1: \documentclass[a4paper,11pt]{article}
2: \pdfoutput=1
3: \usepackage{graphicx,mathptmx,helvet,courier,amsmath,mike}
4:
5: \renewcommand{\textfraction}{0.05}
6: \renewcommand{\floatpagefraction}{0.95}
7: \renewcommand{\topfraction}{0.95}
8: \renewcommand{\thefootnote}{\fnsymbol{footnote}}
9:
10: \oddsidemargin 0mm
11: \evensidemargin 0mm
12: \topmargin -15mm
13:
14: \textwidth 160mm
15: \textheight 25cm
16:
17: \parindent0pt
18:
19: \begin{document}
20:
21: \title{A model of dissociated cortical tissue}
22:
23: \author{Michael Stiber \and Fumitaka Kawasaki\\
24: Computing \& Software Systems \\
25: University of Washington, Bothell \\
26: Bothell, WA, 98011-8246 USA \\
27: stiber@u.washington.edu\\
28: fumik@u.washington.edu \\[2\parskip]
29: Dongming Xu \\
30: Linear Technology \\
31: 15100 Weston Parkway Suite 202 \\
32: Cary, NC 27513 USA \\
33: dxu@linear.com}
34:
35: \date{}
36:
37: % \maketitle
38: % Save space by making the titles myself
39: \begin{center}
40: \textbf{\Large A MODEL OF DISSOCIATED CORTICAL TISSUE\footnote{A more
41: complete version of this paper is available at http://faculty.washingon.edu/stiber/Public/NC07.pdf.}}\\[0.15in]
42:
43: \vskip 16 pt
44:
45: \begin{tabular}{p{3in}p{3in}}
46: \centering
47: \textit{Michael Stiber}\footnotemark \qquad \textit{Fumitaka Kawasaki}\\
48: Computing \& Software Systems Program \\
49: University of Washington, Bothell \\
50: Bothell, WA, 98011-8246 USA \\
51: stiber@u.washington.edu -- http://faculty.washington.edu/stiber \\
52: fumik@u.washington.edu &
53: \centering
54: \textit{Dongming Xu} \\
55: Linear Technology \\
56: 15100 Weston Parkway Suite 202 \\
57: Cary, NC 27513 USA \\
58: dxu@linear.com
59: \end{tabular}
60: \end{center}
61:
62: \footnotetext{To whom correspondence should be sent.}
63:
64: \section*{ABSTRACT}
65:
66: A powerful experimental approach for investigating computation in
67: networks of biological neurons is the use of cultured dissociated
68: cortical cells grown into networks on a multi-electrode array. Such
69: preparations allow investigation of network development, activity,
70: plasticity, responses to stimuli, and the effects of pharmacological
71: agents. They also exhibit whole-culture pathological bursting;
72: understanding the mechanisms that underlie this could allow creation
73: of more useful cell cultures and possibly have medical
74: applications~\cite{vanpelt-etal04,wagenaar-etal05a}.
75:
76: \vskip 11pt
77:
78: This paper presents preliminary results of a computational study of
79: the interplay of individual neuron activity, cell culture development,
80: and the network behavior. We investigate whether bursting can occur
81: in an initially unconnected ``network'' that develops connections
82: according to an experimentally-verified model of cell culture
83: connectivity growth.
84:
85: \paragraph{Neuron Model}
86: An integrate-and-fire neuron model with dynamical synapses that
87: exhibit activity-dependent facilitation and depression was
88: used~\cite{tsodyks-etal98,tsodyks-etal00,maass-etal02}. It includes
89: synaptic, constant, and noise currents, with a reset of its membrane
90: voltage to a fixed value upon exceeding threshold and generating a
91: spike and a fixed absolute refractory period thereafter.
92:
93: \vskip 11pt
94:
95: The synapse has four state variables: three that govern the fraction
96: of synaptic resources in particular states --- $x$ (recovered state),
97: $y$ (active state), and $z$ (inactive state) --- and one, $u$, that
98: represents synaptic efficiency (see the more complete
99: paper\footnotemark[1] for detailed equations). See~\cite{maass-etal02}
100: for parameter values used.
