1:
2: \documentclass{article}
3: \usepackage{times}
4: \include{psfig}
5: \usepackage{epsfig}
6:
7:
8: %\bibliographystyle{apsrev}
9:
10: \begin{document}
11:
12: \title{Morphological Complex Networks: Can Individual Morphology
13: Determine the General Connectivity and Dynamics of Networks?}
14:
15: \author{Luciano da Fontoura Costa, \\
16: Instituto de F\'{i}sica de S\~ao Carlos. \\
17: Universidade de S\~ao Paulo, S\~{a}o Carlos, \\
18: SP, Caixa Postal 369, 13560-970, \\
19: Fone +55 16 3373 9858, FAX +55 1633 71 3616, \\
20: Brazil, luciano@if.sc.usp.br}
21:
22:
23: \date{18th March 2005}
24:
25:
26: \maketitle
27:
28: {\bf ABSTRACT:} This article discusses how the individual
29: morphological properties of basic objects (e.g. neurons, molecules and
30: aggregates), jointly with their particular spatial distribution, can
31: determine the connectivity and dynamics of systems composed by those
32: objects. This problem is characterized as a particular case of the
33: more general shape and function paradigm, which emphasizes the
34: interplay between shape and function in nature and evolution. Five
35: key issues are addressed: (a) how to measure shapes; (b) how to obtain
36: stochastic models of classes of shapes; (c) how to simulate
37: morphologically realistic systems of multiple objects; (d) how to
38: characterize the connectivity and topology of such systems in terms of
39: complex network concepts and measurements; and (e) how the dynamics of
40: such systems can be ultimately affected, and even determined, by the
41: individual morphological features of the basic objects. Although
42: emphasis is placed on neuromorphic systems, the presented concepts and
43: methods are useful also for several other multiple object systems,
44: such as protein-protein interaction, tissues, aggregates and polymers.
45:
46:
47: \begin{quotation}
48: '...the functional superiority of the human brain is \\
49: intimately linked up with the prodigious abundance \\
50: and unaccustomed wealth of forms of the so-called \\
51: neurons with short axons.' \\
52: \hspace{2cm} \emph{(Santiago Ram\'on y Cajal, Recollections of my life)}
53: \end{quotation}
54:
55:
56: \section{Introduction}
57:
58: Look around and all you see is geometry. Although string theory
59: suggests the universe is 10-dimensional, our perception is constrained
60: to 3 spatial dimensions which, jointly with time, accounts for the 4
61: traditional dimensions. By contrast, a non-geographical graph (i.e. a
62: graph where the nodes do not have defined positions) lays in an
63: infinite dimensional space. One of the most important consequences of
64: finite-dimensional spaces is the fact that the adjacencies imposed by
65: such spaces, jointly with the presence of spatial constraints
66: (e.g. impenetrability) and the fact that real objects are finite, tend
67: to severely limit long range connections and interactions. For
68: instance, the only way to eat a pineapple is by cutting the fruit
69: somehow. Even when interactions are allowed to go through space, as
70: is the case of electromagnetic fields and waves, such effects tend to
71: diminish steadily with distance. At the same time as such constrains
72: limit our abilities, they are simultaneously responsible for
73: controlling the rate of dynamical changes in the universe, possibly
74: contributing to the enhancement of complexity.
75:
76: One of the consequences of the spatial constraints and distance
77: metrics typically found in finite-dimensional spaces is the emergence
78: of shape and morphology (e.g.~\cite{Thompson:1917, Bookstein:1991,
79: Small:1996, Dryden_Mardia:1998, Pete:2000, CostaCesar:2001}). For the
80: purposes of this work, a \emph{shape} will be henceforth understood as
81: any spatially limited and connected object~\cite{CostaCesar:2001},
82: while \emph{morphology} will stand for the study of shapes. Systems
83: obtained by combining, through connection or not, objects or shapes
84: will be understood as \emph{morphic multiple object systems --- MMOS}.
85: Examples of shapes include the geographical space occupied by objects
86: such as molecules, cells, organs, and so on. Combinations of
87: molecules give rise to multiple object systems such as materials,
88: combinations of cells to organs, while combinations of organs and
89: tissues produce living beings. Observe that the definition of shape
90: is relative and somewhat arbitrary. For instance, we may be
91: interested in the own shape of MMOS, such as the shape of the heart or
92: lungs, or in the shapes of each constituent object.
93:
94: The \emph{shape-function} paradigm, whose origin is lost in the
95: history of science (going back to ancient Greece, especially
96: Pythagoras, at the very least), suggests that there is a close
97: interplay between shape and function in nature. For instance, the
98: shape of proteins, together with field interactions, is essential for
99: defining their docking potential. At the same time, the wing of an
100: airplane or a bird is essential for obtaining proper aerodynamics,
101: while the effectiveness of our lungs is directly related to the
102: intricacy of the bronchial channels, which allows enhanced surface
103: area of contact with air. Although the shape-function paradigm can be
104: considered for the investigation of almost every natural phenomenon,
105: the current article concentrates attention on multiple object systems
106: which rely on the geometry of its basic components in order to obtain
107: proper behavior. An immediate example of such systems is provided by
108: the skeletons of vertebrates, which are intrinsically suited to
109: perform the mechanical activities of such animals. A more
110: sophisticate example, which will be central in our discussion, are the
111: biological neuronal networks such as those found in the mammals brain.
112:
113: The brain is composed by a myriad of cells --- the neurons, whose
114: morphology is highly specialized to make \emph{selective} connections.
115: Interestingly, neurons do not connect indiscriminately between
116: themselves, but form an intricate system of short to long-range
117: connections with specific targets~\cite{Kandel:1995}. In spite of
118: continuing efforts, the complete understanding on how neuronal
119: connections are established remains a considerable challenge. Current
120: knowledge indicates that neuronal wiring is mediated by every sort of
121: field interactions, including electrical, neurotrophic factors,
122: gradients of ionic and molecular concentrations, gravity, among other
123: possibilities. To any extent, the neuronal milieu represents a nice
124: example of how prolongations (i.e. dendrites and axons) are required
125: from neurons in order to overcome the adjacency constraints imposed by
126: the finite 3 spatial dimensions. At the same time, the degree of
127: spatial occupancy and complexity of a dendritic tree is directly
128: related to its potential as a synaptic target for growing axons. The
129: immediate consequence of such effects is the fact that neuronal cells
130: tend to have a most diverse and specialized overall morphology,
131: ranging from simple bipolar cells to highly elaborated Purkinje
132: cells.
133:
134: The complexity of neuronal shapes is immediately substantiated by the
135: fact that there is no current agreement between scientists regarding
136: the number of morphological classes of neurons, with competing
137: alternatives suggesting from 2 to thousands of types. It is therefore
138: hardly surprising that the diversity of neuronal cells identified by
139: Ramon y Cajal in the beginning of the 20th century would be understood
140: by he, the father of modern neuroscience, as a possible explanation
141: for the complexity of human behavior and
142: intelligence~\cite{Cajal:1989}. Since then, several evidences about
143: the relationship between neuronal shape and function have been
144: provided, including the congruence between the eletrophysiological
145: (i.e. functional) and morphological classes of retinal ganglion
146: cells~\cite{Boycott_Wassle:1974} as well as the correlation between
147: the structure of the receptive fields of those cells and the
148: respective morphology~\cite{Peichl_Wassle:1983}. However, despite
149: such cumulative evidence about the interplay between neuronal shape
150: and function, most artificial neural networks, such as the
151: perceptron~\cite{Rosenblatt:1958} and Hopfield
152: models~\cite{Hopfield:1982}, have not incorporated the neuronal
153: morphology.
154:
155: The mammals brain thus provides what is possibly the most compelling
156: illustration of a morphic multiple object system whose dynamics is, to
157: a substantial extent, the byproduct of the shapes of the individual
158: elements (e.g.~\cite{L_sibgrapi:1997}). In other words, the shape of
159: the individual components becomes essential for obtaining patterns of
160: connectivity which are required for producing, through
161: electrophysiological means, function and behavior of great complexity.
162: Although \emph{complexity} remains a somewhat ellusive
163: concept~\cite{Gell_Mann:1995}, the primate brain is characterized by a
164: continuum of complexity flowing from the sophistication of the
165: \emph{shape} of its basic elements (i.e. the neuronal cells) to the
166: complexity of the emerging \emph{dynamics}, passing through the
167: complexity of \emph{connections} between such cells. As illustrated
168: in Figure~\ref{fig:basic}, the geometrical properties of the
169: individual neuronal cells define the connectivity between those cells,
170: which ultimately produces complex dynamics. Interestingly, such
171: effects are bilateral. For instance, the shape of neuronal cells is
172: changed both as a consequence of stimuli presentation, which promotes
173: the appearance of new neuronal processes and synapses, as well as by
174: evolutionary forces acting during long periods of time.
175:
176: \begin{figure}
177: \begin{center}
178: \includegraphics[scale=.45]{diagram.eps}
179: \caption{The individual morphological features of basic
180: objects can affect the overall dynamical properties of
181: multiple object systems to which they belong either
182: directly or by determining the overall connectivity
183: between the objects. Very often, as in neuronal systems,
184: the influences are two-sided, in the sense that the overall
185: dynamics may also affect the connectivity and shape of
186: the individual constituent objects.}~\label{fig:basic}
187: \end{center}
188: \end{figure}
189:
190: Interesting as it can be, biological neuronal systems are but an
191: example of how the geometrical properties of individual components can
192: affect, or even determine, the emerging connectivity and dynamics of
193: multiple object systems. This article provides an overview of such an
194: important and relatively overlooked paradigm from the following five
195: perspectives: (a) how to measure relevant geometrical properties of
196: shapes; (b) how to obtain stochastic models of classes of shapes; (c)
197: how to simulate systems of multiple morphological objects; (d) how to
198: characterize the connectivity and topology of such systems in terms of
199: complex network concepts and measurements; and (e) how the dynamics of
200: such systems can be ultimately influenced and determined by the
201: individual morphological features of the basic constituent elements.
202: Although emphasis is placed on neuromorphic systems, the problem of
203: protein interaction is also briefly discussed from the above stated
204: perspectives.
205:
206: It should be noted that this work is relative to the oral presentation
207: of the author at the COSIN Final Meeting in Salou, Spain, which took
208: place in March 2005. Emphasis is placed on integrating several
209: related previous developments by the author under the shape-function
210: paradigm, so that this text should by no means be treated as a
211: comprehensive survey to related topics. As a consequence, great part
212: of the supplied references are to developments of the author over the
213: last decade, so that the reader is encouraged to refer to the supplied
214: references for a more comprehensive literature.
215:
216:
217: \section{Shape Measurement and Characterization}
218:
219: Given any shape, it is possible to map its morphological properties
220: into a number of measurements, which can be suitably represented as a
221: \emph{feature vector}~\cite{CostaCesar:2001}, namely a vector in $R^N$
222: where each entry corresponds to each measurement of an object under
223: analysis. For instance, given the shape of a neuronal cell, one may
224: be interested to quantify its volume,
225: orientations~\cite{L_scinst:1995}, fractal
226: dimension~\cite{L_velte:1999,LU101}, lacunarity~\cite{Gefen:1983,
227: Mandelbrot:1983, Allain_Cloitre:1991}, wavelet
228: features~\cite{L_robsproc:1997, L_biocyb:1998}, curvature and bending
229: energy~\cite{L_scientinstr:1997, L_velte:1999,LU101}, excluded volume
230: and autocorrelations~\cite{L_exclvol:2005}, as well as families of
231: Minkowski shape functionals~\cite{PRE_Marconi:2003}, to name but a few
232: among many other possibilities.
233:
234: Given such a plurality of alternatives, a fundamental question arises:
235: what is the best set of measurements? The answer to this question
236: should take into account the purpose of the measurements. For
237: instance, in case we are interested in quantifying the overall cell
238: metabolism, a potentially useful measurement would be the total volume
239: of the cell. Two of the most relevant situations in neuromorphology
240: concerns the choice of sets of measurements capable of: (a)
241: distinguishing between neuronal cell classes and (b) provide a
242: comprehensive representation of the cell geometry, in the sense that
243: the cell shape could be recovered from the measurements, i.e. the set
244: of measurements provides a \emph{complete} representation of the
245: neuronal shape. However, even if the objectives are clearly stated in
246: each of these cases, the optimal set of features is still relative to
247: the specific types of neurons under analysis. For example, in case
248: one wants to separate large from small cells, it may be enough to
249: consider the \emph{diameter} of the cells, namely the largest distance
250: between any pair of points obtained from the cell. Yet, there are
251: some properties of measurements which are generally sought in both
252: situations (a) and (b) mentioned above, especially invariance to
253: translation and rotation. Such properties, which are found in several
254: features such as area, perimeter, curvature, among others, are
255: important because often the shapes in MMOS can appear freely
256: translated and rotated and we do not want the measurement to take this
257: into account. Another relevant issue to be kept in mind is the fact
258: that the set of features should be minimal.
259:
260: As the choice of features for shape representation will be covered in
261: more detail in the next section, we present in the following some
262: general reasonings on the selection of features for the purpose of
263: shape discrimination or geometrical pattern recognition (see, for
264: instance,~\cite{Duda_Hart:2001, CostaCesar:2001}). Here the task
265: consists in, given one or more shapes, to organize then into classes
266: such that an object inside each class tends to be more similar to
267: other objects in that same class than to objects from the other
268: classes. Observe that this general problem of pattern recognition has
269: some interesting similarities with the problem of community detection
270: in complex networks. Note that we may or may not know the number of
271: classes, and that we may have or not have examples of objects from
272: each class. In case we do have examples of objects from all the
273: classes, or information about the propeties of those objects, we say
274: we have a \emph{supervised} pattern recognition problem; otherwise, it
275: becomes a \emph{unsupervised} pattern recognition
276: case~\cite{Duda_Hart:2001, CostaCesar:2001}. Needless to say, the
277: first type of problem is generally easier to be solved than the second
278: type.
279:
280: Unfortunately, the choice of features for pattern recognition is still
281: largely subjective in the sense that there are no definite rules which
282: can be applied for that end. The problem is complicated by the fact
283: that there is an infinite number of features allied to the fact that a
284: large number of measurements do not necessarily lead to better
285: discrimination (e.g.~\cite{CostaCesar:2001}). In practice, one has to
286: rely on previous experience about the problem under analysis, as well
287: as with the properties of the several available features. The
288: remainder of this section presents some brief remarks about some
289: representative shape measurements and their specific interpretation
290: and characteristics.
291:
292: {\bf Hierarchical angles and lengths:} Provided the shape can be
293: organized along hierarchical levels, namely as a branching pattern, an
294: interesting set of measurements can be obtained by taking into account
295: the angles and arclenghts of the constituent segments
296: (e.g.~\cite{L_regina:1997, L_regina:2002}). The hierarchical
297: decomposition, which is more viable when applied to branching
298: patterns, can be obtained by applying skeletonization approaches as
299: discussed in~\cite{L_santorini:1999, L_eletters:1999,
300: L_andetal:2000,L_falcao:2002,L_skel:2000, Costa_Muti:2003,
301: L_jneu:2005} or by curvature-based methods as described
302: in~\cite{L_dendro:1999}. In the case of neuronal cells or shapes
303: which have a body, it is important to detect such a
304: body~\cite{L_jneu:2005} in order to have the correct identification of
305: the beginning of each process (i.e. a dendrite or axon).
306:
307: {\bf Curvature:} Quantifying the rate of change of the tangent angle
308: of a given curve, the curvature provides invariance to translation and
309: rotation and is complete (i.e. inversible) up to such transformations.
310: The identification of high curvature points provide valuable resources
311: for identifying extremities along shapes~\cite{L_scinst:1995,
312: L_dendro:1999}. An effective approach to numerical curvature
313: estimation has been described~\cite{L_robcurv:1996,
314: L_scientinstr:1997, L_dendro:1999, L_velte:1999,
315: CostaCesar:2001,L_dsp:2003} which estimates the shape derivatives
316: required for curvature calculation by using the derivative property of
317: the Fourier transform, combined with Gaussian regularization
318: (i.e. low-pass filtering). The histogram of curvatures of a shape, as
319: well as its respective statistical moments, may also provide valuable
320: information for shape characterization and discrimination.
321:
322: {\bf Bending energy:} Given a differentiable two-dimensional shape,
323: its bending energy can be shown to be proportional to the sum of the
324: squared curvature values. Therefore, the more intrincated the shape
325: is, the higher its bending energy will result. The bending energy has
326: been used with encouraging success for classification of neuronal
327: cells~\cite{L_scientinstr:1997}.
328:
329: {\bf Fractal dimension:} The several available definitions of fractal
330: dimension provide an interesting alternative for quantifying the
331: `complexity', spatial distribution or spatial coverage of a given
332: shape (see, for instance,~\cite{CostaCesar:2001, LU101}).
333:
334: {\bf Lacunarity:} Introduced in order to complement the fractal
335: dimension~\cite{Gefen:1983, Mandelbrot:1983, Allain_Cloitre:1991}, the
336: lacunarity quantifies the degree of translational invariance of a
337: pattern. Typically, the higher the lacunarity, the less translational
338: invariant the shape is. The traditional way to calculate the
339: lacunarity involves sliding a window along the whole image while
340: estimating the mass comprised by the window. Several sizes of windows
341: are usually considered, yielding the lacunarity as function of the
342: windows size. A recent study~\cite{L_self:2005} has shown that
343: translational and rotational invariance can be achieved by using
344: circular windows centered only at the object points, not through the
345: whole workspace.
346:
347: {\bf Minkowski functionals:} As implied in the name, shape functionals
348: are maps from a given shape to scalar values. Minkowski functionals
349: are a special class of functionals which are additive, motion
350: invariant and continuous~\cite{Michiel1:2000, Michiel2:2001,
351: Mecke:1998}. In the plane, such functionals comprise the perimeter,
352: area and Euler number, corresponding to the number of holes in the
353: object under analysis. Minkowski functionals have been successfully
354: applied to the characterization of neuronal
355: shape~\cite{PRE_Marconi:2003, L_morphalt:2004}.
356:
357: {\bf Critical percolation densities:} Given a shape and an empty
358: workspace, the progressive superposition of this shape at uniformly
359: distributed random positions along the workspace eventually lead to
360: percolation, which can be identified by looking for a cluster of
361: shapes extending from side to side of the
362: workspace~\cite{Percolation:2003}. The average density of shapes
363: observed when percolation occurs, called henceforth the critical
364: percolation density, has been suggested~\cite{Percolation:2003} as the
365: most direct measurement of the potential of the given shape for
366: establishing connections, leading to promising discriminative
367: potential. Interesting previous investigations of percolation in
368: neuronal systems, where the neuronal shape was simplified as circles,
369: have been reported in~\cite{Ooyen:1995}. The critical percolation
370: density can also be estimated for ballistic deposition, as
371: investigated in~\cite{L_ballistic:2005}. In the case the shapes
372: change with time, as is the case with growing neurons, it is possible
373: to define a critical percolation density which is a direct consequence
374: of the growth dynamics and spatial distribution of the
375: cells~\cite{L_regin:2004}. In the case of a collection of not
376: necessarily connected and static objects, it is still possible to
377: force percolation through some imposed growth dynamics, such as
378: parallel dilation~\cite{L_bioinfo:2005, L_active:2004, L_active:2005}.
379:
380: {\bf Relative shape measurements:} While all the above measurements
381: are specific only to the shape of interest, it is also important to
382: consider geometrical features that also take into account
383: environmental constraints such as the presence of other objects, as
384: well as fields and other types of interactions. One example of such
385: measurements would be to consider the angle difference between the
386: tangent field orientation along the object, which is a measurement
387: intrinsically related to the shape, and the electric field orientation
388: at those same points induced by the presence of surrounding objects.
389: Low averages of such ratios would therefore indicate that the shape in
390: question is highly affected by the external objects. Such
391: measurements are particularly important for studying and modeling MMOS
392: whose basic objects are known to be affected by environmental
393: influences, as is the case with neuronal systems.
394:
395:
396: \section{Stochastic Shape Modeling and Synthesis}
397:
398: In this section we discuss the problem of how to produce, by
399: mathematic-computational means, artificial instances of shapes of a
400: given category. For instance, one may be interested in obtaining a
401: collection of neuronal cells which are statistically equivalent to
402: those of a certain biological class (e.g. alpha cat retinal gangion
403: cells). By \emph{statistical equivalence} we mean that the
404: synthetized cells will produce the same statistical distribution of
405: values of geometrical measurements. Observe that, given the almost
406: unlimited statistical variability of the properties of natural shapes,
407: even within the same category of shapes, the problem of shape
408: equivalence in the sense of exactly reproducing a given shape is of
409: minor interest. On the contrary, it becomes more important to devise
410: synthesis methodologies capable of incorporating the stochastic
411: variability found in nature.
412:
413: Two important situations arise: (i) the original shapes evolve without
414: influences from the environment and (ii) the shapes are affected by
415: the environment. The problem of stochastic shape synthesis in these
416: two situations are discussed in the following subsections.
417:
418:
419: \subsection{Absence of environmental influences}
420:
421: In this case, the shapes of interest are the result of a growing
422: dynamics which takes into account only internal constraints imposed by
423: the physics/biology of the object. While there are no natural
424: examples of such growth dynamics, this approximation provides a first,
425: simple approach to modeling and simulating shapes. In the following
426: we illustrate this kind of strategy with respect to the simulation of
427: neuronal outgrowth~\cite{L_regina:1997, L_regina:2002}. Other
428: approaches to neuromorphic modeling include those adopting differential
429: equations to model cytoskeleton outgrowth ~\cite{Hentsch1:1994,
430: Hentsch2:1998, L_dania:2003, L_andneuro:2001}, methods based on
431: Hillman's~\cite{Hillman:1979} set of quantitative anatomical
432: correlations~\cite{GA_hill1:2000, GA_hill2:2001}, activity and
433: competition during process outgrowth~\cite{Ooyen:1995, Ooyen:1999,
434: Ooyen:2001, Ooyen:2004} as well as hidden Markov
435: models~\cite{GA_hipp:2005}.
436:
437: The overall idea of the conditional probability
438: model~\cite{L_regina:1997, L_regina:2002} involves obtaining
439: representative measurements from the real biological cells, such as
440: the angles and lengths along the branching hierarchies, represented as
441: conditional densities, and then Monte Carlo sampling such
442: distributions in order to produce synthetic neuronal cells.
443: Figure~\ref{fig:growth} illustrates three stages along the development
444: of simulated alpha and beta cells. In both cases, percolation is
445: observed between the growth stages illustrated in the figure. Although
446: this example considers identical cells, which reduces the variance of
447: the percolation critical density, simulations assuming different
448: neuronal cells have are also important and have been investigated.
449:
450: Given a collection of adult neuronal cells, each of them can be imaged
451: in 2D or 3D by using several types of microscopy, including optical
452: transmission and confocal. Although the following discussion is
453: limited to 2D neuronal shapes (several neuronal types, such as retinal
454: ganglion cells, are mostly planar), the extension to 3D is immediate.
455: The image of the neuronal cell can then be processed in order to
456: obtain the points belonging to the cell and the points belonging to
457: the background. The contour of such cells is particularly important
458: and can be obtained by using edge detection and contour following
459: algorithms (e.g.~\cite{CostaCesar:2001}). The particularly important
460: points of the neuronal ramification, namely the extremity and
461: branching points, can be obtained by applying a curvature-based
462: procedure~\cite{L_linseg:1995, L_dendro:1999} or skeletonization
463: algorithms~\cite{L_andetal:2000, CostaCesar:2001}. The angles and
464: arclengths along each segment along the hierarchies of the
465: ramifications can then be obtained by using simple analytical geometry
466: concepts and methods, so that conditional probability densities of the
467: angles and lengths can be obtained. As discussed in the previous
468: section, such measurements provide, by themselves, a naturally
469: suitable characterization of the neuronal geometry. Note that the
470: number of hierarchies considered in the conditional densities is
471: limited by the amount of neuronal samples, in the sense that the
472: consideration of several hierarchies implies higher dimensionality of
473: joint densities which require larger numbers of cells in order to
474: obtain statistically representative results (i.e. a statistical
475: sampling problem). Finally, the conditional densities can be Monte
476: Carlo sampled in order to produce a virtually infinite number of
477: synthetic cells whose geometrical features will lead to statistical
478: densities equivalent to those which were sampled.
479:
480: \begin{figure*}
481: \begin{center}
482: % \includegraphics[scale=.4]{alfas2.eps} \hspace{0.5cm}
483: % \includegraphics[scale=.4]{betas2.eps}
484: % (a) \hspace{5.5cm} (d) \vspace{0.5cm} \\
485: \includegraphics[scale=.4]{alfas5.eps} \hspace{0.5cm}
486: \includegraphics[scale=.4]{betas5.eps}
487: (a) \hspace{5.5cm} (c) \vspace{0.5cm} \\
488: \includegraphics[scale=.4]{alfas12.eps} \hspace{0.5cm}
489: \includegraphics[scale=.4]{betas9.eps}
490: (b) \hspace{5.5cm} (d)
491: \caption{Three stages along neuronal growth simulations for alpha
492: (a-b) and beta (c-d) cells. In this example, all cells are
493: identical.}~\label{fig:growth}
494: \end{center}
495: \end{figure*}
496:
497:
498:
499:
500: \subsection{Presence of environmental influences}
501:
502: The more realistic situation of shape development under environmental
503: influcences requires a modeling approach capable of integrating
504: \emph{both} internal and external influences, because the former are
505: always present in physical and biological systems. Often, external
506: influences manifest themselves in terms of \emph{fields} or
507: \emph{waves}. For instance, the growth cones (namely the tips of
508: growing axons) are known to follow gradient fields derived from
509: neurotrophic growth factors, ionic concentrations, electric, and even
510: gravitational fields. Therefore, it is reasonable to incorporate the
511: effect of such fields in neuronal outgrowth by adding to the instant
512: growth velocity (implied by internal influences) a vector component
513: parallel to the external field (e.g.~\cite{L_consul:1998,
514: L_andneuro:2001, L_dania:2003}). It has been
515: verified~\cite{GA_repuls:2003} that the incorporation of a vector
516: component pointing away from the cell soma also accounts for
517: biological realism in neuronal outgrowth.
518:
519: Because the neuronal activity during neuronal development can also
520: have effects in shaping the neuronal morphology, it is interesting to
521: consider such dynamics in simulations. We can use the recently
522: introduced Sznajd complex networks~\cite{L_sznaj:2005}, whose topology
523: is determined by correlations between the neuronal
524: connectivity/activity, in order to implement Hebbian dynamics, where
525: the most active connections are reinforced. Another effect to be
526: considered is the fact that waves of ionic concentration have been
527: verified to induce neuronal spikes~\cite{L_rpozn2:2001}. Therefore,
528: it would be interesting to include such waves in neuronal outgrowth
529: simulations.
530:
531:
532: \section{Synthesis of Morphic Multiple Objects Systems and their
533: Representation and Characterization in Terms of Complex Networks}
534:
535: Provided we have implemented the means to grow individual shapes, they
536: can be combined in order to produce morphic multiple object systems.
537: The most natural means to do that involves starting the growth at $N$
538: seeds distributed spatially according to some desired statistical
539: density. In the case of neurons, it is often reasonable to assume a
540: uniform distribution of cells (see~\cite{Costa_APL1:2005} for a
541: discussion on the measurement of spatial dispersion in neuronal
542: mosaics), which are then grown by using a combination of the
543: techniques discussed in the previous section. Every time an axon gets
544: with a pre-specified distance from a dendrite, it may be connected to
545: that dendrite with some probability which may be a function of the
546: distance as well as other factores (e.g. diameter of the axon and type
547: of cells, among many other possibilities)~\cite{L_regina:1997,
548: L_regina:2002}.
549:
550: Fortunately, MMOSs can be effectively represented in terms of
551: graphs/networks and their respective adjacency or weight
552: matrices~\cite{L_regina:1997, Costa_BM:2003, L_analyt:2003,
553: L_regin:2004}. As such systems grow, it is possible to keep track of
554: every new connection (a synapse in the case of neuronal systems),
555: which can be updated into the adjacency or weight matrix of the graph
556: used to represent the MMOS. In the case of neuronal networks, such a
557: representation is almost complete, except for the fact that it does
558: not take into account fields and waves emmanating from the
559: environment~\cite{L_rpozn2:2001}. However, such effects can be easily
560: incorporated by using geographical networks for representing the
561: system, therefore providing information about the spatial position of
562: the nodes and allowing the determination of the fields at those
563: points.
564:
565: One of the most interesting perspective in applying complex network
566: measurements for characterization and analysis of growing MMOS regards
567: obtaining each measurement as a function of time or development
568: epochs. For instance, following traditional complex network
569: measurements such as the node degree and clustering coefficient in
570: terms of time can provide valuable information about the local
571: connectivity of the objects as the system evolves. Virtually every
572: complex network measurement will have some additional aspect to say
573: about the evolving connectivity. More global information about the
574: system connectivity can be gathered by using the recently introduced
575: hierarchical versions of the node degree and clustering coefficient,
576: as well as several new hierarchical
577: measurements~\cite{Costa_Hier:2004, Costa_Hier:2005, Costa_Gener:2004}
578: and wiring lengths between neuronal modules~\cite{L_topo:2005}. The
579: application of community finding algorithms can also provide
580: particularly interesting information with implication to the
581: functional dynamics of the MMOS under analysis, with important
582: implications to the evolution of synchronization and attractors
583: formation. Another promising perspective is the quantification of the
584: cycles of several lengths in the growing networks, which will also
585: have important implications for the system dynamics.
586:
587:
588: The most direct quantification of the evolution of the overall
589: connectivity of the system is provided by the critical percolation
590: density. We have applied static and dynamics, natural and forced
591: percolations~\cite{L_active:2004} in order to investigate the
592: influence of the individual shape of neurons to the overall
593: connectivity~\cite{Percolation:2003, L_regin:2004}. In investigations
594: considering alpha and beta cat ganglion cells with normalized sizes,
595: we observed that more complex cells tend to percolate sooner.
596: However, the critical percolation density or time (in the case of
597: growing systems with constant density) has been verified to be a
598: function of several other parameters, such as the neuronal cell
599: elongation, the excluded volume, as well as the straightness of the
600: neuronal processes. Further investigations are required in order to
601: establish a clear relationship between the individual object
602: properties, as well as their growth dynamics, with the final overall
603: MMOS connectivity.
604:
605:
606: \section{Consequences on the Overall System Dynamics}
607:
608: Because the individual morphological properties of the objects have
609: been found to determine, jointly with their spatial distribution, both
610: the local and global connectivity of MMOSs, such geometrical features
611: will also be inexorably related to the emmerging dynamics of the
612: system. After investigating the effect of using scale free
613: connectivity in Hopfield networks~\cite{Stauffer_Costa:2003,
614: Costa_Stauffer:2003}, we have been investigating how the individual
615: neuronal shape can affect neuronal dynamics~\cite{Costa_BM:2003}. We
616: have quantified the memory potential of the neuronal systems in terms
617: of the \emph{overlap}, which is related to the number of correct
618: recovered bits, between the originally trained pattern and the pattern
619: recovered (starting from a perturbated version of the original
620: pattern) after $P$ patterns had been trained into the neuronal system.
621: Considering small neurons with the same area, we have
622: found~\cite{L_analyt:2003} that such a memory potential is highly
623: dependent on the individual neuronal shape, with shapes exhibiting
624: ramified and broader distribution of mass (i.e. neuron-like) tending
625: to perform substantially better than simple shapes such as bars and
626: crosses. The distribution of complex eigenvalue of the adjacency
627: matrices obtained with neuron-like cells have resulted less
628: degenerated, which is known to be associated with better recall
629: capabilities~\cite{Haykin:1999}.
630:
631:
632: \section{Conclusions and Future Works}
633:
634: The present text provided an integrated review and discussion on the
635: problem of relating shape and function through connectivity in morphic
636: multiple object systems (MMOS). The main paradigm here is the fact
637: that the geometric features at the individual level can play a
638: decisive role in defining the overall connectivity of the system, and
639: consequently its dynamical operation. While the quantification of
640: geometrical features of the individual objects can be achieved by
641: considering a variety of measurements, we have argued that the
642: resulting connectivity of the respective MMOS can be effectively
643: quantified and modeled in terms of complex network concepts and
644: measurements.
645:
646: Although the shape/function paradigm has been clearly identified and
647: investigated in preliminary works, a series of important problems
648: remain open to research. Of particular interest would be to identify
649: how each of the main shape measurements are correlated to the overall
650: connectivity, as quantified by complex network measurements. It would
651: also be interesting to develop further analytical models such as that
652: presented in~\cite{L_analyt:2003} so that the several topological
653: properties of the MMOS can be foreseen from the geometrical features
654: of the constituent objects. One particularly interesting approach
655: would be to try to relate the fractal dimension of the individual
656: neuronal cells with the fractal dimension of the overall topology of
657: complex networks~\cite{Song_etal:2005}, and then with scale-free laws
658: and other properties of the emmerging connectivity and activity
659: dynamics of the respective networks. Other promising perspectives
660: include the analysis of recall in neuronal systems other than
661: Hopfield, as well as the quantification of synchronization between
662: cells as the MMOS grows. The most realistic situation where the
663: connectivity dynamics is related to the network activity dynamics
664: along time and space remains one of the most interesting and
665: challenging situation for simulations and investigations. We are
666: working on such issues and would highly appreciate to consider
667: collaborations.
668:
669: \vspace{0.5cm}
670: {\bf Acknowledgments:} Luciano da F. Costa is grateful to
671: Regina C. Coelho (UNIMEP, Brazil), who kindly supplied the simulation
672: data shown in Figure 2, and to HFSP RGP39/2002, FAPESP
673: (proc. 99/12765-2) and CNPq (proc. 301422/92-3) for financial support.
674:
675:
676:
677: \bibliographystyle{unsrt}
678: \bibliography{salou}
679:
680: \end{document}
681:
682:
683:
684:
685:
686:
687:
688:
689:
690:
691:
692:
693:
694:
695:
696:
697:
698:
699:
700:
701:
702:
703:
704:
705:
706:
707:
708:
709:
710:
711:
712:
713:
714:
715:
716:
717:
718:
719:
720:
721: