OPTICAL RECORDING AND PHOTO MODULATION IN THE STUDY OF DYNAMICS IN NEURAL CIRCUITS HENG XU. B.Sc., NANJING UNIVERSITY, 2003

Transcription

1 OPTICAL RECORDING AND PHOTO MODULATION IN THE STUDY OF DYNAMICS IN NEURAL CIRCUITS BY HENG XU B.Sc., NANJING UNIVERSITY, 2003 M.Sc., BROWN UNIVERSITY, 2005 SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY IN THE DEPARTMENT OF PHYSICS AT BROWN UNIVERSITY PROVIDENCE, RHODE ISLAND MAY 2010

2 Copyright 2009 by Heng Xu

3 This dissertation by Heng Xu is accepted in its present form by the Department of Physics as satisfying the dissertation requirement for the degree of Doctor of Philosophy Date Arto V. Nurmikko, Advisor Recommended to the Graduate Council Date Barry W. Connors, Reader Date James M. Valles, Reader Approved by the Graduate Council Date Sheila Bonde, Dean of the Graduate School iii

4 VITA Heng Xu was born in Nanjing, China on December 13th, He received his B.Sc. in Physics from Nanjing University in He subsequently started his graduate study at Brown University with a fellowship from the Department of Physics. Since 2004 he has been supported by a research assistantship, and he received his Sc. M. in Physics from Brown University in PEER REVIEWED PUBLICATIONS 1. Heng Xu, Arto V. Nurmikko, Carlos D. Aizenman, Visual experience-dependent maturation of correlated neuronal activity patterns in a developing visual system, manuscript in preparation. 2. Wei Dong, Ryan H. Lee, Heng Xu, Shelley Yang, Kara G. Pratt, Vania Cao, Yoon-Kyu Song, Arto V. Nurmikko, Carlos D. Aizenman, "Visual avoidance in Xenopus tadpoles is correlated with the maturation of visual responses in the optic tectum", Journal of Neurophysiology. 10(1152), (2008). 3. Heng Xu, Jiayi Zhang, Kristina M. Davitt, Yoon-Kyu Song, Arto V. Nurmikko, "Application of blue-green and ultraviolet micro-leds to biological imaging and detection", Journal of Physics D 41(9), (2008). 4. Heng Xu, Kristina M. Davitt, Wei Dong, Yoon-Kyu Song, William R. Patterson III, Carlos D. Aizenman, Arto V. Nurmikko, "Combining multicore imaging fiber with matrix addressable blue/green LED arrays for spatiotemporal photonic excitation at cellular level", IEEE Journal of Selected Topics in Quantum Electronics 14(1), 167 iv

5 (2008). 5. Sowmya Venkataramani, Kristina M. Davitt, Heng Xu, Jiayi Zhang, Yoon-Kyu Song, Barry W. Connors, Arto V. Nurmikko, Semiconductor ultra-violet light emitting diodes for flash photolysis, Journal of Neuroscience Methods 160(1), 5 (2007). 6. Jiayi Zhang, Sowmya Venkataramani, Heng Xu, Yoon-Kyu Song, Hyun-Kon Song, G. Tayhas. R. Palmore, Justin Fallon, Arto V. Nurmikko, Combined topographical and chemical micropatterning of neural template for cultured hippocampal neurons, Biomaterials 27(33), 5734 (2006). 7. Sowmya Venkataramani, Kristina M. Davitt, Jiayi Zhang, Heng Xu, Yoon-Kyu Song, Barry W. Connors, Arto V. Nurmikko, Compact wemiconductor light-emitting diodes for dynamical imaging of neuronal circuitry, IEEE Journal of Selected Topics in Quantum Electronics 11(4), 785 (2005). v

6 ACKNOWLEDGEMENTS I would like to express my sincere gratitude to my advisor, Prof. Arto V. Nurmikko for his invaluable guidance, inspiration, encouragement throughout my Ph.D. study. I would like to appreciate Prof. Carlos D. Aizenman and Prof. Barry W. Connors for their important guidance in my research in neuroscience I would like to thank Prof. James M. Valles for his time in reading and commenting on this thesis, and Prof. Leon N. Cooper and Prof. Elie Bienenstock for his time in commenting on my research. I would like to give special thanks to Dr. Sowmya Venkataramani for her selfless mentoring at the beginning stage of my research; Dr. Kristina M. Davitt and Dr. Yoon-Kyu Song for their help and technical support in microelectronic fabrication; Dr. Wei Dong from Aizenman group for his mentoring and help of tadpole electrophysiological recording technique; Dr. Scott Cruikshank from Connors group for his tutoring of brain slice preparation and electrophysiological recording techniques; Dr. Hayato Urabe for his time and help in optogenetics project. I would like to thank all previous and current members of Nurmikko group: Dr. Ilker Ozden, Dr. Yiping He, Dr. Hongbo Peng, Dr. Qiang Zhang, Dr. Tolga Atay, Dr. Jiayi Zhang, Dr. Hyunjin Kim, Dr. Yanqiu Li, Dr. Ming Yin, Dr. Juan Aceros, Mr. Sufei Shi, Mr. Cuong Dang, Mr. Dave Borton, Ms. Jing Wang, Ms. Sunmee Park, Mr. Andy Blaeser, Mr. Fabien Wagner, Mr. Naubahar Agha, and Mr. Jacob Komar for their help in my research at Brown. I would like to thank Michael Jibitsky, Sandra Van Wagoner, Sandra Spinacci, Irina vi

7 Sears, and Saundra Patrick for their excellent technical and administrative support. My thanks also go to all of my friends here for their support and motivation throughout my life in Providence. Last but not least, I would like to dedicate my thesis to my family - my parents, grandparents and all other relatives, who are always behind me and supporting me with their endless love. vii

8 To my family To life 谨以此文献给我亲爱的家人 并纪念我在布朗度过的日日夜夜 viii

9 TABLE OF CONTENTS CHAPTER 1 INTRODUCTION... 1 CHAPTER 2 RETINOTECTAL SYSTEM DEVELOPMENT Introduction Visual experience independent mechanism Visual experience dependent mechanism of development and neural plasticity Modeling of neural plasticity CHAPTER 3 OPTICAL RECORDING OF NEURONAL NETWORK ACTIVITY Introduction Calcium indicators Modeling of calcium dynamics Other optical imaging techniques Acquisition of optical signals CHAPTER 4 FAST CALCIUM IMAGING STUDY OF TADPOLE VISUAL SYSTEM DEVELOPMENT Introduction Experimental method and set-up Development of cross-correlation of evoked neural activity Development of visual response timing Development of the reliability of neural response ix

10 4.6 Development of the spatial-correlation and spatial-temporal relations Development of cross-correlation of spontaneous neural activity Modeling of activity dependent tadpole retinotectal system development Discussion CHAPTER 5 10x10 MICRO-LED ARRAYS BASED IMAGE PROJECTION DEVICE AND ITS APPLICATION IN TADPOLE VISUAL SYSTEM DEVELOPMENT STUDIES Introduction Micro-LED design and fabrication Micro-LED array performance Device packaging and control Tectal visual receptive field mapping of developing tadpoles CHAPTER 6 OPTICAL NEURAL MODULATION AND OPTOGENETIC TECHNIQUE Introduction Optogenetics Photostimulation of ChR2 expressing acute mouse brain slices CHAPTER 7 ALL-OPTICAL STIMULATION/RECORDING STUDY OF THE CORTICAL UP STATE OF MOUSE PREFRONTAL CORTEX IN ACUTE BRAIN SLICES Introduction Persistent activity, cortical UP state and the mechanism of working memory x

11 7.3 All-optical stimulation/recording of neural activity in mouse brain slices Photostimulation of MDN axon projection for the study of cortical UP state in acute mouse PFC slices CHAPTER 8 CONCLUSION SUPPLEMENTARY CHAPTER COMBINED TOPOGRAPHICAL AND CHEMICAL MICRO-PATTERNS FOR TEMPLATING NEURONAL NETWORKS S.1 Introduction S.2 Material and methods S.3 Results and discussion S.4 Conclusion BIBLIOGRAPHY xi

12 LIST OF TABLES Table 3.1 Extracelluar and cytosolic concentration of major ion types Table 3.2 Parameters for multi-compartment calcium dynamics model Table 6.1 Summary of viral vectors for gene delivery in nervous system xii

13 LIST OF ILLUSTRATIONS Figure 2.1 The early development of the brain Figure 2.2 Tadpole retinotectal projection development guided by chemoaffinity cues Figure 2.3 Chemoaffinity cues in the retinotectal projection Figure 2.4 Forward and reverse signals communicated downstream of the Eph-ephrin complex Figure 2.5 Experimental observation of Hebbian plasticity Figure 2.6 AMPA and NMDA receptors Figure 2.7 Biophysical mechanisms of Hebbian plasticity Figure 2.8 STDP and BCM laws Figure 2.9 Biophysical model of Hebbian plasticity Figure 3.1 Structures of Ca 2+ chelators and some calcium indicators Figure 3.2 Optical properties of different calcium sensitive dyes Figure 3.3 The schematic structures and the calcium sensitive mechanism of cameleons Figure 3.4 Schematic of the cell loading process Figure 3.5 Recording schemes of calcium indicator fluorescence Figure 3.6 Schematic drawing of the single-compartment model Figure 3.7 Schematic drawing of the multi-compartment model Figure 3.8 Intracellular distribution of free calcium ions and calcium bounded indicator molecules Figure 3.9 Flurescence signal of calcium indicator and the spike reconstruction techniques Figure 3.10 Voltage sensitive dye recording of neural activity Figure 3.11 Development of a fast reporting VSFP Figure 3.12 Typical experimental set-up for optical recording of neural activity Figure 3.13 Transmission spectra of the Oregon Green 488 filter set xiii

14 Figure 4.1 High-speed calcium imaging of visual responses in tadpole optic tectum in vivo Figure 4.2 Spike train reconstruction scheme (data from stage 46 animals) Figure 4.3 Accuracy of spike train reconstructions Figure 4.4 Developmental increase in correlated neural activity of tectal neurons visual response Figure 4.5 Developmental increase in visual response synchrony among tectal neurons Figure 4.6 Developmental increase in response reliability in single neurons is activity dependent Figure 4.7 Development of spatiotemporal properties of visual responses Figure 4.8 Developmental increase in correlated spontaneous tectal neural activity Figure 4.9 Schematic of the computational model for retinotectal system development Figure 4.10 Computational modeling of retinotectal system development Figure 5.1 The typical InGaN/GaN LED wafer structure Figure 5.2 Design of the matrix addressable microarray LED Figure 5.3 Schematic of the fabrication Process Figure 5.4 Performance of typical array elements Figure 5.5 Device packaging and control Figure 5.6 Schematic of tadpole visual response recording Figure 5.7 Visual receptive field mapping of optic tectal neurons Figure 6.1 Photorelease of caged compounds Figure 6.2 Photoswitchable molecules modulation of potassium channel conductance Figure 6.3 Light sensitive cation channel ChR Figure 6.4 Photostimulation of ChR2 transfected mouse brain slices Figure 7.1 Cerebral cortex of the brain Figure 7.2 Typical firing patterns of different cortical cell types xiv

15 Figure 7.3 Monkey working memory test Figure 7.4 A simple recurrent network model to generate delay-period persistent activity Figure 7.5 Field theory of large neuronal networks Figure 7.6 Standing pulse solutions of a lateral inhibitory neural field Figure 7.7 Activity of different brain areas during the delayed response tasks Figure 7.8 The distributive model of working memory Figure 7.9 3D structure of medial dorsal nucleus axon projection to prefrontal cortex in mice Figure 7.10 All-optical modulation/recording scheme Figure 7.11 Photostimulation of ChR2 expressing MDN afferent axons Figure 7.12 Calcium imaging of mouse PFC brain slices Figure 7.13 Ratiomatric calcium imaging of multiple PFC neurons Figure S.1 Fabrication procedure and photomask geometry for hybrid template Figure S.2 Fluorescent image of hybrid template using phase-contrast microscope Figure S.3 Hippocampal neurons growing on hybrid templates Figure S.4 Hippocampal cultures of 13 DIV neurons Figure S.5 Typical voltage responses of hippocampal neurons xv

16 CHAPTER 1 INTRODUCTION The brain, as the major part of vertebrate nervous system, contains a vast number of multiply interconnected neurons (10 11 neurons with synaptic connections for human brain) [17], each of which acts as a basic signal processing unit to transmit and modulate information coming from other parts of the brain or the body. The complex synaptic connections between different types of neurons, or the so-called neuronal network, induce the extremely complex and nonlinear dynamical properties of the system, which are believed to be the microscopic basis of many macroscopic phenomena, ranging from animal behavior to human cognition. The central goal of the study of brain dynamics involves the quantitative understanding of the brain s electrical and biochemical activity at every different scales of analysis (from molecular level, cellular level, all the way up to the system level), as well as the corresponding anatomical and physiological structures that account for the dynamics. The latter part mainly relies on different types of static imaging techniques, while the former part requires the dynamic monitoring and manipulation of brain activity. In this thesis, I will exploit newly developed optical methods for recording and modulaton of neural activity, and apply these to acquire insight in two specific important problems in neuroscience/physics, namely the visual system development of Xenopus laevis tadpoles, and the modulation of cortical UP states in mouse prefrontal cortical slices by inputs from the medial dorsal nucleus of the thalamus. To access single (or a few) neural cells directly, conventional micropipette based elec- 1

17 trophysiological recording and manipulation of neural activity has been a standard tool of cellular and circuit level neuroscience studies [17, 30], with the advantage of high temporal resolution, high signal-to-noise ratio, and small side effects (when used for modulation). The limited number of recording/modulation channels (normally 3), however restricts its application in most large scale (> 100 neurons) systems. Extracellular multi-electrode array (MEA) based recording and stimulation strategy, which is designed for large scale systems (especially large in vivo models, such as primates), suffers from a small signal-to-noise ratio, large side effects, non-selective activation (for stimulation), and mechanical (i.e. surgical) invasiveness [31-33]. Global level noninvasive tools such as EEG, fmri, and TMS solve the problem of mechanical invasiveness, i.e. piercing the skin envelope and the skull for in vivo work, but further reduce the signal-to-noise ratio and spatial resolution [33-36]. Optical recording and modulation methods, as a relatively new approach, have the potential of achieving very high spatial/temporal resolution and signal-to-noise ratio [8, 9, 13, 37] for reasonable large populations of neurons (> 1000) [37], thereby enabling access to neural microcircuits, while minimizing the mechanical invasiveness and non-specificity (for modulation), and therefore are very promising tools for neuroscience studies, especially for studying neural circuits in vitro models such as in rodent brain slices, or in vivo cases of animals where such circuits are readily accessible by visible light. Tadpole models are an excellent match with optical imaging tools and their study forms the neurophysics core of this thesis. In the first part of this thesis, we focus on neurophysics associated with the early development of retinotectal circuit of Xenopus laevis tadpoles visual system by developing methods using the calcium sensitive dye and high speed CCD camera based fast optical 2

18 imaging. As the most important visual processing center in amphibians, the optic tectum is part of the midbrain, which receives a large proportion of the total retinal axon input. Adult amphibians have a complex circuitry of their retinotectal system [38, 39]. Retinal input triggers polysynaptic activity in this system [40], which then integrates with inputs from other sensory modalities and drives motor output (e.g. swimming direction change of the animal, etc.) [41]. The initial development of this system is mainly a contribution from molecular cues, which guide the retinal ganglion cells axons to targets on the corresponding tectal area, and form a coarse topographic map [24, 42]. The subsequent refinement process, on the other hand, requires neural activity in the visual system [43, 44], which eventually results in a topographically more precise retinotectal projection [45] with a more refined and functional local tectal circuitry [45, 46]. A well-known result of this process is an increasingly focused visual receptive field (RF) of optic tectal neurons over development [24, 47]. However little is known about the development within the microcircuitry of the tectum of different temporal relations between different tectal neurons, which might be important for the visual information processing and the subsequent decision making and motor [48-50]. This is partly due to the difficulty of simultaneously recording from multiple tectal neurons with high temporal resolution (< 100 ms). We apply fast calcium imaging with a temporal deconvolution-based spike train reconstruction technique for population recording of optic tectal neurons of animals between different developmental stages, defines as stages 46 and 48/49, during which the retinotectal system undergoes a period of rapid growth. The fast population recording illustrates the difference of visual response neural synchrony between early and late developmental stages. Furthermore, by comparing normal development with results from animals reared in the 3

19 dark (i.e. without any external activation of the visual system) and animals treated with NMDA receptor antagonists, the mechanism of this developmental change in neurophysical terms of circuit synchrony is discussed. A spike-timing dependent plasticity (STDP) based computational model is then used to reproduce all of the experimental data. The work of this thesis in which new insight has been obtained to the neurophysics of the visual system development in tadpoles begins in CHAPTER 2, which gives a basic overview of neural system development. Tadpole retinotectal system is used as a model to demonstrate different types of activity dependent and independent mechanisms of the development. Two types of activity dependent developmental rules (STDP and BCM) are then presented in detail with their underlying biophysical mechanisms being discussed and compared. CHAPTER 3 reviews various optical recording techniques, in particular the principle of calcium sensitive dye imaging. Its comparison with other candidates for optical neural microcircuit reporters such as voltage sensitive dye imaging is also discussed. This phase of the thesis culminates in the calcium imaging study of developing retinotectal systems is presented in CHAPTER 4. In particular, detailed temporal properties of tectal neuron visual responses are compared respectively over the development. Mechanisms of this developmental effect are then discussed based on various experimental manipulations and computational modeling. As another technical advance in the research for this thesis, a novel micro-led array based visual stimulator and its application in tectal neuron visual receptive field mapping is discussed in CHAPTER 5. The second major theme of this thesis applies the newly developed, so-called "optogenetic" technique (especially the photo sensitive cation channel Channelrhodopsin-2) for fast and targeted photon-triggered neural modulation. Here, the biophysical system of in- 4

20 terest is a particular neural circuit in a rodent model, involved in the medial dorsal nucleus (MDN of thalamus) modulation of prefrontal cortical UP state in mouse brain slices. In mammalian brains, cerebral cortex receives most of its sensitive input from thalamus, which modulates the cortical activity in many senses [17]. As an example, the prefrontal cortex (PFC), which is involved in many kinds of high-level cognitive functions, receives its thalamic input from MDN [51-53]. Working memory (WM) is one of those high-level cognitive functions, which enables the animals to instantly keep and modulate the information from various sensory modalities for a short period of time [54, 55]. Instead of the neural rewiring/growth mechanism of long term memory, WM is thought to be based on specific types of collective neural activities. In particular, the in vivo recording of many different animal species in the past few decades clearly showed that the persistent neural activity of PFC is closely correlated to the neural basis of WM [25, 56, 57]. As the thalamic input of PFC, the role of MDN in this process remains unclear, due to the difficulty of selective modulation of MDN input and simultaneous high spatial-temporal resolution recording of PFC. The recent development photo-modulation tools shed light on this field. This technique, also known as optogenetics, is based on the photo-sensitive ion channels or ion pumps, which can be genetically targeted to specific parts of the brain by localized viral transfection. Transfected neurons, including their axons are then expressing these ion channels and/or ion pumps, and therefore become photo-modulatable. In this work we transfect Channelrhodopsin-2 (ChR2), a blue-light gated cation channel, into MDN neurons. Shining blue light to transfected neurons or their axons opens ChR2, and generates inward current, which induces depolarization of the membrane poten- 5

21 tial and may initiate action potentials. This allows us to selectively stimulate MDN axons on brain slices without the need of preserving the whole projection pathway, which is technically very challenging due to anatomical constraints. Combining this with calcium imaging technique, the modulation of MDN input to the cortical UP state (the in vitro analogue of persistent neural activity) of acute mouse PFC brain slices is then studied all optically with selective axon stimulation and population neural recording. The organization of the research into the MD-PFC neural pathway by photomodulation is organized as follows: CHAPTER 6 reviews the optogenetics and other photo modulation techniques. CHAPTER 7 gives an introduction of the structure of MDN-PFC system, and the basic theory of working memory, and discusses the results of the all optical modulation and recording study of MDN-PFC system on acute mouse brain slices in vitro. CHAPTER 8 concludes my work of optical recording and modulation for the study of brain dynamics, and discusses the perspective of optical tools for neuroscience study. The SUPPLEMENTARY CHAPTER presents my early work of micropatterning of cultured neurons. 6

22 CHAPTER 2 RETINOTECTAL SYSTEM DEVELOPMENT 2.1 Introduction Most multicellular animals start from a single zygote. During the development, cell division as well as differentiation and migration occur to allow the formation of different biological structures (morphogenesis) with different functions for the survival and regeneration of an organism. Particularly, the vertebrate central nervous system (CNS) develops from a flat sheet of cells of the ectoderm called the neural plate, at the early embryonic stage. The neural plate cells first grow rostrally and caudally to form the neural groove, whose walls subsequently move together and fuse dorsally, forming the neural tube (Figure 2.1 (A)). The rostral end of the neural tube then expand and form three swellings or primary vesicles, which eventually form the entire brain (Figure 2.1 (B)). The rostral-most vesicle, or the prosencephalon (forebrain), differentiates two secondary vesicles (Figure 2.1 (C)), the optic vesicles (which ultimately become the optic nerves and the retinas) and the telencephalic vesicles (the source of the cerebral cortices) on both sides, connected by the remaining part, the diencephalon, which develops into the thalamus and hypothalamus. Similarly, the mesencephalon, or the midbrain, differentiates into the tectum and the tegmentum; while the rhombencephalon, or the hindbrain, differentiates into the cerebellum, the pons, and the medulla oblongata, each with their own functions (Figure 2.1 (D)) [17]. Besides the macroscopic or the anatomical view, the developmental changes happen in all different spatial scales. For instance, on the network level, neurons grow their 7

23 Figure 2.1 The early development of the brain. (A) The primitive embryonic CNS begins as a thin sheet of ectoderm (left); the formation of the neural groove (middle); the formation of the neural tube (right). (B) Three primary vesicles develop from the rostral end of the neural tube. (C) The secondary brain vesicles of the forebrain. (D) The differentiation of the brain (Figure reproduced from [17]). dendrites and axons as inputs and outputs to communicate with other parts of the brain, forming functional networks for information transmission and processing. On the cellular and subcellular level, neurons gradually differentiate into different types, each having different physiological properties and expresses different types of ion channels/receptors on their membranes [58]. The developmental process is in general controlled by animals genomes, which store their entire hereditary information, and the particular signaling molecules (also known as morphogenes ), whose concentrations encode the spatial information of cells. Spreading from localized sources, these signaling molecules diffuse over the whole emb- 8

24 ryo, forming concentration gradients, that modulate the expression of different target genes at distinct concentration thresholds, and result in the differentiation of cells at different spatial locations and the generation of secondary signaling molecules [58, 59]. The brain as a special organ processes most of the sensory information from the outside world, and modulates animal behavior consequently. Its precise and optimal functioning is vital for animal survival and reproduction. Therefore besides the gene and morphogene regulation, the brain development is also influenced by the environment, with the former one (the activity independent mechanism) coarsely structuring the system according to the internal information, while the latter one (the activity dependent mechanism) fine tunes the system into its best performance using the external information [58]. The latter process is ongoing in many parts of the brain throughout the animals life, which helps them adapt to different conditions, and becomes the neural basis of learning and memory [17]. In this chapter I will review the study of the development of Xenopus laevis tadpoles retinotectal system as an example of the brain development. The sensory systems of the brain form very organized structures, so that the information coming from the peripheral sensors can be mapped in an orderly way into the brain. The corresponding order in the visual system is called the topography, for which neighboring points in the visual field are represented by neighboring cells in the brain regions that process visual information [17, 58, 60]. The structural basis of this order is the highly organized neuronal projections from the retinal ganglion cells (RGCs) to the visual center of the brain and the proper recurrent network within the visual center, both of which form gradually in the brain development. Retinotectal system in fish, amphibians, reptiles, birds, and its analo- 9

25 gue retinocollicular system in mammals, is a widely studied model of visual system development, which forms the topographic map under the regulation of both the activity-independent mechanism and the activity-dependent mechanism [24, 43, 61]. In non-mammalian animals, the main visual center is the optic tectum, which receives direct input from the retina. With its output directly linked to the motion center for visually-guided behaviors, this system is ideal to study the functional development of neural circuit. In the subsequent sections, we discuss both the activity-independent mechanism (section 2.2) and the activity dependent mechanism (section 2.3) using the developing tadpole retinotectal system as an example. The quantitative modeling of the activity dependent mechanism and neural plasticity is presented in section Visual experience independent mechanism The visual experience independent mechanism, also known as the chemoaffinity cue mechanism, mainly implies the neural development regulated by signaling molecule concentrations. It was first proposed and demonstrated by Roger Sperry in his pioneering studies on the regenerating retinotectal projection in goldfish and frogs [42, 62-67], and was later proved to be a common rule in other systems. After the initial anatomical development of the eye and the brain, the synaptic connections within the retinotectal system start to develop at the early stage of the tadpole (developmental stage 39, [44]), when the RGCs differentiate from the retinal progenitor cells in the inner layer of the optic cup (controlled by signaling molecules, [28, 68-70]). As the only retinal cell type that projects and conveys visual information to the brain, 10