101:
102: \paragraph{Network Model}
103: Simulations were conducted by constructing networks with model neurons
104: on a rectangular grid. Connectivity was determined by incorporating a
105: model of cortical cell culture connectivity
106: development~\cite{vanooyen-etal95} that model's neurite outgrowth as a
107: radius of connectivity that changes at a rate inversely proportional
108: to a sigmoidal function of cell firing rate:
109: \begin{align}
110: \fderiv{R_i}{t} &= \rho G(F_i) \label{eq:outgrowth} \\
111: G(F_i) &= 1 - \frac{2}{1 + \exp((\epsilon - F_i )/\beta)}
112: \end{align}
113: where $R_i$ is the radius of connectivity of neuron $i$, $F_i$ is
114: neuron $i$'s normalized firing rate, $\rho$ is a rate constant,
115: $\epsilon$ is a constant that sets the ``null point'' for outgrowth
116: (the normalized \emph{target} firing rate that causes no outgrowth or
117: retraction), and $\beta$ determines the slope of $G(\cdot)$. One of
118: the parameters varied in these simulations was $\epsilon$.
119:
120: \vskip 11pt
121:
122: Synaptic strengths $W_i$ were computed for all pairs of neurons that
123: had overlapping connectivity regions as the area of their circles'
124: overlap. The bulk of the neurons in the network were excitatory; a
125: small number were chosen to be inhibitory. Similarly, most neurons
126: were not spontaneously active, but a few had lowered firing thresholds
127: to produce spontaneous firing at a rate of between 0.02 and 6
128: spikes/sec. To produce more consistent simulation results, a set of
129: standardized layouts was chosen to maximize spacing among inhibitory
130: and spontaneously active cells and reduce edge effects. The fraction
131: of cells that were inhibitory was the other simulation parameter
132: varied.
133:
134: \paragraph{Computer Implementation}
135: We used CSIM (A Neural Circuit SIMulator) version 1.1 for the
136: simulations. The original code was pared down to a small core that was
137: linked to a stand-alone C++ program to run on Linux, Windows, and
138: Macintosh computers. Generally speaking, each simulation took between
139: two and 20 hours on computers with 2--3GHz microprocessors.
140:
141: \vskip 11pt
142:
143: Simulations consisted of networks of 100 neurons in a 10x10
144: arrangement simulated for 30,000--60,000 seconds. While the rate of
145: neurite outgrowth was greatly sped up compared to the living
146: preparation, numerical investigation indicated that this did not
147: introduce instability in the simulation.
148:
149:
150: \paragraph{Analysis Methods}
151: To examine global behavior, average neuron firing rate and
152: \emph{burstiness index} (BI)~\cite{wagenaar-etal05a} were calculated
153: and plotted versus the two simulation parameters. BI was computed by
154: first calculating a spike count vs. time histogram for the entire
155: network during the last 5,000sec of the simulation. The fraction,
156: $f_{15}$, of the total number of spikes contained by the 15\% most
157: populous bins was then normalized to produce the burstiness index, BI,
158: as $\mathrm{BI} = (f_{15} - 0.15)/0.85$. Detailed examination of
159: single simulations involved plots of neurons' connectivity radii and
160: firing rate versus time.
161:
162: \begin{figure}
163: \centering
164: \begin{tabular}{rlrl}
165: \raisebox{1.75in}{\textsf{\large A}} &
166: \includegraphics[width=2.5in]{norm-rate-25k-30k} &
167: \raisebox{1.75in}{\textsf{\large B}} &
168: \includegraphics[width=2.5in]{BI-25k-30k}
169: \end{tabular}
170: \caption{Firing rates normalized relative to target rate (A);
171: burstiness index (B). Both computed for 25,000-30,000sec and
172: plotted versus the two simulation parameters.\label{fg:rate-BI}}
173: \end{figure}
174:
175: \paragraph{Results}
176: An initial set of 50 simulations was performed with target rates in
177: the range 0.1--1.9 (inclusive, in 10 steps) and fraction of excitatory
178: neurons 0.9--0.98 (inclusive, in five steps).
179: Figure~\ref{fg:rate-BI}(A) shows normalized firing rates for the final
180: 5,000sec of each simulation. Only the simulations with higher target
181: rates and fewer inhibitory neurons showed the great increase in firing
182: rate that might be associated with bursting. Longer (60,000sec)
183: simulations exhibited some bursting type of behaviors with the
184: fraction of bursting as low as 0.9 for the higher target rates.
185: Figure~\ref{fg:rate-BI}(B) indicates that the fraction of excitatory
186: neurons in this range has only a modest effect on burstiness index
187: (for 30,000sec simulations), and that moderate target rates produce
188: the highest BI values. This apparent conflict with the previous
189: observation of bursting at higher target rates can be explained by
190: examining the detailed behavior of individual simulations.
191:
192: \begin{figure}
193: \centering
194: \begin{tabular}{ccc}
195: \includegraphics[width=1.9in]{tR_0_1--fE_0_90--60000_fig2} &
196: \includegraphics[width=1.9in]{tR_0_9--fE_0_94--60000_fig2} &
197: \includegraphics[width=1.9in]{tR_1_9--fE_0_98--60000_fig2}
198: \end{tabular}
199: \caption{Detailed simulation results for simulations with parameters
200: (target rate, fraction excitatory cells) of (0.1, 0.9) (left),
201: (0.9, 0.94) (middle), and (1.9, 0.98) (right). Data for edge
202: neurons are green, non-edge (and non-inhibitory, non-spontaneously
203: active) neurons are black, inhibitory neurons are red, and
204: spontaneously active neurons are blue.\label{fg:detail}}
205: \end{figure}
206:
207: Figure~\ref{fg:detail} shows detailed information for simulations with
208: three sets of parameters: (target rate, fraction excitatory cells) =
209: (0.1, 0.9), (0.9, 0.94), and (1.9, 0.98). These include the parameter
210: extremes and a central value and both bursting and non-bursting
211: activity. In these cases, 60,000sec simulations were performed. Note
212: that some of the simulations that weren't bursting at 30,000sec were
213: bursting shortly thereafter, as evidenced by the (0.9, 0.94) one.
214:
215: \vskip 11pt
216:
217: Nevertheless, this confirms that the low BI values for low target
218: rates correspond to non-bursting behaviors (and that the connectivity
219: radii had not stabilized for cells that were not inhibitory or
220: spontaneously active). Low BI values for high target rates were a
221: possible result of the very broad or frequent bursts. For the bursting
222: behaviors, connectivity radii have stabilized, excepting small
223: variations during bursting. In all simulations, connectivity radii
224: for edge neurons are larger than others, inhibitory neurons had
225: moderate radii, while spontaneously active neurons had a wide range of
226: different connectivities, likely due to the variability in their
227: firing thresholds.
228:
229: \vskip 11pt
230:
231: In either bursting or non-bursting behaviors, spontaneously active
232: neurons tended to be the most active. This is not surprising, as their
233: lowered thresholds would make them more excitable. The next higher
234: firing rates belonged to the inhibitory cells, then non-edge cells,
235: then edge cells.
236:
237: \begin{figure}
238: \centering
239: \begin{tabular}{ccc}
240: \includegraphics[width=1.9in]{tR_0_9--fE_0_94--50000-41000_fig2} &
241: \includegraphics[width=1.9in]{tR_0_9--fE_0_94--60000-40000_fig2} &
242: \includegraphics[width=1.9in]{tR_1_9--fE_0_98--50000-41000_fig2}
243: \end{tabular}
244: \caption{Effects of neurite outgrowth on bursting for simulations
245: with parameters (target rate, fraction excitatory cells) of (0.9,
246: 0.94) (left, middle) and (1.9, 0.98) (right). Growth was
247: ``frozen'' either during (left, right) or between (middle)
248: bursts.\label{fg:growth-burst}}
249: \end{figure}
250:
251: In the results in figure~\ref{fg:detail}, it seems possible that the
252: mechanism for burst initiation and termination is the variation in
253: connectivity. Figure~\ref{fg:growth-burst} presents simulations in
254: which connectivity was fixed either during bursts (left, middle) or in
255: between bursts (right). For the lower target rate/lower fraction of
256: excitatory cells simulation, it does indeed seem that bursting is
257: controlled by connectivity. However, for a higher rate/higher fraction
258: of excitatory cells, bursting can continue even in the absence of
259: variation in connectivity (this simulation also produced bursts when
260: growth was stopped during a burst).
261:
262: \paragraph{Discussion}
263: Bursting occurred with sufficiently small inhibition and high target
264: firing rate. One might expect the latter to produce greater
265: connectivity for every neuron, which in turn would be the mechanism
266: for whole-culture bursting. However, in the low-target-rate,
267: non-bursting simulations, such as figure~\ref{fg:detail}(left),
268: non-inhibitory, non-spontaneously active cells grow large connectivity
269: radii. Spontaneously active cells, on the other hand, tend to have
270: large connectivity radii in the bursting simulations. Presumably,
271: lowering these cells autonomous firing rate would result in bursting
272: at lower target rates.
273:
274: \vskip 11pt
275:
276: In previous investigations of bursting with randomly connected
277: networks~\cite{tsodyks-etal00}, the model synapses' depression and
278: facilitation were, neglecting the influence of noise, the mechanism
279: underlying burst initiation and termination. Our preliminary results
280: indicate that this is possibly the case under certain circumstances,
281: but not all. For some regions of parameter space, it may be the case
282: that the mechanism is a hysteresis effect involving changing
283: connectivity radii. There are a number of possible reasons for this
284: difference:
285: \begin{itemize}
286: \item A number of parameters were set arbitrarily or not fully
287: explored. These include fraction of excitatory cells, scaling of
288: synaptic weights from area of connectivity overlap, and no
289: differential scaling based on type of synapse (i.e., inhibitory
290: vs. excitatory).
291:
292: \item In the current simulations, only spontaneously active cells had
293: any parameter variability; all other cells of a given type
294: (inhibitory or excitatory) had identical parameters.
295:
296: \item The overriding issue here is likely the small network size. Edge
297: effects were great (edge neurons' connectivity radii were always the
298: greatest of all neurons and 36/100 of the cells were edge neurons)
299: and there were small numbers of inhibitory and spontaneously active
300: cells. The final networks were almost completely
301: connected. Increasing network size to, say, 100x100, could have
302: little impact on final connectivity radii but with each neuron
303: having connections to less than 10\% of the network.
304:
305: Increasing network size will have computational consequences that
306: must be addressed: in its current form, a 60,000sec simulation of a
307: 100x100 network would take at least 2,000 hours (83 days).
308: \end{itemize}
309:
310: There are also fundamental differences between the connectivity
311: patterns generated by this model (perhaps most similar to radial basis
312: functions) and many other models of cortex or recurrent networks (in
313: which either network topology is irrelevant or a power law-type
314: distribution is used that produces mostly local connections with a few
315: long-range ones). It will be instructive to investigate the detailed
316: correlation structure of inter- and intra-burst neuron firing.
317:
318: \vskip 11pt
319:
320: \textbf{\textit{Keywords:}} cortical cultures, network development, bursting.
321:
322:
323: \begin{thebibliography}{9}
324:
325: \bibitem{vanpelt-etal04}
326: J.~V. Pelt, M.~Corner, P.~Wolters, W.~Rutten, and G.~Ramakers, ``Longterm
327: stability and developmental changes in spontaneous network burst firing
328: patterns in dissociated rat cerebral cortex cell cultures on multielectrode
329: arrays,'' {\em Neurosci. Lett.}, vol.~361, pp.~86--89, 2004.
330:
331: \bibitem{wagenaar-etal05a}
332: D.~A. Wagenaar, R.~Madhavan, J.~Pine, and S.~M. Potter, ``Controlling bursting
333: in cortical cultures with closed-loop multi-electrode stimulation,'' {\em J.
334: Neurosci.}, vol.~25, pp.~680--8, Jan. 2005.
335:
336: \bibitem{tsodyks-etal98}
337: M.~Tsodyks, K.~Pawelzik, and H.~Markram, ``Neural networks with dynamic
338: synapses,'' {\em Neural Comp.}, vol.~10, pp.~821--35, 1998.
339:
340: \bibitem{tsodyks-etal00}
341: M.~Tsodyks, A.~Uziel, and H.~Markram, ``Synchrony generation in recurrent
342: networks with frequency-dependent synapses,'' {\em J. Neurosci.}, vol.~20,
343: no.~RC50, pp.~1--5, 2000.
344:
345: \bibitem{maass-etal02}
346: W.~Maass, T.~Natschl{\"{a}}ger, and H.~Markram, ``Real-time computing without
347: stable states: A new framework for neural computation based on
348: perturbations,'' {\em Neural Comp.}, vol.~14, pp.~2531--60, Nov. 2002.
349:
350: \bibitem{vanooyen-etal95}
351: A.~Van~Ooyen, J.~Van~Pelt, and M.~Corner, ``Implications of activity dependent
352: neurite outgrowth for neuronal morphology and network development,'' {\em J.
353: Theor. Biol.}, vol.~172, pp.~63--82, 1995.
354:
355: \end{thebibliography}
356:
357: \end{document}
358:
359: % LocalWords: SAO PSP AP PSPs APs Gestrelius washington edu Pottorf Holderman
360: % LocalWords: presynaptic postsynaptic gabaeric neurite burstiness undemanding
361: