Spontaneous dynamics and information transfer in sensory neurons
A dissertation presented to
the faculty of the College of Arts and Sciences of Ohio University
In partial fulfillment of the requirements for the degree
Doctor of Philosophy
Hoai T. Nguyen
August 2012 © 2012 Hoai T. Nguyen. All Rights Reserved. 2
This dissertation titled Spontaneous dynamics and information transfer in sensory neurons
by HOAI T. NGUYEN
has been approved for the Department of Physics and Astronomy
and the College of Arts and Sciences by
Alexander B. Neiman
Associate Professor of Physics and Astronomy
Howard Dewald Dean, College of Arts and Sciences 3 Abstract
NGUYEN, HOAI T., Ph.D., August 2012, Physics and Astronomy Spontaneous dynamics and information transfer in sensory neurons (101 pp.)
Director of Dissertation: Alexander B. Neiman Peripheral sensory systems convey information about the outside world to the central nervous system and are organized in a feed-forward networks passing information through
a series of layers. Of high interest are basic problems of how sensory information is encoded in dynamical states of a network and how correlated activity progresses from one
network layer to another. Often peripheral receptors are characterized by spontaneous noisy oscillatory activity which introduces temporal and spatial correlations in neuronal
spike trains. This dissertation uses computational methods to elucidate the role of oscillations in shaping of spontaneous dynamics and information transfer in sensory receptors. In particular, we concentrate on hair cell - sensory neurons types of receptors
which mediate the senses of hearing, balance and electroreception. We develop a conductance-based Hodgkin-Huxley type model in which spontaneous noisy inputs from
sensory cells modulate stochastic synaptic transmission. The model is tuned to reproduce experimental data from electroreceptors of paddlefish and from vestibular utricle sensory
neurons of a turtle. Using the model we study how the interplay of ionic currents and synaptic noise affects the statistics of spontaneous dynamics and response properties of the system. In particular, we isolate a region in the parameter space of the model which maximizes information transmission of the system. We contrasted linear and non-linear responses of the model and show that coherent oscillations from epithelial cells enhance significantly stimulus-induced synchronization of neural responses and information transfer. We develop a network model for peripheral electroreceptors to study how correlated noises common to several sensory neurons influence the overall dynamics of the network and transmission of sensory stimuli. In contrast to previous studies on noise 4 correlations which concentrated on weakly correlated white noise, we consider strongly correlated narrow-band fluctuations. We describe in detail transformation of input correlated signals through the network. Furthermore, we show that while coherent epithelial oscillations enhance information transmission for a single sensory neuron, the presence of spatially correlated noise introduces redundancy reducing stimulus coding efficiency and information rate on the network level. Interestingly, this information reduction can be minimized if the frequency of noisy oscillations is about one half of the
firing rates of neurons in the network. Although the model is set to mimic electroreceptors in paddlefish, it can be viewed in a more general context of a networks of oscillatory neurons processing correlated rhythmical signals.
Approved: Alexander B. Neiman
Associate Professor of Physics and Astronomy 5 Acknowledgements
Almost five years have passed since I began my PhD studies in Ohio University, and I could not have come this far without the guidance, support, help, advice, and encouragement of many people present along my journey. First and foremost, I would like to express my deepest gratitude to my PhD advisor Professor Alexander B. Neiman, for his excellent guidance, caring, patience, and providing me with an excellent atmosphere for doing research. In addition, I also appreciate his support and understanding in my life outside of physics. Without it, I could not have spent an incredible and beautiful time with my husband Anh Ngo and my son Chi Kien Ngo when we moved to Baton Rouge, Louisiana, one year ago and now in the lovely city Madison, Wisconsin. At the same time, I am thankful for all insightful discussion, support from my PhD degree committee: Professors Nancy Sandler, Michael H. Rowe, and David F. J. Tees. I would never have been able to finish my dissertation without the guidance of my committee members. I specially thank Professors M.H. Rowe and D.F. Russell for providing experimental data. Special thanks to Professor Sergio
Ulloa. I really appreciate all kindness and encouragement he gave to not only me but also my small family. I would also like to thank Department of Physics and Astronomy for the immense support I have had all along. Last but most certainly not least, I would like to call attention to my family: my Mom Pham Thi Than and my Dad Nguyen Van Khang, my husband Anh Ngo, my wonderful son Chi Kien Ngo. Thank you for your unconditional love and support. You are the world to me. I acknowledge financial support from the National Institutes of Health under Grant
No. DC05063 and by the Biomimetic Nanoscience and Nanotechnology program of Ohio University. 6 Table of Contents
Page
Abstract ...... 3
Acknowledgements ...... 5
List of Figures ...... 8
1 Introduction ...... 13
2 Spontaneous dynamics and response properties of peripheral sensory receptors . . 19 2.1 Statistical measures of spontaneous activity ...... 19 2.2 Characterization of response dynamics of a single neuron ...... 23
3 Spontaneous dynamics of vestibular utricle and electroreceptor afferents ..... 30 3.1 Vestibular afferents ...... 30 3.2 Electroreceptors afferents ...... 34 3.3 Conclusion ...... 38
4 Single neuron models of stochastic spontaneous activity ...... 41 4.1 Conductance-based Hodgkin-Huxley type model ...... 41 4.2 Stochastic synaptic model ...... 43 4.3 Numerical simulations of the modified Hodgkin-Huxley model: compari- son with experimental data ...... 47 4.4 Simplified neuronal model: theta neuron ...... 49 4.5 Conclusion ...... 52
5 Effect of temporal correlations on spontaneous and response dynamics of sensory neurons ...... 54 5.1 Spontaneous dynamics of the Hodgkin-Huxley model of electroreceptors . 55 5.2 Linear response and stimulus encoding ...... 62 5.3 Nonlinear response ...... 64 5.4 Conclusion ...... 70
6 Effect of spatial and temporal correlations on spontaneous dynamics and stimulus coding by small-scale networks of peripheral sensory neurons ...... 72 6.1 Network model ...... 73 6.2 Input – output correlation measures ...... 75 6.3 Stimulus encoding by a network ...... 78 6.4 Transformations of noise correlations by the network ...... 79 6.5 Stimulus encoding and information transmission by the network ...... 85 7
6.6 Conclusion ...... 90
7 Conclusion and Outlook ...... 92
References ...... 95 8 List of Figures
Figure Page
1.1 Schematic drawing of different classes of vestibular afferents. Shown are two types of sensory hair cells (I and II) and three classes of afferents: pure-calyx (C), dimorphic (D) and pure button (B). Mechanoelectrical transduction ion channels in stereocilia of hair cells are labeled as Met. Possible locations of spike initiation zones in afferents are labeled by asterisks. Grey unlabeled structures are efferents, which are neurons transferring feedback signals from the brain to hair cells. Modified from a courtesy figure obtained from Dr. Ruth Anne Eatock (see also [1])...... 14 1.2 The organization of an electroreceptor on the rostrum of paddlefish. Shown is a cluster of skin pores each leading to a canal. Sensory cells in epithelial layer at the bottom of canals innervated by a few afferent neurons. Modified from [2]. 15
2.1 Schematics of neuron perturbed by the stimulus s(t)...... 24
3.1 Spontaneous dynamics vestibular utricle afferents. (a) Raw extracellular recording (lower panel) and a sequence of identified spike times of a vestibular afferent. (b) Probability density function (PDF) of interspike intervals for three representative vestibular afferent discharges...... 31 3.2 Serial correlation coefficients (SCCs) (a) and power spectral density (PSD) for three representative utricle afferents. The data is the same as in Fig 3.1: black, red and blue colored lines show SCCs and PSD for the same afferents as in Fig 3.1. Grey lines in panel (a) show minimum and maximum SCCs values for corresponding renewal spike trains. In panel (b) grey lines show PSDs of renewal spike trains...... 32 3.3 Dynamical entropies for ISIs sequences of two VAs. Red line and symbols shows dynamical entropies calculated for the original sequence. Blue line correspond to corresponding renewal surrogate ISIs. Errorbars we estimated by calculating hn for 10 surrogate sequences...... 33 3.4 Spontaneous dynamics of paddlefish electroreceptor afferents. (a) Raw extracellular recording (lower panel) and a sequence of identified spike times ofanafferent. (b) Probability density function (PDF) of interspike intervals for three representative EAs...... 35 3.5 Power spectral density (PSD) (a) and serial correlation coefficients (SCCs) (b) and for three representative electroreceptor afferents. Black, red and blue colored lines show SCCs and PSD for the same afferents as in Fig 3.1(b). In panel (a) grey lines show PSDs of renewal spike trains. Grey lines in panel (b) show minimum and maximum SCCs values for corresponding renewal spike trains...... 36 9
3.6 ISIs correlation time calculated using Eq.(2.3) versus the frequency ratio of epithelial to afferent oscillators, w, for the sample of 66 paddlefish electroreceptor afferents...... 37 3.7 (a): Dynamical entropies for ISIs sequence of an electrorecepor afferent. Red line and symbols shows dynamical entropies calculated for the original sequence. Blue line correspond to corresponding renewal surrogate ISIs. Errorbars we estimated by calculating hn for 10 surrogate sequences. (b,c): Source entropies estimated for words of n = 6 ISIs for the sample of 66 EAs vs the correlation time (b) and the coefficient of variation (c). Red dashed lines show linear regression. Values of the correlation coefficient are indicated on the panels...... 39
4.1 Deterministic dynamics of the modified Hodgkin-Huxley model. (a) Periodic 2 spike train for Iext = 1 µA/cm . (b) Firing rate of the neuron vs Iext for the 2 indicated values of the AHP conductance (gahp is units of mS/cm ...... 43 4.2 Power spectrum density of synaptic conductance. The red and blue lines correspond to simulation and analytical result, respectively. The parameters are τs = 1 ms, λ = 1000 1/s, σ = 0, m = 0.4, B = 0.018, fe = 26 Hz, Q = 20, 2 2 2 gahp = 5 mS/cm , g0 = 7.378 10− mS/cm ...... 46 4.3 Comparison between the experiment× and simulations for a vestibular afferent. Three panels show probability density of interspike intervals (ISIs), the serial correlation coefficients (SCCs) and the power spectral density of corresponding spike train. Experimental data are shown by black circles. Results of numerical simulations of the model are shown by the red lines. The parameters of the 2 model were: gahp = 1 mS/cm , τs = 2 ms , λ = 700 1/s, σ = m = 0, 2 5 2 2 gs = 0.0032 mS/cm , var(gs) = 3.6 10− [mS/cm ] ...... 48 4.4 Comparison between the experiment× and simulations for an electroreceptor afferent. Three panels show probability density of interspike intervals (ISIs), the serial correlation coefficients (SCCs) and the power spectral density of corresponding spike train. Experimental data are shown by the black circles. Results of numerical simulations of the model are shown by the red lines. The 2 parameters of the model were: gahp = 3 mS/cm , τs = 2 ms, λ = 1000 1/s, 2 m = 0.45, B = 0.018, fe = 27.1 Hz, Q = 22, < gs >= 0.0032 mS/cm , 5 2 2 var(gs) = 10 10− [mS/cm ] ...... 49 4.5 Dynamics of× the theta neuron model on a circle. Modified from G.B. Ermentrout, Scholarpedia, 3(3):1398 (2007)...... 50 4.6 Statistical measures of spontaneous dynamics from the theta neuron model Eq.(4.20). (a) Probability density of ISIs. (b) Serial correlation coefficients (SCC). (c) Power spectral density (PSD). Red lines and symbols show the model with epithelial oscillations generated a non-renewal spike trains. Blue lines and symbols correspond to the model with oscillations turned off, but with other parameters tuned to match the probability density of ISIs of the non-renewal model. Modified from [3]...... 52 10
5.1 Probability densities of ISIs. The red lines and blue lines correspond to the original and renewal models,respectively. Parameters of the original model 2 2 2 3 were gAHP = 7 mS/cm , g0 = 8.68 10− mS/cm , B = 5.86 10− , × 1 2 × 2 m = 0.4, Q = 20, τs = 2 ms, λ = 1000 s− , gs = 9.27 10− mS/cm , 5 2 2 × var(gs) = 2.3 10− [mS/cm ] . Parameters of the renewal model were × 3 2 2 gAHP = 0, m = 0, g0 = 7.9 10− mS/cm , B = 1.62 10− , m = 0, 2 2 × 4 2×2 gs = 2.413 10− mS/cm , var(gs) = 1.316 10− [mS/cm ] ...... 56 5.2 The serial correlation× coefficients (SCC) of ISIs× for the original (red line) and renewal (blue line) models. The parameters are the same as in Fig. 5.1 ..... 57 5.3 Power spectrum densities (PSD) of spontaneous spike trains generated by the original and renewal models, fa,afferent oscillator peak, fe epithelial oscillator peak ...... 58 5.4 Fano factor for the original (red line) and renewal (blue line) models for the same set of parameters as in Fig. 5.1. Horizontal dashed line shows theoretical limit for the renewal process...... 59 5.5 Statistical properties of spike trains generated by the original model (with coherent epithelial oscillations) and by the no oscillation model whereby the narrow-band noise was substituted with a short-correlated Ornstein-Uhlenbeck noise. Probability density of ISIs (a), serial correlation coefficients (b) and power spectral density (c) of spontaneous stochastic dynamics. Solid red lines correspond to the original model with epithelial oscillations. Dotted green lines refer to the model with no epithelial oscillations. The parameters 2 4 2 are: gAHP = 6mS/cm , τs = 2ms, λ0 = 10 Hz, gs = 0.081mS/cm , 5 2 2 var(g ) = 3 10− [mS/cm ] , m = 0.5, Q = 20, f = 27.5Hz. These parameters s × e resulted in the mean firing rate fa = 55 Hz and CV = 0.185...... 60 5.6 First serial correlation coefficient, C(1) (a) and correlation time tcor (b) of spontaneous ISIs sequences versus the ratio of fundamental frequencies of EO to AO,w = fe/ f a for the indicated values of the quality factor Q of epithelial oscillations. Other parameters are the same as in Fig. 5.5...... 61 5.7 Stimulus - response coherence for Gaussian band-limited stimulus with the standard deviation σ = 0.5 and cutoff frequency fc = 200 Hz (a) and fc = 20 Hz (b). Red line corresponds to the original model with epithelial oscillations. Green line refer to the model with incoherent epithelial fluctuations (No oscil.). Both models had identical spontaneous mean firing rates and CVs. The parameters are the same as in Fig. 5.5...... 63 5.8 Lower bound estimate of the mutual information rate (a) and the information gain ∆I (b) versus the ratio of fundamental frequencies of EO to AO, w = fe/ fa for the indicated values of the quality factor of narrow-band epithelial oscillations. Stimulus parameters were σ = 0.5, fc = 20 Hz. Other parameters were the same as in Fig. 5.5...... 65 11
5.9 Cross-trial response variability. The original model and the no oscillation model were stimulated repeatedly with identical realization of Gaussian noise band limited to fc = 20 Hz. (a) Short segment of the stimulus. (b -d) Raster plots of models responses to 200 trial presentations of the stimulus. Vertical axis is the trial number; horizontal axis is the time with respect to the stimulus onset. Dots indicate appearance of a spike in a trial relative to the stimulus onset. Panel (b) is for the stimulus standard deviationσ = 0.5 (weak stimulus). Panels (c,d) is for the case of strong stimulus with σ = 2.0 applied to the original model with narrow-band noise (c) and to the no oscillation model with broad-band OU noise (d). Other parameters were the same as in Fig. 5.5. . . . . 67 5.10 Stimulus-response (SR) and response-response coherence for simulations shown in Fig. 5.9. (a) SR (dashed line) and RR (solid line) coherence functions for the original model with narrow-band noise. (b) RR coherence for the original model (oscil., solid line) and for the no oscillation model (no oscil., dotted line)...... 68 5.11 Lower and upper bounds of information rate (a) and information gain ∆I (b) versus the standard deviation σ of Gaussian stimulus. normalized to the standard deviation of epithelial oscillations. The stimulus cutoff frequency is fc = 20 Hz. The parameters of the model are the same as in Fig. 5.5...... 69
6.1 Layered structure of the feed-forward sensory network in paddlefish electrore- ceptors ...... 73 6.2 Coherence function between two inputs, Eq.(6.4). (a) Input coherence for τc = 0.1, fe = 26 Hz, c = 1, α = 5 and indicated values of the quality factor. (b) Input coherence for Q = 20 and indicated values of α. Other parameters are the same as in panel (a)...... 77 6.3 Transformations of noise correlations for two identical neurons driven by independent broad-band noises and a common narrow-band epithelial noise. (a) Input coherence (gray line) and output coherences for the indicated values of the frequency ratio w. (b) Contour lines of the coherence sensitivity χ( f ), Eq.(6.6), versus frequency and w. fe = 26 Hz indicates the location of epithelial oscillations frequency. Other parameters are A = 0.5, D = 0.2, λ = 0.02, β = 0.3, c = 1...... 80 6.4 Transformations of noise correlations for two identical neurons driven by independent broad-band noises and a common narrow-band epithelial noise for the indicated values of quality factor Q. (a) Output coherence. (b) Coherence sensitivity χ( f ). Other parameters are w = 0.5, A = 0.5, D = 0.2, λ = 0.02, β = 0.3, c = 1...... 82 6.5 Transformations of noise correlations for two identical neurons driven by independent broad-band noises and a common narrow-band epithelial noise for the indicated values of standard deviation D of spatially uncorrelated broad- band noise. (a) Output coherence. (b) Coherence sensitivity χ( f ). Other parameters are w = 0.5, A = 0.5, Q = 20, λ = 0.02, β = 0.3, c = 1...... 83 12
6.6 Noise correlations of a pair of non-identical neurons. Neurons are character- ized by different firing rates, fa1 = fe/w1 and fa2 = fe/w2, where fe = 26 Hz and w1,2 are indicated in the figure. (a) Input and output coherences for w1 = 0.5 and indicated values of w2. (b) Low-frequency averaged output coherence coh as a function of neuronal firing rates parametrized by w1,2. Other param- eters are A = 0.5, Q = 20, λ = 0.02, β = 0.3, c = 1...... 84 6.7 Information density for a network of M = 5 neurons stimulated by Gaussian band-limited stimulus with σ = 0.2 and cutoff frequency fc = 20 Hz, for the indicated values of the quality factor of epithelial oscillations Q (a) and the partial correlation parameter c (b). Firing rates of neurons were uniformly distributed from 42 to 62 Hz. Other parameters are A = 0.5, λ = 0.02, β = 0.3. . 86 6.8 Information density for a network of M = 5 neurons stimulated by Gaussian band-limited stimulus with σ = 0.3 and cutoff frequency fc = 20 Hz. (a) Network of non-identical neurons with firing rates uniformly distributed within interval [52 ∆ fa/2, 52 + ∆ fa/2]. Values of ∆ fa are indicated in the figure. Other parameters− are A = 0.5, Q = 20, D = 0.2, λ = 0.02, β = 0.3. (b) Network of identical neurons with the mean firing rates fa = 52 Hz, ∆ fa = 0. Values of independent broad-band noise intensity D are indicate in the figure. . 88 6.9 Lower bound of the mutual information rate versus the intensity of independent broad-band noise D for networks of identical neurons and indicate size M. Other parameters are w = 0.5, A = 0.5, Q = 20, λ = 0.02, β = 0.3...... 89 13 1 Introduction1
Oscillatory activity is observed in various regions of the central nervous system [5].
Rhythmical activity is well-known to mediate walking and breathing. However, the concept that many peripheral sensory systems undergo intrinsic oscillations instead of being just passive transducers of external stimuli has started to emerge only in recent decades. Self-sustained rhythms have been observed in several types of peripheral sensory receptors. For example, peripheral cold thermoreceptors produce periodic or chaotic bursts of spikes when chilled [6] which arise due to slow ion currents in terminals of sensory neurons [7, 8, 9]. Another example is oscillatory responses observed in olfactory
receptor neurons [10, 11]. This dissertation research focuses on 2-stage hair cell - primary afferent receptors.
Such sensory receptors have stimulus-transducing hair cells or similar receptor cells synaptically coupled to sensory neurons (afferents) that convey transduced stimuli to the central nervous system. 2-stage receptors are found in the auditory, vestibular, lateral line,
and electrosensory systems. Recordings of spontaneous firing from auditory afferent fibers in turtles, lizards and birds showed a series of similarly spaced peaks in interspike
interval histograms, indicating ”preferred intervals” [12, 13]. Some primary vestibular afferents demonstrate extremely regular periodic firing [14, 15]. Isolated hair cells of
auditory and some vestibular receptors of lower vertebrates undergo ringing (damped membrane potential oscillations) in response to step changes in membrane potential [16, 17, 18]. Such hair cells may undergo spontaneous oscillations of membrane potential
[19, 20, 21], or spontaneous periodic motions of ciliary hair bundles [22, 23, 24]. These spontaneous oscillations of hair cells are reflected in periodic sequences of post-synaptic
excitatory potentials and in periodic firing of primary afferents [21]. Intrinsic (spontaneous) oscillations have been found in both sensory cells and afferent neurons in
1Part of this section appeared in Nguyen and Neiman (2010) [4]. 14
electroreceptors (ERs) [25, 26, 27]. In particular, it was demonstrated that paddlefish ERs have a 30- 70 Hz oscillator in each afferent terminal, along with a population of 26 Hz oscillators in the sensory epithelia [28]. Interactions of these two types of oscillators cause
biperiodic afferent firing with two fundamental frequencies. The figures below sketch peripheral hair cell – afferent receptors. Fig. 1.1 shows
vestibular receptors and Fig. 1.2 represents ampullary electroreceptors in paddlefish.
Met Met Met Met Met
I I I II II Kv7x
*Na K * * *
C D B
Figure 1.1: Schematic drawing of different classes of vestibular afferents. Shown are two types of sensory hair cells (I and II) and three classes of afferents: pure-calyx (C), dimorphic (D) and pure button (B). Mechanoelectrical transduction ion channels in stereocilia of hair cells are labeled as Met. Possible locations of spike initiation zones in afferents are labeled by asterisks. Grey unlabeled structures are efferents, which are neurons transferring feedback signals from the brain to hair cells. Modified from a courtesy figure obtained from Dr. Ruth Anne Eatock (see also [1]).
Both examples show sensory hair cells transducing external mechanical (in the case of vestibular receptors) or electrical (in the case of ERs) signals in voltage variations across their basal membranes. These membrane voltage variations are then transmitted via 15
Pore Canal Sensory wall Epithelia
Excitatory Canal hair cell- to-afferent synapses Afferent nerve (2−4 axons) Figure 1.2: The organization of an electroreceptor on the rostrum of paddlefish. Shown is a cluster of skin pores each leading to a canal. Sensory cells in epithelial layer at the bottom of canals innervated by a few afferent neurons. Modified from [2].
excitatory synapses to the afferent neurons which encode external signals in sequences of
action potentials. Cellular oscillators, including sensory hair cells and sensory neurons are examples of non-equilibrium, nonlinear, open, complex systems and as such require approaches from
complex system theory, nonlinear dynamics, information theory and modern methods of signal processing. The main goal of this thesis is to gain better understanding of how hair
cell – afferent receptors utilize their intrinsic dynamics to optimize responses and information encoding. In this study we use computational approaches to study the role of intrinsic oscillatory dynamics in the encoding of external signals.
What are possible functions of oscillations in peripheral receptors? 1) Frequency tuning. The existence of a natural frequency promotes frequency selectivity of a receptor. Examples include mechanical and electrical resonances in mechanosensory hair cells [16, 17, 18].
2) Nonlinear amplification. Dynamical systems are most sensitive to external perturbations at bifurcations, where a system’s mode of operation changes abruptly as a control parameter exceeds some critical value. For example, one model for the high
sensitivity and selectivity of mammalian hearing proposes that auditory mechanisms are 16
normally poised near a Hopf bifurcation, such that a weak stimulus of appropriate frequency is greatly amplified [29, 22, 23]. 3) Precise coding. External analog stimulus information is encoded by an individual
peripheral sensory receptor into a time series of afferent spikes, sent to the brain. Regularly firing neurons may potentially provide precise sampling (coding) of external
stimuli [30, 15]. 4) Higher order correlations. Multi-modal oscillations may generate correlations in
sequences of action potentials and of receptor responses to external signals. These correlations provide a higher degree of ordering of neural responses and may enhance information transmission and detection performance of receptors. Serial correlations of
afferent interspike intervals have received recently much interest (for review see [31, 32]) as the times between sequential spikes of some auditory afferents and all electroreceptor
afferents tend not to be independent. Instead, a long interspike interval tends to follow a short interval, and vice versa. Such anticorrelations are a type of ordering, and result in reduced low-frequency noise in afferent firing. This ”noise shaping” yields higher
signal-to-noise ratios and increased discriminability and information transfer, at low frequencies. Paddlefish ER afferents show long-lasting serial correlations, that occur due to self-sustained epithelial oscillations [33, 28, 2]. Using simple theta neuron and integrate and fire models, it was shown that oscillatory internal noise, as found in paddlefish ERs,
can augment the information transfer through a sensory receptor [34, 3]. Here we extend these studies using a conductance-based, Hodgkin-Huxley type model with stochastic excitatory synaptic input.
Experimental validation of the functional role of oscillations in peripheral receptors involves several difficulties. An ultimate hypothetical test would involve abolishing oscillations and then analyzing how the sensitivity, encoding and detection characteristics of a sensor change as compared to normal conditions, i.e., when oscillations are present. 17
Abolishing of oscillations is achieved by pharmacological blockage of certain ion channels in the cell membrane. Such procedures may also change other properties of a sensor, such as mean firing rate. Furthermore, pharmacological procedures are usually
done in in vitro preparation, which often leads to the question of whether a receptor system is still working normally. That is why modeling studies and methods of advanced
signal processing are crucial in understanding the functional significance of coherent rhythmic activity in sensory systems. These methods allow the investigator to impose
clear constraints on a system while varying its parameters and studying its responses to various signals. The majority of previous theoretical and computational work studied the influence of oscillations and serial interspike interval correlation on responses to weak
stimuli, i.e. linear responses [35, 36, 3, 37]. The question of how internal oscillations and correlations affect nonlinear neural responses has not been studied in full details. We note,
that strong stimuli may induce distinct neural responses in the form of bursts, as in paddlefish ERs [38, 39] or as mode locking as in vestibular afferents [40]. Peripheral receptors in many sensory systems are organized in limited scale,
feed-forward networks passing information through a series of network layers, and ultimately to the central nervous system (CNS). A classic example is the retina, a
feed-forward network serving as a preprocessor of visual stimuli [41, 42]. Often peripheral receptors are characterized by spontaneous, noisy oscillatory activity which
may introduce temporal and spatial correlations across elements of a neural network. Problems of how these correlations are processed by the network and how they influence information transmission is of high and continued interest
[42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55]. Examples include spontaneous stochastic oscillations in hair cells that are innervated by several primary sensory afferents in auditory, vestibular and electro sensory receptors. Unlike the majority of theoretical works mentioned above that study weakly correlated, white inter-neuronal noise, we study 18
the influence of strongly correlated noise on spontaneous activity and information transmission in a model of a limited-scale network of electroreceptors. The structure of this work is as follows. Chapter 2 introduces measures of
spontaneous neuronal activity as well as stimulus-response measures, including estimates of the information rate. Chapter 3 is devoted to the analysis of spontaneous dynamics of
vestibular and electroreceptor afferents. Chapter 4 introduces two classes of computational models used in this work. A modified Hodgkin-Huxley model is used in Chapter 5 to
study the effect of epithelial oscillations on linear and nonlinear stimulus encoding and information transfer by a single sensory neuron. Finally, Chapter 6 is devoted to study of a small network of oscillatory neurons subjected to correlated internal noise. 19 2 Spontaneous dynamics and response properties of peripheral sensory receptors
2.1 Statistical measures of spontaneous activity
Experiments in sensory neurophysiology analyze the sequence of action potentials resulting from spontaneous activity of a neuron or the response of the neuron to external stimulation. Neuronal action potentials are believed to be the main carrier of information in the nervous system [56]. The generation and propagation of action potentials is
inherently noisy. Noise inevitably present in any dissipative physical system. In biological neurons the main sources of noise are: (i) fluctuations of ionic currents due to
stochasticity of ion channels and pumps and (ii) synaptic noise arising from the inputs to a neuron from other cells. This inevitable noise results in two kinds of neuronal variability.
First, spontaneously firing neurons are characterized by fluctuating firing rates. Second, responses of a neuron to identical stimuli vary from trial to trial. Theory of stochastic point processes [57] is a powerful tool widely used in
computational neuroscience. In the framework of this theory a spike train is described by
a stochastic point process, i.e. discreet sequences of time events, t1, ..., tm, where tm is the time of arrival of the m-th spike. The intervals between neural firing, called interspike
intervals (ISIs), are defined as I = t + t . The variability of neuronal firing can be m m 1 − m expressed in terms of the statistical properties of ISI distributions. The mean firing rate and the coefficient of variation are among most often used statistical measures of neuronal firing. Both these measures can be extracted from the probability density function (PDF) of ISIs, P(I). In neuroscience literature this measure is often called the interspike interval histogram (ISIH). Given a sequence of ISIs the mean firing rate, f¯ and the coefficient of
variation (CV) are defined as
f¯ = 1/I¯, CV = std(I)/I¯, (2.1) 20
where I¯ is the mean ISI and std(I) is the standard deviation. From (2.1) it is clear that the CV is the standard deviation in units of the mean. It is instructive to consider a Poisson spike train as a reference model for a spike
train. The ISIs distribution for the Poisson spike train is exponential, P(I) = λ exp( λI), − where λ is the intensity (rate) of the Poisson process. The mean ISI is I¯ = 1/λ and the CV
of the Poisson spike train is 1, CV = 1. At the other extreme, a deterministic periodic firing is characterized by zero variability, CV = 0.
The spiking activity of a neuron can be described as either a renewal or a non-renewal point process. The renewal process is one in which all memory is erased after each action potential so that consecutive ISIs are statistically independent. The Poisson
process is one example of a renewal process that is often used to describe the firing of cortical neurons [58]. However, many types of neurons in a variety of sensory and motor
systems display non-renewal spiking statistics with extended correlations between ISIs [31, 32]. These neurons display a memory in their spiking activity that persists over multiple ISIs. This memory can be quantified in terms of the normalized autocorrelation function of ISI, and by serial correlation coefficients (SCCs) C(k), The serial correlation coefficients (SCCs) are given by
2 I I + I¯ C(k) = m m k − , (2.2) var(I)
where represents averaging over ISIs and var is the variance of ISIs I2 I¯2. The serial m − correlation coefficient measures the statistical dependence of pairs of ISIs separated by the intervals (lag). For a renewal process C(k) = 0 for values of k. Positive serial correlations (C(k) > 0) on average, a long ISI is followed by even longer ISI. Negative serial correlations (C(k) < 0) mean that ISI tend to be followed by a shorter ISIs and vice versa. The temporal extent of serial correlations is characterized by the correlation time [4],
∞ t = I¯ C(k) , (2.3) cor | | k=1 21
where the sum is taking over lags with SCCs significantly different from 0. A fundamental problem of neuroscience is to understand how neurons encode, decode and transmit information [30, 56]. Perhaps the simplest model for neural coding is
the so-called rate code where a stimulus is encoded in spike counts in a succession of non overlapping time windows, each of the length T. These spike counts n(T) can be also used to characterize the variability neuronal dynamics. The variability in the spike count distribution is most conveniently characterized by the Fano factor defined as [59],
F(T) = var[n(T)]/ n(T) , (2.4) where var[n(T)] and n(T) are the variance and the mean of the spike count, respectively. The Fano factor is 1 for Possion process, i.e., F(T) = 1 at any T. At short time scales, i.e. for T I¯ any process approaches to this Poissonian limit. Consequently F(T) 1 as ≪ → T 0. For a renewal process F(T) CV2 as T . We note also that the Fano factor → → → ∞ can be used to characterize the discrimination capacity of sensory neurons: smaller values of the Fano factor (i.e. smaller variability) indicate better discrimination between spike count distributions with and without stimulus [34, 59]. The power spectral density (PSD) is used to study variability of neuronal firing in the
frequency domain. If a particular shape of the action potential does not matter, then the spike train can be represented as a sequence of delta functions centered at the spike times with subtracted mean firing rate as,
M x(t) = δ(t t ) f¯. (2.5) − m − m=1
The PSD, Gxx( f ), of a stationary stochastic process x(t) is defined as [60],
G ( f ) = X( f )X∗( f ) , (2.6) xx
where X( f ) is the Fourier transform of a realization of x(t), asterisks indicates complex conjugate and the averaging is performed over the ensemble of realizations of x(t). In 22
practice a single long realization of x(t) is partitioned into overlapping windows to mimic an ensemble of realizations. Fast Fourier transforms (FFT) are performed for each window and then the squared amplitude of the Fourier transforms are averaged to obtain
an estimate of the PSD [60]. For the spike train given by (2.5) the PSD has units of (spikes/sec)2/Hz or simply Hz. A Poisson spike train with the rate λ is characterized by a
uniform PSD, Gxx( f ) = λ. As we noted above, any point process approach Poissonian at short time scales. Thus, the PSD of any spike train should approach to the mean firing rate
f¯ at high frequencies, f f¯. A purely periodically firing neurons, by contrast,is ≫ characterized by a delta-peak in its PSD centered at the mean firing rate, δ( f f¯). − To characterize nonlinear correlations in sequences of ISIs, we employ the so-called
dynamic entropies. These measures characterize correlations not just between pairs of ISIs, but also among various ISIs patterns [61, 62]. In this work we used these high-order
correlation measure to inquire whether they can be used as additional classifiers of variability for utricle and electroreceptor sensory neurons. Given an ISI sequence I , we { m}
introduce a binary alphabet as follows: S m = 0 if Im < I¯ and S m = 1, otherwise. Then for
the resulting binary sequence we calculate probabilities of words containing n letters, Pn, and then the n-block Shannon entropies,
Hn = Pn log2 Pn, (2.7) −
where the summation is carried out over all n-words with positive probabilities, Pn > 0.
The dynamical entropies hn are defined as
h = H + H , h H , (2.8) n n 1 − n 0 ≡ 1 and indicate the average information necessary to predict the n-th +1 symbol in a sequence, given knowledge about preceding n symbols. hn decreases with increase of n and saturates to the so-called source entropy in the limit of large n. For a renewal process, hn rapidly saturates to a positive number determined by the symmetry of ISI distribution. 23
In particular for the Bernoulli string, hn = 1 for all n, i.e. the Bernoulli symbolic sequence
is completely unpredictable. For a periodic ISI sequence hn rapidly drops to zero, so that only a finite number of symbols must be observed to predict with certainty the next
symbol. Correlation and spectral measures of a non-renewal stochastic process can be
compared with those of an equivalent renewal process.For such a comparison we need to generate a sequence of events which will retain certain properties of the original process,
but otherwise will lose serial correlations between ISIs. In this work we used a shuffling of ISIs to generate such renewal surrogates as follows. A given series of ISIs, I , { m} (obtained either from experimental data or from numerical simulations) was randomly shuffled, i.e. the positions of I were randomly exchanged, to obtain a new sequence I˜ . m { m} Although the probability distribution, and thus the mean and the coefficient of variation of
ISIs, are retained by this procedure, all serial correlations are removed. The spike times of
the shuffled ISIs are obtained as t˜m+1 = t˜m + I˜m. The statistical measures introduced above were used to analyze data from numerical
simulations and experimental data from utricle and electroreceptor afferents. For experimental data, all computations were performed in Matlab with a custom made
package using Matlab’s signal processing toolbox. A similar package was developed in Fortran to process spike trains obtained in numerical simulations.
2.2 Characterization of response dynamics of a single neuron
Understanding of how a sensory stimulus is represented in neural response and mechanisms by which sensory neurons encode information remains important problems in
neuroscience. In this section we introduce and discuss various measures of response dynamics which will be used further on in the dissertation. Figure 2.1 shows 24 spike train stimulus neuron stimulus estimate
Figure 2.1: Schematics of neuron perturbed by the stimulus s(t).
schematically a procedure often used in the studies of response properties of a neuron. To probe the response dynamics a time varying stimulus, s(t), is administered to a neuron. In response the neuron generates or modulates a spike train. Often a broad band Gaussian
stimulus s(t) is used. The response is then measured as a sequence of neuron spike times t and the response spike train x(t) is represented by a sequence of delta functions { n} centered at t , Eq.(2.5). In the case of weak stimulus, one can then use linear response { n} theory to estimate the sensitivity of the neuron. The transfer function (also called the sensitivity) is calculated as [60]: G ( f ) H( f ) = | xs |, (2.9) Gss( f )
where Gxs( f ) is the cross-spectral density of the spike train and the stimulus, and Gss( f ) is the PSD of the stimulus. The units of the transfer function depend on the units of the
stimulus. For example, the stimulus for an electroreceptor afferent is time-varying external electric field with units of [V/m] and the transfer function has units of [(spk/s)/(V/m)]. For a mechanosensory neuron, e.g. vestibular afferent, the stimulus is displacement [m], so that the transfer function has units [(spk/s)/m]. Thus the transfer function measures by how much the firing rate changes per unit of external stimulation. The advantage of using broad-band Gaussian noise as a stimulus compared to ,e.g., a sinusoidal stimulus is that Eq.(2.9) allows estimation of the frequency response of a system within the whole
frequency band of the stimulus at once instead of at the single frequency of the sinusoidal stimulus. 25
The stimulus-response (SR) coherence is another linear input-output measure defined as [60] 2 Gxs( f ) CSR( f ) = | | . (2.10) Gss( f )Gxx( f ) The SR coherence is essentially the normalized cross-correlation coefficient in the frequency domain, i.e. it measures linear correlations between Fourier components of stimulus and response. As such the SR coherence is a dimensionless quantity varying from 0 (no stimulus-response correlations; stimulus and response are incoherent) to 1 (perfect correlations; stimulus and response are completely coherent). For an ideal linear system with no internal noise, the SR coherence is 1. Since a real physical system is contaminated by noise actual SR coherence values are typically < 1. Any non-linearity in a system will also result in coherence values < 1. Thus, SR coherence can be used to assess both the internal noise and the nonlinearities of a system. More importantly, SR coherence measures the linear encoding properties of a neuron. That is, high SR coherence values in a frequency band indicate faithful encoding of stimulus in the firing rate of a neuron as we discuss further on.
Response-response (RR) coherence is a measure that does not require the assumption of response linearity [63, 30, 64, 38]. A neuron is stimulated by the same realization of a
stimulus repeatedly and then the variability of responses xk(t) is characterized using a
coherence function between pairs of responses xk(t) and x j(t):
2 K 2 K(K 1) k=1 j