26 Figure 2.2 Tadpole retinotectal projection development guided by chemoaffinity cues. (A) Time Course of RGC axonal arborization in the Xenopus visual system revealed by transverse diagram of the Xenopus brain with one eye depicting the contralateral retinotectal projection (Figure reproduced from [5, 6]). (B) The detailed RGC axon guidance process involves multiple signaling molecules: In the retina, axons are repelled from the periphery by chondroitin sulfate. At the optic disc, RGC axons exit the retina into the optic nerve using a mechanism based on attractive netrin/dcc -mediated action. Within the optic nerve, RGC axons are kept within the pathway through semaphorin distribution and by inhibitory Slit/Robo interaction. Slits also contribute to positioning the optic chiasm by creating zones of inhibition. Zic2-expressing RGCs in the VT retina project EphB1-expressing axons, which repel ephrin-b2 at the optic chiasm and terminate ipsilateral targets. (C) The topography of retinotectal projection: RGC axons from temporal retina mainly map to the anterior part in optic tectum, while RGC axons from nasal retina project more to the posterior part. On the other hand, RGC axons from ventral retina mainly map to the medial part in optic tectum, while RGC axons from dorsal retina project more to the lateral part ((D) Dorsal; (V) ventral; (N) nasal; (T) temporal; (A) anterior; (P) posterior; (L) lateral; (M) medial) (Figure reproduced from [28]). RGCs then extend their axons outside the retina, form the optic nerve and chiasm, and target to the optic tectum (Figure 2.2 (A)) [28, 71]. The process involves multiple signaling molecules with concentration gradients on different directions (Figure 2.2 (B)) [28, 72-74]. The next important event is the topographic mapping of RGC axons to their corresponding area of the optic tectum, meaningly the nasal RGC axons to the caudal tectum, the temporal RGC axons to the rostral tectum, the dorsal RGC axons to the ventral tectum, and the ventral RGC axons to the dorsal tectum (Figure 2.2 (C)) [28]. As an example of the chemoaffinity cue dominated process, two types of signaling molecule-receptor systems, epha/ephrin-a and ephb/ephrin-b, are regulating this process [24, 75, 76]. The ephrin is a family of membrane-bound proteins found in lipid rafts of the cell plasma 11

27 Figure 2.3 Chemoaffinity cues in the retinotectal projection. Connectivity observed in Sperry s original studies of regenerating goldfish retinotectal fibers originating in retinal halves. Black regions of the retina project to black sites in the tectum, and likewise, gray projects to gray. (A) Ephrin-A and the EphA receptors act as chemoaffinity molecules for rostrocaudal mapping in the optic tectum. (B) Ephrin-B and EphB signal dorsoventral mapping. The graded distributions of ephrins and Eph receptors are indicated next to their appropriate structures. (C) Retinotopic organization of inputs imaged in the optic tectum of a live Xenopus tadpole. Temporal and nasal retina were labeled, respectively, with dii (red) and FITC-dextran (green). Orientation as in (A). Scale bar = 20 µm (Figure reproduced from [24]). membrane. Their receptors, eph, are a group of receptor tyrosine kinases found in developing and mature tissues [2]. The interactions between these ligands and receptors are involved in the regulation of a variety of life processes, such as development and cancer [77, 78]. At the beginning of this step, the eph receptors form concentration gradients in the RGC axons (epha: temporal RGC > nasal RGC, ephb: ventral RGC > dorsal RGC), with their ligands forming the matched gradients in the tectum (ephrin-a: caudal tectum > rostral tectum, ephrin-b: dorsal tectum > ventral tectum, Figure 2.3 (A), (B)) [24, 28]. These concentration gradients or the spatial patterns of eph receptors and the corres- 12

28 ponding ephrins are due to the regulation of their upstream signaling molecules some transcription factors [2, 79-81] that carry the spatial information by their concentration gradients, and act at earlier stages of the development. The RGC axons grow and arbor according to the eph/ephrin concentration gradients. Activation of epha receptor by ephrin-a leads to axon repulsion [82] and inhibition of axon branching [83]. Therefore the axons from RGCs in the temporal retina with high expression of epha receptors are inhibited from innervating the caudal tectum with peak level ephrin-a. On the other hand, activation of ephb receptor by ephrin-b leads to axon attraction [84, 85], so that dorsal axons with high expression of ephb receptors innervate mainly the ventral tectum with peak level ephrin-b. Combining with the effect of axon-axon competition for available target space [86, 87], this finally results in a topographic retinotectal map formation, with the nasal RGCs mainly projecting to the caudal tectum, the temporal RGCs mainly projecting to the rostral tectum, the ventral RGCs mainly projecting to the dorsal tectum, and the dorsal RGCs mainly projecting to the ventral tectum (Figure 2.3 (C)) [24, 28]. The underlying biophysical mechanisms of this regulation system is mainly based on the phosphorylation of two juxtamembrane tyrosine residues and the conformational change in the cytoplasmic portion of the eph receptors during the engagement with ephrins [2, 88, 89]. This phosphorylation process activates the downstream signaling system and eventually regulates the cell shape and movement through the Rho family (Rho, Rac and Cdc42) GTPases [2], which promote the formation of stress fibers (Rho), lamellipodia (Rac) and filopodia (Cdc42). The phosphorylation of the epha receptors facilitate their interaction with the exchange factor ephexin [90], which activate Rho GTPases by 13

29 Figure 2.4 Forward and reverse signals communicated downstream of the Eph-ephrin complex. Some of the known EphA ephrin-a or EphB ephrin-b signaling pathways are highlighted (Figure reproduced from [2]). catalyzing the replacement of GDP with GTP. In neurons, Rho activation inhibits neurite outgrowth and promotes growth cone collapse and axon retraction [91-93]. The phosphorylated ephb receptors, especially ephb2, associate with different exchange factor intersectin [94] and kalirin [95], and activate Cdc42 GTPases and Rac GTPases respectively, which promote branching and elongation of actin filaments and the axon growth [96, 97]. Additional reverse signaling pathways also exist mainly through the phosphorylation of transmembrane ephrin-b ligands [2]. Figure 2.4 shows the schematic of eph/ephrin interactions. 2.3 Visual experience dependent mechanism of development and neural plasticity Although the chemoaffinity cues guide the RGC axons to project topographically onto 14

30 the optic tectum, this retinotectal mapping is too coarse compared to the mature map. The subsequent refinement of the map requires the RGC spiking activity, which fine-tunes the system into its best performance to facilitate animal survival and reproduction. First proposed by Donald Hebb ([98]), and observed by Timothy Bliss etc. [99], patterned neural activity is found to play an important role in instructing adaptive changes of neural systems in many different ways. Synapses between neurons that are coordinately active are persistently strengthened and physically stabilized, a phenomenon known as long-term potentiation (LTP) [100, 101]. In contrast, synapses between asynchronous (or less synchronous) neurons are progressively weakened with presynaptic and postsynaptic structures being retracted, known as long-term depression (LTD) [102] (Figure 2.5 (A)). The timing of the spiking activity is also found to be important and de- Figure 2.5 Experimental observation of Hebbian plasticity. (A) High frequency firing of presynaptic neurons effectively stimulates the postsynaptic neurons to fire action potentials, therefore induces the increase of EPSP amplitude (LTP). Low frequency firing of presynaptic neurons is not likely to trigger postsynaptic firing, which causes the decrease of EPSP amplitude (LTD) (Figure reproduced from [21]). (B) STDP rule: synapses that always trigger the postsynaptic spikes (and therefore the presynaptic spikes are always before postsynaptic spikes) get strengthened (LTP), while presynaptic spikes that happen after the postsynaptic spikes weaken the synapses (LTD) (Figure reproduced from [29]). 15

31 termines the direction of the change of the synaptic strength and neuronal morphology in many cases [29, 103, 104]. The repetitive pre/post spike pairs, which preserve causality with presynaptic spikes occur within a few tens of milliseconds before the corresponding postsynaptic spikes, induce LTP; while repetitive anti-causal spike pairs with presynaptic spikes occur within a few tens of milliseconds after the postsynaptic spikes cause LTD (Figure 2.5 (B)). The asymmetrical synaptic weight change depending on the relative spike timing of the pre/post synaptic neurons is therefore called the spike-timing dependent plasticity (STDP). In tadpole retinotectal system, the visual experience induced patterned RGC activity is known to cause various types of Hebbian plasticity, which induce competitions between the growth of different neurites and sculpt the retinotectal projections to be more point-to-point like, resulting in sharper tectal neuron RFs [45, 61]. The subsequent tectal neural activity also reshapes the tectal recurrent network to facilitate intra-tectal visual information processing [46]. Although still under investigation, the biophysical mechanisms of Hebbian plasticity Figure 2.6 AMPA and NMDA receptors. (A) Scheme of AMPA receptor and the single channel recording showing channel openning and closing (Figure reproduced from [12]). (B) Scheme of NMDA receptor showing glutamate and magnesium binding site (figure reproduced from 16

32 are thought to involve different synaptic receptors and their downstream signaling systems, particularly the two main subclasses of the ionotropic glutamate receptors, namely the α-amino-3-hydroxyl-5-methyl-4-isoxazole-propionate (AMPA) receptors, and the N-methyl D-aspartic acid (NMDA) receptors and the dendritic calcium ion (Ca 2+ ) concentration. AMPA receptors are fast ligand gated cation channels, which open and generate inward current when bonding to glutamate or AMPA (Figure 2.6 (A)) [17, 19]. NMDA rceptors, on the other hand, are slower ligand gated cation channels, whose ion gating are blocked by magnesium ions (Mg 2+ ) at the resting membrane potential (V m ~ -65 mv), and require membrane depolarization (V m > -40 mv) in addition to glutamate or NMDA bonding to remove the Mg 2+ block and open the channel (Figure 2.6 (B)) [17, 19]. In adult animals, AMPA receptors are the main source of excitatory postsynaptic current (EPSC), which initiates action potentials (APs). However at the early stage of brain development, most glutamate receptors on the synaptic site of the tectal neural membrane are NMDA receptors, while AMPA receptors are mainly located intracellularly close to the synapses and not functional [ ]. Incoming APs cause the release of neurotransmitter (glutamate) from the presynaptic neurons, which by diffusion bind to the postsynaptic NMDA receptors. At resting potential, this will not open the ion channel due to the Mg 2+ block. However, if the postsynaptic neurons also fires at that period, their membrane potential will get depolarized. The transmission of this depolarization to the dendritic end, either by back-propagating action potentials (BPAPs) or simple conduction, will help release the Mg 2+ ions from the NMDA receptors and open the channel. The high calcium ion (Ca 2+ ) permeability of NMDA receptors will cause an increase of local Ca 2+ concentration around the synaptic 17

33 Figure 2.7 Biophysical mechanisms of Hebbian plasticity. (A) High concentration of calcium ions activate CaMKII, and lead to the delivery of intracellular AMPA receptors and AMPA receptors on outside the synaptic areas to the synaptic site, inducing LTP (Figure reproduced from [15]). (B) Moderate concentration of calcium ions activate calcineurin and PP1, which result in to endocytosis of synaptic AMPA receptors and generate LTD (Figure reproduced from [26]). site, which then triggers the downstream calcium dependent signaling systems. Correlated activity of pre- and postsynaptic neurons or causally related pre- and postsynaptic neuron AP pairs generate high enough calcium concentration around synaptic sites. This then activates Ca 2+ /calmodulin dependent protein kinase II (CaMKII) and the phosphorylation of other proteins, eventually resulting in the delivery of AMPA receptors to the synaptic sites and the increase in the synaptic strength (Figure 2.7 (A)) [ ]. Uncorrelated activity of pre- and postsynaptic neurons or anti-causally related pre- and postsynaptic neuron AP pairs generate moderate calcium concentration around synaptic sites, which activates calcineurin and protein phosphatase 1 (PP1) and the subsequent phosphorylation/dephosphorylation of other proteins; eventually inducing the endocytosis of synaptic AMPA receptors and the decrease in the synaptic strength (Figure 2.7 (B)) [ ]. The activity of NMDA receptors is also found to modulate Rho family of GTPases, which regulate the neurite growth and the morphology of the retinotectal system [108, 109]. Neural activity does not only impact activated synapses, but also causes cell-wide 18

34 changes. Neurons can adjust or rescale the strength of their total synaptic inputs while maintaining the relative differences between individual synapses [110]. Besides, the intrinsic excitability of neurons is also modulated by their spiking activity [111]. Both of these changes, known as homeostatic plasticity, are caused by the activity induced intracellular Ca 2+ concentration changes in soma, through the activation of voltage-gated calcium channels (VGCCs) or NMDA receptors, and the subsequent calcium dependent signaling pathways, which eventually affect global cell properties, sometimes by modulating gene expression [107, 112]. Combined with the competitive Hebbian plasticity, homeostatic plasticity (sometimes also called metaplasticity) stabilizes the cell activity and network growth during brain development. 2.4 Modeling of neural plasticity The mathematical formulation of the neural plasticity is an important step for physical and quantitative understanding of the neural system development as well as functioning. In the past few decades, different types of neural plasticity in many different neural systems have been modeled and fitted to the experimental results. In this section, I will briefly review some of the most common models, based on Wulfram Gerstner and Werner Kistler s book [19] Modeling of STDP According to the Hebbian plasticity, the change of a synaptic strength w ij depends on the activities of the presynaptic neuron j and the postsynaptic neurons i. With u i and u j 19

35 representing the membrane potential of the postsynaptic neuron i and the presynaptic neuron j, a general and phenomenological formulation of the Hebbian plasticity can be written as d w post pre ij ( t ) = F [ w ij ( t );{ u i ( t ' < t )},{ u j ( t " < t )}] (2.1) dt where F is a functional of the time course of pre- and postsynaptic membrane potentials, and t < t, t < t guarantee the causality [19]. By expanding the right-hand side of Eq. (2.1) about the resting state u post i = u pre j = u rest in a Volterra series [113, 114] with u rest reset to be 0, we find d w pre pre ij = a 0( w ij ) + α 0 1 ( w ij, s ) u j ( t s ) ds dt post + α ( w, s ') u ( t s ') ds ' post 1 ij i + α ( w, s, s ') u ( t s) u ( t s ') ds ' ds + corr pre post 2 ij j i (2.2) with the higher order terms of u post i and u pre j being neglected [19]. Eq. (2.2) may be used to provide the general framework for the formulation of the STDP. We now consider the time course of the pre- and postsynaptic membrane potentials in detail. The axon terminals of the presynaptic cell stay at the resting potential most of the time, except when an AP arrives. Since the duration of an AP is short (~ 1 ms), it can be approximated as δ functions, so that u t t t (2.3) pre ( f ) j () = δ ( j ) f where t represents the spike times of the presynaptic terminal [19]. ( f ) j The situation of the postsynaptic dendritic site is more complicated in the sense that both the effect of the direct synaptic responses and the back propagating signals from the 20

36 soma should be considered. Therefore post u () t = η( t tˆ ) + h() t (2.4) i i i with t ˆi representing the postsynaptic spike times, and η, h i standing for the BPAPs and the direct synaptic responses respectively [19]. The eq. (2.4) can be further simplified by neglecting the direct synaptic responses and assuming that the BPAPs are short pulses, which can be modeled as δ functions. Therefore the membrane potential of the postsynaptic neuron i can be written as u t t t (2.5) post ( f ) i () = δ ( i ) f where t denotes the firing times of the postsynaptic neuron, with the transmission ( f ) i delay of BPAPs from the postsynaptic soma to the synaptic site on dendrites being neglected [19]. By applying eq. (2.3) and (2.5) to eq. (2.2) and omitting the w ij dependence, we get dw dt (2.6) ij pre ( f ) post ( f ) corr ( f ) ( f ) α0 α1 t tj α1 t ti α2 t ti t ti f f f f = + ( ) + ( ) + (, ) + Typically a macroscopically significant modification of a neural system requires long term neural activity, much longer than single spike duration so that the exact time pre post corr course of α 1, α 1 and α 2 are not important. We thereby assume that the synaptic weight changes are instantaneous, so that α ( t t ) = a δ( t t ) (2.7) pre ( f ) pre ( f ) 1 j 1 j α ( t t ) = a δ( t t ) (2.8) post ( f ) post ( f ) 1 i 1 i α a ( t t ) δ ( t t ), t < t pre, post ( f ) ( f ) ( f ) ( f ) ( f ) corr ( f ) ( f ) 2 j i j i j 2 ( t ti, t tj ) = post, pre ( f ) ( f ) ( f ) ( f ) ( f ) a2 ( ti tj ) δ ( t ti ), ti > tj (2.9) 21

37 pre pre pre pre where a1 = α1 () s ds and a1 = α1 () s ds. The correlation term is divided into 0 0 two parts pre, post corr a ( τ) = α ( s τ, s) ds and post, pre corr a () τ = α (, s s τ) ds to fit the asymmetrical timing response of the STDP. Thus, the sharply peaked BPAPs and the instantaneous synaptic weight changes lead to the general forms of STDP [19] d w t a S t a a s S t s ds dt pre pre, post ij () = 0 + j ()[ () i ( ) ] post post, pre i( )[ ( ) j( ) ] + S t a + a s S t s ds (2.10) where S t = δ t t and ( f ) j() ( j ) f S t t t are the pre- and postsynaptic ( f ) i() = δ ( i ) f pre neural spike trains. a 0 represents an activity independent term, while a 1 and a 1 post only depend on the pre- or postsynaptic activity. Since a and pre, post 2 a are the post, pre 2 terms responsible for Hebbian plasticity, modeling of STDP sometimes is simplified as d w pre, post post, pre ij () t = S j () t a 0 2 () s S i ( t s ) ds + S i () t a 0 2 () s S j ( t s ) ds dt with the parameters being chosen as (2.11) post, pre a2 ( s) = A+ exp( s/ τ1), s< 0 W() s = pre, post a2 ( s) = A exp( s/ τ 2), s > 0 (2.12) Here W(s) is called the window function, with A + > 0, A - < 0 respectively (Figure 2.8 (A)) [19] Rate based Modeling and BCM theory In the high firing frequency region, the average membrane potential u over a short period of time can be written as a function of the spike rate ν [19], and therefore eq. (2.1) can be rewritten as 22

38 Figure 2.8 STDP and BCM laws. (A) learning window W of STDP as a function of the time difference t = t j (f) - t i (f) between presynaptic spike arrival and postsynaptic firing with A + = -A - = 1, τ 1 = τ 2 = 20 ms; (B) Bidirectional learning rule. Synaptic plasticity is characterized by two thresholds for the postsynaptic activity. Below ν 0 no synaptic modification occurs, between ν 0 and ν θ synapses are depressed, and for postsynaptic firing rates beyond ν θ synaptic potentiation can be observed [19]. d w ( t post pre ij ) = F [ w ( t ij );{ νi ( t ' < t )},{ ν j ( t " < t )}] (2.13) dt In an over-simplified case, the history dependence of the functional is neglected, and therefore it becomes a function of the instantaneous firing rates d w post pre ij () t = F ( w ij (); t νi, ν j ) (2.14) dt which can be further expanded about ν i = ν j = 0, d w post pre corr ij ( t ) = c 0( w ij ) + c 1 ( w ij ) νi + c 1 ( w ij ) νj + c 2 ( w ij ) νν i j + (2.15) dt corr where c 2 is the Hebbian term [19]. Since ν i and ν j are always positive, the Hebbian term itself lacks the ability to induce synaptic weakening, and other terms [115] or additional global mechanisms (such as synaptic scaling [116]) must be involved in the formula to ensure competition and stability, which in many cases makes the whole model nonlinear. Here we only consider a famous condition, where the postsynaptic firing rate is 23

39 thought to determine the sign of the plasticity d w dt = ηφ( ν ν ) ν γ w (2.16) ij i θ j ij φ is chosen to be a nonlinear function, that is negative for ν i < ν θ, positive for the opposite case, and bounded between certain values (Figure 2.8 (B)). In many cases φ is set to be zero for ν i < ν 0, to ensure the synapses are not affected by the ultra-low frequency activity. The gating threshold ν θ is chosen to be a running average of the postsynaptic firing rate ν i within the recent history, in order to stabilize the model and restrict the synaptic weight from increasing out of bounds. Eq. (2.16) is known as the Bienenstock-Cooper-Munroe (BCM) law [117], which combines the Hebbian plasticity with the homeostatic effect (the sliding gating threshold ν θ ) to gain the metaplasticity and the global level self-modulation of the synaptic modification rate. It has been observed in many different neural systems development, such as rat visual cortices [118], and Xenopus laevis tadpole retinotectal systems [119] Biophysical models The detailed biophysical mechanisms of the plasticity can also be described quantitatively. In a simple case of calcium control hypothesis, the changes of synaptic weights is thought to depend only on the postsynaptic Ca 2+ concentration [11, 19] d w dt ij Ω Ca wij = (2.17) ([ ] ) 2+ ([ ] j ) 2+ τ Ca j with Ω the asymptotic calcium dependent modification function, and 2 ([ Ca + ] j ) τ 2+ ([ Ca ] j ) the calcium dependent time constant. The setting of Ω needs to depict the 24

40 experimental results of the constant postsynaptic membrane potential case. For a Ca 2+ concentration below θ 0, Ω is assumed to be a constant value w 0 = 0.5, which leads to a resting synaptic weight. For an intermediate calcium concentrations in the range θ0 [ ] m 2+ < Ca < θ, the weight tends to decrease, while for the high concentration range, where 2+ [ Ca ] θm >, the weight tends to increase. The time constant τ 2+ ([ Ca ] j ) is set to be τ τ ([ Ca ]) = [ Ca ] (2.18) where τ 0 = 500 ms and [Ca 2+ ] is in µmol/l. At a low level of [Ca 2+ ], the time constant is in the range of hours, while for high level [Ca 2+ ], the synaptic weight changes within several hundred millisecond. In particular, this makes the effective time constant for the induction of LTP shorter than that for LTD [19]. The postsynaptic intracellular [Ca 2+ ] can be described by a first order differential equation [11, 19] 2+ d Ca () t Ca 2+ () t I Ca () t dt = τ Ca (2.19) with τ Ca = 125 ms. We assume that NMDA receptor opening generates the main Ca 2+ current, which can be modeled as [11, 19] I t = g t t ut E But (2.20) ( f ) Ca () Caα ( j )[ () Ca ] [ ()] where g Ca is the maximal channel conductance and E Ca is the calcium reversal potential. The dynamics of the glutamate bonding is described by ( f ) ( f ) ( ) ( ) ( f t tj )/ τ f ( f ) ( t tj )/ τs j j j α( t t ) = [ aθ( t t ) e + bθ( t t ) e ] (2.21) 25

41 Figure 2.9 Biophysical model of Hebbian plasticity. (A) STDP. For post-pre conditions (region I: -30 ms < t < -5 ms), LTD is induced. For pre-post conditions (region II: 0 < t < 45 ms), LTP is induced. For larger pre-post intervals (region III: 45 ms < t < 100 ms), LTD is induced. (B) Pairing of presynaptic spikes with postsynaptic depolarization. The weights w ij that are obtained after several hundreds of presynaptic spikes as a function of the depolarization of the postsynaptic membrane (Figure reproduced from [11]). with a + b = 1, and the right-hand side terms representing a fast and a slow dynamics respectively [11, 19]. 1 Bu ( ) = (2.22) 0.062u e describes the magnesium blocking/unblocking dynamics, with u representing the postsynaptic membrane potential in millivolt. In a simple case, the BPAPs are the main part of u, which is described as t/ τ fast t/ τ slow ut ( ) = u (0.75e e ) (2.23) BPAP where u BPAP = 100 mv is the amplitude of the BPAP and τ fast and τ slow representing the time constant of a fast and a slow decay respectively [11, 19]. Apply Eq. (2.18) (2.23) to Eq. (2.17) with the appropriate Ω wave form, we get a window function of STDP similar to the experimental one (Figure 2.9 (A)). The postsynaptic spikes that happen after the presynaptic spikes effectively remove the magne- 26

42 sium block on the NMDA receptors, and generate a big calcium influx, which then dramatically increase the intracellular [Ca 2+ ], and induce LTP. On the other hand, due to the short time course of BPAPs, postsynaptic spikes that happen before the presynaptic spikes cannot effectively remove the magnesium block for the later coming glutamate, therefore only cause a moderate increase of the postsynaptic [Ca 2+ ], and induce LTD. The main disagreement with the experiment appears at the long-time positive branch, where the [Ca 2+ ] is also moderate. This disagreement can be revised by considering a more detailed Ω that does not only depends on [Ca 2+ ] [104]. In the high firing frequency range, the averaged membrane potential (Figure 2.9 (B)) within a short period of time is a sigmoidal function of the cell firing rate. Hence, the figure of Ω can be transformed easily to the BCM learning curve. The spike-timing rule and the spike-rate rule are therefore unified. 27

43 CHAPTER 3 OPTICAL RECORDING OF NEURONAL NETWORK AC- TIVITY 3.1 Introduction Most of the brain s functions are processed in a collective way, such that input information which was coded as temporal patterns of action potentials flows through and distributes into the network of interconnected neurons during the processing with each neuron interacting with others via synaptic connections between them, and generates outputs as firing patterns of neurons in the networks. A fundamental goal of neuroscience is to understand the spatiotemporal features of the collective neuronal network activity, which requires monitoring and analysis of activity of populations of neurons. The conventional patch clamp based electrophysiological recording technique can precisely collect the neuronal transmembrane electrical signals down to sub-millivolt, sub-nanoampere, and sub-millisecond range [30], and has become one of the most popular neuronal recording techniques. The ultra high signal-to-noise ratio makes it easy to detect the small fluctuations of membrane potentials, such as the subthreshold postsynaptic potentials/currents, and is perfectly suited for the study of single cell level neuroscience. However, the limited number of recording channels, namely the number of cells for simultaneous recording, excludes its application in the study of large neuronal networks. The complementary approaches including the multi-electrode array (MEA) recording technique, the optical recording technique, and the global noninvasive recording 28

44 techniques like fmri, were developed and employed to monitor population activity with different emphases [33]. In the multi-electrode array (MEA) scheme, arrays of conducting extracellular electrodes are used to record and/or stimulate network activities [32, 120]. The temporal resolution can be as high as the patch clamp technique and therefore single-spike resolvable. The spatial resolution depends on the number of electrodes, which typically ranges from 10 to 100. It is widely used in studies of large neural systems with spatial dimensions greater than 1 millimeter. The global noninvasive monitoring techniques, such as fmri, detect the side effects of population neuronal activity (e.g. the blood oxygen level for fmri [34]) with limited spatial and temporal resolutions (about 1 millimeter and 1 second for fmri [33]), and therefore only good for noninvasive whole-brain level studies. Optical recording techniques, on the contrary, usually rely on indicators (except the intrinsic optical imaging technique) that are sensitive to various neural parameters (such as the membrane potential V m and the concentration of ions), and change their optical properties (e.g. emission spectra) accordingly [9]. Combining with appropriate hardware, they can provide very high spatial resolution (< 0.5 µm for 2-photon microscopy) and temporal resolution (< 1 ms for fast CCD camera imaging) with relatively little invasiveness, which makes them powerful tools for medial scale network ( 1000 neurons) recordings [9, 37]. Moreover, some non-electrical signals (such as the concentration of particular ion types) can only be detected using optical recording techniques. In this chapter, we focus on the properties of the calcium sensitive probes (section 3.2) and the corresponding calcium dynamics during the imaging (section 3.3). Other optical recording methods (such as the voltage sensitive dye imaging) are discussed in 29

45 section 3.4. At the end, the recording hardware is presented (section 3.5). The aim of this chapter is to review the basic principle of the calcium imaging technique as a background of the subsequent studies of neuronal network dynamics. 3.2 Calcium indicators As shown in the last chapter, the calcium ion serves as an important messenger in many of the cellular level life processes. In addition, its concentration changes dramatically when a neuron fires, due to the extraordinarily high concentration difference between the extracellular and cytosolic media (Table 3.1), and therefore can be used as a coincident measure of neuronal spiking events. These features make it possible and worthwhile to monitor the dynamics of the intracellular calcium ion concentration ([Ca 2+ ]) during the neural activity. The calcium ion is usually detected by calcium indicators, which are designed to bind calcium with high affinity, and change their optical properties accordingly. The first calcium indicator Aequorin was discovered as a photoprotein from luminescent jellyfish Aequorea victoria and a variety of other marine organisms by Osamu Shimomura in Table 3.1 Extracelluar and cytosolic concentration of major ion types [17] Ion Extracellular concentration (mm) Intracellular concentration (mm) Concentration ratio (Outide/Inside) K :20 Na :1 Ca ,000:1 Cl :1 30

46 1962 [121]. The history of the exogenous calcium sensitive dye, the most popular category of calcium indicator today, however started two decades later, from the synthesis of calcium chelator based fluorescent molecules, Quin-2, Fura-2, and Indo-1 by Roger Y. Tsien s group [122, 123]. Recent development of advanced molecular biology techniques turns the trend back to protein based indicators by genetically encoding them into cells for expression without exogenous loading [22]. In this section, I will give a brief review of the structures, mechanisms and the applications of each category Exogenous calcium indicators Exogenous calcium indicators, which are also called calcium sensitive dyes, are based on 1,2-bis (o-aminophenoxy)ethane-n,n,n',n'-tetraacetic acid (BAPTA, Figure 3.1) [9], a ph-insensitive metal ion chelator, and homolog of EGTA. It has four carboxylic acid functional groups which can bind to two metal ions with high selectivity for Ca 2+ (k d 100 nm at ph 7.0) in the presence of competing millimolar concentrations of Mg 2+. The ion bind/unbinding process is fast as a result of aromatizing the aliphatic nitrogens [9]. The binding site of BAPTA-based indicators can be further modified to change its affinity Figure 3.1 Structures of Ca 2+ chelators (EGTA and BAPTA) and some calcium indicators (indo-1 and oregon green 488) [9]. 31

47 (sensitivity to Ca 2+ ) over a wide ranges (from nanomolar to millimolar) by either addition of electron-donating/electron-withdrawing groups to the aromatic rings (such as NO 2 ), addition of modifying groups that sterically alter Ca 2+ binding (such as CH 2 CH 2 -), or reduction in number of coordinating ligands for Ca 2+ chelation [9]. The optical property of an indicator comes from the fluorophore part that is linked to BAPTA. Ca 2+ binding to the indicator is coupled to changes in fluorescence by two types of mechanisms. The first generation of the calcium sensitive dyes (quin-2, fura-2, and indo-1) integrates the Ca 2+ binding site with small UV-excitable fluorophores (such as benzofurans and indoles). Direct conjugation of the chelating aminodiacetate moiety, -N(CH 2 CO - 2 ) 2 with the planar fluorophores results in shifted excitation spectra and/or emission spectra upon Ca 2+ binding. More specifically, in the absence of Ca 2+ the aminodiacetate group lies planar to the aromatic ring system, thereby maximizing electron donation by the nitrogen. Ca 2+ chelating twists this group out of planarity and decreases in the electron density of the fluorophore, therefore blue-shifts the excitation wavelength, which also affects the emission spectrum (Figure 3.2 (A)) [9]. The second generation of the calcium sensitive dyes (fluo-3, rhod-2, calcium green family, etc.) incorporates the more fluorescent fluorescein and rhodamine fluors, which operate at visible wavelengths. In these molecules, the BAPTA moiety is twisted out of plane with the fluorophore to avoid direct electronic coupling, so no excitation or emission spectra shifts are observed. On the contrary, the fluorescence intensity changes as a result of the quenching effect of the electron-rich nitrogen of the aminodiacetate group. In the absence of Ca 2+, an electron can be transferred from the nitrogen to the excited 32

48 Figure 3.2 Optical properties of different calcium sensitive dyes. (A) The emission spectrum of Indo-1 AM (excitation: 338 nm) blue-shifts in response to high calcium concentration. (B) The fluorescence intensity of Oregon Green 488 BAPTA-1 AM (excitation: 488 nm) increases with the concentration of calcium ion ( fluorophore before photo emission, decreasing the fluorescence intensity. The binding of Ca 2+ to the aminodiacetate group decreases the nitrogen s electron density, and therefore inhibits the quenching effect, resulting in higher fluorescence intensity (Figure 3.2 (B)) [9]. Compared to the first generation, the long wavelength calcium indicators operate at visible wavelengths and therefore cause less photo damage and cellular autofluorescence Genetically encoded calcium indicators The natural calcium indicator Aequorin is composed of two distinct units, the apoprotein apoaequorin, and the prosthetic group coelenterazine [121]. The apoaequorin contains three EF-hand motifs, which can bind to Ca 2+ and cause the conformational change of the protein with the coelenterazine being oxidized into the excited coelenteramide and CO 2. The excited coelenteramide then relaxes to the ground state and emits blue light (λ = 469 nm) [124]. 33

49 Figure 3.3 The schematic structures and the calcium sensitive mechanism of cameleons. (A) Domain structure of cameleons showing sequences of the boundaries between the donor GFP and Xenopus calmodulin (XCaM) and between M13 and the acceptor GFP. (B) Scheme showing how FRET between GFPs can measure Ca 2+ [22]. Most artificially designed genetically encoded calcium indicators (such as cameleons) are fluorescent proteins derived from green fluorescent protein (GFP) or its variants (e.g. circularly permuted GFP, YFP, CFP), fused with calmodulin (CaM) and the M13 domain of the myosin light chain kinase, which is able to bind CaM (Figure 3.3 (A)) [22]. Calmodulin is a calcium-binding protein expressed in all eukaryotic cells, which contains four EF-hand "motifs" that can bind Ca 2+ ions and change the conformation accordingly. The two fluorescence protein parts on the indicator molecule are usually chosen in such a way that the emission peak of one matches the excitation peak of the other (CFP and YFP for cameleons). Once excited, the short wavelength fluorescence protein can transfer its energy to the long wavelength fluorophore, increasing the long wavelength emission. Such process is known as the Förster resonance energy transfer (FRET), which is extremely sensitive to changes in the distance ( nm) and the orientation of the fluorophores [125]. In low Ca 2+, the more extended conformation of CaM disfavors FRET and gives a low ratio of long wavelength to short wavelength emission, whereas high Ca 2+ results in more retracted conformation of CaM and more FRET with high ratio of long wavelength to short wavelength emission (Figure 3.3 (B)) [9, 22]. 34

50 3.2.3 Cell loading of calcium indicators The BAPTA based calcium indicators are hydrophilic and therefore cannot penetrate the bilipid layer of cell membrane. The measurement of the intracellular calcium concentration thus requires special loading techniques. The hydrophilic dye molecules can be loaded directly via a perforated micropipette for single cell studies. However loadings of neuron populations usually involve the modification of carboxylic acids residue of the dye molecule with acetoxymethyl (AM) ester groups, which results in an uncharged hydrophobic molecule that can permeate cell membranes. Once inside the cell, the AM ester groups are cleaved by nonspecific esterases, preventing the hydrophilic dye molecules from leaking out of the cell (Figure 3.4) [9]. Genetically encoded calcium indicators do not need to be loaded onto cells. Instead the genes encoding them are transfected into the tissue for the expression of the indicator Figure 3.4 Schematic diagram of the processes involved in loading cells using membrane permeant acetoxymethyl (AM) ester derivatives of hydrophilic fluorescent indicators, exemplified by indo-1 ( 35

51 proteins [1]. It is also possible to create transgenic animals expressing the indicator in all cells or selectively in certain cellular subtypes Calculation of calcium concentration The calcium concentration of the imaged sample can be calculated from the recorded fluorescence signals. In general, the fluorescence intensity F arising from a volume V is proportional to the number of indicator molecules in the volume, the illumination intensity I 0, the indicator absorption coefficient α, the quantum yield of the indicator Q F, the photon-collection efficiency Φ of the optical set-up, and the quantum efficiency of the detector Q D [9]: F=Φ QQα In= S n (3.1) D F 0 with n the concentration of indicator molecules in V, and all factors that depend on indicator properties or the experimental setup being lumped together as a constant factor S. In case of a calcium indicator, the indicator molecules switch between the calcium binding state and the free state, with different fluorescence properties. We note the indicator concentration of the calcium binding state as n b with its fluorescence constant S b, and that of the free state as n f with the fluorescence constant S f. The calcium concentration and fluorescence intensity of the sample can thus be expressed as [9] 2+ d b [ Ca ] i n f Kn = (3.2) F = S n + S n = F + ( S S ) n = F ( S S ) n (3.3) f f b b min b f b max b f f where K d is the dissociation constant of the indicator. F min and F max are the fluorescence at zero and saturating Ca 2+ concentration respectively. 36

52 For the second generation calcium sensitive dye, the calcium binding rescales the fluorescence intensity without showing appreciable spectral shifts. Therefore the calcium concentration is calculated from the fluorescence intensity collecting at a single wavelength with single excitation wavelength (Figure 3.5 (A)). By combining eq. (3.2) and (3.3), we get [9] Kn n ( S S ) + F F [ Ca ] K K 2 d b b b f min i = = d = d nf nf ( Sb Sf ) Fmax F (3.4) For a real experiment, the value of F min and F max might differ from place to place (cell to cell) due to the nonuniform dye uptake and illumination light intensity, which makes the direct application of eq. (3.4) unrealistic. The equation can transform into an applicable form by expressing the signal as relative fluorescence change, so that F F F = 0 (3.5) F F 0 Figure 3.5 Recording schemes of calcium indicator fluorescence. (A) Single wavelength measurement: the dye is excited at a single wavelength and the emission intensity F is collected in a spectral window around the peak of the emission spectrum. (B) Ratiometric measurement: spectral shifts allow measurements of the fluorescence intensities F 1 and F 2 at two different wavelengths (the dashed line indicates the spectrum for high calcium concentration) (Figure reproduced from [9, 10]). 37

53 where F 0 denotes the background-subtracted prestimulus fluorescence level [9]. Eq. (3.4) thus becomes F F0 Fmin F0 ( F / F) Fmin F0 F F ( F / F) F F [ Ca ] = K = K 1 F F ( F / F ) max max 0 i d d Fmax F0 F F0 ( F / F) 0 0 F F ( F / F) ( F / F) K + K Ca + K = = ( F / F) ( F / F) 1 1 ( F / F) ( F / F) min 0 2+ d d [ ] rest d Fmax F0 ( F / F) max ( F / F) max max max max (3.6) with [Ca 2+ ] rest the resting-calcium concentration, and ( F/F) max the maximal change upon dye saturation [9]. By assuming that ( F/F) is much smaller than ( F/F) max, eq. (3.6) can be linearized to provide an estimate of the change in [Ca 2+ ] = 2+ [ Ca ] i Kd ( F / F) ( F / F) max (3.7) The typical value of ( F/F) in response to a single action potential is 0.5% ~ 2%. For the first generation calcium sensitive dye, the calcium binding induces the excitation and/or emission spectral shifts. In addition to the single wavelength recording, the calcium signal can be recorded in a ratiometric way with two emission wavelengths or a single emission wavelength of two excitation wavelengths (Figure 3.5 (B)) [9], such that F = S n + S n 1 2 f1 f b1 b F = S n + S n f2 f b2 b (3.8) By re-expressing everything in terms of the fluorescence ratio R = F 1 /F 2, eq. (3.4) can be transformed to Kn S F S F + R R [ Ca ] K K 2 d b f1 2 f2 1 min i = = d = eff nf Sb F 2 1 Sb F 1 2 Rmax R (3.9) 38

54 with the ratios at zero-ca 2+ concentration f1 f2 Rmin = ( R / R ), at saturating- Ca 2+ concentration Rmax = ( Rb / Rb ), and an effective binding constant 1 2 K = K ( S / S ) [9]. eff d f b 2 2 Compared to the single wavelength recording, the ratiometric rocording is more practical for exact calcium concentration measurement, since R is independent of the indicator concentration and is not affected by the bleaching effect. To conclude, calcium imaging technique monitors the cell activity by measuring the emission levels of calcium sensitive fluorescent (or luminescent) molecules. Since the intracellular calcium level change is highly related to the neuronal spiking activity, calcium imaging can be used for spike detection. Its typical signal level is larger ( F/F ~ 10% in many cases) than most of other optical imaging techniques, and therefore is widely used in population recordings of neuronal network activity. 3.3 Modeling of calcium dynamics The simple relationships between the Ca 2+ concentration and the fluorescence intensity of the calcium indicator only exist in the equilibrium conditions, where the Ca 2+ flow is neglected. However Ca 2+ flow is an important biophysical event (the basis of many types of information transmission) and a coincident process of the neural membrane action potential, and therefore the dynamics of intracellular Ca 2+ need to be considered from the fluorescence signals. Various biophysical models have been developed to extract [Ca 2+ ] information from the dynamical fluorescent data The single-compartment model 39

55 We start with a simple case, where the intracellular Ca 2+ gradients and diffusion are neglected on the time scale of interest, and the whole cytosolic volume is treated as a single homogeneous compartment. This is only valid for small structures such as the presynaptic nerve terminals and the postsynaptic dendrites (r 1µm), for which the characteristic diffusion time t = r 2 /6D is small (for D Ca 200 µm 2 /s, t d 1 ms) comparing to the d Ca recording time bin (usually 10 ms). In this case, Ca 2+ influx, buffering, and clearance will be considered as three main processes. Based on Fritjof Helmchen and David Tank s work [9], we present the detailed model as follows. Ca 2+ influx can result from different events, such as the voltage dependent calcium channel (VDCC) opening in response to an action potential, and the NMDA receptor opening in response to the presynaptic release of glutamate and postsynaptic depolarization [17]. The channel opening is usually brief (< 1 ms for VDCCs, and < 100 ms for NMDA receptors) compared to the time required for calcium clearance (typically on the order of 1 second). Therefore the Ca 2+ influx of a single opening event can be described by a delta function and the increase in total calcium concentration ( [Ca 2+ ] T ) [9] 2+ QCa jin = [ Ca ] T δ( t tp ) = δ( t tp ) (3.10) 2FV where F is Faraday s constant, V the volume of the cellular compartment, and Q Ca the calcium charge injected. Most of the calcium ions entering the cytosol bind to endogenous calcium-binding proteins and exogenous calcium indicators (or the genetically encoded calcium indicators). In this model, the binding process is assumed to be instantaneous, so that the initial changes in free-calcium ion concentration depend on the buffering capacity of these Ca 2+ buffers. Usually, a pool S of endogenous buffers and a pool B of indicators are considered 40

56 Figure 3.6 Schematic drawing of the single-compartment model. Ca 2+ enters the cell via ion channels (e.g. VDCCs, NMDA receptors, etc.) and is cleared/extruded from the cytosol via the plasma membrane, uptake into intracellular organelles, and binding to slow endogenous buffers. Two pools of endogenous rapid Ca 2+ buffers (S) and exogenously introduced buffers (B) are considered [9]. (Figure 3.6), with their Ca 2+ binding ratios (κ S and κ B ) being defined as the ratio of the changes in buffer-bound Ca 2+ over the free Ca 2+ changes κ κ ( S ) [ SCa] [ S] TKd [ S] T 2 + ( S ) S = = Ca 2 2 ( S) 2 ( S) i K + + d [ Ca ] i ([ Ca ] i + Kd ) Kd ([ ] ) ( B) [ BCa] [ B] TKd [ B] T 2 + ( B) B = = Ca 2 2 ( B) 2 ( B) i K + + d [ Ca ] i ([ Ca ] i + Kd ) Kd ([ ] ) (3.11) where we assume small free calcium concentration, so that the binding ratios become [Ca 2+ ] i independent, and the changes in free calcium concentration are now proportional to changes in total calcium concentration. The extra calcium ions are either sequestered into intracellular organelles or extruded via the plasma membrane (Figure 3.6) until decaying back to a resting level (around nm). A linear decay rule is usually used to describe this process [9, 126] j = γ Ca Ca (3.12) out ([ ] i [ ] rest ) with γ the clearance rate. Combine eq. (3.10), (3.11), and (3.12), we get the equation for the calcium dynamics following a single brief calcium influx [9] 41

57 d 2 [ Ca + ] T = j in + j out dt d Q Ca + + = dt 2FV t t Ca Ca 2+ Ca [ ] i (1 κs κb ) δ( p ) γ([ ] i [ ] rest ) (3.13) Eq. (3.13) can be easily solved by integration to get ( t tp )/ τ [ Ca ] i [ Ca ] rest A θ ( t t p ) e = + (3.14) with θ representing the step function and the constants QCa A = 2 FV (1 + κ + κ ) (1 + κs + κs) τ = γ S S (3.15) Due to the linearity of eq. (3.13), the calcium response to a sequence of calcium influx is the superposition of the response to each individual calcium event ( t t )/ τ i = rest + θ p (3.16) p p [ Ca ] [ Ca ] A ( t t ) e Eq. (3.14) is also the Green s function for the problem of arbitrary time-dependent calcium input j m = j m (t), such that t ( t ξ)/ τ i = rest + m ξ [ Ca ] ( t) [ Ca ] A j ( ) e dξ (3.17) The multi-compartment model The single-compartment model breaks down when dealing with larger structures, such as the neuron cell body (r 10 µm), for which the diffusion is slow (for D Ca 200 µm 2 /s, t d 100 ms) and the spatial distributions are nontrivial. In a general case, both the ingredient diffusion process and the finite calcium binding/unbinding rate induced buffering process need to be considered. This leads to a complex reaction-diffusion problem 42

58 [127], which is usually solved numerically by treating the entire volume as multiple compartments. Although the complex internal structures, and inhomogeneous buffer distributions of the real neural cell body (or soma) are facts that have to be considered in detailed models, the general idea of the multi-compartment calcium dynamics modeling can be demonstrated via a simplified case, where the soma is considered as a homogeneous structureless ball of cytosolic solution covered by a spherical membrane (Figure 3.7 (A)). Here we study this case based on some previous works [ ]. By assuming the calcium influx across the membrane is uniform (the action potential propagates very fast, and the voltage-sensitive calcium channels respond to an action potential within 1 ms), the 3D diffusion equation can be reduced to 1D φ φ = D φ = D r 2 t r r r (3.18) with D the diffusion constant. The cytosolic solution contains Ca 2+, calcium indicator F (exogenous buffer), and Figure 3.7 Schematic drawing of the multi-compartment model. (A) Calcium dynamics: Ca 2+ fluxes/leaks through the cell and binds to different intracellular buffers; part of it is extruded from the cytosol. (B) Modeling of a neuron cell body using concentric spherical shells. 43

59 two types of endogenous buffer a diffusible endogenous buffer B, and a uniformly distributed fixed endogenous buffer C. The binding and unbind reactions of Buffers and Ca 2+ are assumed to be first order, such that B k+ [ F] + [ Ca] [ CaF] B k B k+ [ B] + [ Ca] [ CaB] B k B k+ [ C] + [ Ca] [ CaC] B k (3.19) with k + and k + indicating the calcium association and dissociation rate respectively. These rate constants also define the equilibrium constant (or calcium affinity of the buffer) k d = k - /k +. The calcium influx is described by a current I Ca across the cell membrane in response to action potentials, while different calcium extrusion mechanisms are lumped as a voltage-independent Michaelian kinetics [129] acting on the membrane J out = ν max [ Ca] [ Ca] + K m (3.20) where ν max is the maximal extrusion rate (in moles per surface and time unit) and K m is the half-maximal activating concentration. The high calcium concentration difference across the cell membrane also generates a calcium leakage flux, which can be written as a constant [129] J leak = ν max [ Ca] 0 [ Ca] + K 0 m (3.21) with [Ca] 0 representing the equilibrium free Ca 2+ concentration for a quiet neuron. Combining with eq. (3.20), this leads to a steady state at [Ca] = [Ca] 0. For computation, a finite difference method is used by dividing the ball into multiple 44

60 concentric shells of uniform thickness (Figure 3.7 (B)), and assuming that diffusion and flux only happens across the interfaces of these shells. Diffusion terms can thus be rewritten as [128] D t [ φ] = ( A([ φ] [ φ] ) A ([ φ] [ φ] )) D φ i i i+ 1 i i 1 i i 1 Vi r (3.22) where [ φ] D i is the concentration change of component φ (F, B, C, CaF, CaB, CaC, or free Ca 2+ ) in the ith shell caused by diffusion within a time interval t, r is the shell thickness, and V i, A i indicate the volume and the surface area of the ith shell respectively. The first term on the right-hand side of the equation is the influx through the outer shell, while the second term represents the outflux to the inner shell. There is neither diffusion influx to the outermost shell, nor diffusion outflux from the innermost shell. The buffering process is rewritten as S S S [ Ca] = ( k [ CaS] k [ Ca][ S] ) t i i + i i S [ S] = [ Ca] i S i S [ CaS] = [ Ca] i S i (3.23) with [ Ca] S i, [ S] S i, and [ CaS] S i representing the concentration change of free Ca 2+, free buffer S (could be F, B, or C), and bounded buffer CaS in the ith shell caused by their reaction within a time interval t. The calcium influx term is easily rewritten as [129] ICa t [ Ca] in = (3.24) 2FV Where F is Faraday s constant, and V N is the volume of the outermost shell. For action potential driven calcium influx, we assume that I Ca = I peak sin 2 (πt/τ Ca ) with t ranging between 0 and τ Ca. [126, 130]. N 45

61 Table 3.2 Parameters for multi-compartment calcium dynamics model Parameter Symbol Value Ca 2+ diffusion coefficient D Ca 2x10-6 cm 2 /s Resting Ca 2+ concentration [Ca] 0 50 nm Diffusion coefficient of F D F 2x10-6 cm 2 /s Total concentration of F [F] total 20 µm Calcium association constant of F Calcium affinity of F k + F k d F 5x10 8 M -1 s nm Fluorescence ratio F Ca /F free 14:1 Diffusion coefficient of B D B 0.5x10-6 cm 2 /s Total concentration of B [B] total 200 µm Calcium association constant of B Calcium affinity of B k + B k d B 1x10 8 M -1 s -1 5 µm Diffusion coefficient of C D C 0 Total concentration of C [C] total 500 µm Calcium association constant of C Calcium affinity of C k + C k d C 1x10 8 M -1 s -1 5 µm Peak calcium current I peak 100 pa Width of calcium current τ Ca 5 ms Maximal calcium extrusion rate ν max 8 pmol/cm 2 s Half-maximal concentration for extrusion k M 0.83 µm Simulation time bin t 0.01 ms Shell thickness r 0. 2 µm Cell radius R 10 µm The calcium extrusion and leakage terms are also rewritten as [129] 46

62 [ Ca] = ν out [ Ca] = ν leak max max AN t[ Ca] N V ([ Ca] + K ) N N m AN t[ Ca] 0 V ([ Ca] + K ) N 0 m (3.25) (3.26) The concentration of different components at each shell is given by [129] [ F] = [ F] + [ F] + [ F] (3.27) D F it, + t it, i i [ B] = [ B] + [ B] + [ B] (3.28) D B it, + t it, i i [ C] = [ C] + [ C] + [ C] (3.29) D C it, + t it, i i [ CaF] = [ CaF] + [ CaF] + [ CaF] (3.30) D F it, + t it, i i [ CaB] = [ CaB] + [ CaB] + [ CaB] (3.31) D B it, + t it, i i [ CaC] = [ CaC] + [ CaC] + [ CaC] (3.32) D C it, + t it, i i [ Ca] = [ Ca] + [ Ca] + [ Ca] + [ Ca] + [ Ca] D F B C it, + t it, i i i i + ( [ Ca] + [ Ca] + [ Ca] ) δ in out leak i, N (3.33) where the last three terms in eq. (3.33) only act on the outermost shell. By applying proper parameters (listed in Table 3.2, based on [ ]), we simulate the calcium dynamics of a model cell in response to incoming calcium pulses. As shown in Figure 3.8, once Ca 2+ flows into the cell, it first accumulates and reacts with calcium buffers in the region close to the plasma membrane (the outermost shell), and then diffuses into inner regions. After the influx, the extra Ca 2+ is extruded out of the cell from the plasma membrane, which lowers the Ca 2+ concentration in the outermost shell of the cell. Ca 2+ in inner shells is then sucked out via concentration gradient driven diffusion process, while the calcium storage in the buffer is released via unbinding processes. Most of the incoming Ca 2+ binds quickly with intracellular calcium buffers. In 47

63 Figure 3.8 Intracellular distribution of free calcium ions and calcium bounded indicator molecules (CaF) before ((A) and (D)), 5 ms after ((B) and (E)), and 300 ms after ((C) and (F)) a calcium influx. particular, the binding with calcium indicator (here we use the parameters of Oregon Green 488 BAPTA-1) increases the fluorescence intensity of indicator molecules. By integrating over the whole cell, we plot the evolution of the total fluorescence intensity of a model cell in response to incoming calcium pulses. Each calcium pulses tend to evoke a bump of fluorescence intensity ( F/F), which decays slowly afterwards (Figure 3.9 (A)). The height of the bump is about 1.7%, and the peak of the bump appears 44 ms after the initiation of the calcium pulse. Similar to the case in section 3.3.1, multiple calcium pulses induce superposition of fluorescence bumps (Figure 3.9 (B)), with each bump corresponding to a single calcium event (which also represents the timing of a neural action potential). The fluorescence peak value also shows very good linear relationship with the spike number in the pulse train (Figure 3.9 (F)). Because of the slow time course of the calcium dynamics, indi- 48

64 Figure 3.9 Flurescence signal of calcium indicator and the spike reconstruction techniques. (A) Fluorescence intensity increases in response to a single calcium pulse at t = 0.1 s. (B) Fluorescence intensity in response to multiple pulses at t = 0.1, 0.3, 0.5, 0.7, 0.9 s. (C) Fluorescence intensity in response to pulses at t = 0.1, 0.11, 0.12, 0.13, 0.14 s. Individual bumps cannot be resolved. (D) Derivative of the fluorescence signal of (C). (D) Deconvolution of the fluorescence signal of (C). (F) The relation between fluorescence peak value and the pulse number for different inter-pulse interval case (black: 100 ms inter-pulse interval, red: 10 ms inter-pulse interval). (G) Spike prediction timing error for different reconstruction methods. vidual calcium events of a high frequency pulse train will not be resolved on the fluorescence intensity plot (Figure 3.9 (C)), and extra data analysis procedures are needed to extract the timing of the calcium pulses out of the fluorescence signals. For example, the derivative of the fluorescence signal (d( F/F)/dt) resolves calcium events much better 49

65 than the original signal, giving individual peaks for each calcium events with timing errors less than 3 ms (Figure 3.9 (D)). An more elaborate strategy is to deconvolve the fluorescence signal (M(t) = F/F) using a unitary fluorescence kernel (K(t)) generated by a single calcium pulse. Assuming that effects of each calcium events sum up linearly, the entire fluorescence signal should be the convolution of the calcium pulse train (S(t)) and the unitary kernel Mt () = St ()* Kt () (3.34) Deconvolving K(t) out of M(t) should therefore restore S(t). As a general platform, the derivative method can also be included as a special case of deconvolution, where K has an exponential form 0, t < 0 Kλ () t = λt e, t 0 (3.35) Applying the property of convolution M(ω) = S(ω) K(ω), we derive the expression of S M ( ω) S t i M Mt M t (3.36) 1 1 λ () = ( ) = (( λ+ ω) ( ω)) = λ () + () Kλ ( ω) where 1 indicates the inverse Fourier transform. When λ 0, we return to the derivative method. In order to get a better prediction of the calcium pulse train, the kernel should approximate the real unitary fluorescence signal better than K 0. As shown in Figure 3.9 (E), we reconstruct the calcium pulse train (or the corresponding spike train) using a simplified kernel ( ( e e ) / 60 ), and get a better accuracy (timing error less than 1.6 ms) 0.15t 100t compared to the derivative method (Figure 3.9 (G)). In reality, noise of the fluorescence signals lowers the reconstruction accuracy and 50

66 the effective temporal resolution of the recording. The variability of cell size and buffer concentration also makes it difficult to use a universal kernel. Extra techniques are therefore needed for accurate reconstruction. 3.4 Other optical imaging techniques Voltage-sensitive probes, also known as potentiometric probes, which change their spectral properties in response to voltage changes, are another major group of the optical probes. Historically they are classified as "slow" and "fast" probes based on the speed of the response to voltage changes. The slow probes are dyes which redistribute between the extracellular and the intracellular medium in response to the membrane potential changes. The redistribution process is very slow (1 ~ 20 s) compared to the time scale of an action potential, limiting Figure 3.10 Voltage sensitive dye recording of neural activity. (A) The mechanism of fast voltage-sensitive dye ( (B) A voltage-sensitive dye (di-8-anepps from Molecular Probes) recording of neural activity in comparison with the simultaneous patch clamp recording. (C) DIC image of the recorded neurons (scale bar: 50 µm). (D) Fluorescence image of the recorded neurons ((B)-(D) reproduced from [8]). 51

67 their application in neuroscience research. Fast indicators are dyes that can follow the millisecond membrane potential changes, and therefore fast enough to monitor individual action potentials (Figure 3.10 (A)). These indicators are amphiphilic membrane staining dyes which usually have a pair of hydrocarbon chains acting as membrane anchors and a hydrophilic group which aligns the chromophore perpendicular to the membrane/aqueous interface. The chromophore is believed to undergo a large electronic charge shift in response to the membrane potential changes, inducing the electrochromism, also known as Stark effect, which shifts the energy level and the corresponding fluorescence spectra of the dye molecules [131]. The typical signal level ( F/F) for single wavelength measurement is 10-5 ~ 10-4 / mv (Figure 3.10 (B)-(D)). Similar to protein based calcium indicators, advanced genetic technology has led to the design of voltage-sensitive fluorescent proteins (VSFPs). Most of the existing VSFPs are FRET based, which rely on a voltage-sensing domain (VSD) that changes its conformation in response to membrane voltage shifts, and two fluorophore domains connected by the VSD, which changes the ratio of FRET and the emission spectra accordingly (Figure 3.11 (A) and (B)) [23]. Some newest version VSFPs contain only one fluorophore domain, but their voltage-sensing mechanism remains unknown (Figure 3.11 (C)) [132]. VSFPs have similar signal level as the fast voltage sensitive dye (VSD), and bigger time constants (~ 100 ms). They are good choices especially for in vivo recordings. The neural signals can also be recorded without using any indicators. Intrinsic optical imaging, as named, is based on the slow intrinsic changes in the optical properties of the active neural tissue. These activity-dependent intrinsic signals include changes in physical properties of the tissue itself which affect light scattering and/or changes in the 52

68 Figure 3.11 Development of a fast reporting VSFP. Top panel: The membrane topology of (A) and (B) a FRET based VSFP at different membrane voltage level. (C) a single fluorophore based VSFP. Underneath are emission spectra recorded from each construct using 440 nm excitation light. The lower panel shows the fluorescence signals recorded in the yellow and cyan channels with the inset showing the fluorescence and DIC images of a PC12 cell expressing VSFP3.1_Cerulean. For VSFP2A(R217Q) a scaled mirror-image of the cyan signal is shown aligned with the yellow signal; note the fast CFP component. For VSFP3.1 the onset of the fluorescence signal is shown on an expanded time scale; note the dramatically faster response of VSFP3.1. Scale bar indicates 20 mm (Figure reproduced from [23]). absorption, fluorescence or other optical properties of intrinsic molecules that have significant absorption or fluorescence. Although intrinsic imaging suffers from low signal level (around 10-3 ), low spatial resolution (> 50 µm), and a poorly-understood mechanism, the independence of the extrinsic fluorescence indicators leads to its application in some of the in vivo studies [133]. 3.5 Acquisition of optical signals 53

69 Figure 3.12 Typical experimental set-up for optical recording of neural activity. The small size of optical recording signals requires low-noise, high-sensitivity recording techniques. A typical experimental set-up for recording calcium indicator signals from a biological preparation includes an excitation light source with a mechanical or electrical shutter that illuminates the specimen for short times, proper filter sets and optics for focused photo excitation of the specimen, and a proper detector for fluorescence signal measurement (Figure 3.12) Light sources Tungsten filament lamps, arc lamps and lasers are three major kinds of light sources used in optical imaging experiments. Tungsten filament lamps give a stable and continuous 54

70 spectrum, with peak-to-peak noise level less than However, their output intensities are relatively low, particularly at the short wavelength range (λ < 450 nm) [9]. This drawback limits their application to absorption measurements, where the output intensities are usually big, while the fractional changes ( F/F) are small. Light sources based on high power arc lamps, such as the xenon and mercury lamps, are brighter illuminators with continuous spectra and fairly low noise level (< 10-3 ). They are the most popular choice for the fluorescence based optical recordings. Mercury-arc lamps exhibit very high light-flux densities, but their use is restricted to the wavelengths of the mercury lines. When the spectral range is not within one of the lines, xenon-arc lamp is the choice, which exhibits a relatively flat spectrum with only a few lines in the blue and the infrared regions. The coherent light output and the ability to generate high intensity diffraction-limited illumination spots make laser systems a perfect choice for various scanning based excitations, where the spot size is crucial [134]. The typical drawbacks of the laser systems are the relatively high noise level (10-3 ~ 10-2 ), limited available wavelengths, and the high cost. Recently the development of the advanced semiconductor growth and fabrication techniques has boosted the performance of the semiconductor light emitting diode (LED) in various wavelength range (from deep ultraviolet to near infrared), and added a new choice of excitation light source. Compared to the conventional lamp based light source, LED shows much higher energy transfer efficiency (~ 30% for Gallium Nitride based blue LEDs) with much more compact size. The high output power density, narrow spectrum, low noise level (10-3 ), high reliability, and low cost make LEDs a very competitive 55

71 Figure 3.13 Transmission spectra of the excitation, emission and dichroic filter for Oregon Green 488 filter set ( choice for fluorescence imaging Optics The light coming out of the wide-band excitation light source needs to be spectrally selected before illuminating the specimen. This is done either by optical filters or a diffraction grating based monochromator. The former has more flexible choices of the wavelength range while the latter holds the ability of continuous selection of the center wavelength. In epifluorescence mode as shown in Figure 3.12, both the excitation light and the emission light pass through the objective. Hence dichroic filters and emission filters are 56

72 needed to guide the excitation light to the specimen, while preventing it from reaching the detector. The selection of these filters is based on the optical properties of the indicators. Figure 3.13 shows the characteristics of the filter set for Oregon Green 488 BAP- TA Detectors Many factors need to be considered in choosing the right detector system for optical recordings. The first feature is the dynamic range, which determines the size of smallest fractional intensity change that can be measured. It is usually specified in db, in bits, or as an exponent (e.g. 100 db = 17 bits = 10 5, The signals of the optical recordings are usually within the range 10-2 ~ In order to achieve high enough signal-to-noise ratio, a detector with the dynamic range higher than 10 bits is often needed. The second feature is the detector noise level. There are two main sources of noise: the dark noise, which comes from the thermal excitation of the detecting part, and the shot noise, which is due to the quantum nature of emitted photons ( ~ I ) ( For low light intensity measurement, the dark noise dominates, and therefore cooling is usually necessary (e.g. cooled PMT and CCD); while for high intensity experiment, the shot noise dominates, which is independent of the detectors. Detector sensitivity, spatial and temporal resolution are other parameters that affects the performance of the detectors. The main types of optical detectors include the photomultiplier tube (array), the photodiode (array), the CCD based imaging devices and the CMOS based imaging de- 57

73 vices. The PMT gives the highest sensitivity, which is good for low intensity level measurement. However the limited ability to form spatial resolution and the saturation problem in high intensity region restrict its application mainly to 2-photon scanning microscopy. Photodiode (array) do not have the problem of saturation, and therefore can detect high intensity signals. The huge dark current of the device however makes it useless for most of the low intensity fluorescent recordings. CCD and CMOS based imaging devices both give good spatial and temporal resolutions. Cooled CCD devices have much lower dark noise and higher sensitivity (high quantum efficiency and high fill factor) with large dynamic range (12 ~ 16 bits) compared to CMOS devices, and are therefore the most popular choice for optical neural recordings. All of the works presented in this thesis are recorded using CCD based cameras. 58

74 CHAPTER 4 FAST CALCIUM IMAGING STUDY OF TADPOLE VISU- AL SYSTEM DEVELOPMENT 4.1 Introduction The brain visual system processes most of the information in a distributed/parallel way, where the information from different visual areas or of different features (color, motion, etc.) are transmitted through different neuron groups of the interconnected network, and processed via the mutual interactions between them. Since the neural information is coded as temporal patterns of neural electrical activities (mainly action potentials), the relative timing of different neurons activity in the network is usually important for the result of information processing [48-50]. Compared to the well-known studies of the development of circuit morphology and the formation of the topographically organized retinotectal map, the initial development of the temporal relations between activities of different neurons in the visual system is poorly understood. This is partly due to the technical difficulty of the simultaneous recording from multiple neurons with millisecond temporal resolution. The development of the high speed imaging devices and the optical recording techniques, especially the calcium imaging technique, shed light on this field. Shown in the last chapter, calcium imaging, which monitors the intracellular Ca 2+ concentration, can be used as an indirect measure of spiking activity of neurons with benefits of less mechanical invasiveness (compared to the multi-electrode array technique), and big signal level (> 10% in many 59

75 cases), which makes it possible to image large networks with individual neuron resolution [37, 135, 136]. Although the time evolution of calcium transients are much slower than that of action potentials, their onset timing always gives a good presentation of the onset timing of the corresponding spike (or a high frequency burst of spikes). Therefore when combined with high speed camera technique and proper data analysis procedures, it can infer neural spike trains with amazing accuracy [ ]. In this chapter, we focus on the early development of retinotectal circuit of Xenopus laevis tadpoles visual system. Over the development, a topographically precise projection forms from retina to the optic tectum [24, 43, 45, 61]. Meanwhile, the tectal local circuit also gets refined [46]. Both of these processes result in a more focused visual receptive field (RF) of optic tectal neurons, but little is known about the temporal relations between different tectal neuron s activities over the development, which might be important for the visual information processing and the subsequent decision making and motor output. By applying wide-field calcium imaging with high speed CCD camera as well as temporal deconvolution based spike train reconstruction technique, we achieved fast population recording of optic tectal neurons of animals between developmental stage 46 and 48/49, during which the retinotectal system undergo a period of rapid growth with retinal inputs being strengthened and focused, and tectal local circuits getting refined. In the experiment, some tectal neurons were recorded simultaneously by patch electrodes and calcium imaging. The calcium signals ( F/F) were deconvolved and compared to the electrical signals to optimize the spike train reconstruction process (section 4.2). Then by applying wide-field imaging and visual stimuli to in vivo preparations of animals, we 60

76 compared the temporal relations (synchronization) among multiple (30-90) recorded tectal neurons for different developmental stages. Dark-reared animals and animals treated with NMDA receptor (NMDAR) antagonists were used as experimental manipulations to test the effect of visual experience on visual development (section ). Moreover, spontaneous tectal neural activity recorded from whole-brain preparations of different developmental stages and rearing conditions was recorded and compared to check the effect of local circuitry on the temporal refinement of tectal neural activity (section 4.7). A Spike-timing dependent plasticity (STDP) based model is presented to simulate the experimental results (section 4.8), and we give a brief discussion at the end (section 4.9). 4.2 Experimental method and set-up In order to measure changes in neural synchrony among tectal neurons during development, we measured simultaneously the intracellular Ca 2+ levels of multiple tectal neurons in response to a whole-field visual stimulus. As shown in the last chapter, the intracellular Ca 2+ concentration, recorded using Ca 2+ sensitive fluorescence indicators, has been used as an indirect measure of spiking activity of neurons. Using a deconvolution algorithm, in combination with a fast sampling rate, these Ca 2+ responses can be used to predict spikes with a high degree of accuracy [138, 139]. In this study, we exposed tectal neurons, whose cell bodies are located on the surface of the tectum, to a membrane-permeant Ca 2+ indicator Oregon Green BAPTA-AM (Invitrogen) for 45 minutes before washing the indicator out from the extracellular media using ACSF. This resulted in a substantial num- 61

77 ber of tectal neurons which were brightly labeled by the indicator (see section for details). Animals were then imaged on a recording chamber with a high-speed CCD camera at a rate of 125 frames-per-second (fps). Compared with conventional two-photon imaging, our method allows for much higher temporal resolution, allowing us to better resolve spatiotemporal activation patterns across several tectal neurons. Whole-field, ON/OFF visual stimuli were delivered using a custom-made blue LED coupled to a 500 µm optic fiber which was placed close (< 200 µm) to the tadpole s eye (Figure 4.1 (A)). An area about 100 x 100 µm with labeled neurons in the contralateral tectum was imaged (Figure 4.1 (B)). Visual stimuli, consisting of 1 second ON and 1 second OFF were delivered after 2 s following the onset of the optical recording. Only the neuronal activity occurring within the ON period was analyzed (Figure 4.1 (C)). The visual stimuli evoked spiking activity in tectal neurons which was accompanied with rapid increases in fluorescence, corresponding to the opening of voltage sensitive Ca 2+ channels and an increase in intracellular Ca 2+ concentration. The fluorescence signal then decayed slowly back to resting levels with a time constant on the order of a second. In order to find the relationship between the fluorescence signal and neuronal spiking, some tectal neurons were loose-patched and action potentials were recorded electrophysiologically, simultaneously with Ca 2+ imaging. As stated in section 3.3.2, we assume that each spike generates a similar Ca 2+ fluorescence transient, which is then linearly added to the existing fluorescence signals. To reconstruct the electrical traces, we deconvolved the fluorescence signals ( F/F) using a kernel derived from Ca 2+ transients resulting from single spike events (see section for details). The resulting traces were then compared with the electrophysiological traces of the same cell, and each individual 62

78 Figure 4.1 High-speed calcium imaging of visual responses in tadpole optic tectum in vivo. (A) Schematic of the experimental setup: fiber-optic guided whole field visual stimuli were projected on to the tadpole s eye, while visual responses in calcium dye-loaded neurons in the contralateral tectum were imaged using high-speed CCD camera. (B) A fluorescence image of bulk-labeled tectal neurons in the surface of the optic tectum, with some cells being selected as examples (circled). (C) Sample calcium traces of selected tectal neurons in response to a visual stimulus. The timing of visual stimulus is represented by the blue bar. (D) Two examples of data analysis performed on the fluorescence traces. Top: raw fluorescence traces from two separate dye-loaded neurons in response to a visual stimulus. Middle: deconvolved fluorescence traces, showing multiple peaks. Bottom: simultaneous electrical recording using loose-patch electrodes shows that spikes correspond to peaks in the deconvolved trace. (E) Relationship between the peak amplitude of the deconvolved trace, and the spike number measured electrophysiologically (linear fit, data from stage 46 animals). (F) Relationship between the inter-spike interval and the spike number during a burst of spikes (data fitting using inversely linear relations, data from stage 46 animals). spike was found to closely correspond to a peak of the deconvolved trace (Figure

79 (D)). For a high frequency spike burst, the peak of the deconvolved trace was larger, and its height correlated with the number of spikes in the burst. Figure 4.1 (E) shows the averaged results from multiple (n = 7) stage 46 tadpole neurons, and demonstrates that the relation between the deconvolution peak height and the spike number is approximately linear. The burst frequency was also shown to be linearly related to the spike number (Figure 4.1 (F), see section for details). Thus we were able to decode the relationship between the fluorescence signal and the electrical activity of the cell, both during single spikes and during high-frequency bursts. This allowed us to translate the fluorescence signals back to electrical signals with a high degree of accuracy (timing error < 16 ms, spike number error < 1, see section for details). The detailed experimental steps are listed below Animal rearing conditions All animal experiments were carried out in accordance with Brown University Institutional Animals Care and Use Committee approved animal protocols. Wild type Xenopus laevis tadpoles were raised in 10% Steinberg s solution (1X Steinberg s in mm: 10 HEPES, 58 NaCl, 0.67 KCl, 0.34 Ca(NO 3 ) 2, 0.83 MgSO 4, ph = 7.4) in C incubators up to developmental stage 46 and stage 48/49 (typically 8-10 days). Animals were kept on a 12-hour light/dark cycle in the normal rearing condition experiments, some animals were maintained in a dark chamber to block any light between stage 46 and stage 48/49 (about a week). In another experimental group tadpoles were exposed to 10 µm of NMDAR blocker MK-801 (Tocris Bioscience) between stages 46 and 48/49. 64

80 4.2.2 Animal dissection Tadpoles were anaesthetized in 0.01% (v/v) MS-222 (Sigma), and pinned to a 11 cm imaging chamber filled with 4 ml HEPES-buffered extracellular saline (115 mm NaCl, 6 mm KCl, 3 mm CaCl 2, 0.5 mm MgCl 2,5 mm HEPES, 10 mm glucose, ph 7.25, 255 mosm) containing 100 µm (+)-Tubocurarine (Sigma) for animal immobilization. We carefully pierced and removed the skin covering the brain, exposing the surface of the tectum. A cut through the dorsal commissure was then made to expose the optic tectal lobes, with the medial surface being flattened to facilitate optical imaging [47, 143]. For the isolated whole-brain preparation, brains were then filleted along the dorsal midline and dissected out of the body. Brains were pinned down onto a block of sylgard in the imaging chamber [46, 143]. For both in vivo recordings and whole-brain recordings, the ventricular membrane was vacuumed off with a broken micropipette tip before the dye loading to help dye diffusion and penetration into the tissue. All recordings were carried out at room temperature (20-22 C) Calcium indicator loading and imaging OGB1-AM was dissolved at a concentration of 2 mm in DMSO with 10% pluronic acid (Invitrogen) and diluted to a final concentration of 10 µm in the extracellular saline in the imaging chamber. Samples (either the in vivo preparation or the whole-brain preparation) were incubated with this solution for 45 min for dye staining, and subsequently washed with fresh extracellular saline 3 times. The imaging chamber was then mounted on the stage of a Nikon E600FN upright microscope with a 40x objective (N.A. = 0.8) and an OGB-1 filter set (Semrock) for imaging. Excitation light coming from a high-power blue 65

81 LED (LXK2-PB14-N00 from Philips Lumileds Lighting Company) through the back port of the microscope illuminated the neurons, generating green fluorescence, which was projected on the CCD surface of a high speed neural imaging camera (NeuroCCD SMQ from Redshirt Imaging LLC). The OEM software Neuroplex (Redshirt Imaging LLC) was used for image acquisition and basic analysis. An imaging area of 100 µm x 100 µm with 80 x 80 pixels was achieved at a frame rate of 125 fps. The imaging time was set to be 4 seconds for evoked response recordings (in vivo preparations), and 16 seconds for spontaneous recordings (whole-brain preparations). For each experiment about fluorescent tectal neurons were imaged on the same field of view. 10 trials with identical conditions were recorded for each group of cells. A minimum of 10 experiments were collected for each developmental stage (stage 46 and stage 48/49) and different experimental conditions (dark reared animals and MK-801 treated animals) Visual stimulation In the evoked response recordings, whole-field visual stimuli were generated by a custom-built blue LED (illumination peak: 470 nm, 30 ma current injection), which was coupled directly to an optic fiber (500 µm core diameter) for light guidance. The fiber end was mounted on a MP-285 micromanipulator (Sutter Instruments) for positional control. The distance between the fiber end and the tadpole s eye was less than 200 µm. The LED was turned on 2 seconds after the start of the imaging trial, and remained on for 1 second. 66

82 4.2.5 Electrophysiology Visualized cell-attached recordings of fluorescent tectal cells were made with micropipettes (8-12 MΩ) filled with K-gluconate based intracellular saline (100 mm K-gluconate, 8 mm KCl, 5 mm NaCl, 1.5 mm MgCl 2, 20 mm HEPES, 10 mm EGTA, 2 mm ATP and 0.3 mm GTP, ph 7.25, 255 mosm). Signals were measured with an Axoclamp 2B amplifier (Molecular Devices), digitized at 10 KHz using a Digidata 1440A A-D board and acquired using p-clamp 10 software. Data was analyzed together with imaging data using a custom written Matlab (Mathworks) program Spike train reconstruction Neuronal spike trains were reconstructed from fluorescence changes resulting from changes in intracellular calcium levels. While calcium signals have a slow decay time, there are several methods to detect rapid increases in fluorescence which correspond to neuronal spiking [135, 138, 139, 141]. Here we used a deconvolution algorithm to reconstruct periods of neuronal spiking from the calcium imaging data. For each optical recording, regions of interest (ROIs) were drawn to cover the fluorescent neuronal cell bodies. Each ROI typically was comprised of about 40 pixels. Averaged intensity signals of each ROI were exported into text files from the Neuroplex software (Redshirt Imaging LLC) for further analysis using custom written Matlab programs. The change of fluorescence intensity of a ROI was calculated as F F F = 0 (4.1) F F0 Where F is the averaged intensity of the ROI in an image frame and F 0 is the fluores- 67

83 cence baseline, chosen as the minimal averaged fluorescence level (over every 15 consecutive frames) of the ROI over the recording. The normalized data were then filtered three times using a boxcar filter (window size: 7 frames) to smooth out small fluctuations most likely caused by shot noise. The filtered signals were then deconvolved with a kernel based on a unitary calcium transient kernel: y e e t/5 s t/0.2s = (4.2) This kernel is the result of fitting calcium transients resulting from single spikes as measured electrophysiologically. For calibration purposes, a subset of tectal neurons (>5 neurons for each developmental stage and rearing condition) were loose-patched to record action potentials while simultaneously being imaged. Each spike (or burst) measured electrophysiologically was correlated with a peak of the deconvolved waveform of the calcium signal. The relationship between the number of spikes in a burst and the height of the corresponding deconvolution peak was plotted in order to generate a calibration curve for each developmental stage and rearing condition. Bursts of different numbers of spikes (including single spikes) produced peak amplitudes with different ranges, and a group of threshold values d N with N representing the spike number (e.g. 1, 2, 3, 4, etc.) was set to infer the spike number of a burst from its corresponding deconvolution peak (Figure 4.2 (A)). Optimization was made on these thresholds to allow the best prediction with minimal error (σ < 1 spike) for each developmental stage and rearing condition (Figure 4.3 (A)). In order to predict the timing of individual spikes from a deconvolved trace, we first determined the relationship between the actual spike onset timing and the timing of the 68

84 Figure 4.2 Spike train reconstruction scheme (data from stage 46 animals). (A) Deconvolution peak thresholding for the prediction of spike number in bursts. (B) The relationship between the deconvolution peak width and its deviation from the corresponding spike onset time. (C) The relationship between inter-spike interval and the spike number in bursts. peak of the deconvolved trace by comparing these two measures. While the two timings were closely related, they did not match exactly. However, we were able to correct this discrepancy by using a complementary measure to compensate for the difference. The timing difference between the spike onset and the peak of the deconvolution peak was found to be linearly proportional to the width (defined as the time interval with deconvolution value bigger than d 1 ) of the deconvolved trace peak (Figure 4.2 (B)). Plotting one versus the other provided a curve by which we could accurately predict spike timing from the deconvolved trace. Thus, prediction of the spike timing (or the onset timing of a burst) was made based on both the peak latency and peak width of the deconvolved trace, yielding an error of less than 16 ms (Figure 4.3 (B)). To predict the timing of individual spikes within a burst, we used the simultaneously recorded electrophysiological data to determine the relationship between inter-spike interval and the spike number in a burst, and plotted this relationship for each developmental stage and rearing condition (Figure 4.2 (C)). The spike frequency of a burst is linearly related to the number of spikes in a train. Thus, we used an inverse linear function to predict the interspike interval within a 69

85 Figure 4.3 Accuracy of spike train reconstructions. The distributions of (A) spike number prediction errors, (B) spike/burst onset timing prediction errors, and (C) inter-spike interval (within a burst) prediction errors for different experimental groups are shown. burst and, by extension, the total duration of the spike burst (Figure 4.3 (C)). Using the above calibration curves for determining number of spikes, spike or burst 70

86 onset, and burst duration from deconvolved traces, we reconstructed spike trains from the calcium imaging data. Thresholds d N were first applied to the deconvolved calcium signals for the detection of spikes (or bursts). The spike number of each burst was predicted from the corresponding deconvolution peak height, while the spike timing (or the onset of a burst) was estimated from the deconvolution peak timing and peak width. The slow decay of calcium transients does not allow the detection of precise timing of each spike in a burst, however the precise timing of each spike is not important for the calculation of cross-correlations values as described below. Therefore, spike trains were reconstructed as square pulses of spiking frequency profiles (using the inter-spike interval inference mentioned above) instead of trains of individual spikes, with its area equal to the predicted number of spikes. For the consistency of data analysis, a frequency was also assigned to responses with individual spikes, which was equal to the frequency of dual spike burst Data analysis In the following sections, the reconstructed spike trains were used to compare temporal properties of neural activity between individual cell pairs within a population. The cross-correlation coefficient of a cell pair was calculated based on the formula [144] Q ρσ ( τ ) = T T 0 ρ() t σ( t + τ) dt T 2 2 ρ () t dt σ () t dt 0 0 (4.3) with ρ and σ representing the reconstructed spike train, and < > the average over different recordings. The maximal value of Q ρσ (τ) over different τ was used for statistical analysis. 71

87 To calculate trial-to-trial variability within individual cells, we computed the trial-to-trial correlation of cell activity over different trials, which was calculated as Q ρρ ' = T T 0 ρ() t ρ () t dt T 2 2 ρ () t dt ρ () t dt 0 0 (4.4) with ρ and ρ representing the different recordings of the same cell, and < > the average over different recording pairs. The Pearson product-moment correlation coefficient was used to calculate the likelihood of the linear relationship between two variables according to the formula [145] r XY = ( X X)( Y Y) 2 1/2 2 1/2 ( X X) ( Y Y) (4.5) Custom written Matlab (Mathworks) programs were used for all data analysis. 4.3 Development of cross-correlation of evoked neural activity Using the calcium imaging and deconvolution method we reconstructed the electrophysiological activity of groups of tectal neurons in response to whole-field visual stimuli. In each experiment we measured the activity of a population of tectal neurons. Only cells which responded reliably to the stimulus (i.e. > 50% of the time) were included in the analysis so as to eliminate the possibility of recording from unhealthy neurons. We first looked at the development synchronous responses across neurons in a population by measuring the degree of cross-correlation of neural activity between cells. As mentioned in section 4.2.7, this was achieved by calculating the average peak cross-correlation coefficient of each pair of cells over 10 trials, and then plotting the distribution of average 72

88 Figure 4.4 Developmental increase in correlated neural activity among large groups of tectal neurons is dependent on visual experience and NMDAR activation. (A) Correlation matrix for pairs of tectal neurons bulk labeled with Ca 2+ indicator dyes in response to a whole-field visual stimulus. Color represents average correlation values for each cell pair over ten trials. One sample experiment is shown for each experimental condition. (B) Normalized distribution of correlation values for pairs of tectal neurons from the same example as in A. (C) Averaged distributions of correlation values across multiple experiments for each experimental condition. Notice the overall shift to higher correlation values between stages 46 and 48-49, as well as the effect of dark rearing and MK-801 treatment in preventing this shift. (D) Average of total mean correlation value for each developmental stage and rearing conditions. Asterisk denotes p < 0.05, see text for exact p values and number of experiments. correlation values for each experiment. Bigger correlation coefficients represent more consistent and coherent neural activity, which may be important for more reliable information processing. Comparisons were made between developmental stage 46 and stage 48/49 Xenopus laevis tadpoles, a period during which substantial changes in tectal cell intrinsic excitability, synaptic transmission, RF properties and visually-guided behavior 73

89 are known to occur [45, 47, 143, 146]. We found a significant increase in the correlation coefficient between stages 46 and 48/49, as seen in the correlation maps and correlation coefficient histograms shown in Figure 4.4 (A) and (B). The averaged results from multiple experiments (n = 9 for stage 46, and n=14 for stage 48/49) confirmed the shift of the correlation coefficient distribution (Figure 4.4 (C)) and a significant increase of the mean cross-correlation coefficient between the two groups (stage 46: ± 0.019, n = 9; stage 48/49: 0.74 ± 0.036, n = 14; p < ; Figure 4.4 (D)). This indicates that the tectal neural response patterns become significantly more coherent and precise over development. This developmental increase in neural correlation is consistent with previous studies showing that during development retinotectal projections become stronger and more focused [24, 61], and excitatory tectal recurrent connections become refined [46], both of which may help increase the accuracy of the relative response timing across neurons as observed in this study. These synaptic changes are known to be experience dependent and some involve NMDA receptor activation [24, 46]. In order to test whether the increases in correlated activity observed here are also dependent on visual experience, we compared normally-reared stage 48/49 tadpoles with tadpoles dark-reared between developmental stages 46 and 48/49 (7-9 days, see section for more details). Strikingly, we found that dark-rearing completely eliminates the developmental increase in neural correlation, leaving the correlation distribution almost unchanged compared to stage 46 (Figure 4.4 (A), (B), and (C)) with a low mean cross-correlation coefficient (0.224 ± 0.021, n = 17; p = 0.859; Figure 4.4 (D)). To test whether the increase in correlation required NMDA receptor activation, tadpoles were reared in NMDAR blocker MK-801 (10 µm). In contrast 74

90 to dark rearing, blockade of NMDA receptors, only partially disrupted development, resulting in a more flat distribution of correlation coefficients (Figure 4.4 (A), (B), and (C)) and a smaller change in the mean cross-correlation coefficient (0.524 ± 0.025, n = 17; Figure 4.4 (D)). Although MK-801 did not completely block developmental changes in neuronal correlation, the changes were significantly smaller than in control stage 48/49 tadpoles (p < ). The incomplete effect of MK-801 on development is likely due to the fact that multiple mechanisms of plasticity, such as those requiring activation of Ca 2+ -permeable AMPARs, also contribute to circuit refinement [46]. Taken together, these findings suggest that as a result of visual experience between stages 46 and 48/49, the correlated activity among tectal neurons significantly increases. This refinement is due, at least in part, to activation of NMDAR. 4.4 Development of visual response timing To further examine the development of temporally coherent responses between tectal neurons, we analyzed the onset latencies of visual responses across neurons. Most neurons responded to light a few hundred milliseconds after the onset of the visual stimulus (see Figure 4.5 (A) for examples). Stage 48/49 animals showed a much shorter mean onset latency (averaged over 10 trials for each cell, and then averaged across cells for each experiment) compared to stage 46 animals (stage 46: ± 0.04 s, n = 9; stage 48/49: ± s, n = 14; p = 0.023; Figure 4.5 (D)). This observation is consistent with a developmental increase in the strength of retinotectal synapses. Dark-reared stage 48/49 tadpoles had similar onset latency values as stage 46 animals (0.367 ± s, n = 17; 75

91 Figure 4.5 Developmental increase in visual response synchrony among tectal neurons. (A) Examples of visual response onset latencies from different developmental stages and rearing conditions. Each example illustrates response onset latencies from eight randomly-selected tectal neurons in the same recording. Error bars indicate variability over ten trials. (B) Sample distributions of visual response onset latency differences between pairs of tectal neurons. One example is shown for each experimental group. (C) Averaged normalized distributions of response onset latency differences for each experimental group. Notice the narrowing of the response onset differences between cells over development, indicating increased synchronous activation. (D) Averaged mean visual response onset latencies for each experimental group. (E) Averaged standard deviations of visual response onset latencies for each experimental group. For each experiment, the standard deviation of the onset latencies across trials was calculated for each cell, these values were averaged to obtain the mean value for each experiment. This measure indicates the variability in response onsets across trials. (F) Averaged standard deviation onset latency differences for each experimental group. The standard deviation of onset latency differences was calculated from the values in B for each experiment and reflects the width of the distribution. Less synchronous activation will result in higher values. Asterisk denotes p<0.05, see text for exact p values and number of experiments. Figure 4.5 (D)), and were significantly slower than control stage 48/49 tadpoles (p < 76

92 0.0001), possibly due to abnormal development of retinotectal inputs, resulting in weak synaptic connections. Rearing in MK-801 resulted in an intermediate effect, consistent with our previous finding (0.266 ± s, n = 17; Figure 4.5 (D)), and this group was also significantly slower than control stage 48/49 tadpoles (p = ). We also looked at variability in the onset latency across trials for individual cells by measuring the standard deviation of the onset latency (over 10 trials for each cell, see Figure 4.5 (A) for examples). This measure was found to drop dramatically between stage 46 and stage 48/49 (stage 46: ± s, n = 9; stage 48/49: ± s, n = 14; p = 0.002; Figure 4.5 (E)), indicating that retinotectal activation becomes more reliable. This trend was completely prevented by dark rearing (0.132 ± s, n = 17; p < ; Figure 4.5 (E)), while MK-801 treatment resulted in a smaller effect but still significantly different from stage 48/49 controls (0.042 ± s, n = 17; p = ; Figure 4.5 (E)). The onset latency differences between cell pairs were also measured as another metric to describe the level of neural synchrony between cells. We first calculated the average onset latency difference, for each cell pair in a given experiment (see section for details), and plotted the distribution of onset difference values (see Figure 4.5 (B) for examples). We then averaged the normalized distributions for each developmental stage and rearing condition (see Figure 4.5 (C)). Developmental stage 48/49 tadpoles had a much narrower distribution of onset latency differences as evidenced by a significantly smaller standard deviation compared to stage 46 tadpoles (stage 46: ± s, n = 9; stage 48/49: ± s, n = 14; p < ; Figure 4.5 (F)), representing the development of more synchronized neural activity. Tadpoles reared in the dark displayed similar distributions and mean value as stage 46 animals, and were very significantly dif- 77

93 ferent from normal reared stage 48/49 animals (0.160 ± s, n = 17; p < ; Figure 4.5 (F)). The MK-801-treated group, in contrast, showed a more modest yet significant disruption of development (0.045 ± s, n = 17; p < 0.034; Figure 4.5 (F)). Taken together, these results indicate that over development, visually-evoked activity in tectal neurons becomes faster and the cell s initial response becomes less variable. Likewise, the onset latency of visual responses between cells becomes increasingly time-locked, indicating a greater degree of neural synchrony over development. These changes are disrupted by visual activity and may require activation of NMDARs. One possibility is that changes in response onset latency reflect maturation of the retinotectal synapses, while changes in onset differences between cells reflects also maturation of intratectal synapses. Because NMDAR activation is only one mechanism affecting development of intratectal circuits [46], it is possible that MK-801 had a larger effect on overall onset latencies than it did on latency differences between cells. 4.5 Development of the reliability of neural response Over development the visual responses of a given cell were also found to become less variable. We tested trial-to-trial variability of single cells across trials as an indicator of response reliability across experimental conditions. This was done by calculating trial-to-trial correlation values for each cell (See Figure 4.6 (A) for examples). Trial-to-trial correlation is defined as the zero-lag correlation coefficient between two recordings of the same cell s response to the same stimuli (see section for details). A larger trial-to-trial correlation value indicates less variable and more reliable neural responses. 78

94 Figure 4.6 Developmental increase in response reliability in single neurons is activity dependent. (A) Distribution of the trial-to trial correlation values from single neurons. One example is shown for each experimental group. Inset shows raw fluorescence traces in response to multiple trials in a single neuron. Notice the decrease in waveform variability in the older tadpoles and its dependence on neural activity. (B) Averaged mean trial-to-trial correlation values for each experiment in each experimental group. (C) Averaged mean evoked spike number during a visual response for each experiment in each experimental group. Notice that single-cell responses become less variable and more robust over development. Asterisk denotes p < 0.05, see text for exact p values and number of experiments. Overall, we observed a large increase of the mean trial-to-trial correlation value between stage 46 and stage 48/49 tadpoles (stage 46: ± , n = 9; stage 48/49: ± s, n = 14; p < ; Figure 4.6 (B)). Dark-reared animals showed a small mean trial-to-trial correlation value, similar to that of stage 46 tadpoles (0.07 ± 0.027, n = 17; p < ; Figure 4.6 (B)), further supporting the hypothesis that visual input is critical for developing reliable tectal responses. NMDA receptor antagonist MK-801 also interfered with the normal developmental increase of trial-to-trial correlation values (0.22 ± 0.036, n = 17; p = ; Figure 4.6 (B)). 79

95 In addition to being more reliable, visual responses tended to become more robust over development as the retinotectal synapses strengthen. We used the number of visually-evoked spikes as an indicator of the strength of retinotectal projection. Cells from stage 48/49 animals fired significantly more spikes than cells from stage 46 animals (stage 46: 1.29 ± 0.03, n = 9; stage 48/49: 2.41 ± 0.27, n = 14; p = ; Figure 4.6 (C)) in response to the onset of the visual stimuli. Dark reared animals and MK-801 treated animals both responded with less spikes when compared to stage 49 controls (dark reared: 1.86 ± 0.07, n = 17; p = 0.07; MK-801 treated: 1.55 ± 0.05, n = 17; p = ; Figure 4.6 (C)), although the effect did not quite reach significance in the dark-reared group. This may be due to homeostatic compensation of intrinsic excitability during dark rearing. Together, these data indicate that visual input and NMDAR activation may be important for the strengthening of the retinotectal projection over development. 4.6 Development of the spatial-correlation and spatial-temporal relations Retinal inputs to the tectum are organized in a topographic manner such that neighboring retinal ganglion cells project to neighboring optic tectal neurons [24, 43, 61]. This topographic map is a product of both guidance by molecular cues of growing retinal axons and a subsequent sorting mediated by neural plasticity over development [24, 43, 61]. One important functional prediction stemming from this retinotopic organization is that neighboring tectal neurons would be more likely to respond together to visual stimuli, resulting in a higher cross-correlation coefficient and a smaller onset latency difference between these cells, even if the visual stimuli is uniform across the field. 80

96 Figure 4.7 Development of spatiotemporal properties of visual responses. (A) Comparison of correlation values between cell pairs and the distance between each pair, shown with a linear fit. One example is shown per experimental group. (B) Comparison of response onset latency differences between cell pairs and the distance between each pair, shown with a linear fit. One example is shown per experimental group. (C) Average r values of correlation vs. distance comparisons for each experimental condition. (D) Proportion of experiments in which this comparison showed a statistically significant p value (p < 0.05). A significant p value indicates increased correlated activity among nearby neurons. (E) Average r values of onset latency differences vs. distance comparisons for each experimental condition. (F) Proportion of experiments in which this comparison showed a statistically significant p value (p < 0.05). A significant p value indicates increased synchronous evoked activity among nearby neurons. We examined this hypothesis by measuring the spatial distance between each cell pair (see section for details) and plotted it against the cross-correlation coefficient and the onset latency difference between the cells (see Figure 4.7 (A) and (B) for examples). To measure the strength of linear dependence between the different variables we 81

97 calculated the Pearson product-moment correlation coefficient (PMCC, or r-value) for each experiment. PMCC gives a value between -1 and 1, with the negative values representing a negative linear dependence (Y decreases with X) and the positive values representing a positive linear dependence (Y increases with X). The higher the absolute value of r, the stronger the two variables are linearly correlated. If r is close to zero (-0.1 < r < 0.1), two variables are said to be linearly independent (see section for the formula). A statistical table was used to determine if a given r value was statistically significant. Stage 48/49 animals, compared to stage 46 animals, showed overall more negative r values for distance vs. correlation coefficient comparisons (stage 46: ± 0.033, n = 10; stage 48/49: ± 0.039, n = 14; Figure 4.7 (C)) and a more positive r value for distance vs. onset latency difference comparisons (stage 46: ± 0.040, n = 9; stage 48/49: ± 0.047, n = 14; Figure 4.7 (E)). This means that in older tadpoles, cells close together are more likely to exhibit correlated activity and the onset of their responses to occur within a shorter time difference. The r values were shown to be statistically significant in a greater proportion of experiments in the older tadpoles (11 of 14 experiments for distance vs. correlation, Figure 4.7 (D); 13 of 14 experiments for distance vs. onset latency difference, Figure 4.7 (F)) compared to stage 46 animals (4 of 9 experiments for both comparisons, Figure 4.7 (D) and (F)). Taken together these functional data are consistent with an increased refinement of the topographic organization of the retinotectal projection during development. Dark rearing and to a lesser degree NMDAR blockade both disrupted developmental changes in the relationship between distance and both variables (correlation, dark reared: r= ± 0.031, n = 17; correlation, MK-801 treated: r= ± 0.045, n = 17; Figure 82

98 4.7 (C); onset latency difference, dark reared: r= ± 0.037, n = 17; onset latency, MK-801 treated: r= ± 0.045, n = 17; Figure 4.7 (E)), indicating a more poorly developed map as a result of these manipulations. Dark reared animals showed a smaller proportion of experiments, relative to controls, which showed statistically significant correlations (6 of 17 groups of recordings for spatial-correlation test, Figure 4.7 (D); 5 of 17 groups of recordings for spatial-temporal test, Figure 4.7 (F)). In contrast MK-801 treatment only affected the relationship between distance and onset timing, but not between correlation and distance (15 of 17 groups of recordings for spatial-correlation test, Figure 4.7 (D); 10 of 17 groups of recordings for spatial-temporal test, Figure 4.7 (F)). This is consistent with the view that MK-801 may be having a greater effect on retinotectal synapse development rather than on intratectal synapses. Together, these findings directly support previous studies [24, 43, 61] which had suggested that visual experience and NMDAR activity may be important for the refinement of topographic maps in the tectum. 4.7 Development of cross-correlation of spontaneous neural activity The correlation of evoked neural responses is determined by both the feedforward retinotectal projection and the recurrent local connections within the tectum. In order to test whether local tectal circuits are involved in mediating the developmental increase in correlated activity between cells, we used an isolated whole-brain preparation. Here, tadpole brains were dissected out and directly placed on the recording dish, isolated from all sensory input [146]. During each experiment, a group of fluorescently-labeled tectal 83

99 Figure 4.8 Developmental increase in correlated spontaneous neural activity among large groups of tectal neurons is dependent on NMDAR activation but not visual experience. (A) Correlation matrix resulting from spontaneous neural activity for pairs of tectal neurons that were bulk labeled with Ca 2+ indicator. Color represents average correlation values for each cell pair. One sample experiment is shown for each experimental condition. (B) Normalized distribution of correlation values for pairs of tectal neurons from the same examples as in A. (C) Averaged distributions of correlation values across multiple experiments for each experimental condition. Notice the small but significant shift to higher correlation values between stages 46 and (D) Average of total mean correlation value for each developmental stage and rearing condition. While MK-801 rearing disrupted the developmental increase in correlated spontaneous activity, dark rearing did not. Asterisk denotes p < 0.05, see text for exact p values and number of experiments. neurons was imaged over 10 trials, each lasting 16 seconds. Cells were included for analysis if they exhibited fluorescence transients in more than 40% of trials. In the isolated brain preparation, spontaneous activity was low and therefore overall correlation values were low (Figure 4.8 (A) and (B)). However, we found that between stages 46 and 48/49 there was an increase in the amount of correlated spontaneous activity (Figure 4.8 (A) - (C); stage 46: ± , n = 29; stage 48/49: ± 0.001, n = 53; p < 84

100 0.0001; Figure 4.8(D)), consistent with increased maturation of local tectal circuitry. Dark-rearing had no obvious effect on the development of correlated spontaneous activity (Figure 4.8 (C) and (D); dark reared: ± 0.002, n = 74; p = ). One explanation for this is that the tectum is known to receive inputs from a variety of sensory modalities [147], and it is possible that these non-visual modalities are still generating sufficient patterned activity to organize local circuitry in the absence of visual input. MK-801 treatment partially prevented the developmental increase in correlated spontaneous activity (MK-801 treated animals: ± 0.002, n = 53; p = ; Figure 4.8 (D)), consistent with its known partial blockade of plasticity in intratectal synapses [46]. 4.8 Modeling of activity dependent tadpole retinotectal system development Our experimental results suggest that the developmental increase in correlated activity among tectal neurons requires visual activity and is in part dependent on NMDAR activation. Prior studies had implicated that spike-timing dependent plasticity (STDP) in retinotectal synapses and in local tectal synapses may be important for mediating the maturation of these circuits [29, 46] and our results are consistent with these claims. To test whether STDP driven by patterned visual activity could be sufficient to organize the maturation of these circuits, we created a computational model of the retinotectal circuit consisting of a two-layer neural network based on Song and Abbott s work [148, 149]. The model was comprised of an input cell layer (representing the retina), and an interconnected output cell layer (representing the optic tectum with recurrent synaptic connections), with feedforward synaptic connections between them (Figure 4.9). Period 85

101 Figure 4.9 Schematic of the computational model for retinotectal system development. An input cell layer and an interconnected output cell layer represent the RGCs and tectal neurons respectively. The input layer fires action potentials in a spatially organized way, and affects the output cells via excitatory feedforward connections. An STDP rule was applied to all excitatory connections, and a topographic organization can evolve over the simulation. boundary conditions were assumed in order to remove the possible boundary effects. A STDP rule was imposed between the retinotectal synapses and between local tectal connections. Random patterned activity was generated in the input layer through several iterations. The technical details of the model are listed as follows Model details The output layer composed 100 integrate-and-fire neurons following a standard implementation, where the membrane potentials obey dv τ m = V rest V + g ex ( E ex V ) + g in( E in V ) dt (4.6) with τ m = 20 ms, V rest = -74 mv, E ex = 0 mv, and E in = -74 mv. Each synaptic input is modeled as a instantaneous conductance increase (g exm or g inm ) in response to an input spike, with exponential decay, where g ex (t) = g exm exp(-t/τ ex ), and g in (t) = g inm exp(-t/τ in ), with τ ex = τ in = 5 ms. The threshold for firing an action potential was set to be -54 mv, and the membrane potential was reset to -60 mv after firing. 86

102 The input layer contained 500 neurons that fired through Poisson processes in response to a point stimulus as ( s a) /2 σ ( s+ 500 a) /2 σ ( s 1000 a) /2σ r ( ) a = R0 + R1 e + e + e (4.7) with R 0 = 2 Hz, R 1 = 25 Hz (based on data from tadpole retinal recordings, Demas and Cline personal communication), σ = 50, and s and a being the position of the stimulus and the input layer neuron respectively (three terms in the bracket preserved the periodicity). The point stimulus s was chosen randomly between 1 and 500 from a uniform distribution, and kept stable for a time interval chosen from an exponential distribution with a mean of τ r = 50 ms. Each output layer neuron received excitatory synaptic inputs from input layer neurons, other output layer neurons, as well as from an additional background Poisson input at 500 Hz through a synapse of strength 0.096, which mimicked the effects of a large number of afferents. The STDP rule was modeled as a change of the synaptic strength g exm = g exm + g max F( t) for the corresponding pre- and postsynaptic spike pair, where t = t pre t post and A+ exp( t/ τ + ), t < 0 F( t) = A exp( t/ τ ), t 0 (4.8) with g max = 0.015, A + = , τ + = τ - = 20 ms for feedforward connections and 40 ms for recurrent connections. To avoid uncontrolled synaptic growth and make sure that feedforward connections dominated recurrent connection, the ratio of the areas under the negative and positive portions of the STDP window function, was set to be B = A - τ - /A + τ + =1.06 for feedforward connections and 1.04 for recurrent connections respectively. g exm was set to change between 0 and g max, and would set to the appropriate limit value once 87

103 reaching the boundary. The initial feedforward connection strengths were set to be 1 d ij max exp( ( ) ), g (4.9) 2 50 where d ij = min( i/5 j, i/5 j ) for input layer neuron i and output layer neuron j. The recurrent connections between output layer neurons (j, k) were limited in the range min( j k, j k ) 25, and were set initially to 0. An all-to-all inhibitory connections with g inm = 0.15 g max were also applied to the recurrent network. After the STDP based development step, the resulting network were compared to the initial network by applying whole field stimuli (r a = R 0 + R 1 for all input layer neurons, and τ r = 100 ms) multiple times. A time delay of the stimuli was randomly chosen each time from a normal distribution with <t delay > = 0.1 s, and σ delay = 0.02 s. Spike trains of 50 consecutive output layer neurons were saved for 1 second and filtered three times using a boxcar filter (window size: 56 ms). The remaining data analysis steps were the same as the experimental data process. Spontaneous activity simulation were done by generating and saving output layer activity for 16 s simulation time with no stimuli (r a = 0) Model results During the development period, the model reproduced most of the experimental findings observed above: increased mean cross-correlation (Figure 4.10 (A)), decreased width of the response onset latency difference distribution (Figure 4.10 (B)), and increased mean auto-correlation across trials (Figure 4.10 (C)). Furthermore, the model developed a more refined retinotopic organization (Figure 4.10 (E)). Using the current input pattern we did not observe an increase in the spatiotemporal refinement of local 88

104 Figure 4.10 Developmental changes in correlation values, neural synchrony and response reliability can be modeled using STDP rules. (A)-(C): Distributions and mean values of (A) cross-correlation values, (B) response onset latency differences, and (C) single cell trial-to-trial correlation vaues, in response to whole field visual stimuli, at the beginning and the end of the simulation. (D) Distributions and mean values of output layer spontaneous activity cross-correlation values at the beginning and the end of the simulation. (E) Feedforward connection matrix (horizontal axis: input layer, vertical axis: output layer) at the beginning and the end of the simulation. (F) Connection matrix of recurrent connections within the output layer at the beginning and the end of the simulation. tectal projections nor an increase in the correlation values of spontaneous activity (Figure 4.10 (D) and (F)). This suggests that it may take more complex visual input patterns to cause refinement of local inputs or that a combination of patterned multisensory inputs to the tectum are required in order to refine local circuitry. Alternatively, refinement of local inputs may be activity independent, although it can be modified by STDP [46]. 89

105 Taken together, the results from the computational model support our hypothesis that patterned visual input (and possibly STDP) is the major mechanism of the development of neural synchrony during visual responses. Furthermore, the model also supports our experimental findings that visual activity may be sufficient for the development of increased spatiotemporal correlations which result from retinotectal map refinement. 4.9 Discussion Using rapid multi-cell Ca 2+ imaging in vivo we directly show, for the first time, how the large-scale functional properties of tectal networks emerge over development, such that their output becomes increasingly spatiotemporally organized and more reliable. We further show that visual activity is critical for the emergence of this functional organization of the visual system. We found that between developmental stages 46 and 48/49, during which time the retinotectal system undergoes an enhanced period of maturation and growth, neurons in the tectal circuit function with increasingly higher temporal correlation, faster responses, shorter response onset latency differences, less trial-to-trial variability, and higher spatiotemporal correlation, in response to a whole field visual stimulus. Increased correlation between tectal neurons is also observed during spontaneous tectal activity in an isolated brain, suggesting that, at least in part, the functional changes are due to refinement of local tectal circuits. These developmental changes require visual activity, since dark-rearing tadpoles between stage 46 and 48/49 prevents this functional maturation. Chronic NMDAR blockade also partially disrupts functional maturation of tectal circuit response properties. Taken together, these observations suggest that deve- 90

106 lopmental maturation of spatiotemporal properties of activity across the tectal network require visually-driven and, in part, NMDA-dependent plasticity. A simulation of the retinotectal circuit in which a STDP rule is imposed shows that this type of plasticity in combination with patterned visual input could be sufficient to organize tectal responses to light. One important question arising from these data is which parts of the retinotectal circuit are changing over development? On one hand it is known that there is substantial maturation and anatomical refinement of feedforward retinotectal synapses between stages 46 and 48/49 [24, 43, 61]. This refinement results in smaller tectal receptive fields and sharper spatial tuning of visually-driven behavior [47]. On the other hand, recurrent intratectal circuits are also known to change during this time, resulting in shorter and less variable visual responses [46]. The large-scale changes in correlated activity across tectal neurons observed here may result from both maturation and refinement of retinotectal and intratectal synaptic circuits, and the different circuits may contribute to different properties of the responses. For example, refinement of the retinotectal map, as well as maturation and strengthening of retinotectal synapses, increase drive from individual RGCs to a given tectal cell [146]. Thus, visual responses become more reliable, show increased temporal correlation with nearby cells, have a shorter response onset delay, less trial-to-trial variability, and higher spatiotemporal correlations. The development of local tectal circuits, on the other hand, may function to improve the relative timing between different tectal neurons, generating more coherent responses with higher synchrony. Moreover, recurrent intratectal connections may also contribute to retinotectal map formation and refinement by limiting spread of activity across the tectum and recruiting lo- 91

107 cal inhibitory circuits [45, 46]. While our data describing spontaneous activity patterns suggest that some refinement is occurring at the level of local tectal circuits, it is likely that the functional effects observed are a result of refinement at both sets of synapses. In our studies we found that visual activity is critical for the proper development of tectal network properties, but it remains unclear how activity mediates this developmental change. It is well known that patterned activity is important for the proper development of sensory systems across a variety of different species [107, 150]. One commonly held view suggests that activity plays an instructive role during development, in which cooperative pre- and postsynaptic activation result in the induction of synaptic plasticity such that coactive inputs will strengthen and stabilize, whereas non-coactive inputs will weaken and fail to become established [100, 102]. In the optic tectum plasticity in retinotectal synapses can be evoked by visual stimulation [29, 119], and this plasticity has been shown to require activation of NMDA receptors and to follow a STDP rule [ ]. Similarly, plasticity among intratectal synapses can also be evoked by patterned visual stimuli and also follows an STDP rule and requires NMDAR activation, although other receptor types may also mediate this plasticity [46]. Our data are consistent with the hypothesis that dark-rearing tadpoles between stages 46 and 48/49 blocks patterned visual activity and by extension plasticity in both retinotectal and intratectal synapses. This prevents activity-dependent refinement of the retinotectal circuitry and therefore interferes with the normal development of the functional network properties of the tectum. Blockade of NMDAR with MK-801 during this time period has a smaller effect due to the fact that multiple types of synaptic plasticity may be present which can compensate for the loss of NMDAR-dependent plasticity. 92

108 However, there are some limitations with our interpretation of the data. The first stems from the observation that dark-rearing did not have a marked effect on the developmental increase in neural correlation occurring during period of spontaneous activity in the isolated brain (Figure 4.8), in contrast to its profound effect on evoked responses. MK-801 similarly had a small effect on spontaneous activity, although there was an increase in the percent of cells showing zero correlation after MK-801 treatment. One possible interpretation is that refinement of local tectal synapses occurs in a largely-activity independent manner. However, the developing tectum is known to receive significant input from non-visual sensory modalities [147, 151], and the patterned activity generated by this mechanosensory input could be sufficient to organize spontaneous activity in local circuits. Thus the effect of dark-rearing becomes much more evident when the tectal network is activated by a visual stimulus. Furthermore, it is also known that plasticity among local tectal synapses can be evoked via several mechanisms [46], and therefore these may substitute for the absence of NMDAR activation. A second caveat could result from the possibility that dark-rearing primarily affects the retinal circuitry, and therefore would disrupt the synchrony of responses observed downstream in the tectum. In the mouse retina, dark rearing from birth disrupts the segregation of ON and OFF layers in RGC dendrites [154]. While we cannot fully rule out an effect of dark-rearing on the Xenopus tadpole retina, we do not think that this is likely occurring. In contrast to the mouse experiments, we begin dark-rearing at stage 46 (~9 days post fertilization), rather than from birth. This allows for significant development of the retina prior to dark-rearing. Furthermore, it has been shown that in Xenopus, dark-rearing does not affect the morphology of RGC dendrites [71]. Finally, we do not observe a difference in the 93

109 number of tectal cells preferentially responding to ON or OFF stimuli both over the developmental time window studied, and after dark rearing. A third issue that might affect our interpretation of the data, is that we are measuring responses to whole-field visual stimuli, therefore the amount of correlated activity is expected to be higher overall than if we had used spatially restricted stimuli. Nevertheless, the observed developmental effects on correlated activity are quite striking, despite the whole field stimulation. Using more spatially restricted stimuli, however, would probably allow us to better detect spatiotemporal correlations (Figure 4.7) as well as to detect the emergence of specific groups of cells with similar response properties. Future studies with spatiotemporally controlled stimuli will allow us to understand the temporal relationships between cells with similar response properties, giving us an even more in-depth view of tectal circuit organization. A final limitation is that within our recordings, we could not distinguish different tectal cell types. Although anatomical and electrophysiological studies suggest that the multiple cell classes present in the adult tectum have not yet fully differentiated at these early developmental stages [155], GABAergic inhibitory neurons are present and are known to play an important role in retinotectal system development [45] and might affect temporal properties of network. Although our results show responses from among all neural types in tecta, the overall wider distributions of some indices, such as temporal correlation, imply the possible effect of the developing inhibitory system, which helps differentiate activities of different parts of the tectum. Our findings are consistent with a range of studies indicating that precise timing of neural responses within a network is important for the network s function [48-50], and that a well-developed neural circuit should exhibit reliable spatiotemporal activity pat- 94

110 terns. In specific sensory systems, such as visual system, these temporal patterns of network activity encode important sensory information for example the moving velocity of a prey or predator and the output of the network can be used by downstream motor areas to generate behavioral responses such as prey catching or predator avoidance. Previous work has shown that the timing of recurrent activity within the tectum correlates with the ability of Xenopus tadpoles to perform a visual avoidance task [47]. This study shows that increased correlation between tectal neurons over development may be responsible for refining the temporal response properties of individual tectal neurons, and therefore important for the maturation of visually guided behavior. More importantly, we show that this maturation is highly dependent on visual activity. This is in contrast to other studies in the tectum of zebrafish embryos in which the development of the functional properties of tectal cells, including direction and stimulus selectivity, were not affected by dark rearing [141]. Similarly, dark rearing did not prevent matching of binocular inputs and response selectivity in individual tectal neurons in an experimentally rewired zebrafish optic tectum [140]. However, both of these studies use slow imaging speeds and do not look for correlations between neurons. Our data suggest that at least for the emergence of fine scale temporal correlations in the tectum, visual input plays a critical role. Our experimental paradigm of rapid Ca 2+ imaging in vivo will be useful for a number of future studies. For example, it will allow us to examine how a population of tectal neurons encodes behaviorally-relevant visual stimuli which elicit an escape behavior [47] and contrast these to behaviorally-neutral stimuli. This method should also prove useful for studying multisensory interactions between networks of neurons in the optic tectum 95

111 using either a reduced preparation [147] or a whole animal. In summary, our results show that the spatiotemporal properties of the retinotectal network become increasingly robust, organized, and less-variable over development. This increase in organization is driven by visual input through a mechanism which is consistent with spike-timing dependent plasticity. This increased organization is likely to result in optimized processing of visual stimuli and visually-guided behavior during development. 96

112 CHAPTER 5 10x10 MICRO-LED ARRAYS BASED IMAGE PROJEC- TION DEVICE AND ITS APPLICATION IN TADPOLE VISUAL SYSTEM DEVELOPMENT STUDIES 5.1 Introduction The family of nitride semiconductor materials that consist of aluminum nitride (AlN), gallium nitride (GaN) and indium nitride (InN) covers the bandgap E g range between 6.2 ev and 0.7 ev [156], and has the potential to produce light emitters with wavelengths ranging from the infrared to the UV. Nowadays, the application of high brightness nitride based light emitting diodes (LEDs) in biological science is increasing rapidly. In many researches, they are beginning to replace the conventional arc lamps and lasers. Compared to other high power light sources, the LED based light sources can achieve comparable power densities with much more compact size, much longer life time, and the lowest cost. Besides, their emission wavelength is tunable in a wide range by varying the indium or aluminum content in the active region, opening up the possibility of optimizing the power spectra for each application. The study of visual system development in Xenopus laevis tadpoles often requires simultaneous delivery of spatiotemporally patterned visual stimuli to the animal s retina and recording from target neurons in the optic tectum. The conventional measurement techniques involving the use of micro-lcd (liquid crystal display) and objective based image projection are limited by the maximal frame rate and the complexity of the optical 97

113 set-up [141, 151]. This chapter discusses our development of a nitride LED based microscale and flexible optical image projection device, and its application as a light source for visual stimuli in the research of tadpole visual system development. With the development of advanced semiconductor device processing techniques [ ], specialized arrays of microscale LED can provide a compact and versatile tool for biological applications. The design and fabrication of the gallium nitride based, multi-pixel and matrix addressable micro-led arrays shown in section 5.2, offer a portable means of projecting spatiotemporally patterned visual stimuli at short visible wavelengths (blue-green) with nice electrical and optical performance (section 5.3). In section 5.4, the packaging procedures of the devices are discussed. The image projection became flexible by integrating LED arrays to multicore image fibers. And a scalable matrix-addressing scheme enables timing-controlled electrical access to individual elements in the densely packed LED array. In section 5.5, we demonstrate the use of this array as a compact, programmable visual stimulator for measuring visual receptive fields (RF) of Xenopus tadpole s optic tectal neurons. Patterned visual stimuli were generated by the device and projected onto the eye through the output end of the image fiber, which can be easily and flexibly manipulated through conventional micromanipulators. We then simultaneously probed the response of optic tectal neurons in the tadpole brain by electrophysiology. The developmental changes in tectal RF structure have been detected 5.2 Micro-LED design and fabrication 98

114 Figure 5.1 The typical InGaN/GaN LED wafer structure. The LED arrays were fabricated from standard InGaN/GaN quantum-well (QW) p-n junction heterostructure wafers, grown epitaxially on sapphire substrates via Metal Organic Vapor Phase Epitaxy (MOVPE), with layering designs typical of high performance blue or green LEDs. The active region was composed of InGaN/GaN multi-quantum wells (MQWs) with the center wavelength at 470 or 530 nm. The details of the material structure are shown in Figure 5.1. The design and fabrication of the device was carried out so that individual emitters can be electrically addressed with sub-microsecond access time if needed. Based on the size of the tadpole's eye and the spatial resolution we need, the prototype device (Figure 5.2) was designed to consist 100 emitter pixel elements (10 columns and 10 rows) within 500 µm 500 µm area (a little bit bigger than the size of the tadpole's eye). The diameter of each circular emitter is 28 µm, with a center-to-center spacing of 50 µm between neighboring elements. The fabrication of the 10x10-element LED array involves standard lithography and etching steps with an additional deep etching step and a planarization step (Figure 5.3). 99

115 Figure 5.2 Design of the matrix addressable microarray LED. (A) Plannar layout of the design. (B) Cross-sectional view of the device along the direction of the etched trench. Prior to the device processing, a rapid thermal annealing (RTA) step at 750 C for 10 minutes in an N 2 ambient was applied to the wafer material to activate magnesium dopant for p-layer conduction. The sample was then deep-etched down to the sapphire substrate using chlorine based reactive ion etching (RIE) and a nickel based multilayer etch mask to form row trenches for electrical isolation between different rows of the nitride material. The mask consisted of 7 layers of titanium and nickel (Ti/Ni = 100Å/500Å for each layer), and was deposited using electron beam evaporation 7 times with a 30-minute annealing time between each layer deposition to release the thermal stress. After removing the metal mask residue, a shallow-etch (using the standard RIE etching technique and a silicon dioxide (SiO 2 ) etch mask) followed to expose the n-layer at the off-centered area of each nitride row, which were shared by LED elements on that row. Next, we coated the sample surface with a SiO 2 passivation layer using plasma-enhanced chemical vapor deposition (PECVD), and make circular openings for each emitter element. Semi-transparent p-contacts composing nickel and gold (Ni/Au: 50Å/50Å) were deposited on top of the apertures, and then annealed at 600 C for 5 minutes to form an ohmic contact. 100

116 Figure 5.3 Schematic of the fabrication Process. In order to create cross-trench electrical connections for emitters on the same column, the trenches need to be filled before metal line deposition. This planarization step was done using polyimide, a polymer that is widely used in semiconductor industry as an insulating material. First, a thick layer (> 5 µm) of polyimide PI-2771 (HD MicroSystems) was deposited by spin coating and cured at 350 C for 1 hour to form a solid layer that has a plain surface and covers the whole sample (including the trenches). The polyimide was then etched back carefully using oxygen plasma down to the surface of the SiO 2 passivation layer on nitride rows, leaving trenches filled with polyimide. This step was repeated several times in conjunction with AFM scans of the surface morphology, to ensure a surface discontinuity no greater than 2000Å. 101

117 Following the sample planarization, we removed the SiO 2 passivation layer of the n-layer-exposing areas of each nitride row, and deposited aluminum based n-contacts (Ti/Al/Ti/Au: 100Å/600Å/100Å/800Å) on top. Emitter elements on the same nitride row shared an n-contact. Finally, the cross-trench metal lines (Ti/Au: 100Å/1000Å) were patterned and deposited to connect p-contacts of elements on the same column. The device was thus matrix addressable by applying voltage across the corresponding p-metal line and n-contact for individual element controlling. 5.3 Micro-LED array performance Shown in Figure 5.4 are the luminescence-current-voltage (L-I-V) performance and the luminescence spectrum of a typical individual pixel emitter element in an array, where both the light output power density and the voltage are plotted as functions of the current density. We achieved low turn-on voltages (especially for green devices) and robustness of the device under high current injection on both blue and green devices. Other device parameters such as the external quantum efficiency, and tolerance to higher injection le- Figure 5.4 Performance of typical array elements. (A) The L-I-V curves, and (B) the luminescence spectra of blue and green devices. 102

118 vels (~ 1 ka/cm 2 ) were found to be similar to the single element devices of our previous works [157]. In our design, the elements on the same nitride row share the n-contact. Due to the finite conductivity of n-gan, there is an approximately 25 Ω sheet resistance difference between neighboring elements on the same nitride row. We resolved this imbalance by correcting it in the drive electronics so that the turn-on voltages and light outputs of different elements became quite uniform (difference <5%) across the elements. 5.4 Device packaging and control After fabrication, the device was packaged into a home-made housing and butt-coupled (Figure 5.5 (A), coupling efficiency ~ 4%) to a 3-foot long 30,000 pixel multicore image fiber (FIGH S from Fujikura Ltd.) with 600 µm image area diameter, 1.9 µm individual pixel diameter and 3.3 µm intercore distance, which can guide/preserve images from one end to the other end without additional optics. In recent years, various bio-imaging applications of high resolution multicore fibers have been reported by several groups [ ], indicating good image quality and low crosstalk. Compared to the conventional lens-based imaging systems, the multicore image fibers dramatically increase the flexibility and decrease the system size, which are extremely important for in vivo applications. In our experiment, 96 emitter elements were coupled to the fiber (4 corner elements cannot be coupled due to the size of the fiber), and spatio-temporal patterns of light from these elements at the output end of the fiber were observed (Figure 5.5 (B)). 103

119 Figure 5.5 Device packaging and control. (A) Schematic of the array-fiber butt-coupling. (B) Images at the other end of the multicore image fiber. top: patterns of LED array illumination; bottom left: an image of the array elements (no illumination) projected through the fiber; bottom right: images of the packaged device and the close look of the fiber end on the setup. (C) Schematic of the device control. To drive the device, we designed and built a matrix addressing circuit that controls the on/off of each emitter element based on the signals from a Digital I/O card (DIO-Card-6534, from National Instruments). A Labview program was custom made to send control signals for the generation of patterned illumination Figure 5.5 (C). Two scanning modes were available for patterned illumination: (i) line mode (in which elements are accessed row by row), and (ii) pixel mode (in which elements are accessed one 104

120 by one). Limited by the data transfer rate of the DIO card, the maximum scanning rate of a pattern was 10 khz, which is sufficient for a broad range of biomedical applications in providing high speed spatiotemporal photo-excitation patterns. With a faster circuit and control scheme, the devices are expected to run at a much higher speed, with fire times of each element on the order of a few nanoseconds, which can be applied in the future research for ultrafast microdisplay and imaging purposes. 5.5 Tectal visual receptive field mapping of developing tadpoles As stated in the previous chapters, the development of neural circuits is significantly regulated by their activity. This is particularly true in the visual system, where the onset of patterned visual activity has been shown to lead to refinement of inputs to a number of visual areas in the vertebrate brain. For amphibians, the optic tectum is the primary visual information processing area in the brain, and receives a topographically well organized projection from retinal ganglion cells (RGC) of the contralateral eye through the optic nerve. Initial development of the so-called retinotopic map (the projection from RGC to the optic tectum) is dependent on molecular cues, but further refinement and maintenance requires neural activity. During the activity dependent phase, competition between converging retinotectal inputs ultimately leads to strengthening of some neural circuits and elimination of others, resulting in a progressive refinement and stabilization of the retinotopic map in adult animals. This developmental refinement of the retinotectal projection has been extensively studied anatomically [43, 61, 107, 108], but a thorough functional characterization of this process during development has not been done. This is 105

121 Figure 5.6 Schematic of tadpole visual response recording. Left side shows the experimental set-up with patch clamp electrode on tectum and the multicore image fiber close to the animal s contralateral eye. The inset zooms in the fiber end, with one LED element on. The right side shows a recording trace of a tectal neuron s whole-cell current in response to a visual stimuli (blue region). For ON-response recording, the current within the stimuli period is integrated to get the total charge of synaptic input. of particular importance because it is unclear to what extent anatomical refinement matches the function. The most functionally precise map may not be the one that appears most anatomically refined. Furthermore, various manipulations which affect anatomical segregation of inputs may not always affect the functional properties of the map, due to rearrangement of synaptic weights. Here, we applied our 10x10 matrix addressable LED array based device as a new type of flexible microscopic projection system to deliver patterned visual stimuli directly onto the retina of developing (developmental stage 45-48) Xenopus laevis tadpoles to study of the development of functional retinotectal mapping. The retinal sensitivity is especially pronounced in this animal in the blue-green region of the spectrum, which matches the wavelengths of our devices. The experimental method and the set-up is shown in Figure 5.6. The projection end of the image fiber was positioned very close to the animal s contralateral eye such that the center-to-center spacing between the elements in the array was ~33-38 degrees of 106

122 arc to the optical center of the lens. Functional visual receptive field (RF) of an optic tectal neuron (i.e. a single neuron s response to stimulations from different retinal areas) was mapped by illuminating individual emitters for 1 second during whole-cell recording. Stimuli were presented at 6-second intervals, and every emitter was illuminated at least three times during the procedure. The visual response of optic tectal neurons was probed electrically via a standard electrophysiological patch clamp technique. Whole-cell currents resulting from the ON response were averaged over different trials and integrated over the stimuli period to get the total charge of synaptic input for every spatial location. Results were plotted as image maps, where the brightest color corresponds to the spatial location with the largest response (Figure 5.7 (A) and (B)). We defined the largest response as the peak of the RF, and the radius of the RF as the root-mean-square distance (RMSD) from the peak of RF, with the response amplitude Figure 5.7 Visual receptive field mapping of optic tectal neurons. (A) and (B) show typical visual receptive fields of two optic tectal neurons of a stage 45 tadpole and a stage 48 tadpole respectively. The measurements were done using 10x10 matrix addressable LED array based image projection device providing different patterns of photoexcitations. Each square represents a small portion of the retina, which was stimulated by a single LED element. Different colors refer different response strength (see the colorbar). Here only the central 8x8 elements are used. (C) The averaged width of visual receptive fields of stage 45 and stage 48 animals. The width of a tectal neuron's visual receptive field is defined as the root-mean-square distance from the peak of the visual receptive field, weighted by the visual response amplitudes. Tectal neurons of stage 48 tadpoles show more focused visual receptive field compared to those of stage 45 animals. 107

123 being the relative weight for the corresponding point. The results from different developmental stages showed that the RMSD decays significantly over the development (stage 45: 4.46 ± 0.34, n = 4; stage 48: 3.81 ± 0.16, n = 5; Figure 5.7 (C)), implying the refinement of the retinotectal map. In the future, combining this flexible microscale image projection device with the fast wide-field calcium imaging technique mentioned in CHAPTER 4 will allow us to achieve spatiotemporally organized visual stimuli of the retina and the simultaneous population recording of the tectal neuronal network, which enables the simultaneous measurement of the spatiotemporal visual receptive field of many tectal neurons for the study of their relationship and the general visual processing scheme of the tectum. 108

124 CHAPTER 6 OPTICAL NEURAL MODULATION AND OPTOGENETIC TECHNIQUE 6.1 Introduction Selective and reversible modulation of neural activity in the brain, besides population neural recording, is another fundamental goal of neuroscience. The active and direct impact to the inside of the system (e.g. the direct modulation of brain neurons) can switch the system between different internal states and therefore often provides much more information about its structures and functions than the regular modulation through the input pathways. Moreover, artificial modulation of the neural activity overpassing or substituting the nonfunctional or dysfunctional parts of the brain has potential medical applications as novel treatments of many neurological and psychiatric diseases. The first and the most obvious strategy which came early in the history of neuroscience: direct electrical stimulation of the nervous system. This is done via single electrodes or multi-electrode arrays (MEAs), and is still one of the major stimulation techniques. However, neural stimulation using electrical currents presents several challenges. It is difficult to achieve selective stimulation of only the targeted neurons without activating neighboring neurons. Even with multiple electrodes, the spatial resolution of the device is low. Second, the wide spread electrical current in the brain tissue causes many types of side effects, such as the direct heating of the tissue, which in many cases dominate the electrical stimulation effect [33]. Furthermore, electrochemical reactions at the 109

125 electrode-tissue interface may lead to electrode dissolution or tissue damage [ ]. And lastly, there is no way to introduce the bidirectional control (stimulation and inhibition) of the neural activity, as the electrical currents always tend to activate neurons. Moreover the technique also induces mechanical damage of the tissue for in vivo applications. The alternative idea of replacing the direct electrical current with the magnetic field generated current, as the result of the development of the electrodynamics in late 19 th century [166], lead to the development of the transcranial magnetic stimulation (TMS) technique [167], which non-invasively excite nervous systems in vivo. However, it inherits most of the drawbacks of the direct electrical stimulation method, such as non-specificity, low spatial resolution and significant side effects, as well as the high power requirements for the generation of the magnetic field. The application of the technique is thus limited to the macroscopic region. The photon, as another form of electromagnetic energy, can be spatially and temporally localized very easily within the brain tissue, as the tissue scattering and absorption of near-uv visible near-infrared (NUV-VIS-NIR) light is very small [168]. However, the light modulation of neural activity emerged much later than the above two methods, because most neurons (except the photoreceptors in the retina) in vertebrate brains are not light sensitive. Neurons possess ion channels that are directly activated by voltage, ligands, temperature, and mechanical forces, but not by light. Consequently, the light modulation of the neural tissue normally implies either the transformation of photon energy to other types of signals, or the import of exogenous light sensitive mechanisms into the system. Different optical neural modulation methods will be presented in the following 110

126 paragraphs Direct photostimulation Although the first successful neuronal photostimulation done by Richard L. Fork in 1971 [169] did not involve any signal transform process or exogenous light sensitive mechanisms, direct light modulation of neural activity is tough in general with a few exceptions. Neurons that are naturally photosensitive, such as the light sensitive abdominal ganglion of the marine mollusk Aplysia californica, can clearly be controlled photonically [169]. This however does not apply to other neural systems. Reported by Hirase and colleagues from Rafael Yuste s group at Columbia University, a high-power pulsed (average laser power > 100 mw, pulse width ~ 100 fs) laser beam that is tightly focused (spot size < 1 µm) on the neural plasma membrane can induce cell depolarization and action potentials via multiphoton processes [170] either the photochemical reaction that produced reactive oxygen species adjacent to the cell, or a transient perforation of the membrane that quickly re-sealed after the light discontinued. The similar effects can also be generated via one-photon processes, by a short-pulsed ultraviolet excimer laser with energies at the tissue-damage threshold [171]. In general, the direct photostimulation based on the high-power pulsed laser is usually on the energy level of tissue damage, and therefore is not safe for long term or in vivo purposes. The low-level (~ 1 J/cm2), pulsed infrared (~ 2 µm or 750 nm) laser light was shown by Jonathon Wells etc. to activate peripheral nervous systems [172, 173] via light induced thermal transients (3.8 C C) that directly or indirectly activate the transmembrane ion channels [174]. It however still faces the possibility of tissue overheating 111

127 and thermal damage Photo-modulation via caged molecules The caged molecules are inert molecules which are photosensitive, and can be transformed by photolysis processes into biologically active molecules once exposure to light. A caged molecule contains two parts: the biologically functional part, and a photoremovable protecting part (the cage ) that blocks the function of the first part. Photon absorption excites the molecule into an unstable intermediate state, which then relaxes by the split of two parts and the liberation of the active molecule [9] (Figure 6.1). The active part of the caged molecules can be either an agonist or an antagonist of receptors, and therefore is a convenient way to optically control biological processes [7, 37]. Every neuron has ligand sensitive ion channels (e.g. AMPA and NMDA glutamate receptors) that are controlled by specific ligand binding. Opening and closing of these channels can modify the activity of neurons. Photo illumination of neurons treated with caged molecules can thus modulate neural activity with spatial and temporal specificity. Many types of cellular signaling molecules can be caged. For instance, a widely used caged molecule is caged glutamate, which can release glutamate, the most common excitatory neurotransmitter, for local induction of depolarization and action potentials. Caged GABA, on the contrary, can be used to locally silence neurons, as GABA is an inhibitory neurotransmitter. Photolysis of caged molecules, as one of the major neuronal photo-modulation methods, holds several advantages [9]. The spatial resolution of the modulation or the locality of the photorelease process can be very high using the scaning fluorescence micro- 112

128 Figure 6.1 Photorelease of caged compounds. (A) CNB-glutamate is efficiently photolyzed to release glutamate with single-photon UV light. (B) MNI-glutamate is efficiently photolyzed with two-photon infrared light, enabling glutamate sub-micron spatial control and picoliter volume control (Figure reproduced from [7]). scope with high numerical aperture objectives. When 2-photon microscopy is applied, the photorelease size can be < 20 µm [37, 175]. The temporal precision of the modulation is usually very high, because the photorelease is a fast process ( s, [176]). Moreover, many different types of signaling molecules have been developed, allowing much more modulation modalities than neural stimulation and/or inhibition. The main drawback of the technique is the dependence of the exogenous caged molecules, which prevents most of the in vivo applications Photo-modulation via K + channels attached photoswitchable molecules The ion channels on the neuron plasma membrane can be controlled by attaching photo-isomerizable, tethered ligand molecules (photoswitches) to the channels [177]. Each of 113

129 these photoswitches contain three parts: a ligand part that binds to the channel, a photo-isomerizable part that is light sensitive, and a plug part that can block the pore of the channel. Photo illumination changes the conformation of the photoswitches and the corresponding positions of the plug parts to the pores of ion channels, which modify the channel conductance and the neural activity. Two main types of the photoswitches are the maleimide azobenzene - quaternary ammonium (MAQ), and the acrylamide azobenzene - quaternary ammonium (AAQ), both based on the photo-isomerizable molecule azobenzene (AZO) and the plug quaternary ammonium (QA). In the equilibrium state, the AZO is in its elongated trans form, with the attached QA blocking the channel pore. Shining UV light (380 nm) to the system causes the transformation of AZO into its bent cis form, so that the QA moves out from the pore and the channel conductance increases consequently. The AZO switches back to the trans form by either visible light illumination (500 nm) or spontaneous relaxation process (Figure 6.2 (A), (B)) [18, 177]. Different types of K + channels have been treated with these photoswitches. MAQ is mainly designed for a genetically engineered K + channel SPARK (synthetic photo -isomerizable azobenzene regulated K + channel). One mutant of SPARK has increased activity at typical resting potential, therefore shining UV light (channel unblocked) to the channel will induce neural inhibition (Figure 6.2 (C)) [177]. Another mutant of SPARK decreases the selectivity of the channel and inverts the effect of the photoactivation, such that unblocking by 380 nm light will depolarize rather than hyperpolarize the membrane potential (Figure 6.2 (D)) [178]. AAQ is mainly designed for the endogenous K + channels, so that no exogenous gene expression is needed (Figure 6.2 (E)) [18]. 114

130 Figure 6.2 Photoswitchable molecules modulation of potassium channel conductance. (A) Photoswitchable molecules consist of a photoisomerizable azobenzene group flanked by QA and a covalent attachment group (R). Exposure to 380 nm light isomerizes the azobenzene to its shorter cis form, whereas exposure to 500 nm light favors the trans configuration. The molecular structures of AAQ and MAQ are shown respectively. (B) Cartoon of the photoswitch control of a potassium channel. (C) Spontaneous action potential inhibition and revival of a cultured hippocampal pyramidal neuron (SPARK expressing and MAQ treated) by exposure to 390 nm (white bar) and 500 nm (grey bar) light. (D) UV light (white bar) elicits action potentials on a MAQ treated cultured neuron which is also transfected with the modified version of SPARK. The action potential bursts are halted by 505 nm (black bar) illumination. (E) Same experiment done on an AAQ treated normal hippocampal neuron. UV light is able to block the firing activity, and 500 nm light can reverse it (Figure reproduced from [18]). The dependence of the exogenous photoswitches and the slow photo response (open time ~ 1s [179]) are the main disadvantages of the technique Photo-modulation via neural expression of photosensitive proteins The photosensitivity of cells is based on the specific proteins, which change their properties under illumination. Genetic transfection and expression of the photosensitive proteins in neurons therefore provides photo-modulation strategies without the need of exogenous chemicals. Photoreceptors are light sensing cells in the retina, which transform light signals into 115

131 electrical currents. Expression of the light sensing part of the photoreceptors, the rhodopsin and its conjugated G protein, in brain tissues thus induces light sensitivity to neurons. As the first attempt, Zemelman et al genetically encoded Drosophila photoreceptor gene arrestin-2, rhodopsin, and the heterotrimeric G protein (ChARGe) into cultured hippocampal neurons in 2002 [180], which allowed the direct photostimulation of neural activity via G protein coupled receptors. Since the G protein process is slow (> 100 ms), the temporal resolution of this method is low (a few seconds to minutes), and the light evoked spikes occur in a stochastic manner. The following parts of this chapter will be devoted to the new generation of this strategy, the so-called optogenetic technique, which genetically encode photosensitive ion channels and/or ion pumps (originated from microorganisms) to neurons. Compared to other techniques, the optogenetics gives high spatial and temporal resolution, minimal invasiveness, and cell-type specificity. Section 6.2 gives an overview of the optogenetic technique. Section 6.3 shows our experimental results of optical stimulation of transfected acute mouse brain slices. 6.2 Optogenetics Light sensitivity (or vision) is important for animal survival. Vertebrates and invertebrates evolved different types of photoreception, both of which based on primary photoreceptor rhodopsin, a seven-transmembrane helix protein with covalently linked retinal. In vertebrates, the photo-activation of rhodopsin initiates a G protein-coupled enzyme cascade that ends in the hydrolysis of cgmp and the closure of the cgmp-regulated cation chan- 116

132 nels [181]. In many types of microorganisms however, the rhodopsins are light sensitive ion channels or ion pumps, and therefore directly control the photo-current with much faster photo response [182, 183]. The application of microbial-type rhodopsins in optical neuronal modulation was initially reported by Edward Boyden etc. in 2005 [10], in which the mammalian hippocampal neurons were virally transfected with channelrhodopsin-2 (ChR2), a light sensitive cation channel from green algae Chlamydomonas reinhardtii [181]. Blue light pulses had been demonstrated to be able to repeatedly and precisely trigger the neural spiking up to more than 10 Hz with less than 10 ms delay [10]. Moreover, the genetic encoding scheme can potentially provide cell type specificity, which exceeds that of glutamate uncaging and most other optical modulation techniques. In the past 5 years, this so-called optogenetic technique has involved many types of microbial-type rhodopsins with different activation spectra and cell type specificity in both in vitro and in vivo applications [13, ]. In the following sections we will focus on the most widely used photostimulation tool ChR2, and the corresponding gene delivery techniques Channelrhodopsin-2 Channelrhodopsin-2 (ChR2) is a light sensitive retinal binding protein found in the eyespot of the unicellular green algae Chlamydomonas reinhardtii. It is involved in the photoreception of the algae and leads to its phototactic behavior [181]. Reported by Nagel et al, ChR2 is most sensitive to blue light (Figure 6.3 (A)), and in most cases acts as a cation channel with low ion selectivity (Figure 6.3 (B)). The light activation of ChR2 usually causes membrane depolarization, because the reversal poten- 117

133 Figure 6.3 Light sensitive cation channel ChR2. (A) Schematic of ChR2. Blue light illumination opens the channel, and allows the entry of cations into the cell. (B) Action spectrum of ChR2. The most effective wavenlength is around 460 nm. (C) The photocycle of ChR2 and the correlation between different intermediate states. (D) Flashes of green light at different time intervals after the start of blue light illumination induce channel turn-off and the decrease of photocurrent. (E) Cultured hippocampal neurons expressing ChR2-YFP (scale bar 30 mm). (F) Blue light pulses (dots underneath the voltage traces) drive the ChR2 expressing hippocampal neurons to firing action potential trains (Figure reproduced from [10, 13, 14]). tial of the channel in most expressed cell types (e.g. oocytes from Xenopus laevis, and human embryonic kidney cell line 293, or HEK293) is around 0 V [181]. The single-channel conductance of this protein is about 100 fs with an open time of about 10 ms [14]. The detailed photocycle of ChR2 has also been studied, showing several intermediate states (Figure 6.3 (C)) [14]. Interestingly, the channels can be shut down by a flash of green light (Figure 6.3 (D)) [14]. In addition to its biological functions, ChR2 has emerged as an excellent tool for 118

134 neuronal photostimulation of neurons that are genetically modified to express ChR2 (Figure 6.3 (E), (F)) Gene delivery techniques for optogenetics Minimally three parts need to be transfected into neurons for regular optogenetic expression and functioning. The information of the photosensitive protein (e.g. ChR2) is encoded in a several kilo-base (or base pair) long DNA or RNA sequence. A promoter part (usually cut from the beginning of some cell-type specific expressed proteins) adding in front of the protein coding part is used to control cell-type specific expressions [179]. A sequence of fluorescence marker (e.g. Yellow Fluorescence Protein, or YFP) is usually pasted to the end of the protein coding part, so that the expressed proteins and cells are fluorescence tagged. Various techniques, such as viral vector assisted delivery [179], nanoparticle assisted delivery [201], and electroporation [202] have been applied to deliver the optogenetic sequences into different neural systems. Among them, viral vector assisted delivery is the most popular one, because of its high transfection efficiency (> 50% [10]), low chemical toxicity. Table 6.1 shows a summary of the characteristics of different viral vectors for gene delivery. Most of the existing optogenetic studies used lentiviral vectors, which delivers the genetic information directly into the chromosomal DNA of host cells via retrotranscription, and therefore causes permanent changes of the host cell s genome [1]. The pseudotyped lentiviral vectors are usually produced by co-transfection of 293FT cells (Invitrogen) with optogenetic gene (including the promoter, the coding for the protein, and the 119

135 Table 6.1 Summary of viral vectors for gene delivery in nervous system [1]. coding for the fluorescence marker) and lentiviral genes for retrotranscription and viral packaging using Lipofectamine 2000 [10, 203]. The vectors are harvested and concentrated after its reproductions in 293FT cells. In our experiments, the plasmid plenti-synapsin-hchr2(h134r)-eyfp-wpre DNA (from Karl Deisseroth s group at Stanford) was used for viral production (by Upenn Facilities). This plasmid contains both the ChR2 and the enhanced yellow fluorescent protein (EYFP) as the expression part. The synapsin promoter is used to drive the widespread expression in the cortex. 6.3 Photostimulation of ChR2 expressing acute mouse brain slices As a demonstration of viral transfection and ChR2 expression, postnatal day 12 mice (CD-1 Charles River) were anaesthetized by intraperitoneal injections of ketamine (70 mg/kg) and Dormitor (0.5 mg/kg) cocktail. The head of the animal was placed in a stereotactic apparatus (David Kopf Instruments) and the skin was cut to expose the skull. A 120

136 craniotomy was drilled (antero-posterior: 1.8 mm from bregma; lateral: 0.4 mm), and concentrated lentivirus solution (0.5 μl, ~ 10 6 IU/ml) was injected at the desired stereotactic locus (2.5 mm ventral from the brain surface in the craniotomy) at a speed of 0.05 μl/min using a programmable syringe pump (PHD2000, Harvard Apparatus). The skin on the skull was then sutured up and antisedan (0.5 mg/kg) was given to the animal after the surgery. A week after the lentiviral injection, the animals were sacrificed and 300 μm thick coronal cortical slices were prepared using a vibratome (VT 1000, Leica) in ice cold artificial cerebrospinal fluid (1X ACSF: 126 mm NaCl, 3 mm KCl, 1.25 mm NaH 2 PO 4, 2 mm MgSO 4 7H 2 O, 26 mm NaHCO 3, 10 mm Dextrose and 2 mm CaCl 2, equilibrated with 95% O 2 and 5% CO 2 ). Slices were kept at 32 C for half an hour and 24 C for another half an hour before recordings. A Nikon E600FN upright microscope with a 4X objective (N.A. = 0.4, Nikon Instruments Inc.), an YFP filter set (Semrock), and an Andor ixon 897 EMCCD camera (Andor Technology) was used for low magnification fluorescence images of the whole slice. A uniform EYFP fluorescent region of about 500 μm width was identified in layer II - VI of the prefrontal cortex of the injected mouse brain slices (Figure 6.4 (A)). Individual fluorescent cells were picked up under a 40X objective (N.A. = 0.8, Nikon Instruments Inc.) for electrophysiological recordings (Figure 6.4 (B)). During the recordings, cells were kept at 32 C (SH-27B inline solution heater, Warner Instruments) and visualized in an infrared differential interference contrast (DIC) mode. The whole-cell current clamp recording of neuronal activity was made using micropipettes (3-5 MΩ) filled with K-gluconate based intracellular saline (130 mm K-gluconate,

137 Figure 6.4 Photostimulation of ChR2 transfected mouse brain slices. (A) EYFP fluorescence in a coronal slice of the ChR2-EYFP transfected mouse. (B) 40X fluorescence image showing ChR2-EYFP expressing neurons. (C) Spiking activity of a ChR2 expressing neuron in response to a 1-ms-long blue laser pulse (2 mw). (D) The mean spike latencies (n = 10) of a patched neuron in response to weak, moderate, and strong laser pulses. mm HEPES, 4 mm KCl, 2 mm NaCl, 0.2 mm EGTA, 4 mm magnesium ATP, 0.3 mm tris-gtp, and 14mM tris-phospocreatine, ph 7.25, 291 mosm). Signals were measured with an Axoclamp 2B amplifier (Molecular Devices), digitized at 10 KHz using a Digidata 1440A A-D board (Molecular Devices), and acquired using p-clamp 10 software (Molecular Devices). Blue light from a 50 mw diode pumped solid state (DPSS) laser (CL , Crystalaser) was coupled to a multi-mode quartz fiber (core diameter: 1 mm, N.A. = 0.22, Till photonics GmbH) and guided to the illumination port of the microscope via a specially designed fiber coupler (Till photonics GmbH) to achieve a uniform blue light illumination on an area about 250 µm 250 µm for photo stimulation of 122

138 the neurons. We found that blue illumination from the DPSS laser was able to trigger action potentials of fluorescent neurons. As shown in Figure 6.4 (C) and (D), weak laser pulses ( 2 mw, 1 ms long) can elicit spikes on a ChR2 expressing neuron consistently (> 90% spike generation rate) with millisecond onset latency. We tested the neural responses under different laser power (2 mw, 10 mw, and 50 mw), and found that stronger pulses tend to give less spike latency compared to weaker ones (Figure 6.4 (D)). Neither cellular toxicity nor changes of physiological properties were observed in ChR2 expressing neurons. Moreover, no exogenous retinal was necessary for the functioning of ChR2. 123

139 CHAPTER 7 ALL-OPTICAL STIMULATION/RECORDING STUDY OF THE CORTICAL UP STATE OF MOUSE PREFRONTAL CORTEX IN ACUTE BRAIN SLICES 7.1 Introduction The understanding of complex neural system functioning, in many cases, requires simultaneous population modulation and population recording of multiple neurons activity. In the previous chapters, we have presented the optical recording (calcium imaging) and optical modulation (optogenetics) techniques respectively with their applications in the study of different neural systems. Here, by integrating the two techniques, we developed an all-optical modulation/recording scheme for the study of thalamic modulation of the prefrontal cortical persistent activities an example of large neuronal network dynamics. Different from fish, amphibians, reptiles and birds, the main information processing center in mammalian brain is the cerebral cortex, which is the outermost sheet of neural tissue of the cerebrum. In most advanced species (such as primates) this structure is highly folded to maximize its area and the number of neurons within the cranial cavity [17]. Anatomically the cerebral cortex can be divided into different regions according to their structures and functions (see Figure 7.1 (A)). The posterior part of the cortex is mainly the sensory cortices that receive and process sensory information from different modalities (for instance, the vision, audition, and touch) and send efferent axons to other 124

140 Figure 7.1 Cerebral cortex of the brain. (A) Cartoon showing different regions of the brain ( (B) Thalamus as the input and output relay for many cortical regions ( (C) Nissl stained cortical sections, with layers indicated [16]. (D) Summary of the major connections to and from different cortical layers (Figure reproduced from [16]). parts of the cortex for further processing. The anterior part is mainly the motor areas (e.g. motor cortex, premotor cortex, etc.) and the association areas (e.g. prefrontal cortex, etc.); which are involved in motion control, decision making, memory and other cognitive functions based on the information input from the sensory part. Besides the heavy interconnections between different cortical areas, the cortex is also connected to various subcortical structures, such as the thalamus and the basal ganglia. Particularly, most cortical areas connect to thalamus, which serves as a relay of most sensory input (except for ol- 125

141 faction) and motion output (Figure 7.1 (B)) [17]. Using cell staining techniques (for example, Nissl staining), the vertical structure of the cortex can be laminated into six layers according to the distribution of neural cell types and their synaptic connections with other cortical and subcortical regions (Figure 7.1 (C), (D)). Layer I (the molecular layer) is the uppermost layer, and almost neuron free. Layer II and III contain medium size neurons with moderate density, and project densely to other cortical areas (mainly to layer III and IV). Layer III also receives input from the thalamus. Layer IV is densely packed with small pyramidal neurons, and receives most of the thalamic input. Layer V contains large pyramidal neurons, and projects to many subcortical structures. Layer VI has a wide range of cell sizes and shapes, and mainly sends efferent axons to the thalamus [16]. More than 100 types of neurons have been recognized in the brain according to their morphological and physiological properties [16]. These types can be coarsely divided into two categories based on their dendritic structures and synapse types: spiny neurons that make excitatory glutamate synapses (excitatory neurons), and aspiny neurons that make inhibitory GABA synapses (inhibitory neurons) [16]. The excitatory neurons can further be classified as two physiological types: the regular-spiking (RS) neurons which fire low frequency action potentials (< 50 Hz) with adaptation in response to a super-threshold current pulse (Figure 7.2 (A)) [16], and the intrinsically bursting (IB) neurons which respond to a threshold stimulus with a high frequency burst (150 Hz 300 Hz) of 3 to 5 spikes (Figure 7.2 (B)) [16]. The inhibitory neurons can also be classified as two physiological types: the fast-spiking (FS) neurons which can fire consistently at high frequencies with brief action potentials (<0.5 ms, Figure 7.2 (C)) [16], and the low -threshold 126

142 Figure 7.2 Typical firing patterns of different cortical cell types. (A) Regular-spiking excitatory neuron. (B) Intrinsically bursting excitatory neuron. (C) Fast-spiking inhibitory neuron. (D) Low -threshold spike neuron (Figure reproduced from [16]). spike (LTS) neurons that have low-threshold Ca 2+ current and lower firing frequencies (compared to FS neurons) (Figure 7.2 (D)) [16]. Different types of cortical neurons and their synaptic connections, as well as the inputs and outputs from other parts of the brain form the complex cortical neuronal network that is the structural basis of all cortical functions. Our all-optical modulation/recording technique allows us to study multiple neurons activity of the cortical network with high spatiotemporal resolution and the least invasiveness. In this chapter, we focus on the prefrontal cortex, which is believed to be a key part involved in many high-order cognitive 127

143 functions, such as working memory. The phenomena and the existing models of the working memory are reviewed in section 7.2. Our all-optical stimulation/recording set-up is presented in section 7.3. We discuss some of our experimental data and the future about the project of thalamic modulation of prefrontal cortical UP state in section Persistent activity, cortical UP state and the mechanism of working memory Working memory (WM) is an important cognitive function of mammalian brain, which enables the animal to instantly keep/modulate the goal-related information (from various sensory modules) for a short period of time (several seconds) to guide forthcoming actions [25, 27]. Based on the lesion, in vivo recording, and local inactivation studies of Figure 7.3 Monkey working memory test. (A) Experimental procedure: successive frames illustrate the sequence of events in the oculomotor delayed-response task. Trials begin with the appearance of a fixation point at the center of the screen, which the monkey is required to focus to throughout the trial. A spatial cue is subsequently presented, typically at one of eight locations (left). After a delay period of a few seconds, the fixation point is turned off and the monkey is required to indicate the location of the cue by moving his eyes accordingly on the screen. (B) Recordings from a single prefrontal neuron during the execution of the oculomotor delayed response task show persistent discharges in the delay period. Discharges are arranged as to indicate the location of the cue. The neuron is mostly active during the delay period following presentation of a stimulus in the upper left (135-degree) location [25]. 128

144 many different animal species in the past few decades, the prefrontal cortex (PFC) is believed to be the key brain structure that is linked to WM. In particular, it has been clearly shown that the elevated persistent neural activities of the PFC, following a briefly presented cue, are correlated to the information maintenance during the delayed period of working memory tests, and are the basis to guide correct responses at the end of the delayed period (Figure 7.3). Instead of the neural rewiring/growth mechanism of long term memory, the neural basis of PFC persistent activity and WM is thought to be related to specific types of collective neural activities. The simplest idea states that persistent neural activity is maintained in the PFC network through the recurrent excitation of PFC neurons [27], while the discrete memory items are represented as the synaptic weight matrix of the network and stored/retrieved as fixed-point attractors of the activation dynamics [204]. In this model, a memory item is encoded as an assembly of neurons that are wired together reci- Figure 7.4 A simple recurrent network model to generate delay-period persistent activity. (A) Structure of the model. Two cell assemblies (green and yellow boxes) coding for two different objects are embedded in the symmetric synaptic weight matrix. Neurons within the same cell assembly are connected reciprocally by high synaptic weights (w + ), whereas neurons not belonging to the same assembly are connected by low synaptic weights (w ). Single neurons might participate in more than one cell assembly. In addition to these local recurrent excitatory connections, there is a global feedback inhibition (IN) driven by input from the excitatory neurons that allows only one pattern to stay active at a time, and an external afferent input (I aff ) to each unit in the network. Details of the model can be found in. (B) Activity of a model neuron after the initial afferent input shows sustained activity, which can be cleared by an inhibitory input (Figure reproduced from [27]). 129

145 procally by strong excitatory synaptic weights, whereas neurons that participate in different memory items have weak connections or even mutually inhibit each other [27]. An external input might trigger the activity of one neural assembly and initiate the recurrent excitations, which then maintain the activity of the assembly without the need of additional input. At the same time, other cell assemblies may be inhibited by the active one, forming a stable state of the network activity patterns (Figure 7.4). A working memory would correspond to the activation and maintenance of one of these synaptically stored patterns [27]. The structure of the synaptic connection matrix can further be specified in different memory tasks. For the continuous-valued memory tasks, such as the spatial working memory task, a topographic network structure is needed, where the neighboring neurons are wired and fires together to encode neighboring spatial locations [27]. This idea was first proposed by Hugh Wilson and Jack Cowan in 1973 [205] and remodeled by David Pinto and G. Ermentrout [4, 206] later based on a mean field theory of neuronal networks. The derivation starts from a simple firing rate model dui cm = gu m i + wi (, j) f( U j) (7.1) dt where U indicates the time averaged membrane potential, f( U ) representes the neural firing rate as a function of the mean membrane potential, and wi (, j ) is the synaptic connection strength between neuron j and neuron i. For a large neuronal network, discrete neurons can be simplified as a continuous neural media (Figure 7.5 (A)), so that + j 1 uxt (,) = uxt (,) + dywy ( ) f ux ( yt,) α t (7.2) where u and w become continuous functions of space and the homogeneity of the media 130

146 Figure 7.5 Field theory of large neuronal networks. (A) A large one dimensional neuronal network containing many excitatory and inhibitory neurons can be simplified as a one dimensional neural field u(x). (B) The combinatorial synaptic connection strength w for lateral inhibition scheme [4]. is assumed. For a network containing both excitatory and inhibitory neurons, eq. 7.2 can be rewritten as 1 uxt (,) u(,) x t = dywee ( y) fe u( x y,) t dywie ( y) v( x y,) t αe t 1 vxt (,) + + v(,) x t = dywei ( y) fe u( x y,) t αi t (7.3) with w ie and w ei indicating the inhibitory-to-excitatory and the excitatory-to-inhibitory connections respectively. Here we assume that the inhibitory firing rate is assumed to be linear to their membrane potentials, and neglect the inhibitory-to-inhibitory connections [4]. Eq. (7.3) has a formal solution as u = η [ w η w w ] f u (7.4) e ee i ie ei with η e and η i indicating the Green s functions of the first and the second equations of Eq. (7.3) respectively, * representing the spatial convolution, and being the convolution involving both the space and time [4]. A further simplification can be made by if the inhibitory neurons respond fast ( αi ), and 131

147 u = η [ w w w ] f u = η w f u (7.5) e ee ie ei e with w the combinatorial synaptic connection strength. For a lateral inhibition scheme where the extent of excitatory connections (w e ) is broader than that of inhibitory connections (w i ), w is positive (excitation) for small distance and negative (inhibition) for long distance (Figure 7.5 (B)). In this condition, standing pulse solutions exist on the neural media for a certain firing rate function f (one stable and one unstable), and part of the media can sustain at a high firing rate state without external maintenance. Figure 7.6 shows an example, where w is chosen to be y /2 1 y /4 wy ( ) = e e (7.6) 2π 2 π Figure 7.6 Standing pulse solutions of a lateral inhibitory neural field. (A) The shape of neural fields of two standing pulse solutions for θ = (B) Pulse width and stability for different θ values. (C) and (D) shows the evolution of the neural field after a square pulse input at time t = 0. The system eventually evolves into a standing pulse state. 132

148 and the firing rate function is a heavy-side step function fu ( ) = Hu ( θ ) (7.7) with H(x) = 0 for x < 0, and H(x) = 1 for x > 0. If θ is small enough, such that 1 Φ Φ θ < max{ ( erf ( ) erf ( ))} (7.8) two standing-pulse solutions exist for eq. (7.3), with the wider one being stable. So far, we consider the PFC persistent activity is solely due to the local network structures. This mechanism might not be sufficient however, because other parts of the brain were also reported to sustain at high firing rate in the WM tasks (Figure 7.7) [20, 53]. Specifically, the posterior parietal cortex (PPC) and the medial dorsal nucleus (MDN) of the thalamus, just like PFC, increase their firing rate when a stimulus is first presented, and keep the high frequency firing during the delay period. The caudate nucleus (CD) of the basal ganglia only shows high firing rate at the delay period, while the globus pallidus Figure 7.7 Activity of different brain areas during the delayed response tasks. Data are collected from single-cell recording of monkeys in a variety of delayed response tasks (Figure reproduced from [20]). 133

149 (GP) of the basal ganglia responds to the delay period by decreasing the activity. This supports a distributive scheme of WM, in which the neural activity maintenance during the delay period involves multiple parts of the brain (such as cortex, thalamus, and basal ganglia), each serves for different purposes (Figure 7.8). PPC is the high level visual cortex that projects to PFC as a sensory input, which also receives excitatory afferent from layer II/III neurons of PFC, and is thought to form an excitatory loop with PFC during the delay period. MDN is the main thalamic input of PFC layer IV neurons, which also receives afferent projections from layer VI neurons of PFC and form another excitatory loop with PFC during the delay period for persistent activity maintenance. CD neurons of basal ganglia receive excitatory input from PFC layer V neurons, and send their axons to inhibit GP neurons, which always inhibit the MDN. In this model, the MDN plays a very important role such that without its activation (or the disinhibition by GP), the excitatory loops are not strong enough to sustain the high frequency neural activity in PFC. In the rest of this chapter, we are trying to experimentally verify the role of MDN in Figure 7.8 The distributive model of working memory. (A) Schematic of connections between different brain parts. (B) PFC layer structure with connection to/from other parts of the brain (Figure reproduced from [20]). 134

150 Figure 7.9 3D structure of medial dorsal nucleus axon projection to prefrontal cortex in mice. "ACC" stands for anterior cingulate cortex, part of PFC. "Hippo" stands for Hippocampus, "Th" stands for thalamus, and "St" stands for striatum (Figure reproduced from [3]). the PFC persistent activity or its analog in acute brain slice preparation the cortical UP state, in which neurons fire collectively with high frequency (>10 Hz). The complex 3D axonal pathway of MDN projection in mice (Figure 7.9) [3] makes it extremely difficult to keep the MDN and the PFC neurons on the same mouse brain slice with the functional projection for stimulations. Optogenetics however, allows us to selectively stimulate MDN axon ends on the PFC containing slices without the need of MDN cell bodies. Furthermore, simultaneous calcium imaging allows us to record and compare the activity of multiple PFC neurons (> 50) which might reveal important clues of PFC circuit structure and the information coding in working memory tasks. 7.3 All-optical stimulation/recording of neural activity in mouse brain slices The main components of the experimental set-up (Figure 7.10 (A)) include a Nikon 135

151 Figure 7.10 All-optical modulation/recording scheme. (A) Experimental set-up. (B) Transmission spectra of the dichroic filter above the objective in comparison with that of indo-1 emission filter, the excitation spectra of Polychrome 5000 for indo-1 ratiomatric imaging, the excitation wavelength of 470 nm DPSS laser, and the wavelength of 593 nm DPSS laser (for the possibility to activate halorhodopsin for neural inhibition in the future). E600FN upright microscope with electrophysiology components (listed in the last chapter), a customized fluorescence filter set (Semrock, Figure 7.10 (B)), a monochromator/xenon lamp based light source with fiber coupled output (Polychrome 5000, Till photonics GmbH) for calcium indicator excitation, a 50 mw DPSS laser (CL , Crystalaser) with fiber coupled light output (quartz fiber, core diameter: 1 mm, N.A. = 0.22, Till photonics GmbH) for photostimulation, and an Andor ixon 897 EMCCD camera for calcium imaging. Both light sources are coupled to the microscope via a dual-port fiber coupler (Till photonics GmbH), with their illumination being controlled independently. In this experiment we use a short wavelength calcium dye indo-1 AM (Invitrogen), which is excited by UV (340 nm/380 nm) and emits blue light (415 nm). Since ChR2 is not sensitive to UV light and 470 nm laser illuminations do not affect fluorescence emission at 415 nm, this wavelength difference allows simultaneous and independent optical stimulation and recording. The commercial imaging software Metafluor (Molecular Devices) is used in synchronization with p-clamp 10 software (Molecular Devices) to control all equipments and collect calcium imaging data. 136

152 Figure 7.11 Photostimulation of ChR2 expressing MDN afferent axons. (A) low magnification microscope image of EYFP fluorescence on the PFC part of a mouse brain slice, maximal intensity appears at layer III and IV (scale bar: 500 µm). (B) High magnification image showing the fluorescent axons. (scale bar: 50 µm). (C) and (D) Patch clamp recording of two PFC neurons in response to 100 ms blue laser pulses. Postnatal day 12 mice (CD-1 Charles River) were injected (antero-posterior: mm from bregma; lateral: 0.5 mm, ventral: 3.3 mm) with virus (0.7 μl, ~ 10 6 IU/ml) to express ChR2. A week after the lentiviral injection, the animals were sacrificed and 300 μm thick coronal cortical slices were prepared. EYFP fluorescent axons were seen in the PFC area, with a maximal fluorescent intensity at layer III and IV (Figure 7.11 (A) and (B)), indicating the expression of ChR2 on the membrane of the MDN afferent axons. Using the same procedure mentioned in the last chapter, individual fluorescent PFC neurons were picked up and electrophysiologically recorded to test their photosensitivity. Blue laser light was shown to be able to activate PFC neurons via MDN afferent axons (Figure 7.11 (C) and (D)). For calcium imaging, Indo-1 AM was dissolved at a concentration of 4mM in DMSO with 10% pluronic acid (Invitrogen) and diluted to a final concentration of

153 Figure 7.12 Calcium imaging of mouse PFC brain slices. (A) A fluorescent image of indo-1 loaded mouse PFC brain slice (scale bar: 50 µm). (B) Simultaneous optical and electrical recording of a PFC neuron. µm in ACSF. 100 µl dye solution was dropped into a 0.22 µm syringe filter (Millex-GV sterilizing filter unit, Millipore), which was then fixed to a 1 ml syringe filled with air up to about 0.8ml. After connecting the syringe filter output to a flexible needle (MicroFil 34G, World Precision Instruments), we slowly applied pressure to fill a patch-pipette with dye solution to about ¾ its volume. This pipette was then fixed on the micromanipulator and moved to the surface of the slice in the recording chamber. A picospritzer (General Valve Corporation) was used to puff the solution into the tissue (5 psi, 20 minutes) for dye staining, and the slice was kept in the circulating ACSF for 30 minutes to wash away the excess dye. A substantial number of PFC neurons within 400 µm diameter area were stained (Figure 7.12 (A)). Using the same electrophysiological set-up, individual stained PFC neurons were patched and electrically stimulated and recorded to test the sensitivity of the calcium indicator. Fluorescence signals were shown to increase in response to every action potential (Figure 7.12 (B)). 138

154 7.4 Photostimulation of MDN axon projection for the study of cortical UP state in acute mouse PFC slices Applying the all-optical stimulation/recording technique to acute mouse PFC slices with ChR2 expressing MDN axons, we are studying the effect of MDN input on the activity of PFC neurons. The optical recording of PFC neurons in acute brain slices without stimulation gives two types of neural activities (Figure 7.13 (A)). For the type of activity, most PFC neurons generate big calcium bumps, indicating synchronized high frequency firing activity. For the second type of activity, only a small part of PFC neurons show tiny calcium bumps while the rest of neurons are quiet. The first type of activity represents the cortical UP state in the electrophysiological recordings, in which the membrane potentials of cortical neurons get depolarized and cells tends to fire more frequently than normal. The second type of activity, on the other hand, represents the sparse spontaneous firing of individual neurons in the cortical DOWN state, in which the membrane potentials of cortical neurons are low and cells tend to be quiet. Isolated from other parts of the brain (e.g. MDN, PPC, etc.), the cortical neurons in acute brain slices switch spontaneously between these two states. Previous works in acute brain slices indicates that the recipe of the ACSF can affect the activity of neurons [207]. Cells tend to be quiet in a high [Ca 2+ ] and high [Mg 2+ ] environment. The conventional ACSF recipe that we used in the previous recordings contains more Ca 2+ and Mg 2+ (CaCl 2 : 2 mm, MgCl 2 : 2 mm) than the real cerebral spinal flu- 139

155 Figure 7.13 Ratiomatric calcium imaging of multiple PFC neurons. (A) Different types of spontaneous activity of PFC neurons. (B) A brief photostimulation (blue bar) triggers synchronous neural activity in PFC. id (CSF) in the brain, and therefore tends to hold neurons on the DOWN state. By lowering the concentration of calcium and magnesium ions in the ACSF (CaCl 2 : 1 mm, MgCl 2 : 1 mm, KCl: 4 mm), we mimic the real CSF and increase the excitability of cortical neurons, allowing the cortex to stay on UP state for longer time. We found that a brief photostimulation (1 ms pulse, 50 mw) of layer III/IV of PFC slices with afferent ChR2 expressing MDN axons always induces big fluorescence signal bumps of almost all recorded PFC neurons in layer III/IV (Figure 7.13 (B)), indicating synchronous neural activities similar to cortical UP states observed in the spontaneously activated cortical slices. MDN axon input therefore might be a mechanism for cortical UP state and persistent neural activity initiation. There are still a lot of unknowns in this field, and many of them will have to be discovered with in vivo preparations. In the future, we are planning to optically modulate 140

156 and record the neural activity of multiple brain areas (PFC, PPC, MDN, basal ganglia, etc.) simultaneously in behaving animals, which will help us better understand the neural mechanism of working memory. Moreover, the combination with genetically encoded calcium indicators or voltage-sensitive indicators would allow a full-genetic targeting scheme, without the need of exogenous chemicals (such as calcium sensitive dyes) in preparations. 141

157 CHAPTER 8 Conclusion In this thesis, we have presented the design and development of optical recording and modulation tools and their applications in the study of large neuronal network dynamics. As composed of enormous number of interconnected neurons, the brain develops and operates in a super-complicated manner, which results in various animal behaviors, cognitive functions, and even consciousness in human beings. The discovery of the secrets of this complex dynamical system relies on novel techniques that can record and modulate the activity of this system with cellular and subcellular resolution. The calcium imaging technique serves as a noninvasive way of recording multiple neurons activity in a large neuronal network with millisecond temporal resolution. We applied this technique into the study of the tadpole retinotectal system development in CHAPTER 4, and discovered that the activity of multiple optic tectal neurons become temporally more correlated over the development, which was then proved to involve activity dependent mechanisms and NMDA receptor mediated pathways. This effect cannot be studied in a conventional electrophysiological way, since it requires simultaneous recording of multiple neurons with high temporal resolution. As a novel photostimulation light source, a matrix-addressable micro-led based image projection device and its application in tadpole optic tectal neurons visual receptive fields mapping were presented in CHAPTER 5. In the later chapters of this thesis, the genetics based optical modulation technique (optogenetics) was discussed. Genetic encoding of photosensitive ion channels and ion 142

158 pumps into neurons allows photostimulation and/or inhibition of neural activity with cell type specificity. By using a shorter wavelength calcium indicator, we achieved an all-optical stimulation and recording scheme, where optical stimulations and optical recordings can be done simultaneously and independently, with high spatial and temporal resolution. By applying the all-optical scheme into the study of MDN-PFC interactions, we found that the MDN input can initiate the cortical UP state of PFC neuronal networks in acute brain slices. Currently, the development of the advanced molecular biology techniques has created the possibility of optically recording and modulating neural activity in a genetically targeted way without introducing exogenous toxic chemicals. Meanwhile, the development of solid state lighting/imaging techniques (e.g. LED, semiconductor lasers, high speed cameras, etc.) has enabled faster and more precise (high spatial resolution) imaging and illumination with more compact size. The combination of strengths from both sides is likely to generate a revolution in neuroscience, which will lead to a deeper understanding of the brain. 143

159 SUPPLEMENTARY CHAPTER Combined Topographical and Chemical Micro-patterns for Templating Neuronal Networks This chapter discusses my early project on cultured neurons about creating topological neuronal networks to fit optical devices, which is presented based on the published paper [208]. S.1 Introduction Recently, optical stimulation and recording techniques [7, 10, 177, 209] opened the possibility of a non-invasive manner of probing neuron activities. Patterned assembles of light sources such as light emitting diode arrays[8] further inspired ideas of interaction between neuronal networks and optical devices. The availability of such technique makes it possible to study the functioning of neuronal network and the relationship to its structure. Two-dimensional neuronal networks grown on patterned templates have been investigated previously [ ]. Yet, there are significant challenges of these patterning schemes. One is that the neurites grow across patterned channels with typical width 3-10 µm [212, 213], which tend to misguide the neuron off the pattern. The narrowness of the channels may be one reason for the increasing mechanical tension on the neurites. Even if major fraction of neurons stays within the pattern, there is usually no definite connectivity between the neurons [216, 217]. Although synaptic connectivity can be detected using 144

160 multiple patch clamp technique [218], little is known about the relationship between geometrical structure and synaptic function. Furthermore, there is a technical limit in the number of patch clamps which can be applied in studying multiple neuron networks. However, optical imaging techniques are readily available for tracking neural activities. Topographical control is necessary for better geometrical guidance. Recent study [221] has used semiconductor based caging wells for neurons, but the low transparency of the device in visible range makes it incompatible with optical analysis. It has also been shown that micro pillars as obstacles with different sizes affect the orientation of neurite growth, but not the creation of neuronal networks [222]. Topographical manipulation of invertebrate neurons [223, 224] is much easier than vertebrate neurons due to the vulnerability of the later. Recent publications show early proof-of-principle experiment of integrated neuromicroelectronic interfacing devices, which await for further development [219, 224]. Achievement of integrated bio/electronic interface devices is based on the ability to construct a few neurons network with geographically predetermined connections [ ]. Here, we test the idea of combining patterning by topographical and chemical templates. In brief, we use microelectronic fabrication technology by which the hybrid template is processed in two distinct steps. First, the biocompatible polymer SU8 is patterned using photolithography techniques creating 5 µm high barriers. Then one layer of poly-l-lysine is patterned and aligned with the SU8 polymer template using photolithography techniques, providing both promotion and inhibition cues for neuron growth patterning. This hybrid template demonstrates good control over neuron connectivity and compatibility with integration and probing of micro/optoelectronic on-chip tools. 145

161 S.2 Material and methods Negative photoresist SU8-5 (Microchem) and SU8 developer (Microchem) were used as received for negative photolithography, as well as positive photoresist S1818 (Shipley) and developer MF312 (Shipley). Poly-L-lysine solution (MW = 30,000-70,000, Sigma) were prepared in Borate buffer with filtration of 0.22 µm. The substrate was a 12 mm diameter glass coverslip. Negative photoresist SU8 was spun coated onto the glass substrate at 500 rpm with a ramp of 100 rpm/s for 5 seconds and at 3000 rpm with a ramp of 300 rpm/s for 40 seconds. Followed was soft bake at 65 C for 1minute and 95 C for 3 minutes. The substrate was covered with a photomask with desired pattern and exposed to UV light at CI mode with 5.4 mw power for 12 seconds. The geometrical design for photomask with two different dimensions is shown photolithography SU8 Glass slide 5μm SU8 20/40μm Spin coating S /100μm 200/400μm photolithography neuron Neuron plating liftoff Poly-L-Lysine Figure S. 1 Fabrication procedure and photomask geometry for hybrid template. 146

162 in Figure S. 1, composed of circular nodes and interconnecting channels. The channel width and node diameter are designed such that there is reduced restriction for neurite extension. Post-exposure bake was conducted at 65 C for 1 minute and 95 C for 1 minute. The substrate was then developed in SU8 developer for 1 minute after cooled down, rinsed in isopropanol solution and blown dry with nitrogen. The SU8 template fabrication was finished with hard bake at 165 C for 5 minutes. A second layer of positive photoresist S1818 was then spun coated onto the SU8 template at 5000 rpm with a ramp of 5000 rpm/s for 30 seconds followed by soft bake at 110 C for 2 minutes. A reverse photomask was used for UV exposure at CP mode for 48 seconds. Careful alignment between the photomask and the SU8 template was necessary. Subsequently, the substrate was developed in a mixture of MF312 and DI water (1:1 volume) for 40 seconds, rinsed in DI water and then blown dry. The entire substrate was soaked in filtrated poly-l-lysine solution (1 mg/ml) overnight. The remaining S1818 was then lifted off by acetone, left with a monolayer of patterned poly-l-lysine aligned with patterned SU8. The template was rinsed in DI water and sterilized with 100% ethanol for 1 minute. Prior to cell plating, the substrate was soaked in DMEM (Dulbecco s Modified Eagle Medium, Gibco) with 10% FBS (Fetal Bovine Serum, Gibco) and 1% penicillin/streptomycin (Gibco), and kept in an incubator (5% CO 2, 37 C) for approximately 2 hours. We followed procedures of dissecting rat hippocampal neurons as described previously [229]. The tissue of hippocampus was removed from E18 rat embryos, trypsinized (0.25%) at 37 C for 15 minutes and dissociated by triturating in HBSS solution (Hank s Balanced Salt Solution (Gibco) with 1% HEPES and 1% Penicillin-Streptomycin). Digestion was terminated by changing HBSS solution to DMEM/FBS 147

163 solution. Neurons were plated into 24-well culture dish at a density of 10,000 cells/cm 2 and placed inside the incubator. After 3 hours, the DMEM/FBS solution was changed to serum free medium containing Neurobasal media (Gibco), 2% B27 Serum free supplement, 1% Penicillin-Streptomycin and 0.25% Glutamax and fed twice a week. Control samples were cultured on acid -washed glass coverslips coated with Poly-L-lysine (0.3 mg/ml) with an average plating density of approximately 10,000 cells/cm 2. Whole-cell patch clamp recordings were performed using a setup including Axoclamp 2B amplifier (Molecular Devices), micromanipulator units (Sutter Instruments) and Nikon E600FN phase contrast microscope. The recordings were performed at room temperature in extracellular medium, which contained 145 mm NaCl, 3 mm KCl, 3 mm CaCl 2, 1 mm MgCl 2, 10 mm Glucose, and 10 mm HEPES. The osmolality of the medium was adjusted to 315 mosm/l using 1 M KOH solution. Before the experiment, the extracellular medium was pre-warmed. Borosilicate micropipettes (Sutter Instruments) were pulled using a micropipette puller (Sutter Instruments) with parameters set for pipette resistance of about 4 MΩ. Each pipette was backfilled with intracellular medium, which contained 9 mm NaCl, 17.5 mm KCl, 0.5 mm CaCl 2, 1 mm MgCl 2, mm K-Gluconate, 10 mm HEPES, and 0.2 mm EGTA. The osmolality was adjusted to 310 mosm/l. Micromanipulator units were used for micromanipulation of the pipette. S.3 Results and discussion Direct evidence of the patterned poly-l-lysine in the hybrid template was provided by microscopic imaging of fluorescence from Fluorescein isothiocyanate (FITC)-labeled 148

164 100µm Figure S.2 Fluorescent image of hybrid template using phase-contrast microscope. -poly-l-lysine (Sigma). To make the substrate more hydrophilic, substrate with patterned SU8 polymer and S1818 was treated with X-100 triton (Sigma) and DI water (volume ratio 1:1) as a surfactant at 70 C for 30 minutes. Figure S.2 shows the FITC fluorescent from hybrid template. The color filtering set for fluorescence imaging is composed of an exciting filter peaked at around 480 nm, a dichroic mirror and an emitting filter peaked at around 530nm. Since the fluorescence from SU8 polymer is much smaller in comparison to that from FITC-labeled-poly-L-lysine, a clear intensity contrast is observed in Figure S.2, demonstrating that patterned poly-l-lysine layer is formed within the SU8 microstructure. Dissociated rat hippocampal neurons were cultured on the hybrid template with both topographical and chemical guidance cues. The size of the node was 100 µm in diameter. Due to our choice of low plating density, there was one neuron per circular node on av- 149

165 (A) (B) (C) SU8 SU8 SU8 40 Ì m Figure S.3 Hippocampal neurons growing on hybrid templates. (a) 2DIV culture with neurite growth restricted within the pattern. (b) 8DIV culture with long neurite extension within the pattern. (c) 8DIV culture with short neurite extension outside the pattern. erage. We observed that only a few neurons landed in the channels due to its wide width. On average, healthy neurons tended to stay within poly-l-lysine defined patterns. Very few neurons were observed atop the SU8 polymer barrier. The average number of neurons adhered to the template is counted to be about 50 Cells/mm 2 in microscope images. Figure S.3 shows microscopy images of neurons and their neurite growth at different stages in culture. In Figure S.3 (A), 2 Days-in-vitro (DIV) neurons have landed on the intersection between the node and an adjacent channel and have started short range neurite growth of less than 100 µm. The circled area shows one example where the neurite first grew towards the SU8 polymer barrier, encountered the barrier and was then redirected back to an opposite direction. This indicates that SU8 polymer barrier successfully restricted the neurite growth within the poly-l-lysine defined patterns. As the neurons grow more mature, the neurites grew up to hundreds of microns in length before they reached their counterparts from an adjacent neuron. As an example, Figure S.3 (B) 150

166 shows a neurite length between the two 8 DIV neurons of approximately 500 µm, with single neurite outgrowth from each neuron. As a contrasting example of the effectiveness of our hybrid template, Figure S.3 (C) shows a partially inhibited neurite growth to less than 100 µm exhibited by another 8 DIV neuron, part of which reached atop the SU8 barrier (where poly-l-lysine is absent). In this case, the neuron body shrank to an irregular shape, indicating the disturbance of the neuron maturation process. These examples illustrate how the hybrid template has successfully suppressed the growth of neuron and neurite off the defined patterns, while other studies of patterned hippocampal neurons have shown bifurcations of neurites out of the defined pattern to distant neurons [212, 213]. Such neurite misguidance is undesirable for neuronal network formation for interfacing to and integration with micro/optoelectronic devices which generally require well defined pattered geometrical circuit layout. In order to compare the differences of patterning schemes in neuron growth guidance with and without SU8 polymer barrier, we cultured neurons onto both our patterned hybrid and patterned poly-l-lysine-only templates, choosing the same geometry and di- (A) (B) 40 m 40 m Figure S.4 Hippocampal cultures of 13 DIV on (a) hybrid template with few number of neurites for each neuron (b) poly-l-lysine template with many neurites for each neuron. 151

167 mension for both. For reference, Figure S.4 (A) shows yet another result on the hybrid template where we observed that each neuron had only one or two neurite outgrowth (red arrows), directed preferentially towards a neighboring neuron. The tracks of the neurites are readily seen indicating clear geometrical interconnections between adjacent neurons. By contrast, in Figure S.4 (B) neurons were cultured on poly-l-lysine-only template with same plating density. Here the neurites proliferated for each neuron and made connections with many other neurons. In this case, there are no obvious traces which given neuron is geometrically connected to other specifically identified neurons. Neurons grown on both hybrid template and normal glass coverslip with homogeneous poly-l-lysine were compared for electrophysiology experiments. The size of the nodes on patterned templates was here chosen to be 50 µm in diameter. 12 DIV neurons were mature enough for healthy giga-seal formation in patch clamp measurements. In Figure S.5 (A), the resting membrane potential was adjusted to -51 mv during the recording. Action potentials were recorded under the injection of depolarization pulses of 100 pa, 200 pa and 300 pa for 600 ms for neurons cultured on the hybrid template. The (A) (B) Figure S.5 Typical voltage responses for 600 ms current injection to 12 DIV excitable hippocampal neurons on (a) hybrid template and (b) unpatterned template. 152