DYNAMICAL ARCHITECTURES FOR CONTROLLING FEEDING IN APLYSIA CALIFORNICA
by KENDRICK M. SHAW
Submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy
Dissertation Adviser: Dr. Hillel J. Chiel Dissertation Co-Adviser: Dr. Peter J. Thomas
Department of Biology CASE WESTERN RESERVE UNIVERSITY
January 2014 CASE WESTERN RESERVE UNIVERSITY SCHOOL OF GRADUATE STUDIES
We hereby approve the dissertation of Kendrick Matthew Shaw Candidate for the Doctor of Philosophy degree1.
Robbin E. Snyder Hillel J. Chiel Peter J. Thomas Dominique M. Durand Scott E. Cooper
October 16th, 2013
1We also certify that written approval has been obtained for any proprietary information contained within. Contents
Acknowledgements...... 9
Abstract...... 11
1 Introduction...... 13 1.1 Background and history...... 15 1.1.1 Chain reflexes and central pattern generators...... 15 1.1.2 Mathematical background...... 17 1.1.3 Mathematical models of neurons...... 25 1.2 The choice of Aplysia californica as a model system...... 26 1.3 Mathematical framework and central hypotheses...... 28 1.4 Outline of the remainder of the dissertation...... 31
2 Significance of Dynamical Architecture...... 33 2.1 Introduction...... 34 2.2 Mathematical Framework...... 37 2.2.1 Limit cycles...... 38 2.2.2 Destabilization of fixed points...... 39 2.2.3 Stable heteroclinic channels...... 39 2.3 Model Description...... 41
3 CONTENTS 4
2.3.1 Neural model...... 41 2.3.2 Model of the periphery and load...... 44 2.3.3 Proprioceptive input...... 52 2.3.4 Noise...... 52 2.3.5 Parameter changes used for the limit cycle simulations...... 54 2.3.6 Connection to mathematical framework...... 55 2.4 Materials and Methods...... 55 2.4.1 Intact animals...... 56 2.4.2 Suspended buccal mass preparation...... 56 2.4.3 Isolated buccal ganglion...... 57 2.4.4 Data analysis...... 57 2.4.5 Numerical methods...... 58 2.5 Results...... 59 2.5.1 Tuning the limit cycle...... 59 2.5.2 Mechanisms of adaptation to load...... 65 2.6 Discussion...... 75 2.6.1 Limitation of the model and results...... 78 2.6.2 Larger implications for pattern generators...... 80
3 Phase Resetting in a Phaseless System...... 84 3.1 Introduction...... 85 3.2 The piecewise linear iris system...... 95 3.3 Limit Cycles in the Iris System...... 102 3.3.1 Dynamics Within A Linear Region...... 103 3.3.2 Dynamics Across Regions...... 105 CONTENTS 5
3.4 Effects of a small instantaneous perturbation...... 109 3.4.1 Initial effects of a small perturbation...... 109 3.4.2 Subsequent effects of a perturbation...... 113 3.4.3 Infinitesimal phase response curve...... 117 3.5 Asymptotic phase resetting behavior as a 0...... 119 → 3.6 Isochrons...... 124 3.7 Smooth System...... 125 3.8 Discussion...... 127 3.8.1 Sensitivity and control...... 129 3.8.2 Comparison to the PRC near a homoclinic bifurcation...... 131 3.8.3 Qualitative comparison with a biological model: iPRCs for the Morris–Lecar system...... 134 3.8.4 Stability of synchronous solutions for two iris systems with diffu- sive coupling...... 140 3.8.5 Generalization to higher dimensional systems...... 146 3.8.6 Phase resetting in the absence of an asymptotic phase...... 148
4 Conclusion and Future Directions...... 151 4.1 Review of previous chapters...... 152 4.2 Future directions...... 156 4.3 Conclusion...... 163
Bibliography...... 165 List of Tables
2.1 Model parameters...... 47 2.2 State variables...... 47 2.3 Parameters used for the limit cycle simulations...... 53
3.1 Comparison of iPRCs of homoclinic, QIF, and Iris systems...... 134
6 List of Figures
2.1 Isolated trajectories of the SHC and limit cycle...... 43 2.2 Phases of swallowing behavior in the model...... 45 2.3 Model schematic...... 46 2.4 Timing dependence of the limit cycle...... 60 2.5 Improved efficacy with increased maximum muscle activation...... 62 2.6 Metabolic cost of increased muscle activation...... 63 2.7 Mechanical efficiency of the limit cycle...... 64 2.8 Effects of mechanical load on timing...... 67 2.9 Effects of mechanical load on trajectory...... 68 2.10 Simulation of held seaweed...... 70 2.11 Effects of held seaweed in vivo ...... 71 2.12 Simulation of reduced proprioception...... 72 2.13 Behavior of reduced preparations...... 74 2.14 Skewness of retraction duration in simulations...... 76 2.15 Skewness of retraction duration in vivo ...... 77
3.1 Smooth system phase plot...... 88 3.2 Smooth system time plot...... 90
7 LIST OF FIGURES 8
3.3 Iris system schematic...... 95 3.4 Iris system phase plots...... 96 3.5 Iris system time plots...... 101 3.6 Iris square entry to exit map...... 108 3.7 Iris bifurcation diagram...... 110 3.8 Iris iPRCs...... 120 3.9 Iris isochrons...... 126 3.10 Smooth system iPRCs...... 128 3.11 Morris–Lecar time plots...... 137 3.12 Morris–Lecar iPRC...... 139 3.13 Coupled Iris systems...... 142
4.1 PRC for the limit cycle neuromechanical model...... 157 4.2 PRC for the heteroclinic channel neuromechanical model...... 158 4.3 Model with biological noise...... 160 4.4 Trade off between proprioception and noise...... 162 4.5 Saddle bypass diagram...... 163 Acknowledgments
I would like to thank my research advisor, Dr. Hillel Chiel, for his patience, friendship, and guidance throughout my time in the laboratory. His constant encouragement and relentless striving for perfection has helped me and many others accomplish what we did not think was possible. Without him this research would not have happened, and his wisdom will shape my future work. I would also like to thank my co-advisor, Dr. Peter Thomas, for his enthusiasm and guidance in introducing me to the world of mathematical biology, and his guidance on how to bound a problem well enough that it can be solved with rigor. I would like to thank Dr. Robbin Snyder, for her friendship and advice in navigating academia, Dr. Scott Cooper, for his careful clinical tutorship and advice on blending clinical work and basic research, and Dr. Dominique Durand, whose occasional skepticism has helped drive me to strengthen this work. I would like to thank my lab mates, Hui Lu, Jeffrey McManus, Miranda Cullins, Catherine Kehl, and Jeffery Gill for their friendship and useful feedback. Miranda Cullins, Jeffrey McManus, and Hui Lu also provided the in vivo and in vitro behavior and burst onset and offset times used in chapter2, which I am grateful for. I would also like to thank Andrew Horchler for his friendship, advice, and many good conversations on the behavior of heteroclinic channels and stochastic simulations. I am also indebted to my friends Barry Rountree, Eric Herman, and Stephanie Medlock. Their enthusiasm in tackling difficult problems together for the fun of it carried over into my work in the lab, and their encouragement and patience with me during difficult times Acknowledgments 10
in my research helped me have faith that each obstacle could be overcome. I would like to thank Dr. Cliff Harding, Dr. George Dubyak, and the other members of the MSTP steering committee for their advice and guidance of my training of a clinician– scientist, and how to develop this work in this larger context. I would also like to thank my parents, family, and close friends in Seattle for supporting me in pursing this work. Finally I would like to thank the various sources of funding that have supported this work, including the Case MSTP (NIH T32-GM007250), NIH NS047073, and the Case Innovation Achievement Award Fellowship. Dynamical Architectures for Controlling Feeding in Aplysia californica
Abstract by KENDRICK M. SHAW
For behaviors such as swallowing, walking, and swimming, the nervous system must reliably generate sequences of motor behavior. Two competing models have been proposed for how this task is accomplished - chain reflex theory and central pattern generator theory. Chain reflex theory posits that the nervous system contains a sequence of reflexes, so that the action of one reflex creates the sensory input required to trigger the next. In contrast, central pattern generator theory posits that the nervous system is capable, in the absence of sensory input, of generating motor patterns that closely resemble the motor patterns during behavior. When modeling these behaviors with systems of differential equations, these two ideas correspond to a collection of stable nodes, in the case of the chain reflex theory, and a stable limit cycle, in the case of central pattern generator theory. Many systems can exhibit motor patterns in the absence of sensory input, violating the predictions of chain reflex theory, but those patterns are very distorted compared to in vivo behaviors, violating the assumptions of central pattern generator theory. This dissertation explores a third hypothesis, known as a stable heteroclinic channel, where a trajectory slows dramatically Abstract 12 in small regions as it passes near saddle points, creating local regions of sensitivity. We explore the implications of these dynamics by building a neuromechanical model of swallowing in Aplysia californica which can be changed from a stable heteroclinic channel to a limit cycle by changing a single parameter, and then compare the behavior within these two regimes to the behavior seen in vivo. We find that the stable heteroclinic channel provides a better match for what is seen in vivo, due to its timing sensitivity. We then analytically explore the basis for this sensitivity by studying a tractable heteroclinic channel and deriving a closed-form expression for its infinitesimal phase response curve. We show that the qualitative behavior of the tractable model is present in more complex models, including the Morris-Leccar neuron model as it approaches the homoclinic bifurcation. We then discuss the implications of this research for future work in motor pattern generation. Chapter 1
Introduction
13 CHAPTER 1. INTRODUCTION 14
To generate many forms of adaptive behavior, the nervous system needs to reliably generate sequences of motor activity. For example, as we walk, the nervous system must reliably sequence swing and stance behavior for each leg or risk stumbling (Winter, 1991). Similarly, a failure to generate the proper sequence of motor activity in swallowing, for example, failure to close the larynx before peristalsis carries the food bolus past, can lead to aspiration of food and serious sequellae such as aspiration pneumonia (Denaro et al., 2013). Indeed, aspiration pneumonia is a leading cause of death after a stroke (Henon´ et al., 1995). Generating the correct sequence of activity is often not enough, however; the system must also be sensitive enough to adjust the duration of components of the pattern in response to incoming sensory information. If a cat’s leg encounters an obstacle during its swing, the nervous system will increase the flexion of the leg and extend the length of the swing phase to lift the foot over the obstacle, especially if contact occurs late in the swing phase (Forssberg, 1979). Similarly, the animal may extend the touch-down phase of the walking cycle if the animal steps into a hole, allowing more time for the foot to reach the ground (Halbertsma, 1983; Hiebert and Pearson, 1999). In human swallowing, the oral and pharyngeal stages of swallowing nearly double in duration when the bolus consistency is changed from liquid to semisolid (McHorney et al., 2006). It is not clear how the nervous system achieves this balance between robustness and sensitivity. Changes that make the system more sensitive to incoming sensory information often also allow irrelevant sensory information and noise to corrupt the pattern, and changes that make the pattern more robust to irrelevant sensory inputs and noise often make it less sensitive to relevant sensory information. The challenge of balancing these seemingly contradictory goals will be a recurring theme in this dissertation. CHAPTER 1. INTRODUCTION 15
1.1 Background and history
1.1.1 Chain reflexes and central pattern generators
Much of the early research in motor control focused on the question of sensitivity, specif- ically the ability of the nervous system to adapt the nature and timing of behavior to incoming sensory stimuli. Loeb(1899) suggested that feeding, reproductive, and other behaviors in the frog and other “lower” animals could be thought of as examples of a “kettenreflex”, or chain reflex, where the behavior was composed of very simple reflexes such that the action of one reflex triggered the next reflex. For example, the sight of a fly in the periphery might cause the animal to turn its head towards the fly, the sight of the fly in front of the head might trigger an opening of the jaws and extension of the tongue, and so on throughout the entire behavior. Sherrington later applied the chain reflex theory to “higher” vertebrates in his studies of walking in cats and dogs after a spinal transection (Sherrington, 1910, 1913). As support for this theory, he noted that lifting and dropping the animal’s hind paw was sufficient to trigger a series of stepping movements, and catching the paw part way through the walking cycle would immediately stop the walking cycle. Although often viewed as a champion of chain reflex theory, however, even Sherrington was quick to note its limitations, noting that electrical stimulation of the spinal cord would still sometimes trigger a stepping motion in deafferented limbs and isolated muscles, and thus conceded that this principle of chain reflexes “is of itself not the sole rhythm-producing factor in the reflex,” (Sherrington, 1910). Chain reflex theory enjoyed many proponents throughout the early part of the 20th century, although some still argued for the existence of intrinsic oscillators in the nervous system generating the timing of behavior (Marder and Bucher, 2001). The consensus CHAPTER 1. INTRODUCTION 16
of the scientific community began to shift in the 1960’s, however, to theories involving central oscillators, due to accumulating examples of systems that could generate patterns of activity in the absence of sensory input that resembled the motor activity seen in the intact animal. One of the early examples of such a system was flight in the locust (Wilson, 1961). Wilson was able to selectively cut all nerves containing sensory afferents from the wings, yet he was still able to produce flight-like activity in the wing muscles when air was blown on the head of the animal. In particular, many began to argue not only for intrinsic neural oscillators playing a role in oscillatory behavior, but that those oscillators could produce physiologic motor patterns in the absence of sensory input carrying timing information. Such oscillators were given the name central pattern generators (CPGs) (Marder and Bucher, 2001). In its simplest form, this theory would posit that the central nervous system can generate an “ideal” motor pattern in the absence of sensory input, and that sensory input merely tunes this pattern in a mild way. In this pure form, the central pattern generator provides tremendous robustness, in that it can produce useful patterns under a large range of sensory inputs, and even do so when sensory input is entirely lost. While the patterns observed in vitro did share much of the structure and sequence of in vivo motor patterns they were not the near copies of in vivo activity that the idealized central pattern generator theory would predict. Instead, the in vitro patterns were often distorted and slow compared to the patterns seen in vivo; for example in Wallen´ and Williams(1984) the swimming frequencies observed in an in vitro lamprey preparation were about half of those observed in vivo, and in Chrachri and Clarac(1990) the backward walking frequencies in a reduced crayfish preparation were 10 to 30 times slower than those seen in vivo. These differences in in vitro and in vivo patterns led even some of the early proponents of central pattern generators to sound a note of caution. Robertson CHAPTER 1. INTRODUCTION 17
and Pearson, two of the early researchers of the nervous system controlling flight in the locust, noted “Although now abundantly clear that a central rhythm generator can produce powerful oscillations in the activity of flight motor neurons and interneurons, it is equally clear that the properties of this central oscillator cannot fully account for the normal flight pattern” (Selverston, 1985). Thus the nervous system seems to have neither the dependence on sensory input for rhythmic behavior predicted by the chain reflex theory nor the ability to generate physiological patterns in the absence of sensory input predicted by the idealized central pattern generator theory. What lies between these two extremes? Before exploring this question, we will first introduce a mathematical framework within which we can represent and compare different classes of motor control models.
1.1.2 Mathematical background
In this section we will define what we mean by a dynamical system, and then provide definitions of some structures that a dynamical system may have that we will use later in the paper. We also briefly discuss stochastic dynamical systems.
1.1.2.1 Dynamical systems
A dynamical system is a formalization of the idea of a deterministic process. Formally, it
can be defined as a triple (T,X,φ(x,t)) where T is a number set representing time, X is a state space, and φ : X T X is an evolution operator with the properties φ(x,0) = x and × → φ(x,t +s) = φ(φ(x,t),s) for all x X and t,s T (Kuznetsov, 2004). This can be thought ∈ ∈ of as a set of rules, φ, for calculating the new state of the system φ(x,t) based on an initial state of the system x X and an amount of time that has elapsed t T. The two properties ∈ ∈ of φ capture the intuitions that the system should require a non-zero amount of time to CHAPTER 1. INTRODUCTION 18
evolve, and that the effects of advancing the system by a fixed amount of time should be the same whether the system is advanced in a single step or multiple steps. Although this second rule may at first seem to exclude systems that depend on an absolute time, such as a system forced by an external input, these “non-autonomous” systems can be represented by augmenting the state with a new variable t T containing the absolute time that evolves ∈ at the appropriate rate, i.e. φ((...,t),s) = (...,t + s) for all t,s T. ∈ For this dissertation, we will mostly restrict our discussion to the subset of dynamical
systems where time is a real number (i.e. T = R), the state of the system can be represented n + as a vector of real numbers (i.e. X = R ,n Z ), and the evolution operator φ(x0,t) ∈ represents integration of the ordinary differential equation
dx = f (x), (1.1) dt
with the initial condition x0 over a time t, where f : X X. We will also assume we have → chosen a suitable distance metric on X (for example the L2, or Euclidean, norm).
1.1.2.2 Orbits, limit cycles, and fixed points
In the context of dynamical systems, an orbit is, intuitively, the path traced out by a point as it evolves forward and backward through time. More formally, it is a set of points A X ∈ such that for any two points x0,x1 A, there exists a time t T such that x1 = φ(x0,t). ∈ ∈ Note that this definition implicitly provides a unique mapping from every point x X to ∈ an orbit containing x. Note also that this definition is different than the colloquial use of the word “orbit”, in that an orbit need not form a closed loop. The colloquial use of the word orbit maps a bit more closely onto a different structure, known as a limit cycle. A limit cycle is an isolated periodic orbit. Here periodic is taken in its normal sense, CHAPTER 1. INTRODUCTION 19
namely that there exists a period τ > 0,τ T such that for any point x in the limit cycle ∈ x = φ(x,τ), and τ is the smallest positive value for which this equality holds1. By isolated, we mean that there is some (possibly small) open neighborhood around the orbit that does not contain other periodic orbits; this restriction is present to exclude, for example, the infinite set of concentric circles in a uniformly rotating disc. We can state this restriction
more precisely by saying that there exists a distance ε > 0 such that for every point x in
the limit cycle, for all points x′ within a distance ε of x that are not part of the limit cycle,
the orbit containing x′ is not a periodic orbit. A fixed point is a state of the system that does not change over time. It can be equivalently defined either as an orbit containing a single point, as a point x such that
φ(x,t) = x for all t T, or as a point x such that f (x) = 0 (where f is the function used in ∈ (1.1)). Note that although formally a point is not the same as a set containing a single point, in practice these definitions of a fixed point are used interchangeably with the obvious conversion between a point and a set containing only that point implied.
1.1.2.3 Saddle points, heteroclinic cycles, and heteroclinic channels
There are many different types of fixed points, but we will be particularly interested in a type of fixed point known as a standard hyperbolic saddle point. Assume that f is
differentiable in an open neighborhood around a fixed point. That is, if f (x) = 0, then d f /dx is a well defined continuous function. The fixed point is then a standard hyperbolic saddle iff the set of eigenvalues of the Jacobian of f evaluated at this fixed point contains one or more eigenvalues with a positive real component, one or more eigenvalues with a negative real component, and does not contain any eigenvalues whose real component is zero. Intuitively, this means that the point will attract trajectories along some directions
1The equality will obviously hold for integer multiples of τ due to the previously mentioned constraints on φ. CHAPTER 1. INTRODUCTION 20
and repel them along other directions. There will be a set of points Γ−, known as the stable
manifold, that asymptotically approach the fixed point over time, i.e., if the saddle is xs,
then the stable manifold is the set x X limt ∞ φ(x,t) = xs . There will also be a set { ∈ | → } of points that asymptotically approach the saddle backwards in time, Γ+, known as the
unstable manifold, i.e. the set x X limt ∞ φ(x,t) = xs . { ∈ | →− } + Almost all of the orbits near the saddle will not be contained in either Γ or Γ−, + however. Generically, Γ− and Γ will be curves or (hyper) surfaces and thus have zero “volume” (more formally zero measure); thus almost all of the orbits near the saddle simply pass by the saddle rather than originating or terminating at the saddle. These paths that pass near the saddle will be of interest to us in later sections. If we assume that f is smooth
(infinitely differentiable) within a neighborhood of the saddle xs, then the magnitude of
f will become arbitrarily small in a neighborhood around xs. Intuitively, this means that almost all of the trajectories near the saddle will slow down as they approach the saddle, travel slowly through the neighborhood of the saddle, and accelerate as they leave the saddle. In some systems, an orbit may be contained in both the unstable manifold of one saddle and the stable manifold of a second saddle, such that it intuitively originates at the first saddle and terminates at the second. Such an orbit is known as a heteroclinic orbit. If there
is a sequence of saddle points xs,0,xs,1,...,xs,n and a corresponding set of heteroclinic orbits connecting them into a cycle such that the unstable manifold of one node intersects the stable manifold of the next node in the sequence (with the first node following the last), we refer to the union of the saddles and their connecting orbits as a heteroclinic cycle. For simplicity, we will restrict our attention to the case where each of the saddles has a 1-dimensional unstable manifold. A heteroclinic cycle is structurally unstable, meaning that an arbitrarily small change to CHAPTER 1. INTRODUCTION 21
the vector field described by f can cause the unstable manifold of one saddle to “miss” the stable manifold of the next saddle. Because of this brittle dependence on the structure of f , in the absence of certain symmetries (Guckenheimer and Holmes, 1988) one is unlikely to see heteroclinic cycles in real-world systems. When a heteroclinic cycle is destroyed, however, a “ghost” of sorts remains; for sufficiently small perturbations when f is smooth, there will still be trajectories that pass by each saddle in sequence (Afraimovich et al., 2004a,b). Furthermore, certain categories of perturbations will produce a limit cycle which passes near each saddle. An example of such a category is perturbations that cause the unstable manifold of each saddle to pass “inside” the stable manifold of the next in a planar system (Reyn, 1980). A family of limit cycles such as this is structurally stable, and we will refer to these limit cycles and their associated heteroclinic cycle as a stable heteroclinic channel (Rabinovich et al., 2008b; Afraimovich et al., 2011) 2. To be more precise, we
need to restrict the set of limit cycles to those that pass within some minimum distance ε of each of the saddles; in this case we could say that a heteroclinic channel exists within
the neighborhood ε of the saddles. As with many terms like “short” and “tall”, however, it is often convenient to use these terms in a qualitative sense without defining artificially sharp boundaries. Thus for most of this dissertation, we will use intuitive definition of “close” based on the slow passage near the saddle rather than formally choosing a specific
distance for ε; a corresponding value of ε could easily be found by taking the maximum of the closest distance between the limit cycle and each of the saddles.
1.1.2.4 Attractors and stability
In a dynamical systems context, and invariant set is a set of points where the entire set does not change when evolved through time (even though the individual points might). More
2More generally, a stable heteroclinic channel may also refer to a connected chain of saddles that create a non-cyclic channel. We will only consider the cyclic case in this dissertation, however. CHAPTER 1. INTRODUCTION 22
formally, a set of points A X is an invariant set if and only if φ(a,t) a A,t T = A. ⊂ { | ∈ ∈ } Fixed points, limit cycles, and heteroclinic cycles are all examples of invariant sets. Because fixed points, limit cycles, and heteroclinic cycles are points and curves of zero volume (measure) in space, it is almost certain that an experimental preparation will not begin with initial conditions that happen to lie on a particular example of these invariant sets. Thus it is important to consider the behavior of orbits that lie near these structures, because we are much more likely to see these nearby orbits when observing the system than the fixed point, limit cycle, or heteroclinic cycle itself. In many of these cases, the nearby orbits will approach the invariant set in question and begin to behave like the elements of this set.
We first consider a fixed point, then a limit cycle. If there is some ε > 0 neighborhood B of a fixed point x such that all points in the neighborhood B asymptotically approach
x, i.e. limt ∞ φ(b,t) b B = x , then we consider x to be a stable fixed point. The { → | ∈ } { } same approach can be used to define a stable limit cycle γ, i.e. there exists an ε > 0
neighborhood B of the limit cycle γ such that limt ∞ φ(b,t) b B = γ. { → | ∈ } For a heteroclinic cycle Γ, the definition of stability becomes a bit more delicate. In the planar case, for example, it is clear that any point on the “outside” of the cycle will follow a trajectory leading away from heteroclinic cycle, so in the generic case every neighborhood around the cycle includes points that do not approach the cycle. In some cases, however, almost all of the points on the inside of the cycle may approach the heteroclinic cycle, so the concept of stability still seems useful. To work around this difficulty, we replace the neighborhood B with a set C of finite volume (measure) such
that Γ C and limt ∞ φ(c,t) c C = Γ. If such a set C exists, we refer to Γ as a stable ⊂ { → | ∈ } heteroclinic cycle. We next consider a sufficient condition for the stability of a heteroclinic cycle. A CHAPTER 1. INTRODUCTION 23
hyperbolic saddle with a one dimensional manifold is called dissipative if the minimum of the absolute value of the stable eigenvalues is greater than the unstable eigenvalue (Afraimovich et al., 2011). Intuitively, this means that the saddle compresses incoming trajectories more than it stretches them. If all of the saddles in the heteroclinic cycle are dissipative, the heteroclinic cycle is a stable heteroclinic cycle (Afraimovich et al., 2011), and we will call the associated heteroclinic channel a stable heteroclinic channel.
1.1.2.5 Stochastic systems
In a few cases, we simulate a stochastic system with additive noise. We can represent such a system by combining a probability triple that captures the likelihood of various outcomes with a dynamical system whose state encodes one set of outcomes.
A probability triple (Ω,F ,P) is composed of a set of all possible outcomes Ω, a σ- algebra3 of Ω representing, intuitively, all “reasonable” collections of possible outcomes,
and a function P : F [0,1] representing the probability of observing a given set of → outcomes that obeys basic intuitions of probability4. We will use a version of n-dimensional Brownian motion (also known as a Wiener process) to drive the “noise” of our system. An n-dimensional version of Brownian motion
is a stochastic process (i.e. a family of functions parameterized by ω Ω) where the ∈ change in position from time t to a later time s is independent of any previous changes in position and follows an n-dimensional Gaussian distribution with a mean of 0 and covariance of t s I (where I is the n-dimensional identity matrix). More formally, | − | if we let p be the probability density function of a zero-mean n-dimensional Gaussian
3 i.e. a set F that contains the empty set (i.e. ∅ F ), the complement of each of its elements (i.e. C ∈ S∞ A = (Ω A) F for all A F ), and any finite or countably infinite union of its elements (i.e. Ai F ∖ ∈ ∈ i=0 ∈ for all A0,A1,... F ). ∈ 4i.e. P(∅) = 0, P(Ω) = 1, and the probability of a set of mutually exclusive outcomes is equal to the S∞ sum of their probabilities, i.e. for any countably infinite family of disjoint sets A0,A1,... F , P( i=0 Ai) = ∞ ∈ ∑i=0 P(Ai) CHAPTER 1. INTRODUCTION 24
distribution parameterized by a scalar variance σ 2,
1 x 2 p(x,σ 2) = exp ‖ ‖ , (1.2) (2πσ 2)n/2 − 2σ 2
then a stochastic process Wt is an n-dimensional version of Brownian motion if the
probability of the process passing through a set of intervals F1,...,Fn F at times ∈ t1,...,tn T, t1 < t2 < < tn is ∈ ···
Z P(Wt1 F1,...,Wtn Fn) = p(x1,t1) p(xn xn 1,tn tn 1) dx1 dxn. (1.3) ∈ ∈ F1 Fn ··· − − − − ··· ×···×
We combine a probability triple (Ω,F ,P) with a dynamical system (T,X,φ) to form n a stochastic dynamical system, where the state (u,ω,t) X = R Ω T contains both ∈ × × which outcomes ω Ω that will occur and the absolute time t T. Because the outcomes ∈ ∈ that will occur do not change with time and absolute time should advance in a sensible
manner, we also require φ((...,ω,t),s) = (...,ω,t + s). The evolution of u is then driven by the stochastic differential equation
du = f (u) dt + η dWt(ω), (1.4)
where η is a diagonal matrix representing the magnitude of the noise and Wt(ω) is a version of n-dimensional Brownian motion. The formal definition of a stochastic differential equation is somewhat involved; for brevity, we will limit ourselves to the case of additive
noise and “well behaved” functions f (u) (i.e. f is Lipshitz continuous within the region of integration5). Under these conditions, the solution to the stochastic differential equation
n 5meaning that there exists a finite constant K such that for all u,v R within the region of integration, f (u) f (v) < K u v ∈ ‖ − ‖ ‖ − ‖ CHAPTER 1. INTRODUCTION 25
(1.4) can be thought of as the result of taking many small Euler steps and adding noise at
each step. More formally, if we define Euler integration from the initial condition uN,0 over
the time interval (0,t), t T, in N steps as the final term uN N of the recurrence relation ∈ ,
uN,n = uN,n 1 + f (uN,n 1)(t/N) + η(Wnt/N(ω) W(n 1)t/N(ω)), (1.5) − − − −
then the solution u(t) of (1.4) is the (mean square) limit of uN,N as the number of steps N taken between 0 and t goes to infinity (Kloeden and Platen, 1992). For a more formal and general definition of stochastic differential equations, we refer the interested reader to Øksendal(2003).
1.1.3 Mathematical models of neurons
The membrane potential of a neuron can be modeled as an electrical circuit containing,
in parallel, a capacitor Cm representing the charge on the membrane and several ion
channels, each represented as a conductance, gi, in series with a battery, Ei, representing the electrochemical potential of the ions to which the channel is permeable. In general,
a conductance gi may be dependent on the membrane potential and one or more gating variables ni, j. Under these assumptions, the change in the membrane potential over time can be described by dV = I + ∑gi(V,ni, )(Ei V), (1.6) dt i · − where I represents currents in the cell from extrinsic sources, such as synaptic currents or the injection of a current by an experimenter. The dynamics of the gating variables, in turn,
can be described by a membrane-potential-dependent steady state value, ni, j,∞(V), which is approached at a rate determined by a membrane-potential-dependent time constant, CHAPTER 1. INTRODUCTION 26
τi, j(V). The change in the gating variables can then be described by
dni j ni j ∞(V) ni j , = , , − , . (1.7) dt τi, j(V)
The dynamics of this system of equations describe the behavior of the neuron; for example, tonic firing would appear as a simple limit cycle, tonic bursting as a limit cycle containing a coiled region with many spikes connected to a less convoluted region for the inter burst interval, and a stable fixed point would represent a quiescent neuron (we refer the interested reader to Izhikevich(2010) for an in-depth exploration of the dynamics of this family of neural models). In many cases, it is useful to represent the firing rate of the neuron, meaning the number of spikes per unit time, rather than representing the membrane potential directly. A neuron in such a firing rate model can often be represented by the one-dimensional differential equation da = F(a,I). (1.8) dt
There are multiple methods that can be used to make this transformation; for a review we would refer the interested reader to Ermentrout and Terman(2010a). Firing rate models are often used as a lower dimensional representation of the neurons in a neural circuit, and we will use these for the model in chapter2.
1.2 The choice of Aplysia californica as a model system
To ground our exploration of possible architectures in biological pattern generators, it is important that we choose one or more biological systems with structured behavior to compare our theories and models against. Ideally, such a model organism would have CHAPTER 1. INTRODUCTION 27
extensive existing research available to draw from both on the biomechanics of the behavior and on the neural circuit driving the behavior. The behavior in question should be readily observable, and should also have well defined and physiologically meaningful ways of perturbing the behavior so that the response to the perturbation can be studied. Furthermore, experimental techniques should be available for recording behavior in the intact animal and the corresponding fictive behavior in the isolated nervous system. It should be possible to manipulate individual neurons within the nervous system, and ideally the number of neurons that need to be manipulated to influence behavior should be small. For all of these reasons, we have chosen to use swallowing in the marine mollusk Aplysia californica as the model system whose dynamics we will study. We will briefly describe the behavior, then explain how this behavior meets the criteria above. The feeding apparatus in Aplysia consumes food by protracting a grasper, known as the radula-odontophore, through the jaws, closing the grasper on the food, generally seaweed such as Gracilaria spp., and then withdrawing the grasper with the food into the buccal cavity and towards the esophagus. Ingestive behaviors are divided into bites, where the animal attempts to grasp food and fails, bite-swallows, where the animal does not initially have the food in its mouth but successfully grabs the food and pulls it partially into the buccal cavity, swallows, where the food starts in the buccal cavity and the animal pulls it further in, and rejections, where the animal pushes food out of the buccal cavity by closing on the food and protracting the grasper (Morton and Chiel, 1993a). Of these behaviors, we have chosen to focus on swallows because they are easy to elicit from the animal and have a mechanical component (force on the seaweed) that is comparatively easy to manipulate. Swallowing in Aplysia meets all of the criteria we outlined for an ideal model system above. There is an extensive experimental literature on the feeding system of Aplysia for us to build on, including mapping of relevant motor pools (Church and Lloyd, 1994), CHAPTER 1. INTRODUCTION 28
biomechanics (Sutton et al., 2004a,b; Novakovic et al., 2006), sensory inputs and integra- tion (Chiel et al., 1986; Proekt et al., 2007) neural plasticity (Schwarz and Susswein, 1986; Katzoff et al., 2006), and key neurons involved in pattern generation (Hurwitz et al., 1996; Hurwitz and Susswein, 1996; Susswein et al., 2002; Hurwitz et al., 2003; Nargeot et al., 2002). The feeding behavior is easily visible from outside the animal (Morton and Chiel, 1993a) (unlike, for example, the gastric mill in the crustacean stomatogastric ganglion), and changing forces on the seaweed during swallowing provides a behaviorally relevant perturbation (unlike biting, where the animal does not grasp the food). Techniques exist for recording from the key nerves controlling feeding in vivo (Cullins and Chiel, 2010), in reduced preparations (McManus et al., 2012), and in the isolated nervous system (Lu et al., 2013), which is currently not possible in systems such as C. elegans. The nervous system has comparatively few cells (an estimated 20,000 neurons (Kandel, 2000), compared to an estimated 75,000,000 neurons in a mouse (Williams, 2000)) and small neural pools, so that electrical manipulation of a single cell can cause significant muscle activation (Church and Lloyd, 1994) or changes in timing (Hurwitz and Susswein, 1996; Zhurov and Brezina, 2006). In addition, the somata of many of the motor and interneurons are comparatively large, electrically compact with the synaptic region (unlike most arthropod neurons), occur in stereotyped locations within the ganglia, and can often be identified by their location and electrophysiological “personalities” (Kandel, 1979), making them tractable to experimental measurement and manipulation of individual cells (McManus et al., 2012; Lu et al., 2013).
1.3 Mathematical framework and central hypotheses
We now have reviewed the background that we need to discuss the central hypothesis. Consider the motor behavior of swallowing in Aplysia californica. If we make the (weak) CHAPTER 1. INTRODUCTION 29
assumption that the dynamics of the behavior can be described in a deterministic manner, we can describe the behavior in the context of dynamical systems theory. We assume that the state of the system can be decomposed so that the state is the concatenation of a vector, a, containing the state of the components of the central nervous system, and a second vector, x, containing the state of the periphery; this is reasonable to the extent that we can clearly delineate the boundary between the central nervous system and the periphery. The differential equations for the behavior in the intact animal can then be written in the form
da = f (a) + g(a,x), (1.9) dt dx = h(a,x). (1.10) dt
Here f represents the dynamics of the central nervous system that do not depend on sensory input from the periphery, g represents all of the dynamics of the central nervous system that do depend on sensory input from the periphery, and h contains the dynamics of the periphery including the motor system and its response to activation coming from the central nervous circuit. Note that the additive model of (1.9) arises naturally when, for example, the state a contains the membrane potential of cells and the sensory input takes the form of synaptic currents, as in (1.6). We further make the assumption that, in a reduced preparation, the dynamics of the nervous system can be described as
da = f (a). (1.11) dt
Unlike equations (1.9- 1.10), this is a somewhat strong assumption that in the reduced preparation the nervous system behaves in the way it would in the intact animal if sensory input were removed. We add to this the assumption that dynamics producing the sponta- CHAPTER 1. INTRODUCTION 30
neous activity in the reduced preparation (1.11) are related to those producing the behavior in the intact animal. While this assumption is a source of controversy in experimental neurobiology (Marder and Bucher, 2001), this is the common assumption behind the utility of studying reduced preparations that produce motor output that resembles that seen in vivo and is thus not unreasonable to make. The standard assumption when modeling motor pattern generators is that the isolated nervous system contains one or more intrinsic oscillators driving the pattern (Kopell, 1986), which is generally represented as the isolated dynamics (1.11) having a stable limit cycle. In this dissertation, we will explore the following hypothesis: The dynamics of the pattern generator controlling swallowing in Aplysia californica are better described by a model where the intrinsic neural dynamics6 contain a stable heteroclinic channel than one where the intrinsic dynamics contain a more homogeneous limit cycle.
To test this hypothesis, we will use a number of models of swallowing behavior and intrinsic neural dynamics and compare them with the observed behavior of the animal. We have attempted to use models that are as simple as possible while still reproducing the qualitative aspects of the behavior we wish to examine. Although this type of model can not offer the quantitative predictions that a more detailed model might, there are a number of reasons to favor these less complex models during the early stages of a research program. Striving for simplicity forces one to examine the necessity of each component of the model, and it can be very instructive when a seemingly minor component of the model turns out to critical for the behavior. Simpler models are also generally more tractable to analysis, which we will take advantage of in chapter3, and the phenomena seen are more likely to generalize to other systems. Once the phenomena in a simple model are understood,
6Or, alternatively, the slow subsystem in an appropriate fast slow decomposition of the neural dynamics (e.g. Butera et al.(1996); Nowotny and Rabinovich(2007)) CHAPTER 1. INTRODUCTION 31
additional complexity can always be added. Such an approach has been successful both in our own lab (where simple kinematic models such as Drushel et al.(2002) have led to the development of more quantitative models such as Novakovic et al.(2006)) and in others (e.g. Leon Glass, whose studies of very simple phase oscillator models of the heart, e.g. Guevara and Glass(1982) have provided insights into phenomena seen in vivo, such as entrainment of reentrant rhythms (Glass et al., 2002)).
1.4 Outline of the remainder of the dissertation
In chapter2, in order to find differences between the behavior produced by neural dynamics containing a stable heteroclinic channel and neural dynamics containing a limit cycle, we develop a neuromechanical model of swallowing in Aplysia californica, such that the dynamics of the isolated nervous system can be changed between a stable heteroclinic channel and a similar limit cycle with a simple change in parameters. We then examine the behavior of this model with both parameter sets when subjected to load, and we compare the responses of the model to those seen in the animal, both in in vivo behavior and in fictive patterns in isolated preparations. We hypothesized that the behavior of the animal would more closely match the behavior of the model in the heteroclinic channel regime than in the limit cycle regime. In the three qualitative differences between the models we examined, we found that the comparable data from the animal were much more similar to the heteroclinic channel’s behavior and thus consistent with this hypothesis. In chapter3, we hypothesized that the differences seen in the previous chapter were the result of regions of localized sensitivity. We thus set out to characterize the sensitivity of the system and develop quantitative tools that could be used to help distinguish a pattern generator built on a limit cycle from one built on a heteroclinic channel. We explore the question of sensitivity in heteroclinic systems by constructing an analytically tractable CHAPTER 1. INTRODUCTION 32
planar heteroclinic system. We then show that the tools of phase response curve analysis, which at first might not seem suitable for the study of a non-periodic structure such as a stable heteroclinic cycle, can be applied by taking appropriate limits as a stable limit cycle bifurcates into a heteroclinic cycle. Consistent with our hypothesis, the phase response curve shows small regions where the system is much more sensitive than it is outside of these regions. This analysis shows, somewhat surprisingly, that the point of maximal sensitivity is not near the saddle itself, but instead lies on the approach to the saddle where the orbits start to diverge. We show that the pattern of sensitivity in the tractable planar model appears to generalize to other homoclinic and heteroclinic systems, such as a Morris–Lecar neuron in the homoclinic limit. In the final chapter, we review the results and discuss some potential future directions for research in this area. Chapter 2
The Significance of Dynamical Architecture for Adaptive Responses to Mechanical Loads During Rhythmic Behavior
33 CHAPTER 2. SIGNIFICANCE OF DYNAMICAL ARCHITECTURE 34
2.1 Introduction
Motor behaviors, such as cat running, crayfish swimming, and dog lapping all require the nervous system to reliably generate a sequence of motor outputs. To be efficient, however, a fixed sequence of activity is not enough: a cat that fails to step over an obstacle may lose its footing and fall (Forssberg et al., 1975; Forssberg, 1979) and a crayfish that wanders into a current of cold water must control muscles that may suddenly have become stronger but relax more slowly (Harri and Florey, 1977). Sensory feedback plays a key role in allowing an animal to adapt its behavioral pattern to the circumstances in which it finds itself. The way that this sensory information is integrated into pattern generation to produce adaptive behavior, however, can be difficult to ascertain. As discussed in chapter1, two competing theories, chain reflexes and central pattern generators, have been proposed, with central pattern generators enjoying the greater support in the current literature. It should be noted, however, that patterns generated by the isolated nervous system often are very distorted compared to those seen in vivo. In particular, phases of the motor pattern are often significantly longer than those observed in the intact animal. This has led many investigators to question the descriptive power of central pattern generator theory. In the words of Robertson and Pearson, “Although now abundantly clear that a central rhythm generator can produce powerful oscillations in the activity of flight motor neurons and interneurons, it is equally clear that the properties of this central oscillator cannot fully account for the normal flight pattern” (Selverston, 1985). There is some evidence that slowing of isolated neural patterns may be due to the absence of sensory feedback. In Pearson et al.(1983), cycle-by-cycle stimulation of the appropriate sensory afferents was able to restore wing-beat frequency in fictive flight in the locust. Restoration of the normal pattern by sensory input suggests that biological pattern CHAPTER 2. SIGNIFICANCE OF DYNAMICAL ARCHITECTURE 35
generators may occupy a middle ground between pure central pattern generators and chain reflexes. In some cases, endogenous neural input may control where a systems lies on this continuum. AsB assler¨ (1986) noted when considering a relaxation-oscillator like model of a central pattern generator “Hence, one and the same system can behave either like a CPG or like a chain reflex, depending only on the amount of endogenous input.” These investigators thus warned about the dangers of inferring the mechanism used by a pattern generator in vivo based only on the behavior of a pattern generator in vitro. Despite these hesitations, the empirical data supporting central pattern generator hypothesis led to a focus on providing a mathematical formulation for this theory using the qualitative analysis of dynamical systems, which was becoming an important tool for theoretical neuroscience by the middle of the twentieth century. The behavior of an ideal central pattern generator naturally corresponds to a system of nonlinear ordinary differential equations whose solutions contain stable limit cycle (an attracting isolated periodic orbit). As a result, this structure has played a central role in the mathematical description of central pattern generators (Ijspeert, 2008). In contrast, there have been fewer attempts to model chain reflexes with systems of differential equations. Instead, much of the work modeling these types of sensory- dependent systems use different tools, such as finite state machines (Lewinger et al., 2006). While these models can capture individual phases of the behavior well, they generally do not describe the transitions between the phases, which may be important in understanding some forms of behavior. In contrast, one could view the state of a chain reflex system in terms of a series of stable fixed points. In each phase of the motion, the trajectory would be captured by one of the fixed points until the appropriate (external) sensory input pushed the system out of the neighborhood attracted to that fixed point and into the basin of attraction of the next. CHAPTER 2. SIGNIFICANCE OF DYNAMICAL ARCHITECTURE 36
Between these two extremes of models of central pattern generators and chain reflexes, one may consider systems in which the progress of a periodic orbit is slowed, but not stopped, by passage near one or more fixed points. This behavior arises naturally in a structure known as a “stable heteroclinic channel” (Rabinovich et al., 2008b), where multiple saddle points (fixed points that attract in some directions while repelling in others) are connected in a cycle, so that the unstable manifold of each saddle point brings the system near the stable manifold of the next fixed point. This structure has been used to describe motor behavior such a predatory swimming behavior in Clione (Levi et al., 2004; Varona et al., 2004). To our knowledge, however, these models of pattern generation have not been directly compared to those built with a more “pure” limit cycle that does not pass near fixed points. A potential advantage of a dynamical system that allows trajectories to move close to equilibrium points is that it may spend longer or shorter times in that vicinity, rather than proceeding through the cycle with a relatively fixed phase velocity. In turn, this could allow an animal greater flexibility in responding to unexpected changes in the environment, such as increases or decreases in mechanical load as it attempts to manipulate an object. To examine this range of dynamics, we have created a neuromechanical model based on the feeding apparatus of the marine mollusk Aplysia californica. We examine the model in two parameter regimes which produce similar output under small loads. In the first parameter regime, the isolated neural dynamics form a homogeneous limit cycle, as would be expected for an idealized central pattern generator. In the second parameter regime, the isolated neural dynamics form a stable heteroclinic cycle, moving it closer along the continuum to a chain reflex. We show that in this second regime, the behavior of the model falls between that of an idealized chain reflex and an idealized central pattern generator. We then compare the behavior of the two models to the observed behavior of the animal, CHAPTER 2. SIGNIFICANCE OF DYNAMICAL ARCHITECTURE 37 and show that several of the features of the animal’s behavior are better described by the model with the stable heteroclinic channel than the model with the limit cycle. At the end of the chapter, we reflect on possible general principles suggested by this work.
2.2 Mathematical Framework
In this section, we describe a general mathematical framework we will use for modeling the behavior of a motor pattern generator. We model a central pattern generator receiving sensory input from the body as a system of differential equations specifying the evolution
n m of a vector of n neural state variables, a R , and a vector of m state variables, x R , ∈ ∈ representing the mechanics and periphery (e.g. muscle activation). We assume that an applied load interacts only with the mechanical state variables, so that the differential equations can be naturally written in the following form:
da = f (a, µ) + εg(a,x), (2.1) dt dx = h(a,x) + κl(x). (2.2) dt
Here µ is a vector of parameters which can encode states such as arousal of the animal, f (a, µ) represents the intrinsic dynamics of a motor pattern pattern generator, h(a,x) represent the dynamics of the periphery with the given central input, g(a,x) represents the effects of sensory feedback from the periphery, l(x) represents the effects of an external load or perturbation, and ε,κ R+ are scaling constants, not necessarily small. Note that ∈ here we have added to the framework introduced in chapter1 by adding the forces due to the load as a separate term. We further assume that all of these functions are smooth, infinitely differentiable, and have bounded ranges over the domain of interest. CHAPTER 2. SIGNIFICANCE OF DYNAMICAL ARCHITECTURE 38
2.2.1 Limit cycles
We first consider the case of an idealized central pattern generator, where a part of the nervous system can produce sequences of motor activity that closely resemble those seen in vivo, even when it is not attached to the periphery. Thus we assume that, for some range
of the parameter µ, the dynamics of the isolated neural circuit, da/dt = f (a, µ), contain an attracting limit cycle γ(t) which represents the observed motor pattern. We make the further important assumption that there exists a behaviorally relevant mechanical load
κl(x) with which the complete system contains a corresponding limit cycle ξ(t), such that the neural component of ξ(t) closely matches that of the isolated system. By this we mean that, with appropriately chosen initial conditions,
Z T 2 γ(t) Paξ(t) dt 1, (2.3) 0 || − || ≪
m+n n where Pa : R R is the projection operator onto the a subspace, T is the period of → γ(t), and is the usual L2 (Euclidean) norm. Note that it is possible that these very || · || similar patterns will produce very different behavior in the periphery. Therefore, we will
also assume that this isolated pattern γ(t) produces similar behavior in the periphery, that is that the system dx/dt = h(γ(t),x)+κl(x) contains an attracting periodic orbit ζ(t) such that with an appropriate choice of initial conditions
Z T 2 ζ(t) Pxξ(t) dt 1, (2.4) 0 || − || ≪
m+n m where Px : R R is the projection operator onto the x subspace. Finally, we assume → that gradually adding back sensory input does not distort the pattern or the behavior,
meaning that both γ(t) and ζ(t) persist for small values of ε and remain similar to CHAPTER 2. SIGNIFICANCE OF DYNAMICAL ARCHITECTURE 39
projections of ξ(t). Because the system is smooth, the integral in (2.3) can be made to be small by insuring
that the magnitude of the sensory feedback, εg(Paξ(t),Pxξ(t)) , is small for all t (0,T). || || ∈ Although a careful choice of g could be small along ξ but not elsewhere, we will focus || || on the more generic case where ε 1. ≪ 2.2.2 Destabilization of fixed points
We next consider the chain reflex. In this case, the dynamics of the isolated nervous system,
da/dt = f (a, µ), will contain a set of stable nodes, A, where each node represents a “stage” of the chain reflex that can be destabilized by sensory input. Note that in this case, unlike
the central pattern generator, ε may need to be of O(1) to destabilize a node. The combined dynamics of the nervous system and the periphery, however, would still be expected to
contain a stable limit cycle ξ(t) rather than a series of fixed points. Similar dynamics have been seen in models of other biological oscillators; for example in Novak et al.(1998), the authors created a model of the cell cycle where fixed points in the biochemical dynamics (analogous to the isolated neural dynamics) can be destabilized by changes in cell size (analogous to the periphery) so that the coupled system contains a limit cycle. Because a chain reflex can not explain the fictive motor patterns produced by the isolated nervous system, we will not explore this alternative in this chapter.
2.2.3 Stable heteroclinic channels
We now consider a system that is intermediate between the two extremes of an idealized central pattern generator and a chain reflex. We can construct such a system from a set of n- dimensional hyperbolic saddle points, each with a one-dimensional unstable manifold and an n 1 dimensional stable manifold, arranged in a cycle such that the unstable manifold − CHAPTER 2. SIGNIFICANCE OF DYNAMICAL ARCHITECTURE 40
of one saddle point intersects the stable manifold of the next, forming a heteroclinic orbit. We refer to these saddle points and their connecting heteroclinic orbits as a heteroclinic cycle (Guckenheimer and Holmes, 1988). Under appropriate conditions, this heteroclinic cycle attracts nearby orbits (and thus can be called a stable heteroclinic cycle). In particular, if we define the (positive) ratio of
the least negative stable eigenvalue λi,s and the unstable eigenvalue λi,u of the ith saddle
as the saddle index νi = λi s/λi u (Shilnikov et al., 2002), then the heteroclinic cycle − , , will attract nearby orbits if ∏i νi > 1 (Afraimovich et al., 2004a). This type of dynamics can arise naturally from neural models involving symmetric mutually inhibitory pools of neurons; for example see Nowotny and Rabinovich(2007). An unperturbed trajectory on the heteroclinic cycle will, like the chain reflex model in section 2.2.2, asymptotically approach a fixed point. Unlike the chain reflex model, however, a very small perturbation in the unstable direction will push the trajectory out of the stable manifold, allowing the trajectory to leave the neighborhood of the fixed point (and potentially travel to the neighborhood of the next fixed point). Arbitrarily small amounts of noise can thus insure that the system will almost certainly not remain stuck at a given fixed point (Stone and Holmes, 1990). Thus, rather than the stability of states seen in the chain reflex model, the heteroclinic cycle exhibits metastability (Afraimovich et al., 2011), where the trajectory spends long but finite periods of time near each fixed point (Bakhtin, 2011). Thus, like the chain reflex, the system can spend short or long periods of time in one particular state depending on sensory input, but, like the limit cycle, the system will eventually transition to the next state even in the absence of sensory input. While stable heteroclinic cycles are structurally unstable (i.e. a small change in the vector field will generally break the cycle), small perturbations can result in the creation of a stable limit cycle that passes very close to the saddles. For example, in the planar case, CHAPTER 2. SIGNIFICANCE OF DYNAMICAL ARCHITECTURE 41
any sufficiently small perturbation that pushes the unstable manifold of the saddles towards the inside of the unperturbed stable heteroclinic cycle will result in a stable limit cycle (Reyn, 1980). Similar conditions can be found for higher dimensional stable heteroclinic cycles (Afraimovich et al., 2004a). These families of limit cycles that pass close to the original saddles, known as stable heteroclinic channels (Rabinovich et al., 2008b), are structurally stable, and exhibit many of the same properties of sensitivity and metastability as the original stable heteroclinic cycles. As we will see, this extreme sensitivity can be advantageous for generating adaptive behaviors.
In the next section, we provide an example of model dynamics f (a, µ) exhibiting a limit cycle for µ > 0 and a bifurcation to a heteroclinic cycle at µ = 0. We then investigate the behavior of the full (a,x) system in the “limit cycle” and “heteroclinic cycle” parameter regimes.
2.3 Model Description
2.3.1 Neural model
We wish to explore the effects of different types of neural dynamics on the behavior of the animal. Although detailed, multi-cellular and multi-conductance models of neurons and circuits underlying feeding pattern generation in Aplysia have been described (Baxter and Byrne, 2006; Cataldo et al., 2006; Susswein et al., 2002), the complexity of these models makes it difficult to use them for mathematical analysis. As a consequence, we choose to represent motor pools (which contain neurons that are electrically coupled to one another or have mutual synaptic excitation) using nominal firing-rate models. As discussed in section 2.2, we define the neural dynamics as a combination of
an intrinsic component, f (a, µ), that does not depend on the periphery, and a sensory CHAPTER 2. SIGNIFICANCE OF DYNAMICAL ARCHITECTURE 42
(coupling) component, g(x), which does depend on the periphery. For mathematical tractability, we assume that the intrinsic and sensory drive combine linearly, thus giving the evolution equation of the neural activity.
da = f (a, µ) + εg(x ), (2.5) dt r where a is a vector of the activity of each of the N neurons, ε is a parameter scaling the
strength of sensory input, xr is a biomechanical state variable which we will define in more detail in section 2.3.2, and µ is a parameter that can shape the intrinsic dynamics. Specifically we will consider the following modified Lotka–Volterra model which captures the dynamics of N neural pools:
! ! 1 fi(a, µ) = 1 ∑ρi ja j ai + µ , (2.6) τn − j
for 0 i < N. Here µ is a scalar parameter representing intrinsic excitation, τn is a time ≤ constant, and ρ is the coupling matrix
i j 1 = ρi j = 0 i = j 1 (mod N) (2.7) − γ otherwise, where γ is a coupling constant representing inhibition between neural pools.
When N > 2 and γ > 2 this system contains a stable heteroclinic cycle when µ = 0. In contrast, as shown in figure 2.1, it contains a stable limit cycle for small positive values of
µ, with the distance between the limit cycle and saddles increasing with increasing values CHAPTER 2. SIGNIFICANCE OF DYNAMICAL ARCHITECTURE 43
a_I3
a_I2 a_h
Figure 2.1: When µ = 0, in the absence of sensory input (ε = 0), the intrinsic neural dynamics contain a stable heteroclinic cycle (black line) connecting saddles at (1,0,0), (0,1,0), and (0,0,1). When µ is a small positive number and ε = 0, the heteroclinic cycle is broken and a stable limit cycle arises (shown in light blue for the value of µ used for the limit cycle in this chapter). CHAPTER 2. SIGNIFICANCE OF DYNAMICAL ARCHITECTURE 44
of µ. With the goal of parsimony, we use N = 3 and thus (2.6) can be expanded to
1 f0(a, µ) = (a0(1 a0 γa1) + µ), (2.8) τn − − 1 f1(a, µ) = (a1(1 a1 γa2) + µ), (2.9) τn − − 1 f2(a, µ) = (a2(1 a2 γa0) + µ). (2.10) τn − −
We explain the correspondence of these three neural pools to neural pools in Aplysia in the next section.
2.3.2 Model of the periphery and load
We next couple the neural dynamics to a nominal mechanical model of feeding in Aplysia. During ingestive behaviors in Aplysia, a grasper, known as the radula-odontophore, is protracted through the jaws by a muscle referred to as I2. The grasper closes on food, and then is retracted by a muscle called I3, and then opens again, completing the cycle (see figure 2.2). The timing of closing is often not precisely aligned with the switch from protraction to retraction. Instead, closing usually occurs before the end of protraction, although the amount of overlap varies by behavior, from very little overlap in swallows to a significant overlap in rejection. A general model for biting and swallowing could thus contain four components: protraction while open, protraction while closed, retraction while closed, and retraction while open. For simplicity, we reduce these to three components, each of which corresponds to one of the three neural pools in the neural model: protraction open, protraction closing, and retraction open, as shown in figure 2.3. The protraction
open motor pool (a0) corresponds to the electrically coupled group of neurons B31, B32, B61, B62, and B63, which activate the I2 muscle and are all active during protraction (Hurwitz et al., 1996, 1997; Susswein et al., 2002). The protraction closing neuron pool CHAPTER 2. SIGNIFICANCE OF DYNAMICAL ARCHITECTURE 45
Figure 2.2: The model breaks swallowing into three phases. First, the odontophore protracts while open (lower right). Near the end of protraction, the odontophore begins closing (left) and protracts a small distance while closed. In the last phase, the odontophore retracts while closed (upper right). The protraction muscle (I2) is shown in blue, the grasper (the radula-odontophore) is shown in red, and the ring-like retraction muscle (I3) is shown in yellow, with a section cut away to show the grasper. The green strand is seaweed, with the arrows showing how the seaweed moves within a single cycle. CHAPTER 2. SIGNIFICANCE OF DYNAMICAL ARCHITECTURE 46
Figure 2.3: Schematic of the neuromechanical model of the feeding apparatus in Aplysia. The three neural pools (a0, a1, and a2) control three phases of the behavior shown in figure 2.2: protraction open, protraction closing, and retraction closed. The solid lines and triangles indicate excitatory synaptic coupling with a neuromuscular transform represented by a low pass filter. The dashed line and Θ symbol represent a simple summation and thresholding that control closing in the model. The a0 neural pool represents the B31, B32, and B63 neurons, the a1 motor pool represents these same neurons with the addition of B8 (which experiences slow excitation from B34), and the a2 motor pool represents B64, B3, B6, B9, and B8 (which is excited by B64) activity. CHAPTER 2. SIGNIFICANCE OF DYNAMICAL ARCHITECTURE 47
parameter value description γ 2.4 inhibition strength from next pool ε 2 10 3 sensory feedback strength · − κ 3√3/2 length-tension curve normalization constant µ 0 neural pool intrinsic excitation σ0 -1 proprioceptive direction for protraction open neural pool σ1 1 proprioceptive direction for protraction closing neural pool σ2 1 proprioceptive direction for retraction closed neural pool τn 0.05 neural pool time constant τm 0.05 muscle activation time constant br 0.10 grasper damping constant bsw 0.10 seaweed damping constant c0 1 position of shortest length for I2 c1 1.1 position of center of I3 Fsw 0.01 force on the seaweed resisting ingestion k0 1 I2 muscle strength and direction − k1 1 I3 muscle strength and direction S0 0.5 proprioceptive neutral position for protraction open neural pool S1 0.5 proprioceptive neutral position for protraction closing neural pool S2 0.25 proprioceptive neutral position for retraction closed neural pool umax 1 maximum muscle activation w0 2 maximal effective length of I2 w1 1.1 maximal effective length of I3 Table 2.1: Model parameters
state initial description variable value a0 1.0 activity of I2 motor pool (non-negative) 9 a1 10− activity of hinge motor pool (non-negative) 9 a2 10− activity of I3 motor pool (non-negative) u0 0.0 activity of I2 muscle u1 0.0 activity of I3 muscle xr 0.5 grasper position (0 is retracted, 1 is protracted) xsw 0.0 seaweed position (positive is away from the animal) Table 2.2: State variables CHAPTER 2. SIGNIFICANCE OF DYNAMICAL ARCHITECTURE 48
(a1) corresponds to these same I2 motor neurons with the addition of the B8 motor neurons, which activate the I4 muscle used in closing(Morton and Chiel, 1993b). The retraction
closed pool (a2) contains B8 with the addition of the I3 motor neurons B3, B6, and B9 which are simultaneously active during retraction (Church and Lloyd, 1994). Thus the I2
muscle will be driven by both protraction-open (a0) and protraction-closing (a1) motor
pools, whereas the I3 muscle is driven by a single motor pool (a2). The I2 and I3 muscles are known to respond slowly to neural inputs (Yu et al., 1999); we thus model their activation as a low-pass filter of the neural inputs using the time
constants from the model of the I2 muscle described by Yu et al.(1999). Using ui for the
activation of the ith muscle, τm for the filter’s time constant, we use
du0 1 = ((a0 + a1)umax u0), (2.11) dt τm − du1 1 = (a2umax u1). (2.12) dt τm −
In general, the force a muscle can exert will vary with the length to which it is extended (Zajac, 1989; Fox and Lloyd, 1997). The shape of this curve is typically explained by the sliding filament theory as follows: for some maximal length, the actin and myosin fibers will not overlap and the muscle will be limited to passive forces, but below that length, the force will first rise with the increasing overlap of the actin and myosin fibers, reach a maximum, and then decline as the overlapping fibers start to exert steric effects (Gordon et al., 1966). More recently, changes in lattice spacing between the fibers has also been shown to have a role in the force-length dependence (Williams et al., 2013). We model this length/tension curve using the following simple cubic polynomial:
φ(x) = κx(x 1)(x + 1) (2.13) − − CHAPTER 2. SIGNIFICANCE OF DYNAMICAL ARCHITECTURE 49 where κ is a scaling constant. This equation crosses through zero force at zero length and
again reaches zero at the nominal maximal length of 1. We let κ = 3√3/2 to normalize the maximum force between these two points to 1 (which occurs at a length of 1/√3). Although mechanical advantage plays an important role in swallowing (Sutton et al., 2004b; Novakovic et al., 2006), when combined with the length tension curve, the resulting force resembles a shifted and rescaled version of the original length tension curve over the range of motion used in swallowing. We thus choose position and scaling constants for the length-tension curve to approximate the resultant force curve in the biomechanics, rather than the length-tension curve of the isolated muscle. We assume the tension on each muscle is linearly proportional to its activation, and sum all of the muscle forces giving
xr ci Fmusc = ∑kiφ − ui. (2.14) i wi
Here xr (0,1) is the position of the grasper, ki is a parameter representing the strength ∈ and direction of each muscle, ci the position of the grasper where the ith muscle is at its
minimum effective length, and wi the difference between the maximum and minimum
effective lengths for the ith muscle. The sign of ki determines the direction of force of the
muscle; when ki is negative (as it is for I2) the muscle will pull towards its position of shortest length, and when it is positive (as it is for I3) it will push away from this position (in the case of I3, squeezing the radula-odontophore out of the ring of the jaws. We model opening and closing of the odontophore (and thus holding and releasing the seaweed) as a simple binary function, where the odontophore is closed when certain motor pools are active and open otherwise. Specifically, the radula was considered to be
closed when a1 + a2 0.5, and open when a1 + a2 < 0.5. This threshold can be viewed ≥ CHAPTER 2. SIGNIFICANCE OF DYNAMICAL ARCHITECTURE 50
as a plane dividing phase space into two regions with different mechanics (holding the seaweed and not holding the seaweed). In our experience, the teeth on the radula tend to hold the seaweed very firmly, and the animal tends to let go before the seaweed slips from its grasp. Thus the seaweed and the odontophore are considered to be “locked together” when the odontophore is closed and we do not attempt to model slip. The seaweed is assumed to be pulling back with a constant
force Fsw, which is included in the net force on the odontophore when the odontophore is closed
The seaweed and odontophore are assigned viscous damping constants bsw and br, respectively; thus the full equations of motion are
dx r = v (2.15) dt r dx sw = v (2.16) dt sw dvr Fmusc brvr = − (2.17) dt mr dvsw Fsw bswvsw = − (2.18) dt msw when the odontophore is open, and
dx dx r = sw = v (2.19) dt dt r dvr Fmusc + Fsw (br + bsw)vr = − (2.20) dt mr + msw
vsw = vr (2.21) when the odontophore is closed. Note that we are assuming that the momentum of the seaweed is negligible. CHAPTER 2. SIGNIFICANCE OF DYNAMICAL ARCHITECTURE 51
We have observed that when seaweed is abruptly pulled, animals respond with rapid movements of the radula/odontophore without oscillations. This suggests that the system is at least critically damped under these conditions, if not over damped. Furthermore, since the mass of the buccal mass is very small (a few grams), and the accelerations during movement are typically small (based on MRI measurements, they may be close to zero during most of the motion (Neustadter et al., 2002, 2007)), we choose to use equations of motion that assume a viscous limit. Thus, instead of directly simulating equations 2.15-2.21, we use the following reduced system:
dx F r = musc (2.22) dt br dx F sw = sw (2.23) dt bsw when the odontophore is open, and
dx dx F + F sw = r = musc sw (2.24) dt dt br + bsw when the odontophore is closed. For simulations without a mechanical load, where Fsw = 0
and bsw = 0, we replace this equation with
dx sw = 0, (2.25) dt which leaves the seaweed stationary when the radula-odontophore is open. It is entirely possible that the system is effectively quasi-static, and that positional forces dominate over viscous forces, but this formulation does not assume that from the outset. CHAPTER 2. SIGNIFICANCE OF DYNAMICAL ARCHITECTURE 52
In some of the simulations we wish to simulate seaweed that is held or fixed in place rather than experiencing a constant force. For these simulations we replace the constant
Fsw, with a function modeling the force as a stiff spring using Hooke’s law, i.e.
Fsw(xsw) = (xspring xr)kspring. (2.26) −
2.3.3 Proprioceptive input
Proprioceptive neurons detect the position of and forces within the animal’s body. These mechanoreceptors can take many forms, from the muscle spindles and golgi tendon organs of vertebrates to the muscle organs seen in crustaceans to the S-channel expressing neurons seen in mollusks(Vandorpe et al., 1994). Rather than model these in detail, we have assumed that, as a function of the position of the grasper, the proprioceptive sensory neurons will create a net excitation or inhibition of each neural pool. For simplicity we have used a linear relation for this proprioceptive input as a function of position,
g(xr) = (xr Si)σi, (2.27) −
where xr (0,1) is the position of the grasper, Si is the position where the net proprioceptive ∈ input to the ith neural pool is zero, and σi 1,1 is the direction of proprioceptive ∈ {− } feedback for the ith motor pool.
2.3.4 Noise
All biological systems are subject to noise, and as we will show, this can have important effects on the dynamics. Typical examples of noise in a neural context would include the small fluctuations caused by opening and closing of ion channels (known as channel CHAPTER 2. SIGNIFICANCE OF DYNAMICAL ARCHITECTURE 53
parameter value description β 0.20405 neural pool global time constant µ 1 10 3 neural pool intrinsic excitation · − α0 0.6101 neural pool local time scaling near protraction open α1 0.9201 neural pool local time scaling near protraction closing − α2 0.276 neural pool local time scaling near retraction closed umax 2.9 maximum muscle activation Table 2.3: Parameters used for the limit cycle simulations
noise (White et al., 2000; Goldwyn and Shea-Brown, 2011)), the variable release of neural transmitter vesicles, and stochastic effects from small numbers of molecules in second messenger systems. One can also treat parts of the system that we are not including in the model as “noise” (Schiff, 2012), such as small variations in sensory input from the environment with a mean of zero.
We model this noise as a 3-dimensional Weiner process of magnitude η (i.e. white noise). This form of noise arises naturally when the noise is created by many small identical independent events with finite variance, such as channels opening and closing. Although most biological noise is bandwidth limited, the higher frequencies of the noise are filtered out by the dynamics of the model and can thus be ignored. Noise is added to the
neural state variables ai, but assumed to be negligible for the mechanical state variable xr. For simulations in which noise is used, we thus replace the ordinary differential equation (2.5) with the stochastic differential equation