+\delta)$.
1425: \bbox
1426:
1427: {\it Proof of Theorem~\ref{th-glass2}:}
1428: Since $\wg_4(0)=1$, from equations\rf{eq-hybrid-switch1},\rf{eq-hybrid-switch0},
1429: we know that the earliest possible switching time from 1 to 0 is $\alpha_{\twg_4}^{-1}\lth$.
1430: Applying Lemma~\ref{lm-wg4} with $T=\alpha_{\twg_4}^{-1}\lth$ establishes
1431: that $\wg_4(t)=1$ for $t\in[0,T+\delta)$, with $\delta$ given by\rf{eq-delta}.
1432: Next, applying Lemma~\ref{lm-wg4} with $T=\alpha_{\twg_4}^{-1}\lth + k\delta$, $k\in\N$,
1433: shows that $\wg_4(t)=1$ for $t\in[0,T+(k+1)\delta)$. Since $\delta$ is finite, we can conclude
1434: by induction that $\wg_4(t)=1$ for all $t\geq0$.
1435:
1436: To prove that $\PTC_1(t)\equiv0$, note that $\CIA_1(t)\equiv0$ (Lemma~\ref{lm-wg4},
1437: with $T=+\infty$) implies $\ptc_1(t)\equiv0$. Since $\PTC_1(0)=0$ and $\PTC_1$ cannot
1438: become expressed unless $\ptc_1$ is first expressed, the desired result follows.
1439: \bbox
1440:
1441:
1442:
1443: \section{Conclusions}
1444:
1445: We discussed two alternative methods for modeling gene expression networks:
1446: purely discrete Boolean methods and piecewise linear differential systems, which
1447: combine continuous degradation with discrete synthesis.
1448: For both methods we introduced new techniques for a deeper analysis of the networks
1449: with respect to perturbations in the timescales of the system.
1450: For the piecewise linear system we also studied the effect of the ON concentration thresholds.
1451:
1452: We find that unrestricted variability in the duration of the diverse processes present
1453: in the network may lead to significant deviations from experimentally observed results,
1454: thus suggesting the fragility of the developmental process under severe
1455: perturbations (the asynchronous
1456: algorithm fails to predict the correct pattern with probability 50\%, and the Glass-type
1457: system fails with probability at least 10\%).
1458:
1459: Another set of numerical experiments introduces a separation between timescales of
1460: post-translational and transcription/translation processes, and in practice
1461: ``updates mRNAs later than proteins''.
1462: In this context, the piecewise linear Glass-type system indicates a remarkable robustness
1463: of the Boolean model in predicting the final gene expression pattern.
1464: Indeed, we provide a theoretical proof that the Glass-type system always correctly
1465: generates the wild type development (i.e. the convergence to the
1466: wild type steady state when started from the wild type initial state),
1467: under the separation of timescales assumption,
1468: and appropriate OFF/ON thresholds.
1469: %
1470: The asynchronous model's predictions depend very much on the degree of separation between
1471: timescales. As the intervals $A_{\tProt}$ and $A_{\tmRNA}$ become closer, the asynchronous
1472: algorithm increasingly fails to generate the wild type pattern.
1473: This leads us to conclude that a strong separation between timescales of
1474: post-translational and transcription/translation processes is necessary for establishing
1475: the regular gene expression pattern in the segment polarity network.
1476: From analysis of the piecewise linear model we conclude that the fraction of maximal
1477: concentration above which a protein or mRNA is effectively ON needs to be quite small,
1478: below 50\%. Higher concentration thresholds may disrupt the development process.
1479: %
1480:
1481: The comparison between discrete and continuous models shows clearly that sudden
1482: transitions may happen in discrete systems and lead to a false result; such
1483: sudden transitions are smoothed out in continuous models which prevent generation
1484: of false results (see also~\cite{m05}).
1485: Nevertheless, we conclude that both models agree in predicting the fundamental sequences of
1486: gene expression that irreversibly lead to a deviation in the development towards a
1487: mutant state.
1488: %
1489:
1490: By combining continuous-time techniques
1491: with discrete events, we can with great generality
1492: explore and sample the space of all possible timescales
1493: as well as of effective ON levels.
1494: %
1495: Moreover, as information about the mRNA/protein lifetimes, decay rates or
1496: activation thresholds becomes available, it can be straightforwardly incorporated
1497: by fixing the corresponding inverse scaling factor $\alpha_i^{-1}$. The hybrid model
1498: retains the ease of Boolean models in determining the steady states
1499: corresponding to gene knockouts and perturbed initial conditions.
1500: It is straightforward to calculate which mutant patterns result from each gene knockout.
1501: Here we also studied the the effect of perturbations on the prepattern, by simulating delay
1502: in initial expression of each gene. We find that the system is vulnerable to large delays
1503: (larger than two time units) in expression of any gene -- except for $\ci$ --
1504: and, in such delayed conditions, the mutant state characteristic to that gene knockout is
1505: generated.
1506: For low order delays, the system typically recovers and proceeds through the correct wild
1507: type development.
1508:
1509:
1510: The Glass-type system
1511: with time separation, as a model of the segment polarity gene network,
1512: reflects the conclusion of von Dassow et al. that the topology of the
1513: network is more important than the fine-tuning of the kinetic parameters
1514: \cite{dmmo00}, since its results are robust for a large region
1515: of parameter (scaling factor,
1516: activation threshold)
1517: space. Due to its underlying Boolean structure,
1518: the model also intrinsically incorporates the
1519: recent finding of Ingolia that parameter sets need to
1520: satisfy certain constraints - that ensure the bistability of certain
1521: genes- to lead to correct solutions\cite{i04}. Taken together, the results of the
1522: synchronous~\cite{ao03}, asynchronous~\cite{cas05} Boolean and hybrid models
1523: convincingly demonstrate the Boolean
1524: models' capability for effectively describing the basic structure and functioning of
1525: gene control networks when detailed kinetic information is unavailable.
1526:
1527: \bibliography{drosophila}
1528:
1529:
1530: \clearpage
1531: \begin{figure}
1532: \centerline{
1533: \psfig{figure=csa_fig1.ps,height=4.5in}}
1534: \caption{The solution to the system of piecewise linear equations\rf{eq-glass} (dashed lines).
1535: Each column represents one cell, and each of the 13 rectangles represents the continuous-time
1536: variable for proteins or mRNAs (as labeled at left). In each rectangle, the y-axis ranges
1537: from 0 to 2 units.
1538: The nodes for which the trajectories converge to 1 (middle of the rectangle) are exactly
1539: those expressed in the wild type steady state. (The time units are arbitrary.)}
1540: \label{fig-glass-wt}
1541: \end{figure}
1542:
1543:
1544: \clearpage
1545: \begin{figure}
1546: \centerline{
1547: \psfig{figure=csa_fig2.ps,height=5in} }
1548: \caption{Probability of occurrence of the three most frequent patterns:
1549: wild type (WT), broad stripes (BS), and no segmentation (NS), under variable range of timescales.
1550: Dashed lines/squares represent asynchronous algorithm results, while solid lines/circles
1551: represent Glass-type model results (out of 1000 runs). Results were obtained with $\theta=0.5$.}
1552: \label{fig-interval-length}
1553: \end{figure}
1554:
1555: \clearpage
1556: \begin{figure}
1557: \centerline{
1558: \psfig{figure=csa_fig3.ps,height=5in} }
1559: \caption{Effect of effective ON concentration, $\theta$,
1560: on the probability of occurrence of the three most frequent patterns:
1561: wild type (WT), broad stripes (BS), and no segmentation (NS).
1562: The results are for the Glass-type model with $\alpha_i^{-1}$ randomly chosen
1563: in an interval $[1-\eps,1+\eps]$.
1564: Solid lines represent the case $\eps=0.5$, dotted lines represent the case $\eps=0.1$
1565: and dash-dotted lines represent the case $\eps=0.9$.
1566: (results out of 1000 runs).}
1567: \label{fig-threshold}
1568: \end{figure}
1569:
1570: \clearpage
1571: \begin{figure}
1572: \centerline{
1573: \psfig{figure=csa_fig4.ps,height=4.5in} }
1574: \caption{A solution to the system of piecewise linear equations\rf{eq-glass-a}
1575: (dashed lines), in an example where the steady state corresponds to the broad stripes pattern.
1576: In each rectangle, the y-axis ranges from 0 to 2 units.
1577: Notice that {\it wingless} is expressed in two adjacent cells, as opposed to the wild type
1578: pattern, where {\it wingless} is expressed in only one cell (similarly for {\it engrailed}
1579: and {\it hedgehog}).}
1580: \label{fig-glass-bs}
1581: \end{figure}
1582:
1583: \clearpage
1584: \begin{figure}
1585: \centerline{
1586: \psfig{figure=csa_fig5.ps,height=4.5in} }
1587: \caption{A solution to the system of piecewise linear equations\rf{eq-glass-a}
1588: (dashed lines), in an example where the steady state corresponds to the no segmentation pattern.
1589: In each rectangle, the y-axis ranges from 0 to 2 units.
1590: Notice that {\it wingless}, {\it engrailed} and {\it hedgehog} are not expressed in any cell,
1591: thus no segments are visibly detected in the embryo.}
1592: \label{fig-glass-ns}
1593: \end{figure}
1594:
1595: \clearpage
1596: \begin{figure}
1597: \centerline{
1598: \psfig{figure=csa_fig6.ps,height=5in} }
1599: \caption{Effect of initially delayed expression in occurrence of steady state patterns.
1600: Left: delay in {\it wingless} expression. Middle: delay in {\it engrailed} expression.
1601: Right: delay in {\it hedgehog} expression.
1602: (results out of 1000 runs).}
1603: \label{fig-delay-wg-en-hh}
1604: \end{figure}
1605:
1606: \clearpage
1607: \begin{figure}
1608: \centerline{
1609: \psfig{figure=csa_fig7.ps,height=5in} }
1610: \caption{Effect of initially delayed expression in occurrence of steady state patterns.
1611: Left: delay in {\it patched} expression. Right: delay in {\it cubitus interruptus} expression.
1612: (results out of 1000 runs).}
1613: \label{fig-delay-ptc-ci}
1614: \end{figure}
1615:
1616: \clearpage
1617: \begin{figure}
1618: \centerline{
1619: \psfig{figure=csa_fig8.ps,height=5in} }
1620: \caption{Probability of occurrence of the three most frequent patterns:
1621: wild type (WT), broad stripes (BS), and no segmentation (NS), with separation of timescales.
1622: Dashed lines/squares represent asynchronous algorithm results, while solid lines/circles
1623: represent Glass-type model results. The $x$-axis represents the level of separation,
1624: computed by $\min\{b\in A_{\tProt}\}/\max\{a\in A_{\tmRNA}\}$ (out of 1000 runs).
1625: Results were obtained with $\theta=0.5$.}
1626: \label{fig-time-separation}
1627: \end{figure}
1628:
1629:
1630: \end{document}
1631:
1632: \begin{thebibliography}{10}
1633:
1634: \bibitem{giot03}
1635: L.~Giot, J.~S. Bader, C.~Brouwer, A.~Chaudhuri, and B.~Kuang et~al.
1636: \newblock A protein interaction map of drosophila melanogaster.
1637: \newblock {\em Science}, 302:1727--1736, 2003.
1638:
1639: \bibitem{han04}
1640: J.~D. Han, N.~Bertin, T.~Hao, D.~S. Goldberg, and G.~F.~Berriz et~al.
1641: \newblock Evidence for dynamically organized modularity in the yeast
1642: protein-protein interaction network.
1643: \newblock {\em Nature}, 430:88--93, 2004.
1644:
1645: \bibitem{li04}
1646: S.~Li, C.~M. Armstrong, N.~Bertin, H.~Ge, and et~al. S.~Milstein.
1647: \newblock A map of the interactome network of the metazoan c. elegans.
1648: \newblock {\em Science}, 303:540--543, 2004.
1649:
1650: \bibitem{lee02}
1651: T.~I. Lee, N.~J. Rinaldi, F.~Robert, D.~T. Odom, and et~al. Z.~Bar-Joseph.
1652: \newblock Transcriptional regulatory networks in saccharomyces cerevisiae.
1653: \newblock {\em Science}, 298:799--804, 2002.
1654:
1655: \bibitem{a05}
1656: R.~Albert.
1657: \newblock Scale-free networks in cell biology.
1658: \newblock {\em Journal of cell science}, 118:4947--4957, 2005.
1659:
1660: \bibitem{lbyst04}
1661: N.~M. Luscombe, M.~M. Babu, H.~Y. Yu, M.~Snyder, S.~S. Teichmann, and
1662: M.~Gerstein.
1663: \newblock Genomic analysis of regulatory network dynamics reveals large
1664: topological changes.
1665: \newblock {\em Nature}, 431:308--312, 2004.
1666:
1667: \bibitem{st01}
1668: L.~S\'anchez and D.~Thieffry.
1669: \newblock A logical analysis of the {\it drosophila} gap-gene system.
1670: \newblock {\em J. Theor. Biol.}, 211:115--141, 2001.
1671:
1672: \bibitem{ao03}
1673: R.~Albert and H.~G. Othmer.
1674: \newblock The topology of the regulatory interactions predicts the expression
1675: pattern of the {\it drosophila} segment polarity genes.
1676: \newblock {\em J. Theor. Biol.}, 223:1--18, 2003.
1677:
1678: \bibitem{mta99}
1679: L.~Mendoza, D.~Thieffry, and E.~R. Alvarez-Buylla.
1680: \newblock Genetic control of flower morphogenesis in {\it arabidopsis
1681: thaliana}: a logical analysis.
1682: \newblock {\em Bioinformatics}, 15:593--606, 1999.
1683:
1684: \bibitem{epa04}
1685: C.~Espinosa-Soto, P~Padilla-Longoria, and E.~R. Alvarez-Buylla.
1686: \newblock A gene regulatory network model for cell-fate determination during
1687: arabidopsis thaliana flower development that is robust and recovers
1688: experimental gene expression profiles.
1689: \newblock {\em Plant Cell}, 16:2923--2939, 2004.
1690:
1691: \bibitem{dmmo00}
1692: G.~von Dassow, E.~Meir, E.M. Munro, and G.M. Odell.
1693: \newblock The segment polarity network is a robust developmental module.
1694: \newblock {\em Nature}, 406:188--192, 2000.
1695:
1696: \bibitem{cas05}
1697: M.~Chaves, R.~Albert, and E.D. Sontag.
1698: \newblock Robustness and fragility of boolean models for genetic regulatory
1699: networks.
1700: \newblock {\em J. Theor. Biol.}, 235:431--449, 2005.
1701:
1702: \bibitem{g75}
1703: L.~Glass.
1704: \newblock Classification of biological networks by their qualitative dynamics.
1705: \newblock {\em J. Theor. Biol.}, 54:85--107, 1975.
1706:
1707: \bibitem{eg00}
1708: R.~Edwards and L.~Glass.
1709: \newblock Combinatorial explosion in model gene networks.
1710: \newblock {\em Chaos}, 10:691--704, 2000.
1711:
1712: \bibitem{gk73}
1713: L.~Glass and S.A. Kauffman.
1714: \newblock The logical analysis of continuous, nonlinear biochemical control
1715: networks.
1716: \newblock {\em J. Theor. Biol.}, 39:103--129, 1973.
1717:
1718: \bibitem{hs92}
1719: J.~E. Hooper and M.P. Scott.
1720: \newblock The molecular genetic basis of positional information in insect
1721: segments.
1722: \newblock In W.~Hennig, editor, {\em Early Embryonic Development of Animals},
1723: pages 1--49. Springer, Berlin, 1992.
1724:
1725: \bibitem{gakt00}
1726: A.~Gallet, C.~Angelats, S.~Kerridge, and P.P. Th\'erond.
1727: \newblock Cubitus interruptus-independent transduction of the hedgehog signal
1728: in {\it drosophila}.
1729: \newblock {\em Development}, 127:5509--5522, 2000.
1730:
1731: \bibitem{tek92}
1732: T.~Tabata, S.~Eaton, and T.B. Kornberg.
1733: \newblock The {\it drosophila hedgehog} gene is expressed specifically in
1734: posterior compartment cells and is a target of {\it engrailed} regulation.
1735: \newblock {\em Genes \& Dev.}, 6:2635--2645, 1992.
1736:
1737: \bibitem{cgg94}
1738: K.M. Cadigan, U.~Grossniklaus, and W.~J. Gehring.
1739: \newblock Localized expression of {\it sloppy paired} protein maintains the
1740: polarity of {\it drosophila} parasegments.
1741: \newblock {\em Genes \& Dev.}, 8:899--913, 1994.
1742:
1743: \bibitem{av03}
1744: C.~Alexandre and J.~P. Vincent.
1745: \newblock Requirements for transcriptional repression and activation by
1746: engrailed in {\it drosophila} embryos.
1747: \newblock {\em Development}, 130:729--739, 2003.
1748:
1749: \bibitem{sg04}
1750: D.~Swantek and J.~P. Gergen.
1751: \newblock Ftz modulates runt-dependent activation and repression of segment
1752: -polarity gene transcription.
1753: \newblock {\em Development}, 131:2281--2290, 2004.
1754:
1755: \bibitem{tsi}
1756: D.P. Bertsekas and J.N. Tsitsiklis.
1757: \newblock {\em Parallel and Distributed Computation, Numerical Method}.
1758: \newblock Prentice Hall, Englewood Cliffs, New Jersey, 1989.
1759:
1760: \bibitem{jghpsg04}
1761: H.~de~Jong, J.L. Gouz\'e, C.~Hernandez, M.~Page, T.~Sari, and J.~Geiselmann.
1762: \newblock Qualitative simulation of genetic regulatory networks using piecewise
1763: linear models.
1764: \newblock {\em Bull. Math. Biol.}, 66:301--340, 2004.
1765:
1766: \bibitem{gedeon}
1767: T.~Gedeon.
1768: \newblock Attractors in continuous-time switching networks.
1769: \newblock {\em Communications on Pure and Applied Analysis}, 2:187--209, 2003.
1770:
1771: \bibitem{m05}
1772: A.~Mochizuki.
1773: \newblock An analytical study of the number of steady states in gene regulatory
1774: networks.
1775: \newblock {\em J. Theor. Biol.}, 236:291--310, 2005.
1776:
1777: \bibitem{i04}
1778: N.T. Ingolia.
1779: \newblock Topology and robustness in the drosophila segment polarity network.
1780: \newblock {\em PLoS Biology}, 2:0805--0815, 2004.
1781:
1782: \end{thebibliography}
1783:
1784:
1785:
1786: