High-Density Microelectrode Array Platform in CMOS Technology

Research Collection

Doctoral Thesis

Author(s): Muller, Jan

Publication Date: 2015

Permanent Link: https://doi.org/10.3929/ethz-a-010554762

Rights / License: In Copyright - Non-Commercial Use Permitted

This page was generated automatically upon download from the ETH Zurich Research Collection. For more information please consult the Terms of use.

ETH Library DISS. ETH No. 22,625

High-Density Microelectrode Array Platform in CMOS Technology

A thesis submitted to attain the degree of

Doctor Of Sciences of ETH Zurich

(Dr. sc. ETH Zurich)

presented by

Jan Muller¨

MSc ETH in Electrical Engineering and Information Technology Born January 7th, 1984 Citizen of Matzendorf (SO), Switzerland

accepted on the recommendation of

Prof. Dr. Andreas Hierlemann Prof. Dr. Tobias Delbruck Dr. Douglas J. Bakkum

All rights reserved. No part of this book may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, without the prior written permission of the copyright holder.

Cover page: The image shows a CMOS die of the developed microsystem bonded to a green printed circuit board. The images on the back page are taken from the ‘Major Results’ Section 1.6 on page 8.

Printed by Druckzentrum ETH Zurich.

Published by: Bio Engineering Laboratory, BEL ETH Zurich Mattenstrasse 26 4058 Basel Switzerland Contents

1 Introduction1 1.1 Motivation...... 2 1.2 Applications...... 3 1.3 Considerations for the measurement setup...... 4 1.4 Scope and structure of the Thesis...... 4 1.5 Author contributions...... 6 1.6 Major Results...... 8

2 A 1024-Channel CMOS Microelectrode Array With 26,400 Elec- trodes for Recording and Stimulation of Electrogenic Cells In Vitro 11 2.1 Introduction...... 13 2.2 System Design...... 15 2.3 Switch Matrix...... 17 2.4 Readout...... 20 2.5 Stimulation units...... 27 2.6 Chip implementation...... 28 2.7 Measurement...... 30 2.8 Comparison To State-of-the-art and Conclusion...... 34

3 High-resolution CMOS MEA platform to study neurons at subcellular, cellular and network levels 39 3.1 Introduction...... 40 3.2 Materials and methods...... 42 3.3 Results...... 49 3.4 Discussion and outlook...... 60

iii Contents

4 Selection of best recording sites for optimizing spike-sorting yield 65 4.1 Introduction...... 67 4.2 Performance assessment of recording conﬁgurations...... 69 4.3 Example...... 74 4.4 Electrode selection algorithms...... 76 4.5 Technology application...... 79 4.6 Discussion and conclusion...... 81

5 Sub-millisecond closed-loop feedback stimulation between arbitrary sets of individual neurons 85 5.1 Introduction...... 86 5.2 Methods...... 89 5.3 Evaluation and Results...... 93 5.4 Discussion...... 103 5.5 Conclusion...... 105

6 Conclusion 107 6.1 Improvements over previous designs...... 108 6.2 Measurement setup...... 109 6.3 Signal processing and spike-sorting considerations...... 109

7 Outlook 113

Appendix 131

A Glossary 131

B Publications 135

C Acknowledgements 143

D Curriculum Vitae 145

iv Abstract

This thesis presents the design, implementation, and application of a high- density microelectrode array (MEA) platform based on complementary-metal- oxide-semiconductor (CMOS) technology, for bi-directional interaction with electrogenic cells. Studying how networks of neurons process information and whether they exhibit plasticity requires access to many neurons in parallel for extended time periods. MEAs are useful tools for such studies, as they provide non-invasive access to many neurons. They are used to measure the extracellular potentials that are generated by the action potentials of the individual neurons. Using CMOS technology to implement MEAs brings several advantages. A high integration density of electrodes can be achieved through routing of the electrode signals by using the multiple metal layers that are available in standard CMOS technology. The electrode density can be further increased by employing electronic switches to implement multiplexing techniques. Detrimental effects of parasitic off-chip connections can be reduced by integrating amplifier circuitry on the same substrate as the sensing electrodes. By additionally integrating analog- to-digital (A/D) converters on the same substrate, only digital data, which are almost immune to interference noise, are transmitted off chip. Finally, complex integrated systems can be devised that include additional units, such as electrical stimulation buffers, which obviates the need for external components. The device presented in this work features a 3.85×2.10 mm2 electrode array with 26,400 Pt-microelectrodes, arranged in a grid-like configuration with a center- to-center pitch of 17.5 µm. By means of an analog switch matrix, an almost arbitrary subset of the electrodes can be connected to 1024 low-noise readout amplifiers, as well as 32 dual-mode current- and voltage-stimulation units, all of which reside at the periphery of the electrode array. Instead of integrating the amplifiers right next to the electrodes (as it is done for the pixel concept), the switch matrix concept allows for decoupling electrode pitch from needed area for circuits and amplifiers at each electrode. Thus, the trade-off between high electrode density on the one hand, and limited amplifier area and resulting noise on the other hand is relaxed. The amplifiers exhibit very low readout noise levels — 2.4 µVrms in the action potential signal band (300 Hz – 10 kHz) — and provide programmable gains of up to 78 dB. On-chip ADCs sample the data at 20 kHz. The combination of switch-matrix architecture for the circuitry and high-resolution electrode array renders the device suitable for recording signals of individual neurons and even axonal arbors of single cells with signal amplitudes as small as a few microvolts. A custom printed-circuit board (PCB) was developed to provide reference voltages for the analog circuitry of the CMOS device and to interface with the system digital core for data transfer off chip. The data are streamed to an FPGA to enable real-time signal processing, such as band-pass filtering and spike sort-

v Abstract ing. The data are then packaged into UDP (user datagram protocol) frames and transmitted over Ethernet to a host computer. To handle the large amount of data (24 MB/s), a custom UDP accelerator was developed to bypass the slow network stack running on a microcontroller on the FPGA. To operate the setup, software for the host computer was developed to configure the CMOS device, as well as for online analysis and visualization of the recorded data. Based on the theoretically expected spike sorting error under the assumption of optimal recording conditions, a framework was established to evaluate different suitable electrode configurations for given recording scenarios. An algorithm and objective measure to automatically determine optimal recording electrodes was also developed. Cultures of rat cortical neurons were grown on the electrode array for months, and various experiments were performed to validate the proper functioning of the device. Spike-sorting of the neuronal activity, recorded by all 26,400 electrodes, revealed more than 2000 distinct extracellular field potentials that could be attributed to individual neurons. Images of fluorescently labeled cells correlated well with the electrically identified cell positions, which confirmed the suitability of the system and setup to characterize neuronal networks. By computing the spike-triggered averages ofthe extracellular signals of a specific neuron while at the same time scanning through all available electrodes, large axonal arbors of single cells were revealed, which extended over large distances within the 8 mm2 array area. Recruiting of the recording electrodes right below the axonal pathways allowed for the study of axonal signal propagation at high spatiotemporal resolution. Parameters, such as the axonal action potential progagation velocity were analyzed in detail and found to vary between 0.9 m/s and 0.4 m/s along a single branch of the same axon.

vi Zusammenfassung

Die vorliegende Arbeit beschreibt die Entwicklung, Realisierung und Anwen- dung eines CMOS-basierten Mikroelektrodenarrays mit hoher räumlicher Au- flösung fur¨ extrazelluläre Messungen und Stimulation von elektrogenen Zellen. Um Informationsverarbeitung und Plastizität in Netzwerken von Nervenzellen zu erforschen, braucht es gleichzeitigen Zugang zu möglichst vielen solchen Ner- venzellen uber¨ einen längeren Zeitraum. Mikroelektrodenarrays sind nutzliche¨ Werkzeuge, um die Interaktion von Nervenzellen zu untersuchen, da sie extrazel- luläre Potentiale von vielen Zellen gleichzeitig nicht-invasiv messen können. Mikroelektrodenarrays in CMOS Technologie bieten einige Vorteile. So kann z.B. eine hohe Integrationsdichte von Mikroelektroden erreicht werden, da man die Signalleitungen der Elektroden uber¨ mehrere zur Verfugung¨ stehende Met- allschichten fuhren¨ kann. Eine weitere Verdichtung der Elektroden kann erreicht werden, wenn man mit Hilfe von elektronischen Schaltern die Signalleitungen zeitversetzt ausliest. Je weiter die Verstärkungselektronik von der Signalquelle entfernt ist, desto mehr kann das Signal durch Rauschen gestört werden. Darum ist es vorteilhaft, die Verstärker so nah wie möglich an den Elektroden zu platzieren. Zusätzlich kann man die analogen Signale noch auf demselben Chip in Digitalsig- nale umwandeln, auf dem sich Elektroden und Verstärker befinden. So verlassen nur digitale Signale den CMOS Chip, die sehr unempfindlich gegen Rauschen und Störungen sind. Ausserdem können komplizierte voll-integrierte Systeme entwickelt werden, die weitere Einheiten, wie z.B. elektrische Stimulationstreiber auf demselben Chip enthalten, so dass der Bedarf an externen Komponenten reduziert wird. Das in dieser Arbeit vorgestellte CMOS System umfasst auf einer Fläche von 3.85 × 2.10 mm2 26’400 Platin-Mikroelektroden. Die Elektroden sind in einem Raster angeordnet mit einem Abstand von 17.5 µm zwischen den Zentren zweier Elektroden. Mit Hilfe einer Schaltermatrix kann eine fast beliebige Auswahl von Elektroden mit 1024 Verstärkern oder 32 Stimulationstreibern verbunden werden. Diese Verstärker und Stimulationstreiber befinden sich an der Peripherie, des Elektrodenarrays. Durch dieses Konzept, nicht jede Elektrode mit einem eigenen Verstärker auszustatten (wie es z.B. im Pixelkonzept“ implementiert ” ist), aber stattdessen eine konfigurierbare und beliebige Auswahl von Elektroden mit Verstärkern zu verbinden, wird die pro Verstärker zur Verfugung¨ stehende Fläche von der Elektrodenfläche entkoppelt. Dieser Ansatz erlaubt es relativ grosse rauscharme Verstärker ausserhalb der Elektrodenfäche zu platzieren, die entsprechend niedrigeres Rauschen aufweisen. Die Verstärker in diesem CMOS Chip haben eine programmierbare Verstärkung von bis zu 78 dB und erreichen einen Eingangsrauschpegel von 2.4 µVrms (Bandbreite: 300 Hz – 10 kHz). Die Signale werden auf dem CMOS Chip mit 20 kHz abgetastet und digitalisiert. Die Kombination eines flexiblen dichten Elektrodenarrays mit rauscharmen Ver- stärkern ist geeignet, die Signale einzelner Nervenzellen und sogar einzelner Ner- venfasern (Axonen) zu messen.

vii Zusammenfassung

Um Messungen durchzufuhren,¨ wurde ein System bestehend aus einer Leiter- platine, einem FPGA und einem PC entwickelt. Die Leiterplatine versorgt den CMOS Chip mit den nötigen Versorgungs- und Referenzspannungen und liest die Daten uber¨ die digitale Schnittstelle aus. Die Daten werden dann uber¨ eine serielle Verbindung an das FPGA geschickt, wo sie in Echtzeit verarbeitet und auf Signale der Nervenzellen durchsucht werden. Anschliessend werden die Daten in UDP Pakete gepackt und uber¨ eine Netzwerkschnittstelle an den PC geschickt. Eine spezielle digitale Logikeinheit unterstutzt¨ den langsameren Software-basierten Netzwerkstack auf dem Mikrocontroller des FPGA dabei, die grosse Datenmenge (24 MB/s) effizient zu verarbeiten. Um das Messsystem zu programmieren wurde eine spezielle Software entwickelt. Ausserdem wurde eine Software zur Analyse, Visualisierung und Speicherung der Messdaten entwickelt. Unter der Annahme optimaler Messbedingungen und mit Hinsicht auf den theoretisch erreichbaren minimalen Fehler bei der Klassifizierung von Nervenzellaktivitäten, wurde ein Konzept entwickelt um verschiedene Elek- trodenanordnungen zu evaluieren. Dann wurde ein Verfahren entwickelt um op- timale Elektrodenanordnungen automatisch zu bestimmen. Zellen aus der Hirnrinde von Ratten wurden uber¨ Monate hinweg auf dem CMOS Chip kultiviert. Verschiedene Experimente wurden durchgefuhrt,¨ um die Funk- tionalität des Chips zu verifizieren. Die Analyse der Daten von allen 26’400 Elektroden mit anschliessender Zuordnung zu entsprechenden Nervenzellen er- gab mehr als 2000 individuelle Signalquellen oder Nervenzellen. Zur Bestätigung wurden Fluoreszenzbilder der Zellkulturen mit den gemessenen elektrischen Sig- nalen verglichen, wobei eine gute Ubereinstimmung¨ festgestellt wurde. Durch gestaffeltes Auslesen aller Elektroden bei gleichzeitiger Mittelung uber¨ wieder- holte Messungen der Aktivität einer einzelnen Nervenzelle konnten lange Ner- venfasern, die sich uber¨ das ganze 8 mm2 Elektrodenraster erstrecken, sichtbar gemacht werden. Durch gezielte Selektion von Ausleseelektroden direkt unter so einer Nervenfaser konnte die Ausbreitungsgeschwindigkeit eines Nervenimpulses entlang einer Faser zwischen 0.9 m/s und 0.4 m/s variierend bestimmt werden.

viii Chapter 1

Introduction

Brains are complex organs consisting of billions of deliberately interconnected cells or neurons. Trying to understand how the collective activity of all con- stituent cells contributes to brain functions, such as memory and learning, is a fascinating, though challenging, endeavor. Fundamental activity of neurons is based on the choreographed activity of ion channels to generate so-called action potentials (APs) (Hodgkin and Huxley(1952); Bean(2007)). It is believed that the propagation of these action potentials through networks of neurons is pivotal for information processing in the brain. Understanding the rules underlying these propagation mechanisms and how propagation changes over time may help to advance our understanding of how brains work. Studying the dynamics of single neurons is difficult enough and conventionally done through patch-clamp recordings (Stuart et al.(1997)). However, investigating the behavior of networks of many interdependent cells requires means to interrogate many such individual neurons simultaneously. To cope with this chal- lenge, new sets of tools to record the activity of ever larger networks of neurons are being developed (Marblestone et al.(2013a); Alivisatos et al.(2013)), for example, novel genetically targeted all-optical electrophysiology methods (Hochbaum et al.(2014)), or a technique called light-sheet microscopy allowing to record the activity from 80 percent of all neurons in the brain of a zebrafish with a temporal resolution of 0.8 Hz (Ahrens et al.(2013)). Furthermore, microelectrode arrays (MEAs) featuring high-spatial-resolution recording and stimulation electrodes for in vitro (Eversmann et al.(2003a); Bertotti et al.(2014); Frey et al. (2010); Hierlemann et al.(2011)) or in vivo recordings (Johnson et al.(2012); Seidl et al.(2011)) have recently been developed (Obien et al.(2015)). In this thesis, we present the design, implementation, and application of a complementary-metal-oxide-semiconductor (CMOS)-based planar high-density microelectrode array (HDMEA) for in vitro studies. Recording and stimulation capabilities provide bi-directional interaction with single cells from networks including thousands of neurons. A particular feature of this device is to electron- ically “zoom” into a specific area of interest. Similar to the way in which the

1 1.1. Motivation optical zoom of a microscope can be used to first inspect the preparation at a lower resolution and subsequently zoom into an area of interest with higher magnification, the developed HDMEA features a reconfigurable switch matrix that enables the user to first record from coarsely distributed electrodes and subsequently select electrodes in areas of interest to achieve higher resolution.

1.1 Motivation

Passive microelectrode arrays with metal electrodes on a glass substrate have been successfully used in fundamental as well as in pharmacological research (Stett et al.(2003b); Gross et al.(1995)). They can record extracellular activity from many cells in parallel, over days and months. However, due to their typically limited spatial resolution (>30 µm) as well as low number of electrodes (typically less than 300) it is difficult to resolve single cells. The recent advent of CMOS-based MEAs (Eversmann et al.(2003a)) has the potential to overcome some of these limitations. The extracellular electrical signal of individual neurons is very small, on the order of tens of microvolts. Therefore, management of the noise in the recording electronics is a crucial issue to get reasonable signal-to-noise (SNR) values. Using CMOS technology to locate the amplifier and analog-to-digital conversion circuitry as close as possible to sensing sites greatly reduces the influence of noise on measurements, as disturbances, e.g., through parasitics from off-chip connections, can be avoided. However, there is a trade-off between CMOS circuit real estate and noise inherent to the circuitry: the smaller the available real estate, the harder it is to design low-noise circuitry. Two contradictory requirements exist for CMOS-based microelectrode arrays: Densely spaced small sensor units to achieve high spatial resolution versus comparably large real estate for low-noise readout electronics. The CMOS device presented in this thesis follows a similar approach as in (Frey et al.(2010)) to cope with this dilemma. We designed an analog switch-matrix featuring 26,400 microelectrodes, which allows for connecting a subset of all electrodes to 1024 readout channels and 32 stimulation units residing at the periphery of the array. Such an approach provides two main advantages. First, the amplifier area is decoupled from the electrode pitch. Second, by only routing readout units to electrodes that pick up useful signals, no power and CMOS area is wasted on electrodes, which provide no relevant signals for the current experiment. The design of the switch-matrix is organized in a modular way. A common implementation of a basic unit, a pixel, is repeated for each electrode position. The flexibility of the switch-matrix is then attained by augmenting each basic unit with a unique individual set of wires to implement the desired layout. A computer-aided design (CAD) software supports this process.

2 Chapter 1. Introduction

1.2 Applications

The flexible, as well as high-resolution electrode array makes the device presented in this thesis a suitable platform for various experimental scenarios. Mouse retina: The retina is a thin layer of light-sensitive tissue residing at the inner surface of the eye. Photoreceptor cells transform a visual scene through a cascade of cells into action potentials of the retinal ganglion cell (RGC) layer, which projects long axons into the visual cortex. The retina is a mostly two- dimensional structure and is be placed on the CMOS MEA with the layer of RGCs facing the electrodes. Projecting visual scenes while recording the RGC output allows for the study of how the retina transforms visual scenes into a spike pattern. The mouse as model animal is particularly suited for such experiments due to the broad availability of transgenic variants and the resulting possibility of targeted manipulation of defined cell types (Huberman et al.(2009a); Kim et al.(2008); Munch et al.(2009)). Depending on the specific location in the retina, a mouse retina has an approximate cell density of 2700 RGCs/mm2 (Jeon et al.(1998b)), which is lower than the density of recording electrodes of the presented CMOS device (3265 electrodes/mm2 ), so that single cell resolution is possible. Furthermore, there are approximately 20 different types of RGCs in the mouse retina (as determined by morphological studies), organized in repetitive mosaics over the whole retina (Masland(2012)). The flexibility in electrode selection allows for focusing on a particular cell type and for recruiting electrodes around many of the targeted cells. Also, by having large patches of electrodes, different cell types can be recorded at once in order to investigate how they collectively process visual stimuli. Brain slices: Thin layers of brain tissue can be sliced and placed on the 8 mm2 electrode array (Egert et al.(1998)). Due to the large sensor area, slices from, e.g., mouse brains over large brain areas can be studied. A slice through the mouse hippocampus has a size of about 3 × 1 mm2 Different layers of the cortex can, therefore, be studied at once. The flexibility in choosing recording sites allows for adaptation of the setup to the morphology of the particular brain slice under study. Cell cultures: Dissociated cells of, e.g., cortical brain tissue can be cultured on top of the sensor electrodes. When incubated at 37°C and under 5% CO2, the cells grow processes and start to connect with each other through synapses and to form networks. After 1 to 2 weeks of incubation, the cells exhibit spontaneous action potentials and can be electrically stimulated. Depending on the number of seeded cells, such a culture can be sparsely distributed over the array. Plasticity: How individual neurons process information can be studied by applying targeted stimulation to certain cells and by subsequently observing how stimulation alters their functional connectivity. Spike-timing-dependent plasticity (STDP) is one of the governing rules describing how the connectivity between two cells changes depending on the timing of their spiking activity. Recently,

3 1.3. Considerations for the measurement setup many studies have been conducted to identify STDP and to possibly ﬁnd other rules governing memory and learning in the brain (Dan and Poo(2004); Froemke and Dan(2002); Ikegaya et al.(2004); Song et al.(2000)). However, many of these studies either fail to access multiple cells in parallel (by relying on patch recordings) or cannot assess the synaptic strength between them. The availability of a microelectrode array featuring single-cell resolution and many readout sites to be recorded from in parallel, as well as featuring the possibility of recording from each cell in a culture, oﬀers the potential to render investigations of functional connectivity more accurate.

1.3 Considerations for the measurement setup

When designing a setup that should be used in real experiments one has to keep in mind that the operation of the setup needs to be as simple as possible. To exploit the full potential of the HDMEA developed in this thesis, during an experiment, a subset of electrodes that yield meaningful signals for the given investigation need to be chosen from the large number of available electrodes. The flexibility in electrode selection, however, comes at the expense of added complexity to the experiment. Most experiments involving living biological preparations require great care. If, additionally, the researcher has to deal with complicated software during an experiment, the efficiency and throughput of the experiment may suffer. Therefore, our intention was to reduce user interventions to a minimum, and the setup needed to be as easy to operate as possible. Besides intuitive software, preferably with a graphical user interface (GUI) and simple push-button operation, an online visualization of the recording data was considered to be of paramount importance. Rapid access to experimental parameters, such as spiking frequency or the spatial extension of extracellular field potentials and the axonal arbors of individual neurons boosts productivity during experiments. As the selection of suitable or optimal electrodes is a time-consuming and iterative process, the on-line analysis of neuronal activity can greatly speed up this process. Additionally, software crashes can be detrimental for an experiment and have to be avoided by all means. By employing an FPGA to bridge the communication between the CMOS device and the host computer, signal processing algorithms, such as real-time spike sorting or closed-loop feedback stimulation, can easily be integrated into the setup. Moreover, by using Ethernet as a communication means between the FPGA and the host, the setup can universally be interfaced, independent of the used host computer and operation system.

1.4 Scope and structure of the Thesis

The main focus of this thesis is the development and application of a switch- matrix-based high-density microelectrode array. While adopting concepts of a

4 Chapter 1. Introduction previously developed prototype implementation, a completely new design has been conceived and implemented. The high-density microelectrode array forms part of a larger system with recording and stimulation circuitry, analog-to-digital converters and a digital interface; the system has been fabricated in industrial 0.35 µm CMOS-technology. The CMOS system is suitable to record from a variety of biological preparations, such as retinae, brain-slices, or cell cultures. Three primary recording application scenarios have been considered for the design process. 1) Single-cell resolution recordings in sparsely distributed networks of cultured neurons; 2) tracking of the propagation of axonal signals over several millimeters; 3) recording from complete cell mosaics in retinal patches. Addi- tional requirements included means to dynamically and bidirectionally interact with neuronal networks, for studying information processing in neuronal networks. Fast and precise electrical feedback stimulation capabilities are needed to ask basic research questions concerning synaptic strength and functional plasticity in activity-dependent STDP-like experiments. The thesis includes four publications:

1. A 1024-Channel CMOS Microelectrode Array With 26,400 Elec- trodes for Recording and Stimulation of Electrogenic Cells In Vitro Marco Ballini, Jan Muller, Paolo Livi, Yihui Chen, Urs Frey, Alexander Stettler, Amir Shadmani, Vijay Viswam, Ian Lloyd Jones, David Jackel, Milos Radivojevic, Marta K Lewandowska, Wei Gong, Michele Fiscella, Douglas J. Bakkum, Flavio Heer, Andreas Hierlemann IEEE Journal of Solid-State Circuits, 2014

2. High-resolution CMOS MEA platform to study neurons at subcellular, cellular and network levels Jan Muller,¨ Marco Ballini, Paolo Livi, Yihui Chen, Milos Radivojevic, Amir Shadmani, Ian L. Jones, Michele Fiscella, Roland Diggelmann, Alexander Stettler, Urs Frey, Douglas J. Bakkum and Andreas Hierlemann Lab on Chip, 2015

3. Selection of best recording sites for optimizing spike-sorting yield Jan Muller,¨ Felix Franke, Michele Fiscella, Urs Frey, Douglas J. Bakkum and Andreas Hierlemann in preparation

4. Sub-millisecond closed-loop feedback stimulation between arbitrary sets of individual neurons Jan Muller,¨ Douglas J. Bakkum and Andreas Hierlemann Front. Neural Circuits, 2013

The ﬁrst paper is focused on the circuit-level-design of the CMOS-based HD- MEA. It describes the diﬀerent circuit blocks of the device, such as the analog

5 1.5. Author contributions switch-matrix, the amplifier stages, the single-slope analog-to-digital converters (ADCs), the dual-mode current and voltage stimulation units, as well as the digital control units. Electrical characterization measurements have been performed to show the low-noise characteristics of the recording circuitry and to validate all device functions. The second paper demonstrates the applicability of the device to record from neuronal networks at various levels of spatial resolution. Cell cultures derived from primary rat cortical neurons that have been isolated from E18 embryos were cultivated for several weeks on the devices. Exploiting the flexibility in electrode selection allowed for recording from the same culture first a network-wide overview before then analyzing single cells and subcellular features, such as axonal arbors. In the third paper, two algorithms were developed to automate the process of selecting the most suitable recording sites. We introduced objective metrics to quantify the quality of different recording electrode configurations. The expected error in spike-sorting for different electrode configurations has been computed based on linear discriminant analysis. Such an objective classification helps the experimenter to select the optimum set of electrodes from the large number of available electrodes in an automated way. Also, we analyzed in how far cell density and the number of available electrodes influence the expected error. The last paper describes the implementation and application of a closed-loop system to deliver electrical feedback stimulation pulses upon the detection of action potentials. By using an FPGA to close the feedback loop, the detection of action potentials and the subsequent issuing of stimulation pulses through feedback can all be accomplished within less than one millisecond. Digital hardware has been used to implement an event-engine, a set of simple event-based modules, which can be plugged together in arbitrary ways to implement sophisticated pattern matching. Such pattern matching enables to lock feedback stimulation on the occurrence of spiking neuronal assemblies. Finally, a conclusion and outlook will be given in Chapter6&7.

1.5 Author contributions

The presented work is the result of the collaborative effort of several involved people. Owing to the complexity of the CMOS system the circuit design part of the project (Chapter2&3) required a substantial team effort. Besides developing a system that excels in circuit characteristics, one aim of the project was to make the system work reliably and effectively in experiments and fundamental research. Close collaboration between all designers is inevitable to achieve such a goal. Fortunately, during the experimental verification of the design (mainly Chapter3), different biological preparations were ready to use at the Bio Engi- neering Laboratory. Preparations ranging from cell cultures to acute brain slices and patches of retina were available, and several people listed below, were very helpful with setting up proof-of-principle measurements.

6 Chapter 1. Introduction

The circuit design team involved: Yihui Chen, Marco Ballini, Paolo Livi and Jan Muller.¨ Yihui Chen supervised the circuit design part of the project. Marco Ballini designed and implemented the readout amplifiers and the on-chip ADCs. Paolo Livi designed the stimulation units and drafted the layout for the initial version of the pixels in the array. Jan Muller¨ designed the electrode array, the digital circuits, together with software needed to interface the system for configuration purposes. Douglas Bakkum supervised the experimental part of the project, provided support for the closed-loop experiments, helped with cell culturing and the imaging of fluorescent cell cultures. Milos Radivojevic helped with culturing of dissociated cells. Ian L. Jones provided support with experimental retinal preparations. Michele Fiscella helped with experimental preparations and provided help with data analysis. Felix Franke helped with data analysis techniques and discussion of algorithm implementations. Alexander Stettler did the post-processing of the CMOS wafers. Roland Diggelmann contributed software for online visualization of the recorded data. Amir Shadmani characterized the stimulation circuitry. Vijay Viswam helped with the electrode characterization. Urs Frey provided advice during the initial stage of the project and designed the first version of the switch-matrix (Frey et al.(2010)), which served as a prototype for the presented work. Jan Muller¨ designed and performed the experiments, did the data analysis and developed the measurement setup together with signal processing and feedback stimulation facilities on an FPGA.

7 1.6. Major Results

1.6 Major Results

CMOS-based HD-MEA. A fully integrated CMOS-based microelectrode array platform was developed. By means of a ﬂexible, reconﬁgurable switch-matrix, an almost arbitrary subset of a total of 26,400 Pt- microelectrodes can be connected to 1024 readout-channels, as well as 32 dual-mode current and voltage stimulation units (Chapter 2 on page 11).

Sub-millisecond closed-loop feedback stimulation. A fast and precise closed-loop feedback stimulation setup was developed to deliver electrical stimulation pulses to neurons upon detection of speciﬁc activity. Without introducing latency, a modular, FPGA-based event-engine, implemented through a set of simple building blocks, enabled sophisticated compu- tations. Pattern-matching algorithms can be implemented to identify assemblies of spiking neurons in the neuronal network. (Figure 5.5 on page 98).

Recording setup and online data visualization. A measurement setup for the HD-MEA devices was developed that was capable of online interpretation and visualization of the 40 MB/s data stream that was generated by the CMOS device. An online visualization of the recorded signals, such as action potentials and their spatial spread, helped boosting eﬃciency during experiments, as time-consuming oﬄine analysis was circumvented. (Chapter3 on page 39).

8 Chapter 1. Introduction

Novel algorithms to identify optimal recording electrodes. Based on linear discriminant analysis (LDA), a metric was introduced to evaluate diﬀerent electrode selections in terms of how well the recorded action potentials could be assigned to individual neurons. Two algorithms were developed to automate the search for a set of suitable recording electrodes and were applied to the CMOS HD-MEA developed in this thesis (Chapter 4 on page 65).

Spike-sorting revealed more than 2000 individual neurons in one culture. Rat cortical neurons and glial cells have been cultured for weeks. Spontaneous spiking activity of one such culture was recorded by using fifty different patches of densely packed electrodes. Subsequent spike-sorting revealed an approximate total of 2000 identified single cells. (Figure 3.3 on page 50).

Large axonal arbors of spontaneously active cortical neurons. Scanning through all available electrodes while computing the spike-triggered average of the spontaneous activity of a neuron revealed large axonal arbors. The propagating axonal action potentials could be tracked over distances of more than 4 mm and up to 6 ms after the somatic action potential occurred (Figure 3.6 on page 57).

9 1.6. Major Results

10 Chapter 2

A 1024-Channel CMOS Microelectrode Array With 26,400 Electrodes for Recording and Stimulation of Electrogenic Cells In Vitro

Marco Ballini1, Member, IEEE, Jan Muller¨ 1, Student Member, IEEE, Paolo Livi1, Student Member, IEEE, Yihui Chen1, Member, IEEE, Urs Frey2, Member, IEEE, Alexander Stettler1, Amir Shadmani1, Vijay Viswam1, Member, IEEE, Ian Lloyd Jones1, David J¨ackel1, Milos Radivojevic1, Marta K. Lewandowska1, Student Member, IEEE, Wei Gong1, Michele Fiscella1, Douglas J. Bakkum1, Flavio Heer1, Member, IEEE, and Andreas Hierlemann1, Member, IEEE

IEEE Journal of Solid-State Circuits, 2014

1Bio Engineering Laboratory, ETH Zurich, Switzerland

2RIKEN Quantitative Biology Center, Kobe, Japan

11 Abstract — To advance our understanding of the functioning of neuronal ensem- bles, systems are needed to enable simultaneous recording from a large number of individual neurons at high spatiotemporal resolution and good signal-to-noise ratio. Moreover, stimulation capability is highly desirable for investigating, for example, plasticity and learning processes. Here, we present a microelectrode array (MEA) system on a single CMOS die for in vitro recording and stimulation. The system incorporates 26,400 platinum electrodes, fabricated by in-house post- processing, over a large sensing area (3.85×2.10 mm) with sub-cellular spatial resolution (pitch of 17.5 µm). Owing to an area and power efficient implementation, we were able to integrate 1024 readout channels on chip to record extracellular signals from a user-specified selection of electrodes. These channels feature noise values of 2.4 µV in the action-potential band (300 Hz−10 kHz) and 5.4 µV in the local-field-potential band (1 Hz − 300 Hz), and provide programmable gain (up to 78 dB) to accommodate various biological preparations. Amplified and filtered signals are digitized by 10 bit parallel single-slope ADC at 20 kSamples/s. The system also includes 32 stimulation units, which can elicit neural spikes through either current or voltage pulses. The chip consumes only 75 mW in total, which obviates the need of active cooling even for sensitive cell cultures.

12 Chapter 2. CMOS Circuitry and Architecture

2.1 Introduction

Extracellular Recordings of the electrical activity of neural and cardiac cell networks in organs such as the brain, the retina, or the heart, can provide a wealth of information about the physiology as well as the pathological degenerations that may cause diseases, such as Parkinson’s or Alzheimer’s. Microelectrode arrays (MEAs) have been used for a long time for in vitro extracellular recordings of electrogenic cell cultures and tissues, such as acute or organotypic brain slices and retinae (Gross et al.(1995); Jimbo et al.(1998); Stett et al.(2003b)). They provide simultaneous multisite recording capability, which is essential to study cellular interconnections and network properties that arise from synchronized cellular activity (Marblestone et al.(2013a); Jiang et al.(2013)). However, passive MEAs, which typically include metal electrodes on a glass substrate, are limited in both the number of electrodes (usually less than 300) and the spatial resolution (typically ≥ 30 µm ), features that are needed to reconstruct large neural networks at cellular detail. With CMOS technology, these limitations can be overcome by using multiplexing techniques, which enable access to a large number of closely-spaced electrodes to obtain large sensing areas at high spatial resolution (Hierlemann et al.(2011)). Moreover, the monolithic integration of recording amplifiers and ADCs, on the same substrate with the electrodes, avoids off-chip parasitics and interference and, at the same time, allows for realizing a large number of recording channels with a low number of connections. Whereas many neural acquisition chips, mostly for in vivo applications, have been designed in the last decade (see for example Aziz et al.(2009); Shahrokhi et al.(2010); Wattanapanitch and Sarpeshkar (2011); Muller et al.(2012); Gao et al.(2012); Guo et al.(2012c); Jing Guo (2013); Aziz et al.(2007)), only a few implementations of CMOS microelectrode systems have been realized to date. Neural acquisition systems typically interface with arrays or micro-needle probes consisting of only a few hundred electrodes. CMOS MEAs, on the other hand, feature in excess of several thousand recording sites. It is therefore important to provide the capability of recording from such large arrays, while maintaining good signal characteristics. Currently available CMOS MEAs, however, are limited in either spatial resolution (Heer et al. (2006a); Berdondini et al.(2009c)), noise performance (Berdondini et al.(2009c); Eversmann et al.(2003a, 2011)), or readout channel count (Heer et al.(2006a); Frey et al.(2010)). One category of CMOS MEAs is based on an active-pixel sensor (APS) architecture (Heer et al.(2006a); Berdondini et al.(2009c); Eversmann et al.(2003a, 2011)). Since the area for the analog front-end (AFE) amplifier is limited by the pixel size, this scheme results in a tight trade-off between noise, power consumption and spatial resolution. In addition, as all electrodes, even those without significant neural signals, are scanned, the full-frame rate is typically less than 10 kHz (Berdondini et al.(2009c); Eversmann et al.(2011)), due to power constraints. Higher sampling rates are desirable to reconstruct the fast transient of

13 2.1. Introduction

Figure 2.1: Diagram of the packaged microelectrode array chip (device concept). the spike waveforms. In Berdondini et al.(2009c) and Eversmann et al.(2011) a subset of the array can be scanned at an increased rate. Nevertheless, the selection is not flexible enough to adapt to complex morphologies or regions of interest. Another approach is to employ an analog switch matrix (Frey et al.(2010)) to continuously connect a subset of the electrodes to readout units located outside of the sensing area. Each pixel of the sensing array only contains switches and SRAM cells, leading to a high spatial resolution. The relaxed area constraints for the AFE allow for the implementation of amplifiers with lower noise and anti-aliasing filters. The system proposed in Frey et al.(2010) is limited to 126 channels for simultaneous recordings, and, therefore, renders the analysis of large neural networks difficult and time consuming. As an example, ∼ 100 subsequent acquisitions are required to cover an array area of 1.8 × 2.0 mm2 (Bakkum et al. (2013b)). In this paper, we present a recently developed CMOS MEA system that further exploits the switch-matrix approach. The system preserves sub-cellular spatial resolution over a large sensing area (8.09 mm2) and features 1024 channels for recording at high temporal resolution (20 kS/s). Despite an eight-fold increase in the channel count with respect to Frey et al.(2010), state-of-the-art noise performance has been achieved (2.4 µVrms), owing to an area and power efficient design of the circuitry for amplification and A/D conversion. Moreover, the cross-over distortions and channel-length modulation effects, which had been observed in the design of the stimulation unit in Livi et al.(2010), have been largely eliminated. The routing flexibility, provided by the switch-matrix, has been substantially improved, e.g., the high-density blocks can be 5 times larger than those in Frey et al.(2010). Post-CMOS fabrication of electrodes and biocompatible die-bonding and encapsulation have been used to obtain a device that can be handled like a standard MEA dish (see Fig. 2.1). This paper is organized in eight sections. Section 2.2 presents the system requirements and the proposed architecture. The analog switch-matrix, the readout and the stimulation units are described in Sections 2.3,2.4,2.5, respectively. Section 2.6 describes the chip implementation and fabrication. Measurement results,

14 Chapter 2. CMOS Circuitry and Architecture including electrophysiological recordings, are given in Section 2.7. Section 2.8 compares the chip to the state of the art and concludes the paper.

2.2 System Design

2.2.1 System Requirements

To enable a broad range of experiments, we aimed at realizing a versatile platform, capable of recording from various in vitro and ex vivo biological preparations, such as cultured neuronal networks, brain slices, acute retinae and cardiac- cell cultures. High spatial resolution, down to the cellular or sub-cellular level (<20 µm), is required to facilitate the task of separating individual signal sources (Lewicki(1998b); Jackel et al.(2012)). Such separation is necessary to understand how whole-network properties arise from cellular behavior and inter-cellular connections (Jiang et al.(2013)). In the case of neurons, the cell bodies (somata) have diameters in the range of 5-50 µm , but the neurites cover a much larger area. As an example, the denditric trees of Purkinje cells extend over several tens or hundreds of micrometer (Frey et al.(2009b)). For most cases of neuronal preparations, sub-cellular details of single neurons can be resolved with an electrode pitch of less than 20 µm (Bakkum et al.(2013b); Fiscella et al.(2012)). In addition, it is desirable to record simultaneously from distant regions to be able to study interactions between sub-circuits, e.g., in a brain slice, as far as several millimeters away from each other. As a tradeoﬀ with die size, we opted for a rectangular sensing area of ∼ 4 × 2 mm2. The signal levels can vary signiﬁcantly depending on cell type, distance from the recording electrode, and seal resistance of the cell-electrode cleft (Guo et al. (2012a)). A summary of signal characteristics is reported in Table 2.1. In the case of cardiac myocytes, action potentials (APs) feature amplitudes of up to several tens of mV. In the case of neurons, APs recorded at the soma have amplitudes that are typically in the range 100-500 µV In order to also detect low-amplitude spikes from single axons (<20 µV Bakkum et al.(2013b)), for the readout channels we targeted an input-referred noise of 2 µVrms in the band 500 Hz-3 kHz, where most spike energy is concentrated√ (Jackel et al.(2012)), corresponding to a thermal noise level of 40 nV/ Hz. Further reducing the noise can result in overdesign, at the expense of circuit area or power consumption, since the overall noise performance is√ limited by the neural background activity and electrode noise (e.g., ∼ 80 nV / Hz at 1 ∼ kHz, for Pt electrodes with a 25 µm diameter Guo et al.(2012a)).

15 2.2. System Design

Signal Type Frequency Band Neuronal Local Field Potentials <5 mV 1 Hz - 300 Hz Neuronal Action Potentials <1 mV 300 Hz - 6 kHz Cardiac Action Potentials <50 mV 1 Hz - 1 kHz

Table 2.1: Typical signal characteristics

APs have a −3 dB bandwidth typically around 2 kHz (Jochum et al.(2009)), with signal content up to 6 kHz. A sampling rate of 20 kS/s is sufficient for most applications. Limiting the recording bandwidth to ∼7 kHz reduces the aliased noise from both the electrodes and the circuitry. Local field potentials (LFPs), arising from the synchronized activity of many neurons, can occur simultaneously with APs and exhibit amplitudes of up to a few mV, with frequency components in the range of 1-300 Hz. Therefore, in order to also study LFPs, the readout units must be capable of recording frequencies of a few tens of Hz, while rejecting the large offset and drift of the electrode/electrolyte interface potential (in the range of several hundreds of mV Muller et al.(2012); Heer et al.(2006a)). Furthermore, versatile electrical stimulation capabilities for precisely and reliably eliciting APs are essential for investigating, for example, mechanisms such as learning and synaptic plasticity in a neural network. Since neurons can be stimulated by either voltage or current signals (Livi et al.(2010)), the availability of both modes is desirable. Typical stimulation pulses have durations of 50-900 µs , with amplitudes of 0.1 − 1 V and 50 − 900 µA(Wagenaar et al.(2004a)). Finally, to limit the chip-induced heating to less than 2°C and to avoid active cooling, we aimed for a total power consumption of less than 100 mW.

2.2.2 Chip Architecture

A block diagram illustrating the system architecture is depicted in Fig. 2.2. The chip features a sensing area of 3.85×2.10 mm2 with 26,400 electrodes, placed at a pitch of 17.5 µm (3,265 electrodes/mm2). A matrix of switches, placed below the electrodes, is used to connect an arbitrarily conﬁgurable selection of electrodes to 1024 readout channels and 32 stimulation units, all of which are located outside the electrode area. To adapt to varying experimental requirements, the readout channels provide programmable bandwidth and gain. The full signal band of 1 Hz-6 kHz can be recorded from each channel. Parallel single-slope ADCs, sharing the ramp generator and a 10 bit counter, are used to digitize signals at 20 kS/s and 10 ∼ bit resolution. The stimulation circuits, to deliver both voltage and current stimulation pulses, are grouped in two blocks, each comprising 16 units and three 10 bit DAC. By quickly selecting diﬀerent DAC outputs, complex stimulation patterns with independent bi-phasic or tri-phasic pulses can be generated at

16 Chapter 2. CMOS Circuitry and Architecture

Figure 2.2: System architecture of the cmos microelectrode array chip. each stimulated electrode. Arbitrary waveforms can also be generated, such as sinusoidal waveforms for low-frequency impedance measurements. A digital core, operating with two clock domains, transfers the readout data oﬀ- chip (24 MHz) and receives control settings through an SPI-like interface (up to 50 MHz), used to conﬁgure the array, the readout and stimulation units, and to apply stimulation patterns to the DAC inputs. To ensure data validity, both input and output data streams are protected with CRC checksums.

2.3 Switch Matrix

The electrode array is composed of 220 × 120 pixels. Each pixel includes an electrode, three switches and two SRAM cells. The schematic of a pixel is shown in Fig. 2.3. Two switches (S1 and S2), with dedicated SRAM cells to hold their on/off states, are used to configure the routing path from any specified electrode to the readout and/or stimulation units. S1 is used to connect the electrode to a signal wire. Six vertical and six horizontal signal wires are used in each pixel for routing, which are shielded by bit lines (BL), word lines (WL), and supply and ground tracks to minimize cross-talks. Neighboring electrodes are connected each to a different line. The availability of more lines per pixel improves the routing capability. In Frey et al.(2010), the horizontal lines extend over the whole width of the array and are common to each pixel in a row, so that only up to six electrodes per row

17 2.3. Switch Matrix

Figure 2.3: Schematic diagram of a pixel of the electrode array. can be addressed simultaneously. In this new design, to further improve the routing flexibility and to reduce the parasitic capacitance, the signal wires were cut into segments, which extend only for a length of 24 pixels (418 µm length). These segments can be connected through switch S2 to form a specific path and reach the boundary of the array to connect to readout and/or stimulation units. This mesh provides high flexibility to adapt the electrode selection to the morphology of biological samples, for example in sparsely distributed sets, at points of interests, or in high-density blocks with a 17.5 µm resolution. Due to constraints given by the technology, mainly the minimum pitch between metal wires and the number of metal layers, the largest high-density blocks can contain 23 × 23 electrodes, which is 5 times larger than what was possible with the design in Frey et al.(2010), where high-density blocks were limited to 6 × 17 electrodes. Large switch on-resistance can negatively affect the performance of the recording or stimulation. Transmission gates with around 1 kΩ on-resistance were chosen as switches for the given pixel size. All readout channels can be connected to randomly selected electrodes through an average of 4.4 switches. Only in a few configurations, up to ∼20 switches are required to route some electrodes.√ Even in such cases, the switches contribute a noise density of about 18 nV/ Hz, which is still lower than the targeted noise level of the readout units. The stimulation units can also be directly connected to the electrode through switch S3 for a low- resistance path. This switch can be activated through an SRAM cell residing at one side of the array. Custom-made CAD software was developed to design and implement the switch matrix. The performance of different designs was evaluated, in terms of electrode selection flexibility and the shortest paths between electrodes and readout units, using a mathematical graph representing the array. Wires and switches of the

18 Chapter 2. CMOS Circuitry and Architecture

Figure 2.4: Diagram of a subset of the switch-matrix. The larger rectangles represent the electrodes; the smaller rectangular dots represent the switches S1 and S2, used to connect the electrodes to the signal wires and to connect the signal wire segments.

arrays were mapped to nodes and arcs of the graphs respectively. The connectivity between electrodes and readout channels was modeled as “flow” and the number of overall switches used as “cost” in the algorithm. In order to determine the routing paths and readout channels for all selected electrodes, an algorithm similar to the one in Frey et al.(2010) was used. A max-flow min-cost problem is solved through Integer Linear Programming. A variety of electrode configurations, such as randomly chosen electrodes, as well as specific electrode patterns, like large contiguous blocks were evaluated.

The physical layout was then automatically generated from the graph representation, starting from a template containing the electrode, the switches, the SRAM cells, and the wires. Based on the conﬁguration of the mathematical graph, a diﬀerent set of vias and short-track segments was used in each pixel to join the signal wires of adjacent pixels into 24-pixel-long segments, and to provide the wire-to-switch connectivity. A pattern of 24 × 24 pixels, formed in this way, was replicated to construct the whole array. Fig. 2.4 shows a subset of the switch- matrix. Including the periphery of the array, a total of 86,000 switches controlled by 59,000 SRAM cells are used.

The same software was used during experiments to conﬁgure the state of the switches and to program a conﬁguration into the SRAM cells.

19 2.4. Readout

2.4 Readout

Good gain uniformity across all channels is desired to reconstruct the actual signal amplitudes and cell positions (Frey et al.(2009b); Holt and Koch(1999a)). Closed-loop amplifier topologies were preferred over open-loop solutions to ensure gain uniformity without the need for calibration. To achieve an overall gain of more than 70 dB, three amplification stages have been employed to reduce the area of passive devices. The schematic of a readout channel is shown in Fig. 2.5. In the first stage, a low-noise amplifier (LNA) provides a gain of 24 dB and high-pass filtering to reject the electrode offset. The second stage is a variable-gain amplifier (VGA), employing a digitally-assisted offset compensation scheme to cancel the output offset of the LNA. Low-pass filtering is implemented in two steps: the VGA limits the noise bandwidth and provides anti-aliasing filtering, whereas a multirate SC filter (SC LPF) further reduces thermal noise and provides precise control over the cutoff frequency. Fully differential structures were employed for the whole readout chain, to improve rejection of power-supply interference and substrate coupling, and to reduce power consumption in the SC LPF (Schreier et al.(2005)). The area and power breakdown of a readout channel are shown in Fig. 2.6.

2.4.1 Low-Noise Ampliﬁer (lna)

AC coupling is employed to remove the offset and low-frequency drifts of the electrode potential. While alternative solutions to reset (Imfeld et al.(2008a)) or compensate (Muller et al.(2012)) the input offset of the front-end amplifier have been proposed, these solutions introduce step-shaped artifacts in the recorded traces, as an abrupt change of the DC level in the output waveforms is introduced upon reset or compensation. To avoid such artifacts in the signals, continuous-time filtering was preferred for our design. An input capacitance of 1.45 pF, implemented with stacked poly-poly and MIM capacitors, was chosen for reduced area usage (196 µm2 for both branches) and high input impedance (110 MΩ at 1 kHz). A small input capacitance also reduces the voltage attenua- tion caused by the capacitive divider formed with the electrode-electrolyte double layer capacitance (Franks et al.(2005a)) ( ∼0.2 pF/µm2 for Pt electrodes Robin- son(1968)). A low high-pass cutoff frequency fHP has been obtained with MOS pseudoresistors (Harrison and Charles(2003b)) in parallel with 89 fF capacitors. The frequency fHP is tunable by adjusting the gate voltage (VMOSR) of transistor MR (Heer et al.(2006a); Mohseni and Najafi(2004)) and can be set as low as 100 mHz. The possibility to tune fHP allows for increasing the dynamic range in experiments in which LFP recording is not required. The switches activated by the Reset signal are employed to quickly recover from amplifier saturation after a stimulation pulse (Heer et al.(2007d)). An alternative scheme is offered

20 Chapter 2. CMOS Circuitry and Architecture

Figure 2.5: Schematic of a readout channel. The timing diagram for the multirate sc circuit in the third stage is also shown for diﬀerent values of the multirate factor (m = 1, 2, 4).

21 2.4. Readout

Figure 2.6: Area and power breakdown of a readout channel. The VGA dac power contribution is for maximum offset compensation applied. The ADC logic includes the digital latches and output buffers, as well as contributions from shared address decoders. by input switches, controlled by the Disconnect signal, used to disconnect the amplifier prior to stimulation, thus preventing saturation.

A telescopic-cascode OTA, with the common-mode feedback (CMFB) loop using transistors in the triode-region, is employed in order to minimize the number of current branches. All bias currents flow through the input transistors contributing to their transconductance. The reduced output swing is not an issue in this case due to the small signal amplitudes. For a given current budget, the thermal noise of both the input transistors (MPI ) and the active load (MNB) is reduced by increasing the gm/ID of MPI . Since operating in weak inversion requires a very large W/L ratio, MPI were sized for moderate inversion, with a gm/ID of 25 V −1, as a tradeoff between transconductance efficiency and area. The noise contribution of MNB was reduced by operating them in strong inversion, with −1 a gm/ID set to 3.8 V . For the input transistors, a relatively short length of 1.1 µm was chosen to limit area usage. With this choice the 1/f-noise corner occurs at 300 Hz, the lower limit of the AP signal band. Higher noise levels can be tolerated in the LFP frequency band, due to larger signal amplitudes (Muller et al.(2012); Heer et al.(2006a); Venkatraman et al.(2009b)). Further increasing the gate area of MPI also results in a larger input capacitance Cin for the OTA. The input-referred noise PSD of the closed-loop amplifier, Sn,LNA, is related to the noise of the OTA, Sn,OT A, by the relation:

2 C1S + C1F + Cin Sn,LNA = Sn,OT A. C1S

22 Chapter 2. CMOS Circuitry and Architecture

Figure 2.7: Schematic of the dda in the variable-gain amplifier. An additional input differential pair is used for offset compensation.

Therefore a too large Cin can degrade the noise in the closed loop conﬁguration (Jing Guo(2013); Venkatraman et al.(2009a)).

2.4.2 Variable-Gain Ampliﬁer (vga)

The offset of the LNA, which is mainly caused by the mismatch in the resistances and the leakage currents of the pseudoresistors, can saturate the amplification chain. Performing high-pass filtering in the second stage with pseudoresistors can result in large harmonic distortion due to their non-linearity at larger signal amplitudes. The distortion can become severe at low frequencies, since the total har- 2 monic distortion (THD) depends on the frequency as THD ∝ 1/ [1 + (f/fHP ) ] (Guo et al.(2012c); Yuan et al.(2008)). To avoid this issue, here we employed instead a DC-coupled amplifier with digitally-assisted offset compensation. A differential-difference amplifier (DDA) with resistive feedback was used to provide high input impedance in a fully-differential structure. Poly-resistors of 10 kΩ/ allowed for high gain in a small area. The gain can be programmed within the range 0-30 dB, with increments of 6 dB. The DDA is based on a folded- cascode topology (see Fig. 2.7). An additional input differential-pair in the DDA, driven by a 6 bit DAC, is used to implement channel offset compensation without degrading the PSRR or the CMRR of the VGA. The offset compensation is performed off-chip using a binary search algorithm (Muller et al.(2012)). At each step of the binary search, the individual bit values of all channels are determined off-chip simultaneously; then, these bit values are programmed into the registers for the compensation DACs sequentially through the SPI-like interface. Since the compensation is only used to reduce the offset of the readout circuits, whereas the electrode offset and drift are removed by AC coupling, the compensation procedure needs to be applied only once per measurement session.

23 2.4. Readout

Figure 2.8: Total harmonic distortion of the vga used in the second stage of the readout. Comparison between the dda-based implementation, used in this design, and an high-pass ﬁlter (hpf) topology using mos pseudoresistors.

Since accurate models for the oﬀset contributed by the pseudoresistors in the LNA are not available, the compensation range can be controlled globally by adjusting the shared bias current of the DACs. An alternative design for the VGA in the second stage, based on the same HPF topology of the LNA, was also implemented for comparison. The THD of the two designs is shown in Fig. 2.8. In the HPF topology, the distortion becomes severe below 300 Hz, where the THD exceeds −45 dB.

2.4.3 Switched-Capacitor Low-Pass Filter (sc lpf)

The bandwidth of the VGA is susceptible to variations in process and bias current, and is inversely proportional to the closed-loop gain. To ensure a precise low-pass cutoff frequency, SC filtering is used in the third stage of the amplification chain. A multirate operation scheme was employed that allows for boosting the gain with a reduced capacitance spread, without impacting noise performance or circuit complexity, and led to a compact implementation based on a single OTA. A low clock rate signal (fF =60 kHz, 80 kHz, 100 kHz) is used to control the switches in the feedback path to obtain a cutoff frequency around 5 kHz with a small C3F /C3I ratio. The input signal is, instead, sampled at a frequency fS = MfF , which can be set higher (M =1, 2, 4) in order to reduce

24 Chapter 2. CMOS Circuitry and Architecture the kT/C noise of the input switches. The transfer function of the SC LPF is given by

M P −i C3Sz i=1 H (z) = −M C3I + C3F − C3I z with z = ej2πf/fS . By sampling and integrating the input signal M times before leaking charge through C3F , a gain of fSC3S/fF C3F is obtained. Such a scheme results in a low capacitance spread between C3S, C3I and C3F . In this design, C3S = 4CU , C3I = 2CU , C3F = 1CU for a total of 14 unit capacitors CU . In contrast, a conventional SC circuit with a clock rate of MfF would require 2×(6M +1) unit capacitors to achieve the same gain, power consumption, cutoﬀ frequency and noise performance (50 CU , for M = 4).

2.4.4 Analog-To-Digital Converter

Recording spikes with amplitudes of tens of µV , superimposed on LFPs with amplitudes up to a few mV, requires a resolution of at least 9 bit. A resolution of 10 bit was chosen for the single-slope ADC, as a trade-off between resolution and clock rate. The comparators consist of three gain stages with auto-zeroing. Despite a larger static power consumption compared to dynamic comparators, continuous-time comparators were chosen to avoid large kickback noise, since the ramp signal is shared among 1024 ADCs. A capacitive neutralization technique is used in the first gain stage to further reduce the kickback (Figueiredo and Vi- tal(2004)). The input signal is sampled on capacitors CC3 during the φF 1-phase of the SC LPF. During the count phase, a continuous ramp signal is produced by integrating a constant current (IRAMP ) onto a 20 pF capacitor (CRAMP ). A schematic of the ramp generator is shown in Fig. 2.9. IRAMP is generated by a current conveyor applying a reference voltage VR across RRAMP . The upper bound of the ADC range is determined by VST ART . The lower bound is determined by the final value of the ramp, vRAMP (TRAMP ) = VST ART − VRTRAMP /(RRAMP CRAMP ) and is, therefore, subject to process and temperature variations in RRAMP and CRAMP , if VR is fixed. These variations are eliminated by means of a negative feedback loop. The difference between vRAMP (TRAMP ) and the target voltage VEND is sampled on CE at the end of each sample frame. The charge on CE is then transferred onto CS, shifting the voltage VR by (CE/CS)[vRAMP (TRAMP )−VEND]. After a few sample frames, vRAMP (TRAMP ) converges to VEND, and the slope of the ramp equals dvRAMP /dt = VEND/TRAMP . The single-ended ramp at the output of the integrator is converted to a differential signal by a capacitive-feedback amplifier, whose gain can be varied in steps of 0.25 from 0.25 to 2.0 for coarse regulation of the ADC range. The timing diagram of the ADC is shown in Fig. 2.10. During the ramp phase, the amplifier performs single-ended to differential conversion by setting φR = 1 and φS = 0. At the end of each sample

25 2.4. Readout

Figure 2.9: Schematic of the shared ramp generator in the adc.

Figure 2.10: Timing diagram of the single-slope adc.

frame, the amplifier is auto-zeroed by setting φS = 1. The output common-mode of the differential ramp is set by a SC CMFB. The simulated power consumption of the ramp generator is 1.14 mW, including the output buffer. The shared counter runs with a 24 MHz clock signal (1200 clock cycles per sample frame). Each count phase lasts 1024 cycles for 10 bit operation. All switching operations in the comparators and ramp generator (auto-zeroing, calibration, CMFB of the output buffer) occur only during the additional 176 clock cycles after the count phases to avoid glitches in the ramp. Gray code is used for the Count signal, to avoid the acquisition of spurious values when the comparator triggers at transitions between two consecutive codes.

26 Chapter 2. CMOS Circuitry and Architecture

Figure 2.11: Schematic of the stimulation unit.

2.5 Stimulation units

Each stimulation unit can be configured to provide either voltage or current stimulation (Fig. 2.11), as was also done in a previous version (Livi et al.(2010)). The core of each unit is a class-AB opamp, capable of driving loads as large as 10 nF, while maintaining a low static power consumption. In the voltage mode, the circuit is configured as an inverting amplifier with low output impedance. In the current mode, it is configured as a type II current conveyor. The input resistance RIN can be set to either 20 kΩ or 200 kΩ, for coarse adjustment of the current range. Cascoded transistors in the output branch enhance the output impedance to keep the output current constant in the presence of a varying electrode voltage. A pre-level-shifter at the output of the opamp eliminates the cross-over distortion, which has been observed in Livi et al.(2010). The stimulation units also include an auto-zeroing scheme for offset compensation. In current mode, low offset is crucial, because an offset current can quickly drive the electrode voltage to either VDD or ground, or can induce undesirable electrochemical processes at the electrodes, whose reaction products, such as oxygen or hydrogen, can harm the cells or tissue samples (Jochum et al.(2009)). Since most components are shared in the two modes, each stimulation unit occupies only 0.055 mm2.

27 2.6. Chip implementation

Figure 2.12: Layout of a readout block (32 channels), chip micrograph, close-up view of the electrode array and packaged device. The die size is 7.6 × 10.1 mm2

2.6 Chip implementation

2.6.1 Layout and Floor Plan

The recording ampliﬁers and the parallel ADCs are grouped in blocks, each comprising 32 channels and shared logic and bias circuits (see Fig. 2.12). In each block, the ampliﬁers are arranged in four rows per stage, to reduce the aspect ratio of the layout and the perimeter/area ratio of the capacitors. Sensitive analog signals are routed via the top metal (MET4), and shielded from underlying circuits by MET3 planes, which carry supply and reference voltages.

2.6.2 Bias and Test Structures

Bias currents are routed to each readout block and fed by programmable bias generators. These currents can be varied independently for each stage. In the LNA, power consumption can be traded oﬀ with noise levels (Wattanapanitch and Sarpeshkar(2011)); in the VGA, it can be used for global tuning of the low- pass corner and to adjust the oﬀset-compensation range. Each readout block can be powered-down independently, to permit long-term continuous monitoring with a small number of electrodes at low power levels.

2.6.3 Chip Fabrication

The chip was fabricated in a 0.35 µm CMOS technology (2P4M). Platinum electrodes were post-processed at wafer level by means of ion beam deposition and

28 Chapter 2. CMOS Circuitry and Architecture

Figure 2.13: Sem image of the chip surface, showing in-house post-processed pt- electrodes, plated with rat cortical neurons.

Figure 2.14: Measured frequency response of the ampliﬁcation chain in the readout channels, for four possible gain settings.

29 2.7. Measurement etching. In the same step, three Pt-resistors were fabricated on top of the CMOS passivation for use as temperature sensors. To protect the underlying circuits from the saline solution, used as biological medium, and to avoid cell contamina- tion by the aluminum contained in the CMOS process, a multilayer SiO2/Si3N4 passivation stack was deposited by plasma-enhanced chemical vapor deposition (PECVD). Openings in the passivation to the platinum, deﬁning the actual electrode areas (9.3 × 5.4 µm), were then obtained in a reactive-ion etching (RIE) step. A shifted-electrode layout (Heer et al.(2004a)) was employed to prevent any leakage of aluminum into the biological medium. Fig. 2.12 shows a micrograph of the chip and close-up view of the fabricated electrodes. The chip was die-bonded on a custom PCB. A polycarbonate ring was used to contain the biological medium and a bio-compatible epoxy was used to encapsulate the bond-wires (Frey et al.(2010); Heer et al.(2007a)). In Fig. 2.13, an SEM image of the chip surface, plated with rat cortical neurons, is shown.

2.7 Measurement

2.7.1 Electrical Characterization

The frequency response of one readout channel for four possible gain settings is shown in Fig. 2.14. The measured maximum gain is 78.3 dB. The spread of the response across all channels was characterized by applying a common signal to all inputs. Owing to closed-loop topologies and SC filtering, very good uniformity in both the gain and the low-pass corner has been obtained, as shown in Fig. 2.15. The input-referred noise PSD of the readout chain, including√ the ADC, is shown in Fig. 2.16. The noise spectral density is 39 nV/ Hz at 1 kHz. The noise integrated over the band 1 Hz-10 kHz is 5.9 µVrms. In the LFP band (1 Hz-300 Hz) the noise amounts to 5.4 µVrms, whereas in the AP band (300 Hz-10 kHz) it is 2.4 µVrms. When filtered in the band 500 Hz − 3 kHz, for spike detection, the noise is 1.8 µVrms. The CMRR of the readout was obtained from measurements on all 1024 channels, resulting in an average of 72 dB. The response of the ADC to a 1.1 kHz sine wave is shown in Fig. 2.17. The ADC achieves an SNDR of 59 dB and an SFDR of 68.9 ∼ dB. Kick-back from the comparators on the shared ramp has been observed. The SNDR degrades by a maximum of 8 dB in the worst-case condition, which occurs when all comparators toggle simultaneously. The performance of the stimulation unit was assessed with typical biphasic waveforms used for eliciting electrical activity in neurons (Fig. 2.18). Loads as large as 10 nF can be driven in the voltage mode with pulse durations of 250 µs. In the current mode, the cross-over distortion is eliminated; channel length modulation effects were also reduced with respect to the design in Livi et al.(2010). To quantify the accuracy of the amplitude of the stimulation pulses, the static linearity has been characterized by sweeping an input DC voltage and extracting the residuals of a best-fit line. In the voltage mode, 10 bit linearity within an

30 Chapter 2. CMOS Circuitry and Architecture

Figure 2.15: Histograms of the gain (at 1 Hz, 10 Hz and 100 Hz) and the low-pass −3 dB cutoﬀ frequency. A nominal gain of 16×16×4, with a multirate factor m = 2 and ff = 100 kHz, was used for this measurement.

Figure 2.16: Input-referred noise psd of the full readout chain, including the adc, with and without electrode contribution.

31 2.7. Measurement

Figure 2.17: Measured ﬀt spectrum of one adc output for an input sine wave at [ 1.1]kHz.

Supply Circuit Blocks Power [mW] AVDD 1 LNA, VGA, RampGen 36.8 AVDD 2 SC LPF, ADC Comparators 19 DVDD 1 ADC Memory, Decoders, Logic Buﬀers 6.4 DVDD 2 Core, I/O 12

Table 2.2: output range of 3 V was achieved. In the current mode, the linearity is 9 bits within an output range of ±50 µA. The measured total chip power dissipation is 75 mW at a supply voltage of 3.3 V. Table 2.2 shows the breakdown for the diﬀerent supply domains. The power consumption of the stimulation units is largely dependent on the applied stimuli, due to its class-AB operation. Also, these blocks can be powered down during experiments that do not require stimulation. Since low power dissipation is crucial for the survival of cells cultured on the chip’s surface, the temperature increase of the device, ﬁlled with PBS and placed inside an incubator, was monitored by means of the on-chip Pt temperature sensors. An increase < 2°C was observed when all channels were powered up (∼ 0.2°C for each group of 128 channels), demonstrating that the chip’s self-heating is low enough and thus suitable for cell cultures, without requiring active cooling.

2.7.2 Biological Measurements

The devices were further veriﬁed with in vitro and ex vivo measurements. Por- tions of acute ex vivo rabbit retina were placed on the chip surface, and spontaneous electrical activity was successfully recorded. A raw trace, as recorded

32 Chapter 2. CMOS Circuitry and Architecture

Figure 2.18: Biphasic stimulation waveforms in voltage and current mode.

33 2.8. Comparison To State-of-the-art and Conclusion

Figure 2.19: Acute ex vivo recordings of action potentials from retinal ganglion cells. A raw trace is shown on the left. On the right, a temporal zoom-in of the same trace is shown; recorded measurement values are represented as dots. by the CMOS MEA without further processing, is shown in Fig. 2.19. Ac- tion potentials from retinal ganglion cells (RGCs) were detected. The recorded samples are marked with dots in the time zoom. Large-scale recordings were performed with cultures of cortical neurons. The neurons were isolated from rat brain and plated on the surface of MEA chips, which were pre-coated with poly(ethylenimine) and laminin. Spontaneous activity was observed, and APs were detected with a threshold 5.5 times above the noise rms. Fig. 2.20 shows a portion of the electrode array, with spike amplitudes obtained from simultaneous recordings of two distant high-density patches, consisting of 23 × 23 and 15 × 15 electrodes, respectively. The average spike shapes of two neurons with overlapping electrical footprints are also shown. The low-noise characteristics of the recording channels, combined with high spatial resolution, allowed for identifying and separating the individual signal sources. The stimulation capability was veriﬁed by applying biphasic voltage pulses (positive ﬁrst) with a unit200µs phase duration and peak-to-peak amplitude of 800 mV. Fig. 2.21 shows an overlay of 26 raw traces, subsequently recorded on an electrode located 288 µm from the stimulation site. Spikes were reliably elicited ∼8 ms after the stimulation pulse.

2.8 Comparison To State-of-the-art and Con- clusion

The performance of the chip has been summarized and compared to that of other CMOS microelectrode arrays in Table 2.3. Our device achieves state-of-the-art noise characteristics (2.4 µVrms in the frequency band of APs) while maintain-

34 Chapter 2. CMOS Circuitry and Architecture

Figure 2.20: Data from cortical neurons, cultured on the chip surface for seven days. Left: spike amplitude map obtained from simultaneous recordings in two regions of interest. Right: close-up of the area within the red rectangle recorded at maximum spatial resolution; average spike shapes of individual neurons are shown; contour curves used to identify neighboring neurons represent the amplitude of the spike negative peak.

Figure 2.21: Response of a cortical neuron culture to voltage stimulation, recorded at 288 µm from the stimulation electrode. 26 traces of subsequent stimulations and elicited spikes are shown (one trace is shown in black, other traces are in gray).

35 2.8. Comparison To State-of-the-art and Conclusion

ing high spatial resolution (17.5 µm pitch) and low power consumption. De- spite the use of closed-loop topologies, which were adopted to ensure high gain uniformity without the need for calibration, each recording ampliﬁers occupies a very small area (0.033 mm2) and consumes little power (31 µW). The area and power eﬃcient design of the readout channels allowed for the integration of 1024 of such units, which is about 8× the channel count of the switch-matrix – based design reported in Frey et al.(2010). This channel count also exceeds that of all neural acquisition ICs reported in literature (e.g., Aziz et al.(2009); Shahrokhi et al.(2010); Wattanapanitch and Sarpeshkar(2011); Muller et al. (2012); Gao et al.(2012); Guo et al.(2012c); Jing Guo(2013); Aziz et al.(2007)), for which the integration of only up to 256 channels has been demonstrated. The simultaneous signal acquisition at many recording sites facilitates the reconstruction of interconnections in neural networks. The presented device also features the largest sensing area (8.1 mm2), which permits the simultaneous recording of large patches in distant regions, to investigate long-range interactions between sub-networks. Stimulation units with both voltage and current stimulation capabilities have also been integrated on chip. All these features make the whole MEA system a versatile platform for numerous biological applications. The chip was used to successfully record activity from a variety of biological preparations, validating the suitability of the device for high-throughput electrophysiological measurements.

Acknowledgement

The authors would like to thank Dr. F. Franke, Dr. T. Russell, ETH Zurich, and Dr. C. Hagleitner, IBM Research, Zurich, for discussions and help with neuronal cultures. The authors also thank M. Duggelin¨ and D. Mathys, University of Basel, for providing the SEM image. This work was supported by the Advanced ERC Grant “NeuroCMOS” under contract number AdG 267351. M. Ballini, W. Gong, and A. Shadmani received individual support through the Marie Curie Research Training Networks “CellCheck” and “EngCaBra”. M. Radivojevic and D. J. Bakkum received funding support from the Swiss National Foundation through an Ambizione Grant (PZ00P3 132245).

36 Chapter 2. CMOS Circuitry and Architecture 2 s m m rms rms rms / 10kHz) µ µ 300Hz) 10kHz) V V V 5 − . µ µ µ V/I − − 35 3265 1024 . 9 4 4 09mm 3.3 V 26400 10 bit . . . . 20kS 17 75 mW 0 5 5 2 8 This work single-slope Switch Matrix (1Hz (1Hz (300Hz ) 2 kHz rms µm µV 10 mm - - kS/s µV 126 [19] V/I − 3150 8 bit 4 11016 . 0.6 20 16 SAR 17.8 135 mW 2 3.50 Hz Switch Matrix (1 5 V (3.3 V digital) 2 rms mm µm mum - - - - µV 567 [16] kS/s 4096 4096 3.3 V 8 42 oﬀ-chip 132 mW 26 7.13 0.35 Active Pixel 2 µm mm - - - - mum kS/s [18] 5 V 4 W 4 9 bit 775 . 32768 12987 32768 60 . . 2 8 0.5 2 128 pipeline Active Pixel Reference Type Technology Supply Voltage Sensing Area No. Transducers Pixel Pitch Transducers/mm2 No. Channels A/D Conversion ADC resolution Frame Rate Total Input Noise LFP band Input Noise AP band Input Noise Stimulation Total Power

Table 2.3: Comparison to other CMOS microelectrode arrays

37 2.8. Comparison To State-of-the-art and Conclusion

38 Chapter 3

High-resolution CMOS MEA platform to study neurons at subcellular, cellular and network levels

Jan Muller¨ 1, Marco Ballini1, Paolo Livi1, Yihui Chen1, Milos Radivojevic1, Amir Shadmani1, Vijay Viswam1, Ian L. Jones1, Michele Fiscella1, Roland Diggelmann1, Alexander Stettler1, Urs Frey2, Douglas J. Bakkum1, Andreas Hierlemann1

Lab on Chip, 2015

1Bio Engineering Laboratory, ETH Zurich, Switzerland

2RIKEN Quantitative Biology Center, Kobe, Japan

Abstract — Studies on information processing and learning properties of neuronal networks would benefit from simultaneous and parallel access to the activity of a large fraction of all neurons in such networks. Here, we present a CMOS- based device, capable of simultaneously recording the electrical activity of over a thousand cells in in vitro neuronal networks. The device provides sufficiently high spatiotemporal resolution to enable, at the same time, access to neuronal preparations on subcellular, cellular, and network level. The key feature is a rapidly reconfigurable array of 26 400 microelectrodes arranged at low pitch (17.5 µm) within a large overall sensing area (3.85 × 2.10 mm2 ). An arbitrary subset of the

39 3.1. Introduction

electrodes can be simultaneously connected to 1024 low-noise readout channels as well as 32 stimulation units. Each electrode or electrode subset can be used to electrically stimulate or record the signals of virtually any neuron on the array. We demonstrate the applicability and potential of this device for various diﬀerent experimental paradigms: large-scale recordings from whole networks of neurons as well as investigations of axonal properties of individual neurons.

3.1 Introduction

To understand how neuronal networks perform information-processing tasks, such as information storage and learning, it is desirable to have means to simultaneously record the electrical activity of many single neurons and to stimulate defined neurons at the same time. There are growing efforts to record from ever larger networks of neurons (Alivisatos et al.(2013)) with the ultimate aim of recording from complete brains (Marblestone et al.(2013b)). Having simultaneous access to the activity of virtually every cell in a network will lead to a much better understanding of the functional connectivity underlying such networks (Gerhard et al.(2013)), and of how the connections within such a network change over time and exhibit plasticity. Besides the collective activity of cell assemblies, cellular features, such as synaptic plasticity (Sjöström et al.(2008)) or intrinsic excitability (Zhang and Linden (2003)), are also relevant for information processing in neurons. Recent evidence suggests that more subtle cellular aspects, such as axonal information processing (Debanne(2004)), changes in propagation velocities (Bakkum et al.(2008c); Izhikevich(2006)), and changes in spike shapes (Alle and Geiger(2006)), may contribute to a rich set of modalities for tuning population dynamics. Thus, information processing in neurons at the network level also depends on properties of individual cells. Commercially available microelectrode arrays (MEAs) are an established technology for recording from networks of neurons (Gross et al.(1995); Stett et al. (2003b)). However, due to their limited spatial resolution (pitch ≥30 µm) and number of electrodes (usually less than 300), such passive MEAs typically do not allow for recording from targeted individual neurons in large networks. Recently, a different class of MEAs, based on complementary metal-oxide-semiconductor (CMOS) technology, has been developed to address some of these issues (Berdon- dini et al.(2009c); Bertotti et al.(2014); Eversmann et al.(2003a); Frey et al. (2010); Hierlemann et al.(2011); Huys et al.(2012)). By integrating circuitry on the same substrate as the recording electrodes, CMOS MEAs can overcome some of the inherent limitations of passive MEAs. Most importantly, CMOS MEAs allow for overcoming the connectivity problem so that thousands of microelectrodes can be arranged at high spatial resolution through using multiplexing techniques, whereby electronic switches are employed to access shared signal wires. This ap-

40 Chapter 3. Neurons at subcellular, cellular and network levels proach drastically reduces the number of required interconnections between electrodes and amplifiers, thus allowing for a more effective use of available routing area. By integrating the amplifiers and analog-to-digital converters (ADCs) on the same substrate as the electrodes, the number of off-chip connections can also be reduced, since the digitized signals can be sent off-chip sequentially through only a small number of connections. Moreover, since the signals are amplified and filtered close to the signal source, the influence of noise picked up during signal transmission is minimized. However, CMOS MEAs developed so far are limited either in noise performance (Berdondini et al.(2009c); Eversmann et al. (2003a, 2011)), spatial resolution (Berdondini et al.(2009c); Heer et al.(2006a)), or suffer from a comparably low readout channel count (Frey et al.(2010)). In this work, we present a CMOS-based high-density MEA (HD-MEA) device capable of recording and stimulating with bidirectional microelectrodes at high spatial resolution and high signal-to-noise ratio. We circumvent the tradeoff between electrode pitch and readout noise performance by further advancing an approach of Frey et al.(2010). Instead of packing readout circuitry underneath each electrode, we use the available area below the electrodes to implement programmable routings of electrodes to readout channels. Placing the readout and stimulation units at the periphery of the sensor array decouples (i) the electrode pitch from area constraints for readout and stimulation circuitry and (ii) the number of readout and stimulation units from the available number of electrodes. This enables us to implement a large sensing area (8.09 mm2), suitable for place- ment of, for example, large acute preparations including retina patches or brain slices. Compared to Frey et al.(2010), the device design here features 8 times more parallel readout channels (1024 total), more than twice as many microelectrodes (26 400), and an increased flexibility to freely choose specific recording sites or subsets of those. Furthermore, fivefold-larger contiguous patches of neighboring electrodes (23 × 23 electrodes ∼=402 × 402 mm2∼= 0.16 mm2) at arbitrary positions can be connected to readout channels, and every individual electrode can be electrically stimulated. The pivotal feature of our new device is the large flexibility in configuring the electrodes, as various biological preparations have distinct requirements in terms of distribution of recording sites and spatial resolution. Some preparations, such as networks of cultured neurons, may require sparsely distributed recording spots, whereas retinal patches with densely packed ganglion cells, for example, will require high-density arrangements to resolve the populations of different cell types that form mosaic-like repetitive structures (Wässle(2004)). We demonstrate that the CMOS-based HD-MEA presented here is well suited to accommodate such different preparations in that it allows for selection of the most suitable electrodes for a particular experiment. The details of the device circuitry have been described Ballini et al.(2014), so that we only briefly abstract the CMOS HD-MEA circuitry here, while we focus on the details of the flexible-system architecture and on demonstrating the related device performance. In particular, we will show how the device can be used to record at different levels of spatial

41 3.2. Materials and methods

Figure 3.1: System Architecture. CMOS MEA system architecture. (a) Micrograph of the CMOS device (10.1 × 7.6 mm2 ). The 1024 readout channels are arranged at the top and bottom of the 3.85 × 2.10 mm2 microelectrode array. The 32 stimulation units (S) are located on the left and right side of the array. (b) Block diagram of the implemented circuitry. A close-up view into the electrode array shows details of one individual pixel including the 9.3 × 5.45 µm2 platinum electrode, a two-bit SRAM cell, and two switches. resolution from neuronal preparations grown or placed over the electrode array. Once a preparation overview has been gained by systematically scanning the full array, all putative single neurons of the preparation can be identified, and global recordings on, e.g., the network level can be performed. It will be demonstrated how the flexibility in recording electrode selection helps to improve spike sorting yield and permits the recording of neuronal activity with fewer electrodes than targeted neurons. Techniques to find and record from subcellular structures, such as axonal arbors of single neurons, will be presented and used to analyze the propagation of axonal action potentials (Bakkum et al.(2013b)). Finally, the array performance will be demonstrated by stimulating an axonal segment with electrical pulses while tracking the evoked neuronal activity over multiple axonal branches of the same neuron at high spatial resolution over a distance exceeding 1.5 mm.

3.2 Materials and methods

3.2.1 2.1. Microelectrode array

The CMOS device as depicted in Fig. 3.1 features an active sensing area of 3.85× 2.10 mm2 with 26,400 platinum microelectrodes. The electrodes are arranged in a grid-like conﬁguration with a center-to-center pitch of 17.5 µm, yielding an electrode density of 3265 microelectrodes per mm2 Below each of the 9.3 × 5.45

42 Chapter 3. Neurons at subcellular, cellular and network levels

µm2 platinum electrodes are two SRAM cells and switches, which can be used to (i) connect the electrode to one out of 12 metal tracks and (ii) connect two different metal tracks together. Below all electrodes, a matrix consisting of a total of 86,000 switches has been implemented, which is controlled by 59,000 SRAM cells (Muller¨ et al.(2013b)). This matrix is instrumental in connecting arbitrary subsets of electrodes to the readout and stimulation units residing at the periphery of the sensing area. In order to increase routing flexibility and to accommodate for potentially arbitrary electrode-to-readout mappings, routing wires are on average 420 µm long. Every wire connects through switches to at least four other wires, so that possible routing options grow exponentially with every additional wire in the routing path. For instance, one electrode has at least 256 (4 × 4 × 4 × 4) different options to transmit signals through a 1680 µm (i.e. 4-wire)-long path. Having many different options for a given electrode’s routing path enables efficient and flexible use of resources (such as switches and wires) for realizing a large variety of potential electrode configurations. Due to the available area for wiring, the largest number of adjacent electrodes in a configuration includes a contiguous block of 23 × 23 electrodes, in an area of 402 × 402 microm2 at arbitrary positions of the 3.85 × 2.10 mm2 array. Multiple such high-density blocks (e.g., 23×23 plus 22×22) can be selected simultaneously.

3.2.2 Reconﬁguration of the microelectrode array

To accommodate different experimental needs, the microelectrode array can be reconfigured within milliseconds. Fig. 3.2 on the following page A shows a subset of the array with its electrodes, switches, and wires used to implement the flexible routing. Fig. 3.2 B shows an example mathematical graph representing a drastically simplified part of the array. In this example, three out of the four available electrodes can be connected to readout channels by closing the appropriate switches. As it is impossible to manually determine the best routing paths from all 1024 selected electrodes to the respective readout units, a custom software algorithm has been developed to calculate optimum routing paths and to set the states for all 86,000 switches. Fig. 3.2 C shows a flowchart of the steps involved to record from a certain, freely selectable subset of all available 26,400 electrodes. First, custom software maps the electrode array into a mathematical graph. Each electrode and wire is represented as a node in the graph, and each switch as an arc between two nodes. Then, an integer linear programming (ILP) algorithm optimizes a max-flow min-cost problem (Frey et al.(2010)), i.e. the number of readable recording electrodes is maximized (max-flow), while the number of used resources (switches, wires) is minimized (min-cost). To this end, the algorithm needs to apply a set of constraints, such as (i) each assigned electrode is a signal source, (ii) every available readout channel is a signal sink, and (iii) no more than one routing signal is allowed per node and arc. Once the algorithm found a solution to the constrained optimization, the serial interface of the CMOS MEA is employed to download the configuration to all switches in

43 3.2. Materials and methods

Figure 3.2: Reconfigurable Electrode Array (a) Schematic of the wiring in a part of the array. Light gray squares represent electrodes; black squares represent switches controlled by SRAM cells. Buses of 6 horizontal and 6 vertical wires are arranged per row and column. The area highlighted with the red square corresponds to one pixel unit. (b) Simplified mathematical graph showing a drastically reduced subset of the array. Three of the four electrodes picking up neural signals are connected through a set of switches and wires to three readout channels by closing the respective switches. (c) Flow chart of a typical experiment run. After choosing 1024 out of 26,400 electrodes, the hardware is mapped into a mathematical graph representing the array. Within this graph, an integer linear programming (ILP) max-flow min-cost problem is solved, and optimal signal routing paths are determined. The corresponding switch configuration is then downloaded into the CMOS MEA, and neural activity can be recorded. Once the activity on all 26,400 electrodes has been analyzed, best-suited recording electrode candidates are determined so that the final experiment can be performed.

44 Chapter 3. Neurons at subcellular, cellular and network levels the chip, and neural data can be recorded. As depicted in Fig. 3.2 C, a typical experimental session consists of iterative execution of the ﬂowchart. The opera- tor chooses initial sets of 1024 electrodes either randomly, or according to some scanning scheme, or manually. Neural activity, as captured by these electrodes, is analyzed for measures such as ﬁring rate, signal amplitude, or response characteristics to a given stimulus. Once an overview over the preparation under study has been attained, the most appropriate electrodes can be selected, and the actual experiment can be performed.

3.2.3 Readout and stimulation units

The 1024 readout channels consist of three amplification stages, providing a programmable total gain of up to 78 dB for the accommodation of a wide range of different signal amplitudes (Ballini et al.(2014)). The first stage includes a 1.4 pF input capacitance to provide AC coupling and to filter the offset and low-frequency drift in the electrode potential. The second stage implements a digitally-assisted offset compensation scheme to cancel output offset of the first stage. Anti-aliasing low-pass filtering is implemented through the second and third stage. The input-referred readout noise in the action-potential signal band (300 Hz – 10 kHz) is 2.4 µVrms (Ballini et al.(2014)). The amplified and band- pass filtered signals are sampled at 20 kHz with 1024 on-chip 10 bit analog-to- digital converters (ADCs). By digitizing the signals on-chip, digital codes can be multiplexed and transmitted off-chip. In contrast to analog signals, digital signals are very resistant to interference noise that may be caused by nearby measurement equipment or culturing incubators. Each packet of data is protected with a cyclic redundancy code (CRC) checksum to detect data corruption.

A total of 32 stimulation buffers (Livi et al.(2010)), capable of providing arbitrary voltage or current-controlled stimulation waveforms, resides at the left and right side of the sensing area. The stimulation units can be connected to an arbitrary selection of electrodes. Multiple electrodes can be connected together to form larger stimulation patches (Behrend et al.(2011); Eickenscheidt and Zeck(2014)). Although featuring low static power consumption, the buffers can drive loads as large as 11 nF, while a signal rise time below 50 µs for a 2.5 V step is preserved. This load corresponds roughly to 200 – 500 connected Pt bright electrodes or 5 – 20 Pt black electrodes, depending on the electrode impedances. In current mode, the deliverable current can be as large as 50 µA at a resolution of 2 nA. Three digital-to-analog converters (DACs) can be programmed to provide independent stimulus waveforms. In an alternative stimulation mode, the output voltages of the DACs can be kept fixed, but the input of the stimulation buffers can be switched between the available DAC voltages. This allows for implementation of independent waveforms for each of the 32 stimulation units with arbitrary phase timing but fixed voltage or current amplitudes.

45 3.2. Materials and methods

3.2.4 Fabrication and post-processing

The CMOS device has been fabricated in a 0.35 µm technology (2P4M) and post-processed at wafer level to (i) produce long-term stable Pt-electrodes and to (ii) further enhance the passivation layer to protect the circuitry against culturing media. During the same post-processing step, a Pt-thermoresistor was fabricated, which allows for measuring the temperature at the chip surface during experiments in order to strictly preserve physiological temperatures. The post-processing steps have been abstracted before (Muller¨ et al.(2013b)). In brief, Si3N4 was first deposited by means of plasma-enhanced chemical vapor deposition (PECVD), and the pads and electrodes were subsequently re-opened through reactive-ion etching (RIE). Next, TiW (50 nm), for promotion of adhesion, and Pt (270 nm), as electrode material, were ion-beam-deposited and then patterned by using an ion-beam etching step. A 4-layer 1.6 µm-thick passivation stack, consisting of alternating SiO2 and Si3N4 layers was deposited by PECVD; finally, a re-opening of the platinum electrodes was achieved through an RIE step. The top-metal layer below the electrode openings is free of features to provide flat structures and to ensure good connectivity and adhesion of the post-processed Pt-layer.

3.2.5 Chip preparation

Once the wafers were post-processed, and diced, the chips were bonded to custom printed-circuit boards (PCBs). A biocompatible epoxy was used to encapsulate the bond-wires and PCB tracks in order to protect them from the culturing media. This packaging makes the devices stable and suitable for experiments involving culturing of cells over periods of months. To reduce the impedance of the microelectrodes, Pt-black was electrochemically deposited on the electrodes (Heer et al.(2007a)). The on-chip Pt thermoresistor was calibrated by using a two-point calibration procedure.

3.2.6 Culturing and imaging

Cortical neurons and glia were grown over the CMOS-MEA to test the performance of the system. Brieﬂy, E18 rat cortices were dissociated enzymatically in trypsin (Invitrogen), followed by mechanical trituration. A layer of poly(ethyleneimine) (Sigma) and a layer of laminin (Sigma) were used to adhere between 20k to 40k cells. The media used for plating consisted of 850 µL of Neu- robasal, supplemented with 10% horse serum (HyClone), 0.5 mM GlutaMAX (Invitrogen), and 2% B27 (Invitrogen). After 24 hours, the plating medium was changed to growth medium: 850 µL of DMEM (Invitrogen), supplemented with 10% horse serum, 0.5 mM GlutaMAX, and 1 mM sodium pyruvate (Invitrogen). Prior to experimentation, cultures matured for 3 to 4 weeks, and experiments

46 Chapter 3. Neurons at subcellular, cellular and network levels were conducted inside an incubator to control environmental conditions (36.5°C and 5% CO2). All animal handling protocols were approved by the Basel Stadt veterinary office according to Swiss federal laws on animal welfare. To visual- ize clusters of neurons grown on the HD-MEA, cells were sparsely transfected with a human synapsin I promoter driving DsRedExpress from the Callaway laboratory, Salk Institute, La Jolla, USA (Addgene plasmid 22909) with Lipo- fectamine 2000 (Invitrogen 11668). After experiments, cell cultures were fixed in 4% paraformaldehyde (Invitrogen) in PBS (phosphate-buffered saline; Sigma) and permeabilized (0.25% Triton X-100 (Sigma) in PBS). Then, the primary antibody to MAP2 (Abcam ab5392) diluted 1:500 in PBS with 1% BSA (bovine serum albumin; Sigma) and 0.1% Tween20 was added and left overnight at 4°C on a shaker at low speed. The secondary antibody containing Alexa Fluor 647 (Invitrogen A21449), diluted to 1:200 was applied for 1 h at room temperature in the dark. A Leica DM6000 FS microscope with a 10x long-working-distance objective lens and a Leica DFC 345 FX camera were used to collected images at room temperature. For further details see Bakkum et al.(2013b).

3.2.7 Single cell identiﬁcation

The extracellular field of a single neuron can be measured up to tens of microm- eters away from the soma (Frey et al.(2009a); Gold et al.(2006a); Mechler et al. (2011)). Due to the densely spaced microelectrodes, multiple electrodes usually record the extracellular field potential of a single neuron. Exploiting this fact facilitates the task of assigning extracellularly recorded field potentials to individual neurons (Franke et al.(2012)). As it is computationally intractable to process all thousand channels at once, prior to analysis, the recording electrodes were clustered into groups of up to 16 neighboring electrodes. Spike identification and sorting was performed within these electrode groups, and, subsequently, spike trains of neurons, present in more than one electrode group, were merged based on (i) the Euclidean distance of the extracellular field potential and (ii) the similarity in the spike train time stamps. Analysis within these electrode groups was carried out as follows. First, traces were band-pass filtered (300–2500 Hz). Then, spikes were identified when the negative voltage peak crossed a threshold, which was set to 4.5 times the standard deviation of the noise (Lewicki(1998b)). Spike waveforms were cut out from 2 ms before to 3 ms after reaching their negative peak values on all electrodes of the corresponding group. The cut out traces were first up-sampled 8 times and subsequently aligned with respect to the negative peaks. Principal component analysis (PCA) was performed on the concatenated waveforms to reduce dimensionality of the data, and principal components were clustered by fitting a mixture of Gaussians through an expectation maximization (EM) algorithm as further described in Kadir et al.(2013). Once, spike times for the individual neurons have been identified, the spike triggered average on all 1024 recorded electrodes (the extracellular electrical signature of one single neuron on all electrodes, on which it is visible, is sometimes denoted

47 3.2. Materials and methods electric ‘footprint’) has been estimated by re-extracting and averaging the events from the raw data, which now have been band-pass filtered between 100 and 9000 Hz. It frequently happens that two nearby neurons spike simultaneously, or very shortly after one another, causing overlapping electrical fields. Depending on the exact relative timing of the overlaps, this causes significant distortions to the recorded traces, and spike sorting becomes more difficult. However, un- less the cells are active multiple times with the exact same relative timing, the overlaps will not produce the same compound waveform shape so that during the clustering process, these overlaps can be identified as outliers and discarded. Although there have been techniques developed to address this problem (Franke et al.(2010); Jackel et al.(2012); Pillow et al.(2013); Pouzat et al.(2002)), we decided to discard such events, as we were primarily interested in reconstructing the accurate extracellular electric footprint for every single cell and not in reconstructing the exact spike train for each cell. To this end, we computed the spike-triggered-average (STA) for single cells by taking the median over many time-aligned AP occurrences. When long recordings were analyzed (on the range of tens of minutes to hours), a slightly different approach was used, since overlaps may occur frequently, to the extent that they form their own clusters and are erroneously identified as neurons. For the first few minutes of the recordings, the clustering process described above was applied, but then template matching was performed on the remaining data using the extracted electrical footprints as templates (Franke et al.(2010); Vollgraf and Obermayer(2006)).

3.2.8 Identiﬁcation of axonal arbors

The large sensing area and the large amount of microelectrodes allow for electrical identification of potentially large axonal arbors of single cells. Compared to the relatively large amplitudes of extracellular somatic signals (hundreds of microvolts), signals from axons are very small and often buried in noise (on the order of a few microvolts). Therefore, a special technique must be applied in order to reveal such axonal signals. Once, somatic APs of a single cell have been identified through spike sorting, the axonal arbor of this cell can be reconstructed by spike-triggered-averaging of multiple instances of single axonal APs to reduce the influence of uncorrelated noise. Since not all electrodes can be read out at the same time, a scheme to scan through all electrodes is employed. To this end, a few electrodes recording with the best available signal-to-noise ratio (SNR) are used to continuously record from the soma of a given cell, while the remaining recording channels are used to consecutively scan the activity on all other electrodes. Template matching is performed on the fixed electrodes to identify spike timings of the single cell (see 3.2.7 Single cell identification). Averaging of all identified events reduces noise and helps to reveal electrical signals originating from small neurites far away from the cell body (Bakkum et al.(2013b)) For robustness and to eliminate outliers (in this case spikes from other neurons), the median was calculated instead of the mean.

48 Chapter 3. Neurons at subcellular, cellular and network levels

3.3 Results

We used the HD-MEA to analyze information processing in neuronal networks at different scales of spatial resolution. First, a coarse overview of the neuronal network was obtained. At the next level of higher resolution, a smaller area of the network was examined and analyzed in more detail, revealing many spatially overlapping different single neurons. A fluorescence image of a MAP2 staining of the respective cell culture provided the correlations between electrical recordings and optical imaging. Finally, at the highest resolution, subcellular properties of single neurons were identified, showing axonal arbors spanning over all the active sensing area. Furthermore, precise and spatially-confined electrical stimulation was used to evoke electrical activity of individual axons. When an axon is stimulated, the elicited AP can be tracked as it propagates down the axonal arbor through several different branches. As outlined above, the HD-MEA architecture is designed in a way that electrodes are not equipped with individually dedicated amplifiers. Instead an arbitrary subset out of all electrodes is connected to 1024 low-noise readout channels (including amplifiers), residing at the periphery of the sensing area (see Methods). Thus, before information processing in the neural network can be analyzed, the most suitable electrodes must be chosen; the respective selection strategies are described below.

3.3.1 Network-wide analysis

Recording of electrical activity on all electrodes gives an activity map of the biological preparation and reveals areas with no activity, which can be excluded from further analysis. Initially, dense blocks of electrodes were scanned through the array, and activity was recorded for 2.5 minutes for each such block. To cover all 26,400 electrodes, 27 different configurations were used. After the data were recorded, a threshold-crossing algorithm, applied offline, identified spike timings. In Fig. 3.3 A the neural spiking activity on each electrode is presented on a logarithmic gray scale between 1 Hz and 2 kHz. Areas with no detectable electrical signals can be excluded from further analysis, and the available recording channels can be focused on areas exhibiting activity. Once areas with active neurons have been identified, the readout channels can be connected to electrodes most suitable for recording from many neurons in parallel. To generate a suitable recording electrode selection, one of two strategies was applied: In the more time-consuming strategy, single electrodes were carefully chosen to record reliably and in parallel from as many cells as possible. The electrodes were selected in a way to maximize the number of recorded neurons and to minimize redundancies in recordings. This approach requires careful analysis of all neuronal footprints and is computationally intensive. It is further described in Section 3.3.4. Alternatively, a more heuristic approach can be employed by first estimating the cell density from the activity map. In the presence of dense

49 3.3. Results

Figure 3.3: Network of cortical neurons grown over the 3.85 × 2.10 mm2 microelectrode array area. (a) Average action potential firing rate as measured by each electrode and displayed on a logarithmic gray-scale between 1 Hz and 2 kHz. Red dots indicate the 1024 electrodes used for recording of the network activity. (b) Representation of all 2000 individual single cells that could be identified through spike sorting the signals of high-density electrode configurations. A circle is drawn around each detectable cell and indicates the level where the amplitude of the electrical signals of a cell footprint exceed -4.5 standard deviations of the electrode noise. The color-coding indicates the maximum amplitude of the most negative peak for each neuronal electrical footprint. The red rectangle indicates the area used for further analysis shown in Fig. 3.5 on page 54. (c) Fluorescence image of transfected cells. Transfection ratio was around 5% (according to the manufacturer) of all cells; therefore, only a subset of all cells lights up; clearly visible are clusters of neurons and the tracks of interconnecting neurite bundles. (d) Raster plot of 100 seconds of activity for all 1024 recording channels. The red marker indicates the time period shown in the close up view to the right. Between 38.7 and 38.8 seconds, waves of activity propagate through the network. The histogram at the top shows the number of spikes per time bin.

50 Chapter 3. Neurons at subcellular, cellular and network levels neurons with spatially overlapping extracellular field potentials, a single electrode picks up electrical activity from more than one neuron. Thus electrodes recording high activity and many APs short after each other, i.e., within less than a single cells refractory period (between 1.5 ms and 3 ms), indicate overlapping extracellular action potentials from multiple cells and may indicate areas with high cell densities. Previous experiments done with tetrodes (Gray et al. (1995)) have shown that the spike-sorting yield of locally-dense electrode clusters is markedly better than distributing the readout sites into isolated spots. Be- cause the extracellular electric fields rapidly decay in amplitude with increasing distance (Rall(1962)), the fields of neighboring neurons can be distinguished on nearby electrodes, which helps to capture shapes of field potentials of individual neurons more reliably. To maximize the number of recorded single cells, active recording sites have, therefore, been arranged in several hundred locally-dense, globally-sparse electrode clusters. The red dots in Fig. 3.3 A represent a configuration of locally-dense, globally- sparse distributed clusters of 1024 electrodes used to record from the neuronal network. Following application of this configuration of electrodes, 1105 cells could be identified after spike sorting. Fig. 3.3 D shows a raster plot of 100 seconds recorded with the 1024 electrodes highlighted in Fig. 3.3A.

3.3.2 Single-cell analysis

The previously recorded data obtained from blocks of densely-spaced electrodes was spike-sorted to yield not only temporal information about individual neuron spike timings but also the spatial distribution of single-cell extracellular electrical action potentials, termed “footprints,” as described in the Section 3.2.7. Two distinct temporal network states were observed: one with the majority of cells spiking sparsely in time and weakly correlated with each other; the second state was one in which most cells of one cell cluster were bursting and thus were synchronously active. It is difficult to identify single-cell activity, when all neurons are active at the same time, due to their strongly overlapping electrical fields. To reduce the number of overlapping extracellular action potentials and to, thereby, obtain best sorting performance, only time periods with no or small cell-cluster bursts involving no more than a few tens of cells were analyzed (for identification of bursts, see also Bakkum et al.(2013d)). For visualizing extracellular AP footprints (Fig. 3.3 B), a contour line was plotted in regions where the electrical activity signals of that cell exceeded 4.5 standard deviations of the noise threshold, to indicate electrodes capable of picking up the signal from that cell. With spike sorting, an estimated total of 2000 neurons could be identified in the culture shown in Fig. 3.3. From Fig. 3.3 A-C, it is apparent that cells in this culture aggregated into clusters of densely-packed neurons. The activity map, as well as the spike-sorted overview (Fig. 3.3 A,B) matches closely to the fluorescence image of the culture in Fig. 3.3 C. Furthermore, it can be seen from

51 3.3. Results

Fig. 3.3 C that these islands of cell clusters are interconnected with tracks of neuronal processes; see also Section 3.3.5.

3.3.3 Correlation of electrical activity with ﬂuorescence imaging

To prove correct identification of single cells, the electrical activity data have been compared to the fluorescence image of a MAP2 immunostaining of the respective cell culture. Through staining, the cell bodies and dendrites become fluorescent and can be imaged on top of the array with a microscope. As can be seen in Fig. 3.4, the spike-triggered averages (STA) of the action potentials of a single cell align with the fluorescently marked cell bodies. The largest voltage signal is not necessarily at the cell body but offset towards the axon initial segment, which has the highest ion-channel density. The cell in the upper left corner (STAs plotted in green) exhibits action potentials with very large amplitudes. The largest peak-to-peak signal of this cell has an amplitude of almost 3 mV and is recorded on the electrode labeled with “5”. The very same same cell exhibits electrical activity with large amplitudes even at electrodes quite distant from the soma (more than 180 µm), which can be assigned to the axon of this cell (see section 3.3.5).

3.3.4 Determination of best recording electrodes

A procedure to determine the smallest number of necessary electrodes that are optimal for recording from a given set of neurons was developed. In order to simultaneously record from as many neurons as possible, only electrodes strictly needed to properly identify neurons should be allocated for recording and connected to readout channels. When zooming into the culture of Fig. 3.3 within the area marked with the red frame in Fig. 3.3 B, electrical footprints of at least five overlapping cells are revealed. Fig. 3.5 A shows these five cells with their spike-triggered average (STA) signal on all electrodes where the amplitude of their action potential exceeds the 4.5 std noise threshold. As the neurons in this cell cluster exhibit overlapping electrical activity in space as well as in time, proper cell identification is difficult, and spike-sorting algorithms are challenged. However, spike sorting is most effective when the activity of a cell is recorded on many electrodes, which is made possible by the high spatial electrode density. Each of the five neurons could be detected, on average, on 45 electrodes on which their signal amplitudes exceeded 4.5 std noise levels. For two neighboring neurons (e.g., the red and the green shapes in Fig. 3.5 A), the recorded signals might differ only marginally on a single electrode. But by taking together many small differences from many electrodes, the cumulative difference gets more significant, and activities of these two cells can more reliably be differentiated.

52 Chapter 3. Neurons at subcellular, cellular and network levels

Figure 3.4: Electrical activity superimposed to a MAP2 staining of the neurons. (a) The electrical activity of three neurons is superimposed to a fluorescence image of a MAP2 staining of the cell culture in the respective area. Spike-triggered averages of signals from 3 different neurons over 50 trials are drawn in green, red and blue. Averaged traces are only displayed for electrodes with a peak-to-peak signal amplitude exceeding 50 µV. The activity of a single cell can be recorded on fairly distant electrodes. Particularly, the green traces exhibit signals of very large amplitudes (almost 3 mV peak-to-peak), and putative axonal signals can be seen at electrodes at more than 180 µm distance from the soma. The red set of signal traces exhibits the largest negative peak value in the 4th row, 9th column and features signals with negative peaks in the first row and signals with positive peaks in the 8th row. (b) Sixty milliseconds of raw data as recorded on the five electrodes marked with arrows and numbers in (a). At least three individual spikes can be identified in this period. The activities of the single cells can be recorded through multiple electrodes.

53 3.3. Results

Figure 3.5: Single-cell resolution. The electrical activity of five neurons has been identified and spike-sorted with 209 electrodes. Subsequently, the performances of all combinations of selecting 3 out of 209 electrodes were analyzed in terms of correctly classified APs. Refer to text for a detailed discussion of the procedure. (a) Spike- triggered averages of five identified neurons with overlapping electrical footprints. For each neuron, a circle is drawn where the amplitude of their electrical signal exceeds a threshold of 4.5 standard deviations of the noise level. Black-yellow circles indicate the three electrodes yielding best sorting performance. (b) Spike-triggered average waveforms recorded with the three electrodes marked in (a). (c) Principal component (PC) projection of 500 AP waveforms recorded with the three electrodes marked in (a). The PC projection is used for clustering. (d) First six PCs of all 500 AP waveforms. Color coding of the neurons is identical for all subfigures. (e) Distribution of performances for all 1.45 million tested electrode combinations. A considerable fraction yields more than 95% correct classifications. (f) Distribution of medians of the silhouette coeffi- cients for all 1.45 million tested electrode combinations for the clustered waveforms as in (c). (g) Comparison of the best achievable spike-sorting performance for different numbers of electrodes. With just one electrode, only about 65% of all APs can be correctly classified. With three electrodes and more, the performance saturates at 100%, thus three electrodes chosen at suitable spots are sufficient to reliably record and distinguish the signals from the five neurons displayed in (a).

54 Chapter 3. Neurons at subcellular, cellular and network levels

Thus, in a first step, all 209 electrodes below these five neurons were used for recording AP activity. Since AP events were recorded with approx. 45 electrodes per neuron, spike sorting could be performed with relatively high reliability, and its results were considered the ground truth for further analysis (Franke et al. (2012); Ruz and Schultz(2014)). The waveforms for 100 AP occurrences per neuron and for each electrode were aligned in time, and data between 2 ms before and 3 ms after the AP peak were cut out. Next, all 1.45 million different combinations of choosing 3 out of 209 electrodes were sampled, and their spike- sorting performances were analyzed. To this end, the 100 AP waveforms for each neuron together with 100 noise waveforms (to accommodate for the case of no AP) were taken from each electrode. Next, PCA was performed on the concatenated waveforms to reduce dimensionality and, subsequently, the first 10 PCs were clustered with the k-means algorithm (MacQueen(1967)). Since it is a priori known that there should be 6 different clusters, i.e., the 5 neurons plus the white noise, k-means was performed with k=6. As the k-means algorithm has a stochastic component, which can mistakenly lead to correct classifications by chance, the clustering was repeated 10 times, and the most frequently occurring solution was chosen. For each electrode combinations, the number of correctly classified AP waveforms that could be clustered was analyzed. If multiple electrode combinations performed equally well, the one that also provided the best cluster separation, as quantified by the median of the silhouette coeffi- cients (Rousseeuw(1987)), was chosen. 3.5 A indicates the three best-performing electrodes, marked with black and yellow circles. In Fig. 3.5 B, the STA for all five neurons on these three electrodes is shown. It is apparent that the shapes provide good separability, which can also be seen from the clustering results in the PCA space in Fig. 3.5 C,D. This procedure was repeated iteratively by choosing 1, 2, 4 or 5 out of 209 electrodes. Instead of analyzing the clustering results in the PCA space, other measures, such as Fisher information, could be used (Cybulski et al.(2014)). Fig. 3.5 G shows the best achievable spike-sorting performance for the different numbers of selected electrodes. It is no surprise that the electrodes yielding the largest peak signals for the respective neurons are chosen upon searching for the 5 optimally placed electrodes (not shown here). This selection result is in accordance with intuition. It is worth noting, however, that upon choosing a subset of electrodes to identify different cells, in particular, when the electrode number is smaller than the number of neurons to be distinguished, the electrodes that record the largest spike amplitudes are not necessarily chosen; rather, those that collectively provide best separability for all neurons are selected. With 3 electrodes and more, performance saturates at 100% correctly classified APs, and, therefore, the three electrodes indicated in Fig. 3.5 A are sufficient to record the activities of these five neurons. Hence, it is possible to reliably record from a number of neurons that exceeds the number of available electrodes.

55 3.3. Results

3.3.5 Subcellular-resolution analysis

The minute extracellular signals of the axonal arbor of a single cell can be identified and tracked over the large sensing area of the HD-MEA by spike-triggered averaging of the electrical signals. The method described in Section 3.2.8 was applied, where three electrodes capturing the electrical signal with the largest amplitude for a defined cell were kept fixed, and all other array electrodes were scanned through. Between 30 and 50 AP occurrences were recorded per configuration. Large axonal arbors of single cells could be revealed with this method; Fig. 3.6 A,B show the axonal arbor of one single cell. The axonal arbor spans over almost the entire electrode array, and signals of individual axons can be recorded at up to 3 mm distance from the soma and up to 6 ms after the somatic spike.The availability of multiple electrodes around one axon enables averaging in space in addition to averaging in time. Averaging in the time-domain, as done previously to get the STAs, reveals the spatial arrangements of axonal segments through their electrical footprints. Once the spatial arrangement of the axon has been identified, spatial averaging can be used to reduce the noise of an individual axonal AP (amplitudes of 5-20 µV ). Fig. 3.6 E shows 25 traces recorded with electrodes that pick up signals of an axonal branch. The spike-triggered average in the time domain (50 averages) is shown at the right, and a single axonal AP buried in noise at the left. Temporal alignment of the signals from neighbored electrodes with respect to the negative peak value of the axonal AP (Fig. 3.6F left) and subsequent averaging of all signals (Fig. 3.6 F right) enhances the correlated signal amplitude much more than that of the uncorrelated noise. Fig. 3.6 G shows how the signal amplitude increases linearly, while the standard deviation (std) of the noise increases according to a square-root function of the number of summed time-aligned electrode signals. The time-aligned summation is shown in Fig. 3.6 G to display the linear and square-root dependence, whereas the time-aligned average is shown in Fig. 3.6 E,F. In the particular case displayed in Fig. 3.6 G, the use of more than four electrodes makes the signal amplitude exceed 4.5 std of the noise floor and thus makes it detectable. Under the assumption of statistically independent noise, the power of noise on multiple electrodes adds up linearly with the number of used electrodes, whereas the power of the signal features a square-law dependence (Ku(1969)). Noise is typically not statistically independent on nearby electrodes (blue curve). However, by whitening the signals prior to summing them, uncorrelated noise can be approximated (Pouzat et al.(2002)). The dashed blue curve shows how uncorrelated noise would add up in the ideal case.

3.3.6 Electrical stimulation of axonal tissue

Electrical stimulation of neural tissue through single microelectrodes reliably elicits time-locked action potentials. Fig. 3.7 shows a conﬁguration where 841

56 Chapter 3. Neurons at subcellular, cellular and network levels

Figure 3.6: Axonal arbors. Identiﬁed axonal arbor of a single cell revealing subcellular features up to more than 2 mm distance from the cell body. (a) All electrodes capturing activity attributed to a single neuron. Color-coding indicates the time of arrival of the AP at the respective electrode. It takes 6 ms for the AP to arrive at the left-most visible axonal segment. (b) The same neuron and electrodes as in (a), this time showing the amplitude of the most negative peak on a logarithmic color-scale. Its putative soma is inside the box indicated with (c). (c) And (d) spike-triggered averages (30 to 50 averages) of the electrical footprint from two areas of the array as indicated in (b). The scale bars of (d) apply to the signals in (c) and (d). (e) Left: 25 traces from electrodes that detect axonal signals. Spike-triggered averaging (50 APs) reduces noise. Right: traces from the same electrodes showing a single axonal AP hidden in the noise. Red dots indicate the timing of the negative peak. (f) Left: all 25 recording traces of one axonal AP overlaid and aligned in time with respect to the negative peak. Right: spatial averaging of traces aligned in time with respect to the negative peak from multiple neighbored electrodes that detect axonal signals improves axonal AP detectability. The Gaussian-like distribution of the noise is shown, and dashed red lines indicate one standard deviation. Green lines indicate 4.5 standard deviations, the detection threshold. (g) The signal amplitude (green curve, right ordinate) scales linearly with an increasing number of electrodes (#electrodes), whereas the noise ﬂoor (blue curve, left ordinate) scales with a square root function. The dashed blue curve indicates the scaling of uncorrelated noise. For more than four electrodes, the signal amplitude is more than 4.5 times larger (70 µV) than the standard deviation of the noise (15 µV), so that the signals can be considered detectable. See text for a more detailed discussion.

57 3.3. Results

electrodes were chosen below an axonal segment. The electrode denoted with “stimulation site” in Fig. 3.7 A was repeatedly stimulated with biphasic, first positive then negative voltage pulses of 200 µs duration per phase and ±400 mV amplitude. Such a stimulation pulse activates the axonal segment passing across the microelectrode so that an action potential is initiated and travels down the axon. It takes 2.1 ms to travel 1.56 mm, i.e., the average propagation speed for this axon is 0.74 m/s. The propagation speed has been calculated by dividing the linear distance between two electrodes marked with the red-white circle by the time it takes for the negative peak of the AP to travel that same distance. While the measured speed is highest close to the stimulation site (0.9 m/s) it decays with increasing distances down to 0.4 m/s. The measured values are in agreement with reported propagation speeds for unmyelinated axons (Debanne et al.(2011)). In Fig. 3.7 B, a stereotypical positive-first waveform of the traveling axonal AP can be seen at three different time points. Close to electrode 3, around 800 µm from the stimulation site, the axon splits into two branches, an upper and a lower branch. Interestingly, it appears that the AP in the upper branch travels faster than in the lower one. At t2 = 1.0 ms the AP in the upper branch is further advanced than that in the lower branch. It is unclear, however, whether this difference in propagation velocity is due to, e.g., a longer axon, i.e., an extra loop, which is not visible, or whether it is due to differences in axonal diameters (Schierwagen et al.(2008)) or other properties. The cause for this difference in propagation velocity cannot clearly be identified without further analysis, such as high-resolution imaging of the axon geometry. An analysis directly on the device is, however, very challenging due to the nontransparent nature and corrugated surface of the CMOS substrate (super-resolution optical methods are not applicable). The amplitude of the axonal signal in Fig. 3.7 on the next page was larger than 20 µV in some locations and varied along the axon. These variations could be a consequence of biological material, such as another neuron or glia residing between the axon and the electrode, the effect of which would locally reduce signal amplitudes. Also, such material could reside on top of the axon towards the solution and then enhance signal amplitudes by spatially confining the area through which charge-carrying ions could spread.

3.3.7 Microelectrode characterization

A series of measurements was performed to characterize the microelectrodes and to quantify electrode-to-electrode variations, as well as their stability over time. Impedance values for 300 randomly selected bright Pt and 300 Pt-black electrodes were measured at 1 kHz (see section 3.2.4 and 3.2.5 for how Pt-black electrodes are fabricated). At this frequency, the electrodes can be approximated as being purely capacitive, Franks et al.(2005a). The electrode capacitances were determined via the measured impedance at 1 kHz according to the formula:

58 Chapter 3. Neurons at subcellular, cellular and network levels

Figure 3.7: Tracking of stimulated signals in axonal arbors. 841 electrodes below an axonal arbor record the stimulation-induced axonal propagation of an AP. (a) The site at the bottom denoted stimulation site has been stimulated 200 times with biphasic 300 mV voltage pulses at 300 ms inter-stimulus intervals. Each square shows the average AP minimum value on a clipped color scale over the 200 trials for each recorded electrode within 3 ms after stimulation. The red circles with white fillings indicate electrodes for which voltage traces are shown in (c). Close to the electrode indicated with 3, the axon splits into two different branches. (b) Same neuron and electrodes as in (a). This time, the propagating AP is shown at three different time points, t1=0.3, t2=1.0 and t3=1.8 ms after stimulation occurred. At t2, the AP has already passed the branch point such that the AP can be seen in both branches. (c) Voltage traces from the 7 electrodes marked with the red-white circles in (a). Shown are voltage traces for single trials (gray) and the median over all trials (black).

59 3.4. Discussion and outlook

Figure 3.8: Electrode impedance measurements. Shown are the capacitances of 300 randomly selected bright Pt and Pt-black electrodes based on impedance measurements at 1 kHz. The capacitances are plotted on a logarithmic scale. The capacitances of Pt-black electrodes are about 50 times larger than those of the bright Pt electrodes. The distribution shown in green indicates the capacitances for 300 randomly selected electrodes after 5 months of culturing cells on top of them.

Zimp = 1/(jωCel), where Zimp is the measured impedance, Cel is the determined electrode capacitance, and ω is the angular frequency. Using the on-chip stimulation buﬀers, a sinusoidal current with amplitudes of ±20 nA was applied to bright Pt electrodes and ±200 nA to Pt-black electrodes. On-chip ampliﬁers were used to measure the voltage drop across the electrodes. The capacitances of bright Pt electrodes were found to be 45.0 ± 10 pF (mean ± std). Capacitance values of Pt-black electrodes amounted to 2.0 ± 0.26 nF (mean ± std). Measurements for a device with Pt-black electrodes, after being exposed to more than 5 months of cell culturing were shifted to 0.65 ± 0.27 nF (mean ± std) (Fig. 3.8). The impedance did not measurably change over tens of thousands of stimulations. Finally, electrodes survive long periods of cell culturing. Two devices exposed to more than 5 months of cell culturing were submerged in PBS, and a test signal was applied to a counter Pt-electrode held in the PBS. All electrodes recorded the test signal without degraded amplitude as compared to an unused device.

3.4 Discussion and outlook

We presented a CMOS based HD-MEA capable of, at the same time, recording from neuronal networks with hundreds of cells as well as of capturing subtle subcellular signal details from the axonal arbor of single neurons. This performance is made possible by the inherent system ﬂexibility as well as a large number of parallel low-noise readout and stimulation channels. We think that the system presented here constitutes a favorable combination of high signal quality and high spatial resolution. The number of 1024 channels that can be simultaneously read out is on the same order of magnitude albeit somewhat lower than that of other devices (Berdondini et al.(2009c); Bertotti et al.(2014); Eversmann et al. (2003a)), whereas the noise in the recordings is considerably lower (2.4 µVrms in the action-potential band between 300 Hz and 10 kHz). The low noise and high

60 Chapter 3. Neurons at subcellular, cellular and network levels recording quality together with the large dynamic range of the on-chip ADCs, enable not only tracking of the spreading of multiple APs over many different neurons, but, at the same time, also their propagation along the axonal arbor of a single cell. Compared to previous implementations, such as Lopez et al.(2014); Seidl et al.(2011) and, in particular, Frey et al.(2010), the switch matrix was advanced through the following features: instead of a few long wires running across the whole array, we implemented hundreds of thousands of comparably short routing wires (420 µm average length) in order to massively increase the number of potential routing paths for all array electrodes; moreover, the number of wires per line and column was increased to 6 horizontal and 6 vertical wires, and the number of switches was increased to two per electrode, enabled through the use of 0.35 µm CMOS technology , to achieve a highly flexible electrode to readout-circuitry routing. This flexible switch matrix structure together with the large overall number of available array electrodes enables the device to record the activity of neuronal networks consisting of hundreds of cells by employing globally-sparse, locally-dense configurations of recording electrodes. The electrode array supports operation in three different modes, or arbitrary combinations thereof: (i) locally-dense electrode clusters of up to 23 × 23 (529) contiguous electrodes at arbitrary positions massively improve the separation of spatially overlapping extracellular action potentials; (ii) sparse distribution of such clusters or single electrodes across the whole array allows for recording from many cells in distant regions at the same time; (iii) by recruiting all available electrodes below a certain axonal arbor, it is also possible to observe how cellular parameters, such as signal propagation velocities (Bakkum et al.(2013b, 2008c); Greschner et al.(2014); Zeck et al.(2011a)), fluctuate over time, and how axons spatially move during the course of an experiment. Signals originating from axons are often on the same order of magnitude as the noise floor, making it difficult to detect such signals without averaging. However, temporal averaging only gives a population average over all observed spikes and does not permit studying single APs. Spatial averaging, enabled by the high spatial resolution of electrodes and the low noise, can be employed to study temporal changes in propagation velocity or branch point failures (Debanne et al.(2011); Kim et al. (2009)) of single APs . Besides spatial averaging, more sophisticated signal processing algorithms will be developed in the future in an effort to more reliably detect axonal signals. Concerning the number of readout channels, it has to be noted that the data volume produced by HD-MEAs can be enormous and amounts to, e.g., 1.4 GB per minute for reading from 1024 channels at 20 kHz. An option to deselect electrodes with no relevant information can be very useful, as exclusion of non-active areas saves disk storage space, as well as data analysis time. To make most efficient use of available resources, recording channels can be allocated only to electrodes in the vicinity of neurons that are producing signals relevant to the ongoing investigation. Electrode-to-electrode variations in impedance values were found to be very small over the whole array. All 26,400 microelectrodes could be reliably used even after long periods of culturing cells

61 3.4. Discussion and outlook on top. The shift in impedance values for Pt-black electrodes, which had cells on top for more than 5 months, can be due to various reasons: (i) mechanical damage to the fine Pt-black structures upon washing/cleaning the chips, (ii) residual adhesion or protein layers from the culturing that cover the dendritic Pt structures, or (iii) cell debris, all of which would cause an increase in impedance. Recording the dynamics of large networks at a spatiotemporal resolution sufficient to simultaneously resolve individual APs of single cells is challenging for conventional techniques and arrays. For example, single cells can be resolved upon imaging calcium activity. However, in addition to being phototoxic, the temporal resolution is limited making it difficult to capture the waveform shape and the temporal-dynamics of single APs (in the range of kHz) (Grienberger and Konnerth(2012)). Genetically targeted all-optical electrophysiology methods have recently emerged that provide better temporal and spatial resolution and hold the promise of all-optical electrophysiology (Hochbaum et al.(2014)), but experiments at resolutions necessary for studying axonal signals have, however, not been demonstrated to date. Intracellular electrical recordings carried out with patch-clamp are well-suited for the study of cellular or subcellular properties of individual neurons; they can be used to record and resolve synaptic currents with high SNR, but fail at recording from populations of more than a few cells. Extracellular recordings based on passive MEAs on the other hand, can access many neurons at the same time with temporal resolutions high enough to resolve individual action potentials. However, due to the comparably large electrode spacing (≥ 30 µm), it is often the case that not all neurons in a population can be captured, that the resolution of overlapping neurons is difficult, or that, for example, axonal signals cannot be reliably detected. Recent develop- ments also include small arrays of mushroom shaped electrodes (Hai and Spira (2012); Spira et al.(2007)) or nanowires on electrodes (Robinson et al.(2012)), which can be engulfed by the cell membrane and allow for pseudo-intracellular recordings. The arrays currently feature, however, only tens of simultaneously readable recording sites. Although extracellular recordings are better suited to investigate network-wide activity than recordings with patch-clamp setups, inferences about neuronal plasticity in extracellular recordings are inherently more difficult, as the postsynaptic potentials are not directly measurable (Muller¨ et al.(2013a)). Exploiting, e.g., homeostasis in neuronal networks (Turrigiano and Nelson(2000)), measuring changes in network functional connectivity estimations could be used as a proxy for quantifying changes in postsynaptic potentials and network plasticity (Stevenson and Koerding(2011)).This approach, will, however, require access to potentially every cell in a network. Possible future research with a configurable HD-MEA, as used in this study, may include constraining a few hundred neurons to grow on the active area of the HD-MEA. By capturing the full activity of this complete neuronal population, under-sampling of the network is avoided, and the number of hidden variables is reduced. In this case, the inference of network parameters, such as functional

62 Chapter 3. Neurons at subcellular, cellular and network levels connectivity, is signiﬁcantly more robust (Gerhard et al.(2010)). Care has to be taken, however, when interpreting data recorded on such a network-wide scale: spikes originating from the same cell but recorded on diﬀerent neighboring electrodes might falsely show up as strongly correlated cells, or two or more cells might be recognized as a single cell, thus underestimating connectivity. Spike sorting of such datasets prior to analyzing network activity is a crucial step, since failing to do so can distort or bias the data.

Acknowledgments

The authors would like to thank J¨org Rothe for support and consulting while developing the electronic components for the setup, David J¨ackel for help with the imaging, Dr. Felix Franke for helpful discussion on spike-sorting algorithms and Dr. Thomas Russell for culturing assistance. This work was supported by the ERC Advanced Grant “NeuroCMOS” under contract number AdG 267351. M. Radivojevic and D. J. Bakkum received funding support from the Swiss National Foundation through an Ambizione Grant (PZ00P3 132245).

63 3.4. Discussion and outlook

64 Chapter 4

Selection of best recording sites for optimizing spike-sorting yield

Jan Muller¨ 1, Felix Franke1, Michele Fiscella1, Urs Frey2, Douglas J. Bakkum1, Andreas Hierlemann1

In preparation

1Bio Engineering Laboratory, ETH Zurich, Switzerland

2RIKEN Quantitative Biology Center, Kobe, Japan

65 Abstract — Recording the activity from large populations of neurons with single-cell resolution requires recording schemes with a large number of recording sites arranged at high spatial resolution. Some recently developed CMOS-based microelectrode arrays provide hundreds to thousands of electrodes, a subset of which needs to be selected for recording. In order to optimize the number of identifiable neurons, recording sites need to be chosen carefully. When recording from networks of neurons, one is typically interested in the exact relative timing of the occurrence of action potentials in different neurons. Thus, we here consider the assignment of recorded action potentials to different neurons as a classification problem. Based on this consideration, we present a metric relying on linear discriminant analysis to compute the expected classification error for a given recording site or electrode arrangement. Using this metric, we pro- pose two different algorithms to automatically find optimal and close to optimal arrangements of recording sites. The first algorithm follows a computationally demanding approach to search for the optimal recording site arrangement. The second algorithm relies on simplified heuristics to find close to optimal arrangements but with significantly decreased computational efforts. Choosing a limited number of recording sites from a larger set of potential recording sites is a combinatorial problem, as the information provided by each site also depends on which other sites are selected. Thus, the problem can quickly become fairly complex when the number of possible sites becomes large. Additionally, potential constraints on the recording site selection imposed by the recording hardware can aggravate the situation. The developed algorithms have been used to find the best recording sites on a recently developed microelectrode array featuring 26,400 Pt-microelectrodes. The high-density electrodes are arranged in a grid- like configuration with a center-to-center pitch of 17.5 µm. From all available electrodes, 1024 can be read out at the same time. We demonstrate how the developed algorithms can be extended to take into account constraints given by the recording hardware. The presented methods can be applied to other neuronal recording technologies that require the selection of a subset of recording sites, for example, random-access scanning beam microscopy techniques.

66 Chapter 4. Selection of best recording sites

4.1 Introduction

In an ongoing quest to understand how networks of neurons perform tasks, such as memory and learning, new technology is being developed to record from ever- larger populations of cells (Stevenson and Koerding(2011); Marblestone et al. (2013a)). In order to reliably identify single cells in dense populations of neurons with spatially overlapping extracellular potentials, and to record their activity over time, multiple recording sites arranged at spacings close to those of the neuronal cells are required (Gray et al.(1995); Hyvarinen and Oja(2000); Franke et al.(2012)). Advances in microtechnology (Hierlemann et al.(2011)) and emerging optical tools (Ahrens et al.(2013)) have enabled the development of devices featuring high recording site densities (Obien et al.(2015)), which pave the way to identification of single cells in dense populations of neurons. Spatial sampling at high resolution allows for accurately reconstructing positions and spike trains of single neurons (Harris et al.(2000a)). However, there still is a trade-off between recording site density and, e.g., recording quality or power dissipation. Thus, it is desirable to perform experiments with the lowest number of required recording sites. One possible experimental paradigm is to first sequentially scan through the neuronal tissue under study with high spatio-temporal resolution to identify the exact locations and spatial spread of single cells. Once cells have been identified, recording resources (electrodes or scanning beam position in case of 2-photon imaging etc.) are assigned to locations providing the most relevant information for the respective experiment. Further recording is then restricted to these locations. Indeed, a variety of different electrophysiological devices has been recently developed to support such an experimental paradigm: an implantable neural probe featuring 455 electrodes, where 52 can be read out at the same time (Lopez et al. (2014)); another implantable probe with up to four probe shafts and 188 electrodes per shaft, where 8 electrodes per shaft can be read out at the same time (Seidl et al.(2011)); planar microelectrode arrays featuring 11,011 electrodes with 126 parallel readout channels (Frey et al.(2010)); or recently, another planar array featuring 26’400 electrodes with 1024 parallel readout channels (Ballini et al.(2014)). Various techniques were developed to automate the process of selecting the most suitable recording electrodes. Van Dijck et al. introduce a measure called PSNR, which penalizes electrodes in the selection process that record similar signals as already selected electrodes. Based on this measure, they develop a greedy algorithm to select the most suitable channels (Van Dijck et al.(2012)). Vysotska et al. select recording sites based on how well they can predict the activity on their neighboring sites through a Gaussian process (Vysotska et al.(2014)). In Cham et al.(2005), the position of the neural probe is mechanically adjusted

67 4.1. Introduction based on maximum peak-to-peak signals of recordings or maximum PCA components.

The problem of choosing a limited number of recording sites from a larger set of potential recording sites is not a trivial one: the information provided by each site also depends on which other sites are selected. Thus, the problem becomes a combinatorial problem and is complex to solve when the number of possible sites becomes large. Additionally, the problem is aggravated by potential constraints on the recording site selection imposed by the recording hardware. Different methods to assess the quality of a particular recording site selection have been proposed. Cybulski et al.(2014) developed a framework to assess the quality of different recording schemes (including electrical sensing, wide-field fluorescence microscopy and 2-photon microscopy) based on the Fisher information matrix (FIM). By assuming a common point-spread function for all neurons, the accuracy with which each neuron can be located can be computed for a certain sensor arrangement. Similarly, Shahram and Milanfar(2006) apply an information-theory-based analysis to assess the resolution limit in microscopy.

Here, we present a quantitative measure to estimate the performance of a particular recording site configuration, once the locations and spatial extents of the signal sources (neurons in our case) are known. To measure performance, we estimate the expected error in assigning action potentials to single neurons. Under the assumptions of colored Gaussian noise and time-invariant signal sources, this error can be approximated via linear discriminant analysis (LDA). The quantifi- cation allows for comparing different recording schemes and establishes an objective criterion to select the recording site configuration with minimal expected error.

In the rest of the paper, we will focus on electrophysiology scenarios and consider electrodes for recording sites. Although the analysis can be applied to many other recording methods and techniques, focusing the discussion on electrode arrays will help to make the presentation clear. We assume that all signal sources have been identified prior to analysis. While finding all present signal sources is certainly a difficult task, various existing techniques can be used (Lewicki (1998b)).

This paper is organized as follows. First, we introduce a metric based on the linear discriminant analysis to assess the quality of a particular recording scheme, given prior knowledge of all present signal sources, including noise. Then, we introduce two algorithms to choose an optimal subset of recording sites. The ﬁrst algorithm ﬁnds a set of recording sites by initially assuming that we can simultaneously record from all sites and by then iteratively removing sites in a greedy way: for each site, we estimate how much the expected error would increase if the site was not included, and the least important site is subsequently removed. The second technique relies on an integer linear programming algorithm and optimizes a linear approximation of the expected error.

68 Chapter 4. Selection of best recording sites

We apply these algorithms to a recently developed electrode array, where a large number of electrodes is available (26’400), while the readout infrastructure provides only 1024 parallel readout channels. Therefore, we aim at selecting those 1024 electrodes that constitute the most favorable recording configuration for a given experimental scenario. The problem is twofold: First, we would like to know which electrodes are optimal to identify all present signal sources, i.e., all neurons. Second, the hardware also needs to support connecting the selected electrodes to sets of readout channels. For example, if two neighboring electrodes form part of an optimal electrode set for a given neuronal preparation, but cannot both be connected to readout channels at the same time, one of the electrodes needs to be discarded, and a different, next-best-suited electrode needs to be chosen instead. Both proposed electrode selection algorithms can be combined with a previously published integer linear programming (ILP) algorithm (Frey et al.(2010)) to obtain a suitable electrode-to-readout-channel routing. Finally, the performance of the developed algorithms is evaluated for different simulated data sets with varying neuronal densities.

4.2 Performance assessment of recording con- ﬁgurations

To find an optimal recording configuration, the performances for different recording electrode selections need to be quantified. A recording configuration of neuronal signals can be judged by its ability to identify and assign signals from/to individual neurons. Based on the assumption that every recorded neuron produces a unique waveform, this problem can be formulated as a classification problem. The prototypical, noise free representation of the waveforms is denoted “template”. In practice, this classification problem is called spike sorting and can be done in a variety of different ways (Lewicki(1998b)) by using template matching (Vollgraf and Obermayer(2006); Franke et al.(2010)) or clustering (Kadir et al.(2013)). Here, we will not focus on the spike sorting itself, but rather quantify the smallest achievable error rate for a given recording electrode selection. When computing the expected error, we assume optimal discrimination in a Bayesian sense, i.e., by taking the prior probabilities into account. Intuitively, a good classification can be achieved, if the templates (the cluster centers) for each neuron are far apart from each other. Here, we present a framework that allows for quantifying classification and that treats the neuronal recordings as a linear discriminant classification problem. Computing the expected error of the classifier then serves as a measure of the recording quality. Below, a brief overview is given, how the expected error of the classifier can be computed, given optimal discrimination in a Bayesian sense, i.e., statistically best achievable dis-

69 4.2. Performance assessment of recording conﬁgurations crimination. For a more detailed derivation of the Bayesian argument see Franke et al.(2015).

We ﬁrst specify Ntot as the total number of available electrodes from which a subset Ω with M electrodes needs to be chosen. We deﬁne a short period of data with L samples recorded on electrode k by:

T xk = [xk,1, ..., xk,L] , k ∈ Ω

For the remainder of the discussion, we choose a more convenient vector notation. To represent data from the subset Ω of all active recording electrodes, M, we concatenate all M vectors into a new vector:

T x = [x1,1, ..., x1,L, ..., xM,1, ..., xM,L]

Similarly, the distinct, unique signal on all electrodes, evoked by an action potential of neuron i, is:

T µi = [µ1,1, ..., µ1,L, ..., µM,1, ..., µM,L] and is called the template of neuron i. Spike sorting compares the newly recorded piece of data x with all available templates µ and chooses, based on a discriminant function, one of the available templates or noise. When computing the expected error, we consider all available pairs of templates individually and then sum up the error for all pairs. We start oﬀ by estimating a probability distribution for each of the templates. It is reasonable to assume that noise on all templates follows a Gaussian distribution with a covariance matrix Σ. Then, the distributions of signal intensities for a given template µi follow a normal distribution N(µi, Σi), with mean µi and covariance Σi. Furthermore, we make the simplifying assumption that the covariance matrix is the same for all templates. We then can compute the discriminant function for two templates by multiplying each probability distribution with its prior probability and setting the two distributions equal. The prior probability for template µi to appear in the data is denoted by πi and can usually be estimated from the same training set that was used to acquire the templates.

πiN(µi, Σ) = πjN(µj, Σ)

Solving this equation for x-dependent and x-independent parts yields:

T πi 1 T −1 T −1 x ~wij = −ln + µi Σ µi − µj Σ µj = thrij (4.2.1) πj 2 where ~wijis the normalized projection vector:

70 Chapter 4. Selection of best recording sites

−1 Σ (µi − µj) ~wij = kµi − µjk2

An estimated noise covariance matrix Σ can be calculated from recordings without any action potential, i.e., noise recordings, and can be used to prewhiten the signals (Pouzat et al.(2002)). Prewhitening transforms the noise covariance matrix into the unity matrix, thus simplifying the discrimination threshold thrij to:

T πi 1 T T x ~wij = −ln + µi µ − µj µ = thrij (4.2.2) πj 2

T ~ A new data sample x is classiﬁed as belonging to neuron i if x wij > thrij for all j. Having determined the discrimination threshold, thrij, we can now integrate the area where both distributions overlap, which yields the expected error for these two templates.

Z ∞ Z thrij T T EError = πi N(µi ~wij, 1)dx + πj N(µj ~wij, 1)dx thrij −∞

This situation is illustrated in Fig. 4.1. The probability distributions for two templates µi and µj are projected on and scaled according to their prior probabilities πi and πj. The position of the dashed line indicates the discrimination threshold thrij. The integrated darker area below the curves corresponds to the expected error. As can be seen from Fig. 4.1, this measure is nonlinear in the distance between the cluster centers. Close-by cluster centers yield larger errors than distant clusters: if two cluster centers are already far away, little is gained by separating them even further. On the other hand, separating two close-by clusters just a little bit more reduces the expected error signiﬁcantly. To illustrate the concepts, this framework is applied to a simple example scenario as given in Fig. 4.2. The objective of this example is to select two arbitrary electrodes along the x-axis between 0 and 10, yielding best separation in the sense of the expected error measure for all available neurons. In this example, three simulated neurons are arranged in a one-dimensional space, with their peak amplitudes at locations E[µ1]=3, E[µ2]=5 and E[µ3]=7 respectively. Instead of considering multiple samples per electrode, only one sample is considered. The amplitudes are modeled according to an exponential law. A is the maximum amplitude (12, 7.5 and 7.5) and is normalized to multiples of one standard deviation of the noise. C is a parameter to control the spatial decay of the amplitude.

−x2 A exp C

71 4.2. Performance assessment of recording conﬁgurations

Figure 4.1: Illustration of the probability distributions for two templates µi and µj projected on ~wij. The two distributions are scaled according to their prior probabilities πi and πj. The dashed line at thrij indicates the discrimination threshold. The integrated area where both distributions overlap corresponds to the expected error for these two templates.

In the example in Fig. 4.2, electrodes can be selected at arbitrary positions along the x-axis. Fig. 4.2 A shows how the amplitude for the three neurons changes along the x-axis. Fig. 4.2 B shows the Euclidean differences between the three templates. The difference between the templates and the “noise template”, i.e., the null-hypothesis, is not shown for better clarity of the figure; the difference would simply coincide with the curves from Fig. 4.2 A. The amplitude of the differences between the templates corresponds to the Euclidean distance between the cluster centers. For this simple recording scenario, all different electrode configurations of two electrodes can be computed explicitly. This analysis is shown in Fig. 4.2 C, where performance was evaluated for every possible arrangement of two electrodes. The color code indicates the expected error; blue stands for little error, whereas red represents large errors. The area indicated by the white circle exhibits the smallest expected error and corresponds to the two electrodes being arranged at locations 3.8 and 6.6. The two-dimensional space spanned by the two selected electrodes, which yield the smallest expected error, is shown in Fig. 4.2 D. 500 noisy events are shown for each simulated neuron and for the noise case (depicted in cyan).

4.2.1 Considering the worst case

When computing the expected error for two diﬀerent templates, it is important to note that the relative timing, at which these two action potentials occur, is not known beforehand. Thus, for robustness of the results, the minimum possible distance between the two templates needs to be considered when computing the projection vector ~wij. This can be achieved by repeatedly shifting one of the templates by τ samples until the minimum distance is reached, at which the

72 Chapter 4. Selection of best recording sites

Figure 4.2: Three simulated neurons arranged in a one-dimensional space of arbitrary units between 0 and 10. Two electrodes at arbitrary locations between 0 and 10 have to be chosen “for recording”. To be able to plot subﬁgure c) and d) on two dimensions, only one sample feature, the maximum amplitude, is considered per electrode. a) The noise-free amplitudes of the three simulated neuronal signals recorded along the x-axis, the only dimension considered here. The simulated neurons exhibit their peak amplitudes at 3, 5, and 7. The amplitudes are scaled to correspond to multiples of one sigma of noise. b) Euclidean distances between all three templates, as measured along the x-axis. The red crosses with the black lines in a) and b) indicate the locations of the optimal recording sites given a limit of two recording electrodes, as determined in c). The electrodes can be placed at arbitrary positions along the x-axis and are not constrained to integer values. c) Analysis of the expected error for each possible combination of two electrodes along the x-axis. The dark blue area, indicated by the white circle, corresponds to the best electrode combination at 3.8/6.6. d) Projection of 500 samples for each neuron corrupted by noise in the two dimensions spanned by the two speciﬁed electrodes. The cyan cluster represents the noise cluster.

73 4.3. Example templates temporarily coincide. This time-shifted version of the template is then used for further analysis:

argminkµi(τ) − µjk2 τ

T µi = [µ1,tau, ..., µ1,L, 01,1, ...01,tau−1, ..., µk,tau, ..., µk,L, 0k,1, ...0k,tau−1]

This approach ensures that the projection for the worst-case distance is analyzed and not an arbitrary projection.

As is apparent from the equation 4.2.2 for the discrimination threshold, thrij, both, the Euclidean distance between two templates (µi − µj), as well as the probability of a template to appear in the data (prior probability πi/j) inﬂuence the location of the decision boundary, thrij. We analyzed how these parameters qualitatively inﬂuence the outcome of the electrode selection.

4.3 Example

A simulation incorporating more realistic templates than used before can be seen in Fig. 4.3 A. The templates of three neurons are recorded from 25 potential electrodes. For each template, a time vector of L = 70 samples (3.5 ms in this case) is plotted, showing the different temporal evolution for each template. The expected errors for all combinations of choosing four out of 25 electrodes were analyzed with different signal-to-noise-ratios (SNR) and noise probabilities (πnoise). Noise probability denotes the probability to find no spike in the recorded data; together with the probabilities for spikes, the noise probability needs to sum to 1. In Fig. 4.3 A, the best performing electrodes for each situation are indicated with red, black and blue markers. In Fig. 4.3 B, one standard deviation of the simulated noise floor for each scenario is indicated by a gray band. First, it was analyzed, which electrodes are optimal for templates with a high SNR (red marker). In both cases of simulated noise probabilities, the selected electrodes are the same. For low SNR templates and a noise probability comparable to the prior probability for spikes, the electrodes indicated by the black marker are chosen. Finally, for the case of low SNR templates and high noise probability, again different electrodes are chosen, in this case electrodes that increase not only the distance between the templates but also the distances between the templates and the noise. It can be seen in Fig. 4.4 that the expected error decreases exponentially when more electrodes are available.

74 Chapter 4. Selection of best recording sites

Figure 4.3: Four electrodes to record from three model neurons arranged over a two- dimensional grid of 25 electrodes. a) Templates are shown for the three neurons, each exhibiting its own characteristic temporal evolution of spike waveforms. All possible combinations of choosing four out of 25 electrodes were exhaustively sampled and analyzed for diﬀerent noise scenarios, as depicted in c). The markers indicate the optimal electrodes for a particular noise scenario. b) Neuron templates from three combinations of four electrodes as indicated in a). The gray area indicates one sigma noise. Darker gray corresponds to high noise probability (low spike probability), lighter gray correspondingly indicates smaller noise probability (higher spiking rate). c) The three diﬀerent noise scenarios that were analyzed. Red-cross: large SNR and low and high noise probability πnoise. Black-diamond: small SNR and small noise probability, blue-cross: small SNR and large noise probability

Figure 4.4: The expected error for two of the noise scenarios from Fig. 4.3 shown on a logarithmic scale. The expected error decreases exponentially, as the numbers of available electrodes increases.

75 4.4. Electrode selection algorithms

4.4 Electrode selection algorithms

We developed two algorithms to automate the process of finding the best electrodes selection. Both algorithms take as an input the templates of all neurons, which need to be identified in a preprocessing step, which is not discussed here (Lewicki(1998b)). The first algorithm directly optimizes the expected error. It works in a greedy manner, by starting out with all available electrodes, and iteratively removes those electrodes that contribute least to minimizing the expected error. This way, a selection of electrodes is obtained, which is optimal in terms of the lowest expected error. However, if the number of electrodes and the number of present neurons is large, this algorithm can take considerable time until yielding results. The second algorithm does not minimize the expected error, but maximizes the Euclidean distance between the templates. It is considerably faster to run. It uses an integer linear programming (ILP) approach to find a subset of electrodes that give a minimum Euclidean distance between templates, and, at the same time, maximizes the Euclidean distance between all templates.

4.4.1 Greedy algorithm

The greedy algorithm is outlined in Fig. 4.5. It starts out by considering all available electrodes. Each electrode is analyzed one after the other, and it is calculated, how much the expected LDA error increases by removing this one electrode. The electrode contributing least towards keeping the error small is removed. This procedure is then repeated until the number of remaining electrodes corresponds to the targeted number of electrodes. The algorithm relies directly on the earlier introduced measure of the expected error, so that any found solution will correspond to a minimum in terms of expected classiﬁcation errors. However, the algorithm is computationally expensive. For N electrodes, and M templates, all templates need to be analyzed M ∗ M + N ∗ (N + 1)/2 times.

4.4.2 ILP algorithm

For the ILP algorithm, the correlations between electrodes are ignored, and each electrode gets one quality value for each template pair, which does not change during execution of the algorithm. An ILP algorithm in general works according to the following scheme: A solution vector x is searched, which maximizes a linear cost function, while, at the same time, it fulﬁlls a set of constraints.

max cT x x

The set of constraints usually can be described as:

76 Chapter 4. Selection of best recording sites

Figure 4.5: Flowchart of the greedy electrode selection algorithm.

77 4.4. Electrode selection algorithms

Ax > b

lb ≤ x ≤ ub

Here, c is the n × 1 cost vector, x is the n × 1 solution vector, A is an m × n constraint matrix, and b is an m × 1 vector of bounds for the constraints. lb and ub are the n × 1 lower- and upper-bound constraints on the solution vector x. Additionally, some or all elements of the solution vector x are constrained to be integer values. To turn our problem of finding the most suited recording electrodes into an ILP, we first need to map the above vectors and matrices into the problem domain. The solution vector x represents the state of all the electrodes. x1 = 1 corresponds to electrode1 being selected and so on. The elements of x are constrained to binary values, either 0 or 1, i.e., lb = 0 and ub = 1, as an electrode can only be either selected or not selected. Thus, for each available electrode, there is one column in the constraint matrix A. For all Euclidean differences between all templates and the noise, a constraint is added to the equations. This is done by adding a row to the matrix A with the norm of the differences on all K electrodes between two templates µ1 and µ2. When kµ1,k − µ2,kk2 is the norm on electrode k for the two templates, the row of matrix A for these two templates looks like:

A = (kµ1,1 − µ1,2k2, kµ1,2 − µ2,2k2, ..., kµ1,k − µ2,kk2)

To reduce complexity of the constraints, differences from templates with virtually no spatial overlap can be ignored. Such differences would essentially be equal to the difference between the respective template and the noise template and thus, are already considered by including the noise template. The constraint vector b determines the minimum required Euclidean distance between any two templates. Each row of A, multiplied with the binary solution vector x, yields the total accumulated difference of two templates. By constraining this accumulated difference, a minimum distance between different templates and the noise is ensured, guaranteeing detectability for all templates. Finally, for the cost function, the elements of vector c quantify, how much each electrode contributes to the overall distance between all templates. The vector can be computed by summing all differences between all templates and noise for each electrode. However, in order to fairly balance all available templates, all templates need to be normalized to the same amplitude prior to summing. The cost function cannot be recomputed during execution of the ILP algorithm. For this reason, the worst-case minimum distance for time-vector templates cannot be computed as it was done for the greedy algorithm. The worst-case distance

78 Chapter 4. Selection of best recording sites might be diﬀerent for every speciﬁc electrode selection, i.e., it is dependent on the solution vector x. For this reason, the minimum distance is computed by using a representative heuristic of ten electrodes exhibiting the largest amplitudes per template.

4.5 Technology application

The algorithms described above were subsequently applied to a recently developed CMOS based high-density microelectrode array (HDMEA) system. This array system features 26’400 Pt-electrodes arranged in a grid-like conﬁguration with a center-to-center electrode pitch of 17.5 µm. Below the electrodes reside a large number of switches and wires used to connect a subset of the electrodes to a total of 1024 readout channels that are located on the same chip at the periphery of the electrode area. The readout channels provide programmable gain up to 78 dB and include signal-conditioning circuitry for band-pass ﬁltering. On-chip ADCs sample the data at 20 kHz. Additionally, the device features 32 stimulation channels. The circuitry of the device is described in more detail in Chapter 2.

4.5.1 Array implementation

The electrode array is described here in more detail, as it is instrumental for applying the electrode selection algorithm. The array consists of 220×120 square pixels. All pixels share a common arrangement, with each pixel including one Pt- electrode, six horizontal and six vertical routing wires together with two switches, each of which is controlled by an SRAM cell. Every pixel covers a 17.5 × 17.5 µm2 area. Six such pixels are shown schematically in Fig. 4.6. The pixels are marked by red dashed squares. Additionally needed circuitry infrastructure to provide power and control signals to program the SRAM cells are also integrated into the pixels but are not shown in the figure for the sake of clarity. The array is designed in a modular approach. The basic pixel design specifies the general layout of the array. A particular realization of a switch-matrix scheme is then determined by (i) how these 6 horizontal and 6 vertical wires are interconnected, (ii) how these wires are connected to the switches, and (iii) how these wires are connected to the Pt-electrodes. The switch-matrix realization defines which electrode-to-readout configurations can possibly be realized once the design has been fabricated. To implement a specific realization of a switch-matrix scheme, the basic pixels need to be augmented by placing vias (connections between metal layers) and short pieces of wires at the appropriate locations within the electrode array. This is schematically shown in Figure 6, where some of the horizontal and vertical wires are interconnected with short pieces of metal to implement long-range wires. To facilitate the array design process, computer

79 4.5. Technology application

Figure 4.6: Pixel and pixel stitching. A small section of the implemented array is shown. A single individual pixel is surrounded by a dashed red square. Every pixel in the array shares a common arrangement of basic infrastructure. By placing small pieces of wires between the horizontal or vertical short wires, the wires can be extended to implement sophisticated routing options. aided design (CAD) software was developed to prototype and simulate different realizations of the array. This software meets three objectives: In the first stage of the design, different array realizations can be simulated, and it can be analyzed how well these realizations support recording from as many neurons as possible. Second, once a suitable realization has been found, the software implements the selected layout of the array by first placing all basic pixels and then placing vias and short-wire segments to connect wires with wires and wires to switches. Finally, once the device has been fabricated, the same software can be used to select recording electrodes and determine which switches need to be opened and which to be closed in order to realize a certain electrode selection or configuration. Due to constraints given by the actual layout of the array, not every arbitrary electrode selection is possible. For example, an electrode-to-readout channel mapping requiring certain routing resources cannot be routed if the respective resources are already used to realize other connections. In other words, resources in the array need to be distributed and carefully allocated to electrode-to-readout paths in order to achieve optimal utilization of all resources. An ILP based routing algorithm optimizes the resource distribution among all routing paths and is used to determine the state of all switches in the array. The resource allocation problem can nicely be combined with the previously detailed problem of finding the optimal recording electrodes. First, out of all available electrodes, those yielding best-expected LDA error are determined. Then,

80 Chapter 4. Selection of best recording sites

object constraint template difference Σ difference > minimum distance electrode Σ inflow - Σ outflow = 0 electrode 0 ≤ Σ inflow ≤ 1 wire Σ inflow - Σ outflow = 0 wire 0 ≤ Σ inflow ≤ 1 readout channel 0 ≤ Σ inflow ≤ 1

Table 4.1: ILP constraints if some of these electrodes cannot be routed due to, e.g., conﬂicting resource de- mands, the algorithm can automatically choose to abandon the less informative electrode and instead choose the next best candidate. Therefore, we now describe how the ILP algorithm to ﬁnd optimal electrodes can be combined with the routing algorithm to determine electrode-to-readout paths (Frey et al.(2010)). The electrode array hardware is modeled as a mathematical graph, where each electrode, wire and readout channel is represented as a node in the graph and switches are represented by arcs between the nodes. Then, the ILP algorithm can solve the routing problem by applying a set of constraints to all nodes. The constraints are summarized in Table 4.1. Essentially, the constraints for the electrodes and the wires ensure that for each incoming signal, there is a corresponding outgoing signal. Furthermore, not more than one outgoing signal is allowed per electrode. The cost vector c has a small cost assigned to each switch. This ensures that in cases of multiple possible routing paths, the path with the least amount of switches gets chosen. To combine these constraints with the electrode selection algorithm, the constraint matrix A can directly be combined with the constraint matrix of the electrode selection algorithm. This combination is visualized in Fig. 4.7. The matrix from the electrode selection algorithm is shaded by a light blue rectangle. The matrix for the ILP routing is marked with an orange background. The columns in matrix A not only correspond to electrodes, as was previously the case, but now all electrodes plus all switches in the array are represented by a column in A.

4.6 Discussion and conclusion

We have presented a framework to quantify the quality of a particular recording configuration. The framework quantifies the error probability when discriminat- ing between all a priori known templates of all included neurons. Two different algorithms were proposed to select an optimum recording configuration out of many different possible configurations. These algorithms were applied to a recently developed high-density MEA that allows for routing a subset of 1024 electrodes out of 26,400 available Pt-electrodes to readout channels.

81 4.6. Discussion and conclusion

Figure 4.7: ILP algorithm for combined electrode-selection and readout-path-routing. a) A schematic representation of a mathematical graph showing a small array with two electrodes (n1, n2) capturing signals from two neurons, one wire (n3) and one readout channel (n4). As only one of the available electrodes can be connected to the readout channel at any given time, the algorithm chooses the electrode capturing the larger diﬀerence between the two neuronal signals. b) Complete constraint matrix A for the simple case depicted in a). The shaded blue part of the matrix stems from the electrode selection algorithm. The shaded orange part comes from the routing algorithm.

82 Chapter 4. Selection of best recording sites

In order to compute the expected error, we made several simplifications. The assumption that a common noise covariance matrix Σ exists for all templates is not necessarily met with real recording data. Usually, the templates are non- stationary and vary over time, causing dispersion in the covariance matrix. It is known that neuronal spike amplitudes decrease when spikes occur within milliseconds after each other (Fee et al.(1996a)). Additionally, we ignored temporarily overlapping action potentials, which can frequently occur in real recordings (Pillow et al.(2013); Franke et al.(2015)). Both of these effects can be incorporated into the framework by using additional templates. To account for the non-stationary templates and amplitude scaling, one can use one template for a large-amplitude and an additional template for a small-amplitude action potential. The expected error will then not be computed between these two templates, as one will not want to discriminate between them. A similar strategy may work for overlapping spikes. However, a careful analysis of which additional templates need to be constructed will be required. Also, the expected error is overestimated by using the presented method. To assess the quality of a combination of electrodes, the expected error is computed for each pair of templates individually and subsequently summed together. By summing the error individually for each template pair, the error gets overestimated, as the same error gets accounted for multiple times for different template pairs. However, as it is our intention to have the clusters far apart, this approximation error will be very small for most cases, almost negligible, as the integrated error for distant probability densities is small. Most often, the error is within or below the numerical precision of the implementation. Our approach is related to earlier work (Van Dijck et al.(2012); Vysotska et al. (2014)), however, by relying on linear discriminant analysis, the computation of the expected error is optimal in a Bayesian sense. Therefore, the proposed method finds the optimal trade-off between detection and discrimination while considering the constraints given by the recording hardware. We do not rely on a simplifying heuristic for the greedy algorithm, as we compute the realisti- cally expected error for all template pairs. By computing the minimum distance between templates for every electrode combination separately, we can compute the expected error with increased accuracy. The approach involving the ILP algorithm is applicable for very large electrode arrays with tens of thousands of candidate electrodes that can be used to record from thousands of neurons. The heuristic ILP algorithm converges quickly to a suitable solution even if the number of candidate electrodes is large. Additionally, by taking hardware constraints into account, solutions will include only valid routing paths. If an optimal electrode selection cannot be routed, a next best one will be chosen automatically instead. Finally, the presented algorithms can be useful for a variety of different scenarios and technologies in which sensors can be chosen with a large degree of freedom, but for which the number of available recording channels is limited. Possible

83 4.6. Discussion and conclusion applications include scanning-beam microscopy or in-vivo implantable probes, as, in both cases, hardware constraints are considerable. These applications could also beneﬁt from the proposed combination of recording-site selection methods with hardware constrained routing techniques.

84 Chapter 5

Sub-millisecond closed-loop feedback stimulation between arbitrary sets of individual neurons

Jan Muller,¨ Douglas Bakkum, Andreas Hierlemann

Frontiers in Neural Circuits, 2013

Bio Engineering Laboratory, ETH Zurich, Switzerland

Abstract — We present a system to artificially correlate the spike timing between sets of arbitrary neurons that was interfaced to a complementary metal oxide semiconductor (CMOS) high-density microelectrode array (MEA). The system features a novel reprogrammable and flexible event engine unit to detect arbitrary spatiotemporal patterns of recorded action potentials and is capable of delivering sub-millisecond closed-loop feedback of electrical stimulation upon trigger events in real-time. The relative timing between action potentials of individual neurons as well as the temporal pattern among multiple neurons, or neuronal assemblies, is considered an important factor governing memory and learning in the brain. Artificially changing timings between arbitrary sets of spiking neurons with our system could provide a “knob” to tune information processing in the network.

85 5.1. Introduction

5.1 Introduction

Different theories describing learning and memory in the brain have been developed, and converging evidence shows that the precise activity timing of individual or groups of neurons may play a paramount role in plasticity of neuronal circuits. The well-known spike timing dependent plasticity (STDP) rule states that if two synaptically connected neurons fire within tens of milliseconds of each other, the connectivity strength of the involved synapses gets potentiated or depressed depending on the firing order. In pioneering studies, STDP rules were discovered (Markram et al.(1997a)) and further characterized (Bi and Poo(1998); Song et al.(2000)) by observing the effect of correlated firing of two neurons either artificially induced by stimulating a pre- and a postsynaptic neuron with two patch clamps or by applying trains of paired-pulse stimuli to one neuron in the network (Bi and Poo(1999a)). Furthermore, computation in a network is likely due not only to the relative timing of two individual neurons but also to the correlated activity of different neurons forming an associated group, i.e. assembly (Chang et al.(2000a); Izhikevich(2006)). In this vein, different studies reported the existence of precise time-locked activity patterns of multiple neurons, both in vivo and in vitro (Abeles and Gerstein(1988); Bienenstock(1995); Ikegaya et al. (2004); Rolston et al.(2007)). Having a system to generate feedback stimulation quickly and accurately to interact with such activity patterns would expand such studies beyond finding rules governing the plasticity between two cells towards finding rules governing the spatio-temporal dynamics of whole networks or assemblies (Froemke and Dan(2002); Izhikevich(2004)). In recent years, different systems to artificially control such feedback stimulation in a closed-loop manner, and thus study neuronal plasticity, have been developed for both in vivo (Jackson et al.(2006c); Bontorin et al.(2007a); Venkatraman et al.(2009a)) and in vitro applications (Bontorin et al.(2007a); Hafizovic et al. (2007e); Novellino et al.(2007); Rolston et al.(2010); Zrenner et al.(2010); Wal- lach et al.(2011)). In turn, activity-dependent feedback stimulation was shown to modify the functional connectivity of neuronal networks, both in vivo and in vitro, as done by reprogramming the motor output of freely behaving primates (Jackson et al.(2006b)), changing the functional connectivity in rat forelimb sen- sorimotor cortex (Rebesco et al.(2010)) or shaping in vitro neocortical networks into predefined activity states (Bakkum et al.(2008a)). In vivo systems usually record from needles inserted into a certain location of the brain and subsequently stimulate the same or another site upon the detection of activity. These systems usually comprise the implanted needles, a head stage to amplify the signals, and some means to transmit the acquired signals to a PC. In the case of closed-loop feedback stimulation, these systems usually feature a dedicated very-large-scale- integrated application-specific circuit (VLSI ASIC) (Chen et al.(2009); Rizk et al.(2009); Lee et al.(2010); Azin et al.(2011)), or use a general purpose microcontroller to achieve the respective goals (Mavoori et al.(2005); Zanos et al. (2011)). Most in-vitro systems, on the other hand, use a data acquisition card

86 Chapter 5. Sub-millisecond closed-loop feedback stimulation

(DAQ) to sample data for analysis on a PC; feedback stimulation is typically returned through a DAQ card as well. In order to accurately control the timing of feedback stimulation loops within the timescales relevant for STDP to occur, the delays introduced by a system must be understood. A generic description is given in Fig. 5.1. Different system implementations will have different sources for and values of delays. Signal processing algorithms introduce an inherent delay in the processing itself. Sys- tems, which rely on general-purpose computers, might introduce latencies and jitter through the presence of data buffers, interrupts, shared resources, or user interactions, etc. In Fig. 5.1, the time points t0-3 and tS specify the occurrence of important events. At t0=0, the trigger neuron emits an action potential, which is recorded by the acquisition system. After entering the signal processing stages, it is ready to be detected as a spike event at time t1. From there, the system emits a stimulation pulse hitting the electrode at time t2. Conventionally, the loop is considered “closed” at this point. The stimulation pulse evokes neuronal activity, frequently activating nearby axons (Bakkum et al., 2008a) whose signals propagate antidromically towards the soma until eliciting an action potential at time t3. In the case depicted in Fig. 5.1, where the trigger neuron is synaptically connected to the elicited neuron, an additional biological time, tS, denotes the duration of an action potential propagation through the axon of the trigger neuron until synaptic activation of the elicited neuron. In case where t0–t1–t2 is faster than t0–tS, that is when the signal propagates faster through the artificial feedback loop than down the axon towards the biological synapse, acausal stimulation, and thus the introduction of long term depression (LTD) according to the STDP rule, is possible. In order to apply closed-loop stimulation feedback precise and fast enough to study plasticity at the timescales of STDP or acausal stimulation, and flexible enough to interact with cell assemblies, we developed a field-programmable gate array (FPGA)-based system, interfaced with a complementary metal–oxide– semiconductor high-density microelectrode array (CMOS-MEA). The CMOS- MEA features a total of 126 readout and 42 stimulation channels, which can be connected to an almost arbitrary subset of 11,011 5 × 7 µm2 electrodes, arranged in a 2 × 1.75 mm2 array. The feedback stimulation loop is closed around the CMOS-MEA using an FPGA that performs signal processing, such as spike detection and feedback generation. The system functionality was verified using cultured networks of cortical neurons and glia. The minimum programmable latency of the closed-loop stimulation feedback (t0–t1–t2) was 400 µs with jitter below 50 µs, suitable to induce STDP. This is faster than many axonal propagation delays (t0–tS), rendering it possible to conduct acausal stimulation experiments. An “event engine” was designed and implemented to trigger feedback stimulation at the occurrence of activity patterns, such as those described in (Ikegaya et al. (2004); Rolston et al.(2007)). Patterns could be of almost arbitrary length and could consist of up to 1000s of individual elements, only limited by the available resources of the FPGA. Configurations for the event engine could be (re)loaded

87 5.1. Introduction

Figure 5.1: Schematic overview of latencies in feedback stimulation systems. a) The different components making up a closed-loop feedback stimulation system are shown. The green circle represents the “trigger neuron” whose action potential initiates the start of the loop. The green line represents an axon connecting to synapses of the elicited neuron drawn in yellow. The black dashed arrow shows the closed-loop feedback stimulation path. Between data acquisition and stimulation feedback, different components, over which the feedback loop can be closed, are possible, including digital signal processing hardware, a real-time host PC, or a general purpose host PC. The time points t0-3 and tS correspond to different events as listed in b), such as the occurrence of the spike; its detection after signal processing; the stimulation feedback; and the antidromic propagation of an action potential back into the soma of the elicited neuron. At time tS, the synapse activates due to presynaptic activity of the trigger neuron. The color of the traces corresponds to the color of the timings of t0-3,S and schematically shows the timeline of the respective signals.

within milliseconds. Unique to this system is the possibility to enable low-latency, high-throughput, STDP-like experiments as well as acausal stimulations across many individual neurons or neuronal assemblies in parallel through the simultaneous application of many feedback stimulation loops. To infer changes in synaptic strengths, correlations between putative mono-synaptically connected neurons (Fujisawa et al.(2008)) can be monitored using extracellular spikes. In the future, high-throughput STDP experiments will be possible by adding a patch electrode to the system in order to monitor changes in intracellular post-synaptic currents.

88 Chapter 5. Sub-millisecond closed-loop feedback stimulation

5.2 Methods

5.2.1 System Architecture

The main design goals were to implement (1) multiple feedback stimulation loops (2) to match arbitrary spike patterns with (3) short latencies (<1 ms) and (4) high accuracy (<50 µs) (5) while still recording from all available 126 channels. A main component of the presented system is an FPGA, used to hijack signals traveling between the analog-to-digital converter on the CMOS device and the host PC. Due to the inherent parallel nature of FPGAs, signal processing and feedback generation using data from additional recording channels can be done without introducing additional delays or jitter. The system consists of three main parts as shown in Fig 5.2. The first is a high- density CMOS-MEA device featuring on chip signal-conditioning, stimulation, and analog-to-digital conversion (ADC) units (Frey et al.(2010)), described in more detail in the next section. It is plugged into a custom printed circuit board (PCB) that provides reference voltages and clock signals. The digital data as provided by the CMOS-MEA are transmitted through a low-voltage differential link to reduce sensitivity to electromagnetic interferences as caused, for example, by a nearby incubator. The second part is an FPGA, which reads in the differential signals and subsequently performs signal processing, spike detection, and feedback stimulation, as well as compression and framing of the data to be sent via TCP/IP over Ethernet to a host PC, the third main part. On the host PC, further data analysis can be performed online or offline. It is also used to program and control the CMOS-MEA device during experimentation with different settings, like amplifier gain or electrode-to-amplifier routing, in order to be adopted for use in different experimental sessions.

5.2.2 CMOS device

The CMOS-MEA includes 126 readout channels with programmable ampliﬁ- cation (0 dB to 80 dB), on chip ADCs sampling at 20 kHz, and stimulation capabilities (see below). It features a sensor area of 2 × 1.75 mm2 with a total of 11,011 electrodes, each with a size of 5 × 7 µm2 and a pitch of 18 µm. Beneath the electrodes resides a sophisticated analog-switching matrix to connect an almost arbitrary subset of the 11,011 electrodes to the 126 readout channels. The readout electronics were placed outside of the sensor array, instead of directly below the electrodes as done in active-pixel sensor devices (APS) (Berdondini et al.(2009c)), to provide space for larger circuitry elements that produce less noise. This scheme also allows for reducing the pitch of the electrodes below the spatial requirements of the readout electronics. See Frey et al.(2010) for more details.

89 5.2. Methods

Figure 5.2: Overview of the presented closed-loop system, implemented with a CMOS-MEA, an FPGA, and a host PC. a) Micrograph of the CMOS- MEA highlighting the electrode array, ampliﬁcation and stimulation units, and the digital core with an inset showing a close-up of the stimulation buﬀer. b) Photograph of the CMOS-MEA plugged into the custom printed-circuit board, which is connected through an LVDS link to the Xilinx Virtex II pro FPGA board from Digilent Inc., Pullman, USA. The host PC running data acquisition and visualization software is connected to the FPGA through Ethernet. c) Schematic diagram of the setup. The diagram shows the acquisition (upper part) and stimulation path (lower part). The feedback stimulation loop is closed around the CMOS-MEA and the FPGA. The components are described in detail in the text.

90 Chapter 5. Sub-millisecond closed-loop feedback stimulation

5.2.3 FPGA

A reprogrammable Virtex II pro FPGA (Xilinx Inc., San Jose, USA) was used as an intermediate signal-processing device between the CMOS-MEA and the host PC to perform real-time signal processing, decision-making and feedback generation. The FPGA acquires digital data coming from the differential link and forwards it to a PC over Ethernet. The Virtex II pro features an embedded PowerPC microprocessor running at 300MHz that operates a Linux kernel with a Busybox operating system. The TCP/IP stack of the Linux kernel handles the network communication and data transfer. As the embedded PowerPC microprocessor is relatively slow, compared to modern CPUs, this provides a bottleneck for fast data transmission. We measured the latency between the TCP/IP stack of the FPGA and the host PC to be 83±21 ms (mean±SD, N =308) at full-frame data transmission, which is larger than the STDP window of up to tens of milliseconds. One solution to this problem might be to stop streaming of the full data readout, while performing a closed-loop experiment and to only route out the data channels strictly needed for the closed-loop feedback stimulation. This would free some of the bandwidth of the Ethernet link and make it available for faster feedback stimulation. Crucially, however, we would lose the possibility to simultaneously monitor neural activity elsewhere in the cultured network by applying such a paradigm. Another option might be to bypass the Ethernet link by streaming the data directly to a DAQ card, attached to the host PC, and to send stimulation information back through a second link to the FPGA. All these methods are less practical than using the universal TCP/IP connection, which plugs into almost every kind of host PC and does not require additional hardware. An attractive alternative for achieving low latencies was to implement all needed signal processing and feedback generation directly on the FPGA. The next paragraphs highlight the different building blocks needed to implement such a scheme. Although the FPGA can be reprogrammed at will, this is time consuming and error prone and, therefore, not suitable during an experimental session. To accommodate reprogramming, a more flexible, module-based design was developed in VHDL and programmed into the FPGA logic together with a software interface to quickly reconfigure the connectivity of the individual modules (see Section: 5.2.5, Event Engine).

5.2.4 Spike detection

One such signal-processing building block is spike detection, which extracts spiking events from the raw voltage traces, recorded at the electrodes. Spike detection is implemented as a threshold crossing. The signals are first digitally band-pass filtered with a two tab Butterworth filter (500Hz-3kHz) to suppress DC offset components and higher frequency noise; this will emphasize the action potential frequency components. The detection threshold level is user-programmable and typically set around 4.5 times the noise standard deviation. During experimen-

91 5.2. Methods tation, this value can be determined by software running online on the host PC. After an identiﬁed spike event, we set a programmable refractory period to 3 ms. After stimulation, detection was disabled for 3 ms as well, to avoid oscillating loops due to feedback stimulation artifacts being falsely classiﬁed as spikes.

5.2.5 Event Engine

To avoid time-consuming reprogramming of the FPGA fabric, a more flexible and modular event-based scheme for feedback generation (Event Engine) was designed and implemented. The event engine consists of small building blocks, called modules, each of which implements a specific simple function. Each module has one or more event sinks as inputs and one event source as an output. By connecting the event sources to the appropriate event sinks, different, almost arbitrary pattern matching and event handling algorithms can be achieved. Table 5.1 summarizes the implemented modules. Fig. 5.3 shows different basic configurations to achieve defined pattern matching. In Fig. 5.3 A, the simplest closed-loop configuration is depicted, where the source of a spike-detection module gets connected to the sink of a delay unit and from there to a stimulation function generator. Whenever the source produces an event (i.e., in this case de- tects a spike), the sink triggers a stimulation pulse after a defined time delay. By means of software, the sources can be connected to sinks dynamically and rapidly within milliseconds while running an experiment such that pattern matching can adapt to ongoing activity in the living culture. One notable property is the lack of time binning. Each spike gets represented as a single pulse with a temporal resolution set by the sampling frequency, i.e., 20 kHz. As a consequence, certain desired operations might not make sense, as the biological neurons have some inherent variability in when they spike. For example, the user might want to match a pattern, where two neurons spike together (see Fig. 5.3 E). To achieve this, a SPREADING module ‘spreads’ the spike pulse in time in order to compensate for jitter. This way, the subsequent AND module can generate an output event whenever the two neurons fire together within a specified range of time. As discussed in Ikegaya et al.(2004); Rolston et al.(2007), 2 ms is suitable for most recurring patterns. Another module can be used to convert the spread- out spike pulse back into a single one-shot event, which then can be used, for example, to trigger the stimulation unit only once per spread-out pulse. The particular selection of implemented modules (as listed in Table 5.1) represents a minimal set, which, if combined in the appropriate way, allows for matching different kinds of events, such as specific spatio-temporal activity patterns, time sequences, network bursts, local bursts etc. In order to keep the event engine as flexible as possible and adaptable to different, possibly unforeseen pattern matching sequences, the implementation of a minimal set of small building blocks has been chosen over the approach, where each envisioned pattern would require a single, but more complex and less flexible building block. Thus, available mod-

92 Chapter 5. Sub-millisecond closed-loop feedback stimulation ules can be combined together in almost inﬁnite diﬀerent ways, limited only by the available FPGA memory that keeps track of all source-sink associations.

5.2.6 Stimulation/Function generator

The CMOS-MEA has 32 on-chip integrated stimulation units, which are driven by two 10 bit DACs. On the FPGA is a function generator implemented to achieve arbitrary stimulation waveforms. A defined waveform has to be programmed at the start of the experiment. We used biphasic, first positive then negative voltage pulses of 200 µs duration per phase and ± 300 or 400 mV amplitude. The stimulation buffers can be chosen to operate in voltage- or current mode (Livi et al., 2010). Whenever the event engine outputs an event, the appropriate stimulation buffer, located on the CMOS-MEA, gets connected, and the function generator starts its operation. Stimulation artifacts on the readout channels could result in falsely detected spikes and cause a reverberation problem for low-latency feedback loops. Therefore, spike detection is blanked during a time period of a few milliseconds after stimulation onset.

5.2.7 Cultures

The performance of the closed-loop system was tested with cortical neurons and glia grown over the CMOS-MEA. Animal handling protocols were approved by the Basel-Stadt Veterinary office according to Swiss federal laws on animal welfare. Briefly, a time-pregnant rat was anesthetized using isoflurane, then de- capitated to gain E18 embryos. Cortices were extracted from the embryos and dissociated enzymatically in trypsin (Invitrogen) followed by mechanical trituration. A layer of laminin (Sigma) over a layer of poly(ethyleneimine) (Sigma) was used to adhere between 20k to 40k cells. Plating media consisted of 850 µL of Neurobasal, supplemented with 10% horse serum (HyClone), 0.5 mM GlutaMAX (Invitrogen), and 2% B27 (Invitrogen). After 24 hours, the plating media was changed to growth media: 850 µL of DMEM (Invitrogen), supplemented with 10% horse serum, 0.5 mM GlutaMAX, and 1 mM sodium pyruvate (Invitrogen). Cultures matured for 3 to 4 weeks prior to experimentation, and experiments were conducted inside an incubator to control environmental conditions (34.5°C and 5% CO2). For further details see (Hales et al.(2010)).

5.3 Evaluation and Results

This section begins with data characterizing the suitability of our setup to perform closed-loop feedback stimulation experiments, using cultures of cortical neurons and glia for validation. First, the process of identifying neurons to be used in closed-loop feedback stimulation will be described. Then the system’s loop speed

93 5.3. Evaluation and Results . p occurred in a occurred occurred. B B B , whenever input c or and A A . t . t , detected a spike. ) an internal accumulator and c back into a single event. B A , however, no event on A to the output or drops it after a Bernoulli- , in order to create inhibitory feedback-loops. t A in time for a defined time, ) or decrements (event by a defined amount of time A A A happened. A Emits an event, when both of the two input events, Emits an event, when either of theEmits two an input event, events when andefined event time on window, distributed pseudo-random variable with a definable probability, Converts the onset of a spread-out event Emits an event, when the specified channel, Single pulse after systemstimulation start-up, protocols. which can be used to start repetitive Delays the event simultaneously. Propagates the event Spreads the event Increments (event Generates a stimulation pulseevent on the specified channel, emits an event afterit a is definable reset threshold, to n, zero. has been reached, after which ) c ) ) ( ) B A ) , ) , ) ) t A A , A B ( ) , B , , t A , , ( p n B c ( A , ( A , ( ( t A ( ( DELAY AND OR INH RAND ACCU SPREAD SPREAD DETECTION STIM START

Table 5.1: A minimal set of modules making up the event engine. Conﬁgurable parameters are represented in italics (t, p, n, c), and input events are denoted in bold letters (A, B).

94 Chapter 5. Sub-millisecond closed-loop feedback stimulation

Figure 5.3: Example configurations of the event engine. Stitching together the appropriate set of modules allows the event engine to be configured to match a variety of patterns in order to trigger feedback stimulation. Different minimal examples are shown. a) A DELAY element is inserted after a DETECTION module to trigger STIMULATION after a programmable delay. This configuration, with the delay set to zero, was used for the experiments shown in Fig. 5.5 and Fig. 5.7. b) Either an event on channel A OR an event on channel B triggers stimulation. c) In a programmable time window before and after an event on channel A, there may not be any event on channel B in order to trigger stimulation (trace C). d) A RAND module propagates or discards the events, in this case with a probability of 1/2. e) Events on channel A and channel B are fed through SPREAD modules into an AND module, which outputs events (on trace C), when both inputs are active. The intermediate trace C is fed into a SPREAD-1 module to trigger stimulation at the onset of the event. f) When the event on channel B happens subsequently to an event on channel A, an event C is generated g) An ACCU module is set to increment, when either an event on channel A OR channel B happened, and to decrement, when a delayed event from channel B (trace C) arrived. In this example, the ACCU threshold is set to three events. Once the threshold is reached, the internal counter gets reset to zero. When the three input events happen shortly after each other, a stimulation event gets emitted. As shown in the example, the delayed channel B (trace C) decrements the accumulator and thus delays or prohibits crossing of the threshold. h) All modules can be combined together to achieve almost arbitrarily complex pattern matching. For example, this configuration was used to match the pattern of Fig. 5.6. The formula describing this pattern is: STIMULATION( 1, SPREAD-1( AND( AND( SPREAD( 2ms, A ), SPREAD( 2ms, B ) ), SPREAD( 2ms, C ) ) ) ).

95 5.3. Evaluation and Results and jitter performance will be quantiﬁed. An example event engine was run to provide stimulation feedback, triggered by an activity pattern. Preliminary data and techniques to analyze the consequences of such stimulation on the functional connectivity between neurons will be presented and discussed. Finally, an experimental session to induce LTD through acausal stimulation will be sketched, and its implications discussed. Data in the ﬁgures demonstrate proof-of-principle experiments from individual cultures, the setup has, however, been successfully applied to many tens of cultures.

5.3.1 Recording/stimulation selectivity

High-density CMOS-MEAs can potentially sample from complete neuronal populations. Due to the high density (18 µm pitch) of the CMOS electrode array, every neuron lying on the 2 × 1.75 mm2 array can be bidirectionally addressed. On the other hand, when stimulating one electrode, a defined subset of neurons is often directly activated in response (Bakkum et al.(2008a)). Fig. 5.4 shows such a scenario. In Fig. 5.4 A, one electrode, marked with a black cross, was stimulated multiple times, and the evoked activity was recorded during a window of 12 ms after stimulation onset. The median calculated over all voltage traces filters out noise and spontaneously spiking neurons/traces. Reliable activity (usually with a jitter on the order of 100 µs or below) is considered due to an antidromic action potential initiated at the neuron’s axon (Lipski(1981)). Since only a subset of 126 out of the 11,011 electrodes can be readout simultaneously, the stimulation sequence was repeated multiple times, each time with a different subset of electrodes, until all electrodes were covered. After recording all sequences, the traces of the individual recordings were aligned in time. To highlight the electrodes that recorded elicited action potentials, the negative peak of the recorded voltage level during 12 ms after stimulation is color-coded and clipped at −100 µV. The red circles around the exemplified 11 spots highlight neurons that fired directly-elicited action potentials. Their traces are individually shown in Fig. 5.4 B, demonstrating that the elicited action potentials were reliably and precisely fired after a given time, and, only in a few cases (traces 2, 4, 6, 9), activity with different timing occurred. These could stem from a different neuron that happened to sit near the same electrode and/or from action potentials occurring within a coincident network burst.

As shown, recording and stimulation with the CMOS-MEA feature high spatial resolution and, therefore, are locally very conﬁned. However, the facts that one electrode can detect signals from more than one neuron, and that the stimulation through one electrode can directly evoke action potentials of more than one neuron have to be considered when planning closed-loop feedback stimulation experiments. In this case, the feedback loop is not closed between two neurons, but includes two sets of neurons.

96 Chapter 5. Sub-millisecond closed-loop feedback stimulation

a b 100 µV 2 ms 9 10 11 10

4 5 8 9 1 8 7 6 3 5 4 2 3 6 2 0 7 1 Voltage (µV) 11 200 µm −100

Figure 5.4: Identification of directly evocable action potentials. a) Data recorded in response to repeated stimulation of one electrode (black cross) from the whole 2 × 1.75 mm2 sensor area of the CMOS-MEA (each pixel is one electrode). Recording electrode configurations were scanned across the array in sets of 126 electrodes at a time. For every configuration, data was recorded for 12 ms after stimulation onset. The amplitude of the negative voltage peak within these 12 ms is color-coded and clipped at −100 µV. Blue indicates the detection of directly evoked somatic action potentials. b) Example traces from 11 somas and the stimulation pulse are shown on the right. Traces from 30 stimulation trials are overlaid, with the median trace highlighted in black. The stimulation artifact was blanked prior to recording. Numbers are ordered by increasing distance from the stimulation site.

97 5.3. Evaluation and Results

t −t 1.25 ms 0 3 40

400100 30 Stimulation Amplitude (mV) stimulation artifact 300 V)

µ 20

0 0

−10

−50

−20 trigger spike elicited spike

Recording Amplitude ( −300 −30

−100 t −t 0.85 ms −400 2 3 t −t 0.2 ms −40 1 2

t −t 0.2 ms −50 −0.5 0 0 1 0.5 1 1.5 2

Figure 5.5: Feedback stimulation performance. 128 traces from a closed-loop stimulation sequence are aligned at the stimulation onset-time and overlaid. Traces in red show the trigger spikes with the median over all trigger traces shown in bold red. The stimulation artifact is greyed-out for better visual clarity. The traces in black show spikes, elicited in all but four cases after stimulation. The median over all elicited traces is shown in bold white. The antidromic propagation delay for the elicited spikes was around 0.85 ms. The diﬀerent timings, detection delay, stimulation delay, and antidromic propagation delay sum up to the full loop delay of 1.25 ms.

5.3.2 Feedback latencies

According to the rules of STDP, the timing window to induce long-term potentiation at synapses is between less than a few milliseconds and up to tens of milliseconds postsynaptic activation before and after pre-synaptic activity. Thus, even though feedback cycles of 5 to 10 ms are fast enough to induce long term potentiation (LTP), we aimed at reaching cycle-times below 1 ms to enable the system to perform acausal stimulations, as explained in the respective section below. Fig. 5.5 shows the overlay of 128 traces of the feedback loop. Here, the event engine was conﬁgured to detect events on only one channel and stimulate im- mediately after detection, i.e., without any further delays in order to test the system performance (cf. Fig. 5.3 A). The traces are aligned at the onset time of the stimulation pulse, and time zero is set to be at the negative peak of the spike of the trigger neuron. In red are the traces from the trigger neuron, and in black, the traces from the elicited neuron. The timing between a trigger neuron spike

98 Chapter 5. Sub-millisecond closed-loop feedback stimulation and the onset of the stimulation pulse was 200 µs, i.e., 4 sampling periods. This delay arises as follows: 50 µs (1 sampling period) was used to buffer the incoming data in the FPGA; 100 µs accounted for the delay of the two-tab Butterworth filter and the last 50 µs account for all other delays, such as synchronizing the stimulation pulse with the recording sampling time. Delays for sending digital data between the CMOS device and the FPGA were on the order of nanoseconds and thus are negligible. When stimulating with biphasic voltage pulses, the steep negative transition, which injects negative current (I = C×dV/dt), is the time point, when a cell is activated (Wagenaar et al.(2004a); Bakkum et al.(2008a)). Thus, this time point was taken to measure the latency between stimulation and an elicited spike. In the case depicted in Fig. 5.5, this timing is 0.85 ms , and the overall latency between trigger neuron activity and a spike on the elicited neuron was 1.25 ms. As can be seen in Fig. 5.5, besides achieving short feedback cycles, another advantage of using digital hardware (in this case FPGAs) for feedback generation is that no additional jitter is introduced, as such a system is fully deterministic. Sources of jitter in other systems (Hafizovic et al.(2007e); Rolston et al.(2010)) that close the feedback-loop around general purpose or real-time personal computers are, for example, system interrupts that might disrupt the data processing, or buffer sizes of the USB, TCP/IP or DAQ cards, which have to be set large enough in order to guarantee full data throughput. Usually these buffers have a size larger than one sample period. Depending on when an event happened inside this buffer, the latency could be larger or smaller and thus introduce jitter. This can be avoided by using digital hardware to hijack the data stream. In our case, the jitter was below ±50 µs and arose from the fact that neural activity is, of course, not aligned to the sampling period of the CMOS-MEA (50 µs). The exact time of the threshold crossing relative to the negative spike peak depends, among other things, on the slope of the spike waveform. Since the recorded signal was not interpolated between samples, this was an unavoidable source for jitter.

5.3.3 Pattern matching

To demonstrate the event engine in operation, feedback stimulation, triggered by an activity pattern, was performed. For the dataset presented in Fig. 5.6, the event engine was programmed according to Fig. 5.3 H and classified spontaneous activity patterns as follows: A neuron recorded on electrode N2 fires an action potential; then an action potential is recorded from a neuron on electrode N3 after 3 ms; finally an action potential is recorded on electrode N1 after another 1.5 ms. Each individual event occurrence was allowed to have a jitter of ±1 ms. After successful identification of such a pattern, a stimulation pulse was emitted to elicit action potentials on a different neuron, NE. The cell cultures under investigations typically expressed bursting behavior, and this was when almost all of the patterns occurred. During bursts, the cells usually fired more than once

99 5.3. Evaluation and Results

N3 5 ms N1 N2

8 ms N2 3.5 ms

N3 N1 2 ms ±400 mV

100 µV 1 ms NE 100 µm NE −8 −5 −3.5 0 1.75 ms

Figure 5.6: Pattern-matching feedback stimulation. Electrode traces were recorded from neurons sitting on three different electrodes N1-N3 while performing pattern matching. The pattern was matched 22 times within 12 minutes, all overlaid and drawn in light-grey color. One arbitrary pattern is highlighted with black traces. The 12 ms before and 4 ms after stimulation pulse are shown. The orange, green, and blue colored boxes represent the spread-out-windows set in the event engine. A yellow box of arbitrary width is drawn around the elicited activity of neuron NE. Above the traces, negative peak times are marked with black vertical bars, showing spikes clustered within the colored boxes. The figure on the right shows electrode locations and the timings making up the pattern to match as well as the antidromic propagation delay of 2 ms to the elicited neuron. at an elevated frequency, and this explains why the neurons on electrodes N1- N3 showed additional spikes ‘outside’ of the detected pattern. Nevertheless, the pattern matching event engine identified 22 activity pattern occurrences during 12 minutes of recording. To assess the connectivity between different neurons and the efficacy of change, induced by the closed-loop feedback stimulation, cross-correlation curves (Perkel et al.(1967b)) were computed between spike trains of the trigger neuron and the elicited neurons. When exceeding a 95 per cent confidence interval (Brillinger (1976)), correlation is considered significant. Fig. 5.7 shows three descriptive cases, comparing the cross-correlation curves from 1.5 hours of spontaneous activity before and after closed-loop feedback stimulation was applied for 1 hour. To evaluate significance of the change, a similar procedure as in (Fujisawa et al. (2008)) was used. Briefly, the two times 1.5 hours of spontaneous activity recordings were divided into smaller bins of 10 minutes duration and were randomly assigned to be before or after the closed-loop stimulation. Cross-correlation from this shuffled data was computed for both “before” and “after” and the difference was evaluated. This procedure was repeated 1000 times to generate a surrogate data set. Points on the x-axis, where the true difference is larger than 95% of the surrogate data, were assigned to be significant and are marked with an orange bar in Fig. 5.7. Assessing the true connectivity of neuronal networks by means of extracellular measurements is difficult, and using the cross-correlation to that end is not ideal, as effects like common inputs or firing rate changes cannot be

100 Chapter 5. Sub-millisecond closed-loop feedback stimulation

a b c 4 6 7 before after significant

Normalized X−Corr 1 0

−20Time (ms) 20 −100Time (ms) 100 −600Time (ms) 600

60 60 60 40 40 40 20 20 20 Time (min) 0 0 0 −20 20 −20 20 −100 100 −100 100 −600 600 −600 600

Figure 5.7: Cross-correlation analysis. Three descriptive cases of changes in correlated firing between trigger neurons and elicited neurons. Spontaneous activity was recorded 1 hour before and 1 hour after the application of closed-loop feedback stimulation. Periods, where the difference exceeded a confidence bound (see text), were assigned to be significant and are indicated with an orange bar. The 95% confidence intervals are indicated with black dashed lines. Cross-correlation is computed based on trains with 9000 to 13000 spikes per neuron. a) Relative probability remained constant, but the timing between trigger neuron and elicited neuron changed and became more synchronous. b) The elicited neuron became more likely to fire in concert with the trigger neuron. c) Relative timing within a network burst changed. easily explained. However, in our context of evaluating the effect of feedback stimulation, we do not necessarily seek to precisely explain the changes in network connectivity, but to rather demonstrate that a change occurred at all and to what extent. One motivation for very short feedback cycles is to open the possibility of acausal stimulation. If the closed-loop stimulation (t0-t2) is faster than the time it takes the action potential to travel along the axon and hit the synapses (t0-tS), acausal stimulation and, therefore, induction of LTD by means of closed-loop feedback stimulation is possible. The time that it takes for an action potential, initiated at the axonal hillock, to propagate down the axonal arbor to the synapses depends on the propagation velocity of action potentials along axons and the length of the axons. Action potential conduction velocities in unmyelinated axons were reported around 0.2 to 0.4 m/s (Debanne et al.(2011)). As demonstrated in Fig. 5.5, the closed-loop stimulation (t0−t2) can be as fast as 0.4 ms, meaning acausal stimulation is possible for trigger neurons (t0) with unmyelinated axons that synapse to an elicited neuron (t3/S) after a minimum axial length of 80 to 160 µm. Fig. 5.8 shows such an acausal stimulation procedure. First, before applying a closed-loop, the activity between different neurons was measured then evaluated by computing the cross-correlation. In the example in Fig. 5.8, the firing activity of the second neuron B with respect to the first neuron A was elevated around a delay of 2.5 ms, implying neuron A has a functional con-

101 5.3. Evaluation and Results

a b c d A

artificial synapse0.4 ms ~2.5 ms

~2.4 ms ±0.4 ms ~1.1 ms 0.7 ms B 100 µm

±400 mV e

before CL during CL after CL

Figure 5.8: Schematic of an acausal stimulation sequence. a) Spontaneous activity before application of the closed-loop. Shown spike traces are the median waveform of several spikes aligned at the negative peak. Top: Spike trace of the trigger neuron, A, in green. Middle: Example spike trace of a correlated neuron, B, drawn in yellow. The time delay between the plotted spikes of neuron A and neuron B was chosen to align with the maximum peak of the cross-correlation curve. Bottom: Cross- correlation curve of spike-times of neuron B with respect to neuron A. 95% confidence intervals are drawn with dotted red lines. Cross-correlations were computed with trains having 2000 to 3000 spikes. Significantly elevated correlated activity of neuron B can be detected around 2.4 ± 0.4 ms after neuron A fired an action potential. b) Same situation as in a) but with a closed-loop feedback stimulation applied. Due to the low-latency loop, the time delay of the yellow spikes with respect to the green ones was reduced by about 1.3 ms. For neuron A, the trace was zeroed at the start of the stimulation pulse. c) Same as a) but after the application of the closed-loop feedback stimulation. The cross-correlation no longer shows a significant peak for latencies larger than zero. The time delay between the plotted spikes of neuron A and neuron B was again chosen to align with the maximum peak of the cross-correlation. d) Geometric sketch of the situation. The trigger neuron A and its axon are shown in green and the elicited neuron B in yellow. e) Comparison of the two cross-correlation curves before (black) and after (red) the acausal stimulation with their 95% confidence intervals.

102 Chapter 5. Sub-millisecond closed-loop feedback stimulation nection with neuron B. Integrating the cross-correlation curve, where it exceeds the confidence intervals around 2-3 ms after the reference time zero, reveals an integral probability of around 40% chance for neuron B to spike 2 to 3 ms after neuron A had fired. Once two such neurons could be identified, closed-loop stimulation can be applied between them with a very short feedback cycle. In the presented example, the delay from the trigger neuron to the elicited spike was around 1 ms, smaller than the average delay between the occurrence of their spontaneous action potentials. The closed-loop feedback stimulation was applied for 20 minutes, and, afterwards, the correlation was measured again. Now, the correlation no longer exceeded the confidence intervals at around 2 to 3 ms after the trigger neuron. Note, however, that Bi and Poo(1998) have shown that LTD can only be induced, if the spontaneous synaptic efficiency is not strong enough to evoke a postsynaptic action potential. Otherwise, the postsynaptic Ca2+ in- flux dominates, and LTP will occur. For the experiment shown in Fig. 5.8, the elicited neuron spiked only a fraction of the time, and provided an intermediary synapse; in all other cases, evoked excitatory postsynaptic currents (EPSCs) remained below the threshold. Further experiments are required before drawing conclusions. Additionally, to explore LTD and LTP in more depth, and, advanta- geously, across many synapses simultaneously, extracellular recordings targeted to many trigger neurons and an elicited neuron on the CMOS-MEA could be combined with an intracellular patch clamp, attached to the elicited neuron and measuring the incoming EPSCs.

5.4 Discussion

With the presented system, capable of applying multiple flexible feedback loops simultaneously, many different experiments will be possible. The dynamic clamp technique proved to be a valuable tool for investigating the membrane dynamics involved in action potential generation (Destexhe and Bal(2009); Economo et al. (2010)). In such systems, intracellularly applied closed-loop-controlled voltage feedback enables the manipulation of cell membrane functions. Similarly, extracellularly applied closed-loop stimulation feedback, as presented in this work, might provide a useful tool for investigating cellular and network level plasticity and enable the manipulation of neuronal network functions. Potential questions include, how information processing and the amount of memory that can be stored in a cultured network are influenced by adding one or more feedback loops. Further experiments might involve more detailed studies of both LTP and LTD of individual sets of neurons by implementing causal and acausal feedback loops between them. Using the pattern matching capabilities of the event engine will allow for extending plasticity studies to the network level. For example, investigations of the temporal order and history of spike trains, similar to those reported by Froemke and Dan(2002); Ikegaya et al.(2004), could be performed, however, in parallel on multiple different neurons and pathways and, in addition,

103 5.4. Discussion the respective pathways could be dynamically altered by targeted closed-loop feedback stimulations. Further rules governing plasticity beyond the classical STDP could be investigated.

An inherent limitation of extracellular recording systems is the inability to directly measure excitatory postsynaptic currents (EPSCs). Conventional plasticity studies rely on patch clamp to directly measure the EPSC to assess synaptic connectivity strength. Since these currents are not accessible with extracellular measurement techniques, indirect methods to assess synaptic connectivity have to be employed. Although cross-correlation seems attractive and is commonly used to assess connectivity, either between diﬀerent brain regions or networks, or even between individual cells, it remains to be investigated to what extent correlation analysis unveils the direct synaptic strength between neurons. A combination of patch-clamp techniques and MEAs would provide a more direct way to measure the EPSC than through the computation of cross-correlation curves. By patching the postsynaptic neuron, EPSC strengths can be directly measured and related to extracellularly recorded presynaptic activity. Combining the advantages of both techniques, i.e., the precise EPSC measurements through patch clamp, and the large-scale parallel, extracellular measurements and stimulations through CMOS-MEAs with ﬂexible feedback loops programmed by the event engine, would greatly expand experimental horizons. One could study the plasticity of hundreds of synapses in parallel. Furthermore, by hooking up the patch clamp system to the event engine through dedicated spike detection and stimulation modules, feedback loops could be applied through the patch clamp between extracellularly recorded and intracellularly stimulated (or vice versa) neurons.

Although, due to the high density of electrodes, potentially all neurons can be read out individually, the recorded signals from two different neurons, located close to each other, are sometimes difficult to separate. A spike-sorting step, incorporated prior to event detection, can help to sort and separate even neurons recorded from with the same electrodes. This holds in particular for using high- density electrode arrays (Franke et al.(2012)). The spike sorting might enable the identification of neurons with smaller spiking amplitudes, close to the noise level, and the identification of more neurons or cell assemblies. However, a drawback of more sophisticated spike sorting algorithms is an additional time delay in the detection phase (t0−t1). Spike sorting, together with intracellular stimulation through patch clamp as described above, could eliminate the aforementioned limitations in Section 5.3.1 “Recording/stimulation selectivity”: Trigger spikes can be assigned to an individual neuron through spike-sorting, and stimulation pulses will only activate action potentials in the patched neuron.

104 Chapter 5. Sub-millisecond closed-loop feedback stimulation

5.5 Conclusion

By using an FPGA to perform signal processing, as well as feedback generation, fast and flexible loop cycles have been realized. Our approach using reconfigurable digital hardware to perform computationally intensive tasks, such as signal filtering, spike identification, decision making, and feedback generation, is a compromise between traditionally employed methods either using a general purpose (micro-) processor, which introduces additional latencies and jitter, and the highly integrated application-specific circuits (VLSI ASICs), which are much less flexible in terms of adaptations to new experimental paradigms. Our achieved closed-loop feedback latencies are lower than many axonal propagation delays and thus enable acausal stimulation. Due to the flexible event engine, high throughput experiments applying many feedback loops in parallel are con- ceivable.

Acknowledgments

We thank Milos Radivojevic and Marta Lewandowska for culturing assistance and Felix Franke, Michele Fiscella, Ian Jones and David J¨ackel for helpful discussions. This work was ﬁnancially supported through the ERC Advanced Grant 267351 “NeuroCMOS” and the Swiss National Science Foundation Ambizione Grant PZ00P3 132245.

105 5.5. Conclusion

106 Chapter 6

Conclusion

A CMOS-based high-density microelectrode array setup for high-throughput extracellular electrophysiology recordings from thousands of neurons in parallel has been presented. At the core of the device is a flexible, reconfigurable electrode array, featuring 26,400 Pt-microelectrodes that can be routed to 1024 low-noise readout channels and 32 stimulation units, all of which are integrated into the same CMOS device. Recordings from cultures of embryonic rat cortical neurons demonstrated the proper functionality of the device. The high spatiotemporal resolution (17.5 µm center-to-center pitch, 20 kHz sampling rate), along with the low-noise readout amplifiers (2.4 µV in the action potential band 300 Hz – 10 kHz) is suited to resolve single neurons. Moreover, subcellular features, such as the axonal arbors of individual cells, can be revealed. The propagation of action potentials along the tiny axons (a few tens of nm in diameter) can be tracked at high resolution by employing recording electrodes right below where the axons pass along. Employing locally dense, globally sparse selections of recording electrodes allows the activity of single neurons from within a larger network of neurons to be recorded in order to study how network parameters, such as functional plasticity, change over time. The architecture of the system supports a two-step approach to experiments. First, by scanning blocks of dense electrode configurations through the array, parameters, such as the spatial extent of the extracellular field potential, or the firing frequency of single neurons, can be identified. Once the biological preparation has been scanned for activity, in a second step, recording sites can be chosen to record from the most informative spots as suited for a particular experiment. For example, in a retina experiment, only certain cell types at specific spatial locations may be of interest. A mathematical framework and two algorithms have been proposed to automate this procedure and make it applicable to different recording systems with tens of thousands of potential recording sites. A closed-loop feedback stimulation system was designed to record the activity of neurons, as well as to implement a two-way interaction with the biological preparation. Much like an artificial synapse, the system can detect action potentials

107 6.1. Improvements over previous designs of a set of neurons and subsequently elicit stimulation pulses to other neurons. Artiﬁcially correlating the spiking activity of two neurons enables the study of Hebbian (Hebb(1949)) and spike timing-dependent plasticity rules (Markram et al.(1997a)). Closing the loop over an FPGA allows very fast and precise feedback loops to be implemented. Loop times below 1 ms with jitter of less than 50 µs (the sampling rate of the CMOS device) have been achieved in this work. This is faster than what has been achieved with feedback loops going through software running on a host computer (Haﬁzovic et al.(2007e); Zrenner et al. (2010)) (on the order of tens of milliseconds). Software, on the other hand, can typically be reprogrammed much faster than FPGAs and can be quickly adapted to new experimental conditions. To compensate for this, a modular event engine was implemented on the FPGA. The simple building blocks of the event engine can be combined to implement sophisticated pattern-matching algorithms. Cultures from dissociated embryonic rat cortical neurons have been chosen as a model system throughout most of the work presented in this thesis. However, the same procedures can easily be transferred to other preparations, such as brain slices or retinae. In fact, researchers at BEL currently use the presented setup routinely to perform high-throughput recordings from patches of acute mice retinae.

6.1 Improvements over previous designs

Compared with the previously developed switch-matrix based HDMEA (Frey et al.(2010)), this device features several significant advantages. The system features eight times more readout channels, more than twice as many electrodes and a twice as large sensing area. The four-times larger dynamic range of the ADCs (10 bit instead of 8 bit) resolves large somatic signals (hundreds of microvolts) together with tiny axonal signals (tens of µV ) with high resolution. Instead of 126 electrode-to-readout channel mappings, now 1024 electrode-to-readout channel mappings have to be implemented, which requires a significantly improved flexibility in the switch-matrix implementation. To increase routing options for individual paths, the wires are interrupted in average after every 20 pixels instead of having long-running wires through the whole array. In addition, larger patches of adjacent electrodes can now be routed (23 × 23, instead of 6 × 18 electrodes), and elongated patches of electrodes to track, e.g., axonal pathways are feasible. The current-mode stimulation buffers now feature auto-zeroing to implement offset compensation. Without such feature, the offset of the stimulation buffer could drive the electrode potential quickly to VDD or ground, rendering current stimulation difficult or even impossible. Furthermore, all stimulation units can be driven to deliver 32 independent biphasic pulses with arbitrary timings but fixed positive and negative voltage peak levels, or they can be driven by three independent DACs to deliver arbitrary stimulation waveforms. Pt-resistors were

108 Chapter 6. Conclusion fabricated on top of the CMOS passivation to monitor the preparation temperature directly at the chip surface.

6.2 Measurement setup

Although the measurement setup does not have its chapter in this thesis, a significant amount of effort and time went into its development. By using an FPGA instead of, for example, a proprietary data acquisition card (DAQ) to interface with the CMOS device, additional complexity is added to the overall system. However, such architecture opens up the possibility to implement advanced signal processing algorithms, such as real-time spike-sorting, in hardware. Furthermore, by streaming the data in UDP packets over Ethernet, the host computer needs only a thin software layer to capture the data stream instead of a potentially bulky and expensive DAQ card. Due to the extensive experience gained with the measurement setup of the first- generation HDMEA, several improvements and simplifications could be implemented in the new setup as well. A dedicated software API exposes all settings of the device to various scripting languages so that experiments can be automated conveniently and be controlled by software. A hardware controlled scheduler ex- ecutes sequences of configuration commands in a precise and deterministic way so that, e.g., precise electrical stimulation protocols can be performed, as they may be necessary for plasticity experiments. Finally, a modular implementation of the data stream infrastructure (on the FPGA and the host computer) allows for conveniently integrating signal processing algorithms. Bandpass filtering or spike-sorting can be implemented either on the FPGA or the host computer, and processed data can be integrated together with the raw data stream for online analysis and visualization. Overall, the system consisting of the CMOS device, an FPGA, and a host computer is portable and self-contained. It allows experiments to be performed ”in the field.”Due to the full integration of the analog signal processing on-chip, there is no need for bulky Faradaic cages and shields as is the case for conventional MEA setups.

6.3 Signal processing and spike-sorting considerations

A major advantage of microelectrode arrays featuring an electrode density comparable to the cell density is the possibility to resolve action potentials of single neurons. Consequently, to fully exploit the potential of such microelectrode arrays, spike-sorting of the recorded data is an important step. A naive approach

109 6.3. Signal processing and spike-sorting considerations would simply perform threshold crossing on all recording channels to identify neuronal spike trains. When the electrode density is comparable to the cell density, this approach is not deemed completely unsuccessful, as every electrode is supposed to record the maximum peak from, at most, one neuron. How- ever, there are at least two problems with this naive approach. First, cells with large extracellular field potentials can be recorded from many neighboring electrodes. Correctly attributing all detected spikes on all these electrodes to the same neuron is non-trivial but important. Otherwise, one might infer strongly correlated cells, whereas, in fact, it is only one cell, so that the cell density can be overestimated. The second problem appears when one electrode records signals from multiple cells. However, this is a rather conventional problem, and different means to solve it have been established. For a review, see Lewicki(1998b). Thus, it makes sense to employ spike-sorting algorithms, which are more complicated than simple threshold crossings, even at the expense of increased computational complexity. Most currently used spike-sorting approaches are optimized for single channel or tetrode recordings (Gray et al.(1995); Lewicki(1998b); Franke et al.(2010)). The large amount of available readout channels to record from densely packed electrodes is a challenging problem in spike sorting and, to some extent, still un- resolved. Increasing the number of readout channels from a few to hundreds or thousands of parallel data channels causes new difficulties to arise. The likelihood to catch temporally overlapping action potentials from different neurons or false positive spikes increases with the number of available recording sites. Nowadays, this problem is conventionally tackled by first subdividing the available electrodes into smaller patches of, e.g., 3 × 3, or 4 × 4, electrodes, and, subsequently, spike sorting these isolated patches. However, this requires an additional, potentially error-prone step of merging detected neurons from neighboring patches. The negative peak of the same action potential can have different timings as detected by the electrodes in different patches (Jackel et al.(2012)). Additionally, if the amplitude in one patch is close to the detection limit, some of the spikes might pass undetected in this patch. Thus, the action potentials from the same neuron, recorded in two different patches, can have different timings and different occurrences, which makes it difficult to assign these action potentials to a single neuron. Once reliable methods are established to simultaneously analyze large groups of channels, the error-prone step of merging neighboring patches can be avoided, and the spike-sorting yield will be increased. On the other hand, it has to be noted that there are additional obstacles that may render accurate spike-sorting a difficult process, even with the many available high-resolution recording electrodes. If a neuron fires multiple spikes within milliseconds, the amplitude decreases for the subsequent spikes (Fee et al.(1996a)). Such amplitude scaling results in non-stationary templates, which are difficult to detect for most spike-sorting algorithms. Furthermore, temporal overlaps of action potentials from different neurons distort the spike shapes. Several techniques have been developed to resolve such temporal overlaps (Pillow et al.(2013)

110 Chapter 6. Conclusion

Franke et al.(2015)), but they remain a difficult problem. Also, it is hard to separate the extracellular field potentials from neurons that exhibit strong spatial overlap. All of these complications render spike-sorting a hard problem. Thus, one has to be careful when interpreting spike-sorting results solely based on extracellularly recorded data. It is advisable to back up the sorting results with additional cross-validations, such as a cell’s expected response to a stimulus. For example, exploiting the fact that neighboring cells in the retina typically include different cell types and respond differently to a visual stimulus can help to distinguish close-by cells.

111 6.3. Signal processing and spike-sorting considerations

112 Chapter 7

Outlook

The high-density CMOS-based MEA presented in this thesis enables various exciting experiments. For example, recording from every neuron in a sparse culture with only a few hundred neurons that were physically constrained to grow exclusively over the array area is possible. Since potentially all spiking activity can be tracked, the number of hidden variables is minimized, and interesting insights into how these cells communicate with each other, and how functional connectivity changes over time can be gained. However, since the survival rate of neurons is much better in higher density cultures, it may be technically difficult to cultivate cells in such low density. Moreover, it would be very interesting to record from large patches in the retina while covering multiple complete sets of distinct retinal ganglion cell types. Combined with other techniques, such as optogenetic manipulations (Hochbaum et al.(2014)), this would yield further insights into how different cell types collectively process visual information and would provide opportunities to test visual restoration techniques (Busskamp et al. (2010)). Besides the conduction of novel neuroscience experiments, the next big frontier that needs attention is the efficient evaluation of the large amount of aggregated data. Developing efficient and robust signal processing algorithms as well as data analysis techniques is required to make sense of large datasets collected during single experiments. Even mundane tasks, such as storing and backing up ter- abytes of recording data, become challenging, let alone the evaluation of the acquired data in a timely manner. Aside from reliable spike sorting algorithms, for example, algorithms to compute network parameters, such as functional connectivity, could be used to track changes in such parameters over time and analyze neuronal plasticity. Future versions of a high-density MEA could also further increase the readout- channel count. Currently, a new version is under development, doubling the number of readout channels.

113 114 Bibliography

M. Abeles and G. L. Gerstein. Detecting spatiotemporal ﬁring patterns among simultaneously recorded single neurons. J Neurophysiol, 60(3):909–24, 1988. URL http://www.ncbi.nlm.nih.gov/pubmed/3171666. Journal of neurophysiology. Misha B. Ahrens, Michael B. Orger, Drew N. Robson, Jennifer M. Li, and Philipp J. Keller. Whole-brain functional imaging at cellular resolution using light-sheet microscopy. Nature Methods, 10(5):413–420, May 2013. ISSN 1548-7091. doi: 10.1038/nmeth.2434. URL http://www.nature.com/nmeth/ journal/v10/n5/full/nmeth.2434.html. A Paul Alivisatos, Anne M Andrews, Edward S Boyden, Miyoung Chun, George M Church, Karl Deisseroth, John P Donoghue, Scott E Fraser, Jen- nifer Lippincott-Schwartz, and Loren L Looger. Nanotools for neuroscience and brain activity mapping. ACS nano, 7(3):1850–1866, 2013. URL http: //pubs.acs.org/doi/pdfplus/10.1021/nn4012847. H. Alle and J. R. Geiger. Combined analog and action potential coding in hippocampal mossy ﬁbers. Science, 311(5765):1290–3, 2006. URL http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve& db=PubMed&dopt=Citation&list_uids=16513983. M. Azin, D.J. Guggenmos, S. Barbay, R.J. Nudo, and P. Mohseni. A battery-powered activity-dependent intracortical microstimulation ic for brain- machine-brain interface. IEEE JOURNAL OF SOLID-STATE CIRCUITS, 46 (4):731–745, 2011. J.N.Y. Aziz, R. Genov, B.L. Bardakjian, M. Derchansky, and P.L. Carlen. Brain- Silicon Interface for High-Resolution in vitro Neural Recording. IEEE Trans- actions on Biomedical Circuits and Systems, 1(1):56–62, March 2007. ISSN 1932-4545. doi: 10.1109/TBCAS.2007.893181. J.N.Y. Aziz, Karim Abdelhalim, R. Shulyzki, R. Genov, B.L. Bardakjian, M. Derchansky, D. Serletis, and P.L. Carlen. 256-Channel Neural Record- ing and Delta Compression Microsystem With 3d Electrodes. IEEE Journal of Solid-State Circuits, 44(3):995–1005, March 2009. ISSN 0018-9200. doi: 10.1109/JSSC.2008.2010997.

115 D. J. Bakkum, Z. C. Chao, and S. M. Potter. Spatio-temporal electrical stimuli shape behavior of an embodied cortical network in a goal-directed learning task. J Neural Eng, 5(3):310–23, 2008a. doi: 10.1088/1741-2560/5/3/004. URL http://www.ncbi.nlm.nih.gov/pubmed/18714127. Journal of neural engineering. Douglas J Bakkum, Zenas C Chao, and Steve M Potter. Long-term activity- dependent plasticity of action potential propagation delay and amplitude in cortical networks. PLOS one, 3(5):e2088, 2008c. URL http://www.ncbi.nlm. nih.gov/pmc/articles/PMC2324202/pdf/pone.0002088.pdf. Douglas J Bakkum, Urs Frey, Milos Radivojevic, Thomas L Russell, Jan Muller,¨ Michele Fiscella, Hirokazu Takahashi, and Andreas Hierlemann. Tracking axonal action potential propagation on a high-density microelectrode array across hundreds of sites. Nature communications, 4, 2013b. Douglas J Bakkum, Milos Radivojevic, Urs Frey, Felix Franke, Andreas Hier- lemann, and Hirokazu Takahashi. Parameters for burst detection. Name: Frontiers in Computational Neuroscience, 7:193, 2013d. URL http://www. ncbi.nlm.nih.gov/pmc/articles/PMC3915237/pdf/fncom-07-00193.pdf. Marco Ballini, J Muller, Paolo Livi, Yihui Chen, Urs Frey, Alexander Stettler, Amir Shadmani, Vijay Viswam, Ian Lloyd Jones, and D Jackel. A 1024-channel cmos microelectrode array with 26,400 electrodes for recording and stimulation of electrogenic cells in vitro. 2014. Bruce P. Bean. The action potential in mammalian central neurons. Na- ture Reviews Neuroscience, 8(6):451–465, June 2007. ISSN 1471-003X. doi: 10.1038/nrn2148. URL http://www.nature.com/nrn/journal/v8/n6/abs/ nrn2148.html. Matthew R Behrend, Ashish K Ahuja, Mark S Humayun, Robert H Chow, and James D Weiland. Resolution of the epiretinal prosthesis is not limited by electrode size. Neural Systems and Rehabilitation Engineering, IEEE Trans- actions on, 19(4):436–442, 2011. URL http://ieeexplore.ieee.org/ielx5/ 7333/5977257/05752253.pdf?tp=&arnumber=5752253&isnumber=5977257. L. Berdondini, P. Massobrio, M. Chiappalone, M. Tedesco, K. Imfeld, A. Mac- cione, M. Gandolfo, M. Koudelka-Hep, and S. Martinoia. Extracellular recordings from locally dense microelectrode arrays coupled to dissociated cortical cultures. Journal of Neuroscience Methods, 177(2):386–396, 2009c. doi: 10.1016/j.jneumeth.2008.10.032. URL ://WOS:000263393300017. Gabriel Bertotti, Norman Dodel, Stefan Keil, Dirk Wolansky, Bernd Tillak, Matthias Schreiter, Max Eickenscheidt, Gunther¨ Zeck, Alfred Stett, and An- dreas M¨oller. A capacitively-coupled cmos-mea with 4225 recording sites and 1024 stimulation sites. In 9th International Meeting on Substrate-Integrated Microelectrodes, pages 247–250, Reutlingen, Germany, 2014.

116 Chapter 7. Outlook

G. Bi and M. Poo. Distributed synaptic modification in neural networks induced by patterned stimulation. Nature, 401(6755):792–6, 1999a. doi: 10.1038/44573. URL http://www.ncbi.nlm.nih.gov/pubmed/10548104. Nature. G. Q. Bi and M. M. Poo. Synaptic modifications in cultured hippocampal neurons: dependence on spike timing, synaptic strength, and postsynaptic cell type. J Neurosci, 18(24):10464–72, 1998. URL http://www.ncbi.nlm.nih. gov/pubmed/9852584. The Journal of neuroscience : the official journal of the Society for Neuroscience. E. Bienenstock. A model of neocortex. Network-Computation in Neural Systems, 6(2):179–224, 1995. URL ://A1995RB18900004. Network-Comp Neural. G. Bontorin, S. Renaud, A. Garenne, L. Alvado, G. Le Masson, and J. Tomas. A real-time closed-loop setup for hybrid neural networks. Conf Proc IEEE Eng Med Biol Soc, 2007:3004–7, 2007a. doi: 10.1109/IEMBS.2007.4352961. URL http://www.ncbi.nlm.nih.gov/pubmed/18002627. Conference proceedings : ... Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering in Medicine and Biology Society. Confer- ence. David R Brillinger. Measuring the association of point processes: a case history. American Mathematical Monthly, pages 16–22, 1976. V. Busskamp, J. Duebel, D. Balya, M. Fradot, T. J. Viney, S. Siegert, A. C. Groner, E. Cabuy, V. Forster, M. Seeliger, M. Biel, P. Humphries, M. Paques, S. Mohand-Said, D. Trono, K. Deisseroth, J. A. Sahel, S. Picaud, and B. Roska. Genetic reactivation of cone photoreceptors restores visual responses in retinitis pigmentosa. Science, 329(5990):413–7, 2010. doi: science.1190897[pii]10.1126/ science.1190897. URL http://www.ncbi.nlm.nih.gov/entrez/query.fcgi? cmd=Retrieve&db=PubMed&dopt=Citation&list_uids=20576849. Jorge G. Cham, Edward A. Branchaud, Zoran Nenadic, Bradley Greger, Richard A. Andersen, and Joel W. Burdick. Semi-Chronic Motorized Micro- drive and Control Algorithm for Autonomously Isolating and Maintaining Op- timal Extracellular Action Potentials. Journal of Neurophysiology, 93(1):570– 579, January 2005. ISSN 0022-3077, 1522-1598. doi: 10.1152/jn.00369.2004. URL http://jn.physiology.org/content/93/1/570. E. Y. Chang, K. F. Morris, R. Shannon, and B. G. Lindsey. Repeated sequences of interspike intervals in baroresponsive respiratory related neuronal assemblies of the cat brain stem. J Neurophysiol, 84(3):1136–48, 2000a. URL http: //www.ncbi.nlm.nih.gov/pubmed/10979989. Journal of neurophysiology. T.-C. Chen, K. Chen, Z. Yang, K. Cockerham, and W. Liu. A biomedical multi- processor soc for closed-loop neuroprosthetic applications. In IEEE Int. Solid- State Circuits Conf. Dig. Tech. Papers, pages 434–435,435a, 2009.

117 Thaddeus R Cybulski, Joshua I Glaser, Adam H Marblestone, Bradley M Zamft, Edward S Boyden, George M Church, and Konrad P Kording. Spatial information in large-scale neural recordings. arXiv preprint arXiv:1402.3375, 2014. Y. Dan and M. M. Poo. Spike timing-dependent plasticity of neural circuits. Neuron, 44(1):23–30, 2004. doi: 10.1016/j.neuron.2004.09.007. URL http: //www.ncbi.nlm.nih.gov/pubmed/15450157. Neuron. Dominique Debanne. Information processing in the axon. Nature Reviews Neu- roscience, 5(4):304–316, 2004. URL http://www.nature.com/nrn/journal/ v5/n4/pdf/nrn1397.pdf. Dominique Debanne, Emilie Campanac, Andrzej Bialowas, Edmond Carlier, and Gis`ele Alcaraz. Axon physiology. Physiological reviews, 91(2):555–602, 2011. URL http://physrev.physiology.org/content/physrev/91/2/555.full. pdf. A. Destexhe and T. Bal. Dynamic-Clamp: From Principles to Applications, volume 1 of Springer Series in Computational Neuroscience. Springer, 2009. M. N. Economo, F. R. Fernandez, and J. A. White. Dynamic clamp: al- teration of response properties and creation of virtual realities in neurophysiology. J Neurosci, 30(7):2407–13, 2010. doi: 10.1523/JNEUROSCI. 5954-09.2010. URL http://www.ncbi.nlm.nih.gov/pubmed/20164323http: //www.jneurosci.org/content/30/7/2407.full.pdf. The Journal of neuroscience : the oﬃcial journal of the Society for Neuroscience. U. Egert, B. Schlosshauer, S. Fennrich, W. Nisch, M. Fejtl, T. Knott, T. Muller, and H. Hammerle. A novel organotypic long-term culture of the rat hippocampus on substrate-integrated multielectrode arrays. Brain Research Protocols, 2:229–242, 1998.

Max Eickenscheidt and Gunther¨ Zeck. Action potentials in retinal ganglion cells are initiated at the site of maximal curvature of the extracellular potential. Journal of neural engineering, 11(3):036006, 2014. URL http://iopscience. iop.org/1741-2552/11/3/036006/pdf/1741-2552_11_3_036006.pdf. B. Eversmann, M. Jenkner, F. Hofmann, C. Paulus, R. Brederlow, B. Holzapﬂ, P. Fromherz, M. Merz, M. Brenner, M. Schreiter, R. Gabl, K. Plehnert, M. Steinhauser, G. Eckstein, D. Schmitt-Landsiedel, and R. Thewes. A 128 x 128 cmos biosensor array for extracellular recording of neural activity. IEEE Journal of Solid-State Circuits, 38:2306–2317, 2003a.

Bj¨orn Eversmann, Armin Lambacher, Thomas Gerling, Alexander Kunze, Peter Fromherz, and Roland Thewes. A neural tissue interfacing chip for in-vitro applications with 32k recording/stimulation channels on an active area of 2.6 mm 2. In ESSCIRC (ESSCIRC), 2011 Proceedings of the, pages 211–214. IEEE, 2011.

118 Chapter 7. Outlook

M. S. Fee, P. P. Mitra, and D. Kleinfeld. Variability of extracellular spike waveforms of cortical neurons. J Neurophysiol, 76(6):3823–3833, 1996a. URL http://www.ncbi.nlm.nih.gov/pubmed/8985880.

P.M. Figueiredo and J.C. Vital. Low kickback noise techniques for CMOS latched comparators. In Proceedings of the 2004 International Symposium on Circuits and Systems, 2004. ISCAS ’04, volume 1, pages I–537–40 Vol.1, May 2004. doi: 10.1109/ISCAS.2004.1328250.

M. Fiscella, K. Farrow, I. L. Jones, D. Jackel, J. Muller, U. Frey, D. J. Bakkum, P. Hantz, B. Roska, and A. Hierlemann. Recording from deﬁned populations of retinal ganglion cells using a high-density cmos-integrated microelectrode array with real-time switchable electrode selection. Journal of Neuroscience Methods, 211(1):103–113, 2012. doi: 10.1016/j.jneumeth.2012.08.017. URL ://WOS:000311018900014.

F. Franke, D. Jackel, J. Dragas, J. Muller, M. Radivojevic, B. Douglas, and A. Hierlemann. High-density microelectrode array recordings and real-time spike sorting for closed-loop experiments: an emerging technology to study neural plasticity. Frontiers in neural circuits, 6, 2012. doi: 10510.3389/fncir. 2012.00105. URL ://WOS:000312557800001.

Felix Franke, Michal Natora, Clemens Boucsein, Matthias HJ Munk, and Klaus Obermayer. An online spike detection and spike classiﬁcation algorithm capable of instantaneous resolution of overlapping spikes. Journal of computational neuroscience, 29(1-2):127–148, 2010. URL http://www.ncbi.nlm.nih.gov/ pmc/articles/PMC2950077/pdf/10827_2009_Article_163.pdf.

Felix Franke, Rodrigo Quian Quiroga, Andreas Hierlemann, and Klaus Ober- mayer. Bayes optimal template matching for spike sorting - combining ﬁsher discriminant analysis with optimal ﬁltering. Journal of Computational Neu- roscience, pages 1–21, February 2015. ISSN 0929-5313, 1573-6873. doi: 10.1007/s10827-015-0547-7. URL http://link.springer.com/article/10. 1007/s10827-015-0547-7.

W. Franks, I. Schenker, P. Schmutz, and A. Hierlemann. Impedance characterization and modeling of electrodes for biomedical applications. Biomedical Engeneering, IEEE Trans on, 52(7):1295–1302, 2005a.

U. Frey, U. Egert, F. Heer, S. Haﬁzovic, and A. Hierlemann. Subcellular- resolution recording of electrical activity using a cmos-microelectrode system. Proceedings of the IEEE Transducers 2009, Denver, Colorado, USA, 2009a. URL http://dx.doi.org/10.1016/j.bios.2008.11.028.

U. Frey, U. Egert, F. Heer, S. Haﬁzovic, and A. Hierlemann. Microelectronic system for high-resolution mapping of extracellular electric ﬁelds applied to brain slices. Biosensors and Bioelectronics, 24(7):2191–2198, 2009b.

119 U Frey, J Sedivy, F Heer, R. Pedron, M. Ballini, J. Mueller, D. Bakkum, S. Haﬁ- zovic, F.D. Faraci, F. Greve, K.-U. Kirstein, and A. Hierlemann. Switch- matrix-based high-density microelectrode array in cmos technology. Journal of Solid-State Circuits, IEEE, 45(2), 2010.

R. C. Froemke and Y. Dan. Spike-timing-dependent synaptic modiﬁcation induced by natural spike trains. Nature, 416(6879):433–8, 2002. doi: 10.1038/ 416433a. URL http://www.ncbi.nlm.nih.gov/pubmed/11919633. Nature.

Shigeyoshi Fujisawa, Asohan Amarasingham, Matthew T Harrison, and Gy¨orgy Buzs´aki. Behavior-dependent short-term assembly dynamics in the medial prefrontal cortex. Nature neuroscience, 11(7):823–833, 2008. URL http:// www.nature.com/neuro/journal/v11/n7/pdf/nn.2134.pdf.

Hua Gao, R.M. Walker, P. Nuyujukian, K.A.A. Makinwa, K.V. Shenoy, B. Mur- mann, and T.H. Meng. HermesE: A 96-Channel Full Data Rate Direct Neural Interface in 0.13 m CMOS. IEEE Journal of Solid-State Circuits, 47(4):1043– 1055, April 2012. ISSN 0018-9200. doi: 10.1109/JSSC.2012.2185338.

Felipe Gerhard, Gordon Pipa, Bruss Lima, Sergio Neuenschwander, and Wulfram Gerstner. Extraction of network topology from multi-electrode recordings: is there a small-world eﬀect? Frontiers in Computational Neuroscience, 5:4–4, 2010.

Felipe Gerhard, Tilman Kispersky, Gabrielle J Gutierrez, Eve Marder, Mark Kramer, and Uri Eden. Successful reconstruction of a physiological circuit with known connectivity from spiking activity alone. PLoS computational biology, 9(7):e1003138, 2013. URL http://www.ncbi.nlm.nih.gov/pmc/articles/ PMC3708849/pdf/pcbi.1003138.pdf.

C. Gold, D. A. Henze, C. Koch, and G. Buzsaki. On the origin of the extracellular action potential waveform: A modeling study. J Neurophysiol, 95(5):3113– 3128, 2006a. URL http://jn.physiology.org/cgi/content/abstract/95/ 5/3113.

Charles M Gray, Pedro E Maldonado, Mathew Wilson, and Bruce McNaughton. Tetrodes markedly improve the reliability and yield of multiple single-unit iso- lation from multi-unit recordings in cat striate cortex. Journal of neuroscience methods, 63(1):43–54, 1995.

Martin Greschner, Greg D Field, Peter H Li, Max L Schiﬀ, Jeﬀrey L Gauthier, Daniel Ahn, Alexander Sher, Alan M Litke, and EJ Chichilnisky. A polyaxonal amacrine cell population in the primate retina. The Journal of Neuroscience, 34(10):3597–3606, 2014. URL http://www.jneurosci.org/content/34/10/ 3597.full.pdf.

120 Chapter 7. Outlook

Christine Grienberger and Arthur Konnerth. Imaging calcium in neurons. Neuron, 73(5):862–885, 2012. URL http://ac.els-cdn. com/S0896627312001729/1-s2.0-S0896627312001729-main.pdf? _tid=0aedfae4-da9f-11e3-9b4c-00000aacb35f&acdnat=1399986417_ 6750b65a7e1de563d29138eee6230462.

G. W. Gross, B. K. Rhoades, H. M. E. Azzazy, and W. Ming-Chi. The use of neuronal networks on multielectrode arrays as biosensors. Biosens. Bioelectron., 10:553–567, 1995.

Jing Guo, Jie Yuan, and Mansun Chan. Modeling of the cell-electrode interface noise for microelectrode arrays. Biomedical Circuits and Systems, IEEE Transactions on, 6(6):605–613, 2012a.

Jing Guo, Jie Yuan, Jiageng Huang, J.K.-Y. Law, Chi-Kong Yeung, and Mansun Chan. 32.9 nV/rt Hz 60.6 dB THD Dual-Band Micro-Electrode Array Signal Acquisition IC. IEEE Journal of Solid-State Circuits, 47(5):1209–1220, May 2012c. ISSN 0018-9200. doi: 10.1109/JSSC.2012.2185590.

S. Haﬁzovic, F. Heer, T. Ugniwenko, U. Frey, A. Blau, C. Ziegler, and A. Hi- erlemann. A cmos-based microelectrode array for interaction with neuronal cultures. Journal of Neuroscience Methods, 164(1):93–106, 2007e. URL http://dx.doi.org/10.1016/j.jneumeth.2007.04.006.

Aviad Hai and Micha E Spira. On-chip electroporation, membrane repair dynamics and transient in-cell recordings by arrays of gold mushroom-shaped microelectrodes. Lab on a Chip, 12(16):2865–2873, 2012. URL http://pubs. rsc.org/en/content/articlepdf/2012/lc/c2lc40091j.

C. M. Hales, J. D. Rolston, and S. M. Potter. How to culture, record and stimulate neuronal networks on micro-electrode arrays (mea). J. Vis. Exp., 39, 2010.

K. D. Harris, D. A. Henze, J. Csicsvari, H. Hirase, and G. Buzsaki. Accuracy of tetrode spike separation as determined by simultaneous intracellular and extracellular measurements. J Neurophysiol, 84(1):401–414, 2000a. URL http: //jn.physiology.org/cgi/content/abstract/84/1/401.

R.R. Harrison and C. Charles. A low-power low-noise CMOS ampliﬁer for neural recording applications. IEEE Journal of Solid-State Circuits, 38(6):958–965, June 2003b. ISSN 0018-9200. doi: 10.1109/JSSC.2003.811979.

Donald O Hebb. The organization of behavior, 1949.

F. Heer, W. Franks, A. Blau, S. Taschini, C. Ziegler, A. Hierlemann, and H. Baltes. Cmos microelectrode array for the monitoring of electrogenic cells. Biosensors Bioelectronics, 20:358–366, 2004a.

121 F. Heer, S. Hafizovic, W. Franks, A. Blau, C. Ziegler, and A. Hierlemann. Cmos microelectrode array for bidirectional interaction with neuronal networks. Solid-State Circuits, IEEE Journal of, 41:1620–1629, 2006a. F. Heer, S. Hafizovic, T. Ugniwenko, U. Frey, W. Franks, E. Perriard, J. C. Perriard, A. Blau, C. Ziegler, and A. Hierlemann. Single-chip microelectronic system to interface with living cells. Biosens. Bioelectron., 22:2546– 2553, 2007a. URL http://www.sciencedirect.com/science/article/ B6TFC-4MBC4YV-3/2/fd483345d8e347f9741aa040596472f3. F. Heer, S. Hafizovic, T. Ugniwenko, U. Frey, B. Roscic, A. Blau, and A. Hier- lemann. Using microelectronics technology to communicate with living cells. 2007 Annual International Conference of the Ieee Engineering in Medicine and Biology Society, Vols 1-16, pages 6082–6085, 2007d. URL ://WOS: 000253467005044. A. Hierlemann, U. Frey, S. Hafizovic, and F. Heer. Growing cells atop microelectronic chips: Interfacing electrogenic cells in vitro with cmos-based microelectrode arrays. Proceedings of the IEEE, 99(2):252–284, 2011. doi: 10.1109/jproc.2010.2066532. URL ://WOS:000286384300003. Daniel R Hochbaum, Yongxin Zhao, Samouil L Farhi, Nathan Klapoetke, Christopher A Werley, Vikrant Kapoor, Peng Zou, Joel M Kralj, Dougal Maclaurin, and Niklas Smedemark-Margulies. All-optical electrophysiology in mammalian neurons using engineered microbial rhodopsins. Nature methods, 2014. A. L. Hodgkin and A. F. Huxley. A quantitative description of membrane current and its application to conduction and excitation in nerve. J. Physiol., 117:500– 544, 1952. G. R. Holt and C. Koch. Electrical interactions via the extracellular potential near cell bodies. J Comput Neurosci, 6(2):169–184, 1999a. URL http://eutils.ncbi.nlm.nih.gov/entrez/eutils/elink.fcgi?cmd= prlinks&dbfrom=pubmed&retmode=ref&id=10333161. A. D. Huberman, W. Wei, J. Elstrott, B. K. Stafford, M. B. Feller, and B. A. Barres. Genetic identification of an on-off direction-selective retinal ganglion cell subtype reveals a layer-specific subcortical map of posterior motion. Neuron, 62(3):327–34, 2009a. doi: S0896-6273(09)00295-5[pii]10.1016/ j.neuron.2009.04.014. URL http://www.ncbi.nlm.nih.gov/entrez/query. fcgi?cmd=Retrieve&db=PubMed&dopt=Citation&list_uids=19447089. Roeland Huys, Dries Braeken, Danny Jans, Andim Stassen, Nadine Collaert, Jan Wouters, Josine Loo, Simone Severi, Frank Vleugels, and Geert Callewaert. Single-cell recording and stimulation with a 16k micro-nail electrode array integrated on a 0.18 um cmos chip. Lab on a Chip, 12(7):1274–1280, 2012. URL http://pubs.rsc.org/en/content/articlepdf/2012/lc/c2lc21037a.

122 Chapter 7. Outlook

A. Hyvarinen and E. Oja. Independent component analysis: algorithms and applications. Neural Netw., 13:411–430, 2000.

Y. Ikegaya, G. Aaron, R. Cossart, D. Aronov, I. Lampl, D. Ferster, and R. Yuste. Synﬁre chains and cortical songs: temporal modules of cortical activity. Science, 304(5670):559–64, 2004. doi: 10.1126/science.1093173. URL http://www.ncbi.nlm.nih.gov/pubmed/15105494.

K. Imfeld, S. Neukom, A. Maccione, Y. Bornat, S. Martinoia, P. A. Farine, M. Koudelka-Hep, and L. Berdondini. Large-scale, high-resolution data acquisition system for extracellular recording of electrophysiological activity. Biomedical Engineering, IEEE Transactions on, 55(8):2064–2073, 2008a.

E. M. Izhikevich. Which model to use for cortical spiking neurons? IEEE Trans Neural Netw, 15(5):1063–70, 2004. doi: 10.1109/TNN.2004.832719. URL http://www.ncbi.nlm.nih.gov/pubmed/15484883. IEEE transactions on neural networks / a publication of the IEEE Neural Networks Council.

Eugene M Izhikevich. Polychronization: computation with spikes. Neural computation, 18(2):245–282, 2006. URL http://www.mitpressjournals.org/doi/ pdfplus/10.1162/089976606775093882.

D. Jackel, U. Frey, M. Fiscella, F. Franke, and A. Hierlemann. Applicabil- ity of independent component analysis on high-density microelectrode array recordings. Journal of Neurophysiology, 108(1):334–348, 2012. doi: 10.1152/jn.01106.2011. URL ://WOS:000306191800029.

A. Jackson, J. Mavoori, and E. E. Fetz. Long-term motor cortex plasticity induced by an electronic neural implant. Nature, 444(7115):56–60, 2006b. doi: 10.1038/nature05226. URL http://www.ncbi.nlm.nih.gov/pubmed/ 17057705. Nature.

A. Jackson, C. T. Moritz, J. Mavoori, T. H. Lucas, and E. E. Fetz. The neurochip bci: towards a neural prosthesis for upper limb function. IEEE Trans Neural Syst Rehabil Eng, 14(2):187–90, 2006c. doi: 10.1109/TNSRE.2006.875547. URL http://www.ncbi.nlm.nih.gov/ pubmed/16792290http://ieeexplore.ieee.org/ielx5/7333/34432/ 01642765.pdf?tp=&arnumber=1642765&isnumber=34432. IEEE transactions on neural systems and rehabilitation engineering : a publication of the IEEE Engineering in Medicine and Biology Society.

Chang-Jin Jeon, Enrica Strettoi, and Richard H. Masland. The Major Cell Pop- ulations of the Mouse Retina. The Journal of Neuroscience, 18(21):8936–8946, November 1998b. ISSN 0270-6474, 1529-2401. URL http://www.jneurosci. org/content/18/21/8936.

123 Xiaolong Jiang, Guangfu Wang, Alice J. Lee, Ruth L. Stornetta, and J. Julius Zhu. The organization of two new cortical interneuronal circuits. Nature Neuroscience, 16(2):210–218, February 2013. ISSN 1097-6256. doi: 10.1038/ nn.3305. URL http://www.nature.com/neuro/journal/v16/n2/full/nn. 3305.html.

Yasuhiko Jimbo, Hugh P.C. Robinson, and A. Kawana. Strengthening of synchronized activity by tetanic stimulation in cortical cultures: application of planar electrode arrays. IEEE Transactions on Biomedical Engineering, 45 (11):1297–1304, November 1998. ISSN 0018-9294. doi: 10.1109/10.725326.

Waichiu Ng Jing Guo. A 200-channel 10µW 0.04mm2 dual-mode acquisition IC for high density MEA. pages C48–C49, 2013.

Thomas Jochum, Timothy Denison, and Patrick Wolf. Integrated circuit ampli- ﬁers for multi-electrode intracortical recording. Journal of Neural Engineering, 6(1):012001, February 2009. ISSN 1741-2552. doi: 10.1088/1741-2560/6/1/ 012001. URL http://iopscience.iop.org/1741-2552/6/1/012001.

L. J. Johnson, E. Cohen, D. Ilg, R. Klein, P. Skeath, and D. A. Scribner. A novel high electrode count spike recording array using an 81,920 pixel tran- simpedance ampliﬁer-based imaging chip. J Neurosci Methods, 205(2):223–32, 2012. doi: 10.1016/j.jneumeth.2012.01.003. URL http://www.ncbi.nlm.nih. gov/pubmed/22266817. Journal of neuroscience methods.

Shabnam N Kadir, Dan FM Goodman, and Kenneth D Harris. High- dimensional cluster analysis with the masked em algorithm. arXiv preprint arXiv:1309.2848, 2013.

I. J. Kim, Y. Zhang, M. Yamagata, M. Meister, and J. R. Sanes. Molec- ular identiﬁcation of a retinal cell type that responds to upward motion. Nature, 452(7186):478–82, 2008. doi: nature06739[pii]10.1038/ nature06739. URL http://www.ncbi.nlm.nih.gov/entrez/query.fcgi? cmd=Retrieve&db=PubMed&dopt=Citation&list_uids=18368118.

Young-tae Kim, Kailash Karthikeyan, Sajal Chirvi, and Digant P Dav´e.Neuro- optical microﬂuidic platform to study injury and regeneration of single axons. Lab on a chip, 9(17):2576–2581, 2009. URL http://pubs.rsc.org/en/ content/articlepdf/2009/lc/b903720a.

HH Ku. Notes on the use of propagation of error formulas. Precision Measure- ment and Calibration, NBS SP 3D0, 1:331–341, 1969.

J. Lee, H.-G. Rhew, D. R. Kipke, and M. P. Flynn. A 64 channel programmable closed-loop neurostimulator with 8 channel neural ampliﬁer and logarithmic adc. IEEE JOURNAL OF SOLID-STATE CIRCUITS, 45(9):1935–1945, 2010.

124 Chapter 7. Outlook

Michael S. Lewicki. A review of methods for spike sorting: the detection and classiﬁcation of neural action potentials. Comput. Neural Syst., pages 53–78, 1998b.

J. Lipski. Antidromic activation of neurones as an analytic tool in the study of the central nervous system. J Neurosci Methods, 4(1):1–32, 1981. URL http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve& db=PubMed&dopt=Citation&list_uids=7253697.

Paolo Livi, Flavio Heer, Urs Frey, Douglas J Bakkum, and Andreas Hierlemann. Compact voltage and current stimulation buﬀer for high-density microelectrode arrays. Biomedical Circuits and Systems, IEEE Transactions on, 4(6): 372–378, 2010.

Carolina Mora Lopez, Alexandru Andrei, Srinjoy Mitra, Marleen Welkenhuy- sen, Wolfgang Eberle, Carmen Bartic, Robert Puers, Refet Firat Yazicioglu, and Georges GE Gielen. An implantable 455-active-electrode 52-channel cmos neural probe. Solid-State Circuits, IEEE Journal of, 49(1):248–261, 2014.

James MacQueen. Some methods for classiﬁcation and analysis of multivariate observations. In Proceedings of the ﬁfth Berkeley symposium on mathematical statistics and probability, volume 1, pages 281–297. California, USA, 1967.

A. H. Marblestone, B. M. Zamft, Y. G. Maguire, M. G. Shapiro, T. R. Cybulski, J. I. Glaser, D. Amodei, P. B. Stranges, R. Kalhor, D. A. Dalrymple, D. Seo, E. Alon, M. M. Maharbiz, J. M. Carmena, J. M. Rabaey, E. S. Boyden, G. M. Church, and K. P. Kording. Physical principles for scalable neural recording. Frontiers in Computational Neuroscience, 7, 2013a. doi: 10.3389/fncom.2013. 00137. URL ://WOS:000326332100001.

Adam H Marblestone, Bradley M Zamft, Yael G Maguire, Mikhail G Shapiro, Thaddeus R Cybulski, Joshua I Glaser, Dario Amodei, P Benjamin Stranges, Reza Kalhor, and David A Dalrymple. Physical principles for scalable neural recording. Frontiers in computational neuroscience, 7, 2013b.

H. Markram, J. Lubke, M. Frotscher, and B. Sakmann. Regulation of synaptic eﬃcacy by coincidence of postsynaptic aps and epsps. Science, 275(5297): 213–5, 1997a. URL http://www.ncbi.nlm.nih.gov/pubmed/8985014.

RichardAˆ H. Masland. The Neuronal Organization of the Retina. Neu- ron, 76(2):266–280, October 2012. ISSN 0896-6273. doi: 10.1016/j.neuron. 2012.10.002. URL http://www.sciencedirect.com/science/article/pii/ S0896627312008835.

J. Mavoori, A. Jackson, C. Diorio, and E. Fetz. An autonomous implantable computer for neural recording and stimulation in unrestrained primates. J Neurosci Methods, 148(1):71–7, 2005. doi: 10.1016/j.jneumeth.2005.04.017.

125 URL http://www.ncbi.nlm.nih.gov/pubmed/16102841. Journal of neuroscience methods.

Ferenc Mechler, Jonathan D Victor, Iﬁje Ohiorhenuan, Anita M Schmid, and Qin Hu. Three-dimensional localization of neurons in cortical tetrode recordings. Journal of neurophysiology, 106(2):828, 2011. URL http://jn.physiology. org/content/jn/106/2/828.full.pdf.

P. Mohseni and K. Najaﬁ. A fully integrated neural recording ampliﬁer with DC input stabilization. IEEE Transactions on Biomedical Engineering, 51(5): 832–837, May 2004. ISSN 0018-9294. doi: 10.1109/TBME.2004.824126.

R. Muller, S. Gambini, and J.M. Rabaey. A 0.013 , 5 , DC-Coupled Neural Signal Acquisition IC With 0.5 V Supply. IEEE Journal of Solid-State Circuits, 47 (1):232–243, January 2012. ISSN 0018-9200. doi: 10.1109/JSSC.2011.2163552.

T. A. Munch, R. A. da Silveira, S. Siegert, T. J. Viney, G. B. Awatramani, and B. Roska. Approach sensitivity in the retina processed by a multifunc- tional neural circuit. Nat Neurosci, 12(10):1308–16, 2009. doi: nn.2389[pii]10. 1038/nn.2389. URL http://www.ncbi.nlm.nih.gov/entrez/query.fcgi? cmd=Retrieve&db=PubMed&dopt=Citation&list_uids=19734895.

J. Muller,¨ D. J. Bakkum, and A. Hierlemann. Sub-millisecond closed-loop feedback stimulation between arbitrary sets of individual neurons. Fron- tiers in neural circuits, 6, 2013a. doi: 12110.3389/fncir.2012.00121. URL ://WOS:000313320300001.

JMuller,¨ M Ballini, P Livi, Y Chen, A Shadmani, U Frey, IL Jones, M Fis- cella, M Radivojevic, DJ Bakkum, and A. Hierlemann. Conferring ﬂexibility and reconﬁgurability to a 26,400 microelectrode cmos array for high throughput neural recordings. In Solid-State Sensors, Actuators and Microsystems (TRANSDUCERS & EUROSENSORS XXVII), 2013 Transducers & Eurosen- sors XXVII: The 17th International Conference on, pages 744–747. IEEE, 2013b.

A. Novellino, P. D’Angelo, L. Cozzi, M. Chiappalone, V. Sanguineti, and S. Mar- tinoia. Connecting neurons to a mobile robot: an in vitro bidirectional neural interface. Comput Intell Neurosci, page 12725, 2007. doi: 10.1155/2007/12725. URL http://www.ncbi.nlm.nih.gov/pubmed/18350128. Computational in- telligence and neuroscience.

Marie Engelene J. Obien, Kosmas Deligkaris, Torsten Bullmann, Douglas J. Bakkum, and Urs Frey. Revealing neuronal function through microelectrode array recordings. Frontiers in Neuroscience, 8, January 2015. ISSN 1662- 4548. doi: 10.3389/fnins.2014.00423. URL http://www.ncbi.nlm.nih.gov/ pmc/articles/PMC4285113/.

126 Chapter 7. Outlook

Donald H Perkel, George L Gerstein, and George P Moore. Neuronal spike trains and stochastic point processes: Ii. simultaneous spike trains. Biophysi- cal journal, 7(4):419–440, 1967b. URL http://www.ncbi.nlm.nih.gov/pmc/ articles/PMC1368069/pdf/biophysj00624-0098.pdf.

Jonathan W Pillow, Jonathon Shlens, EJ Chichilnisky, and Eero P Simoncelli. A model-based spike sorting algorithm for removing correlation artifacts in multi- neuron recordings. PloS one, 8(5):e62123, 2013. URL http://www.ncbi.nlm. nih.gov/pmc/articles/PMC3643981/pdf/pone.0062123.pdf.

Christophe Pouzat, Ofer Mazor, and Gilles Laurent. Using noise signature to optimize spike-sorting and to assess neuronal classiﬁcation quality. Journal of neuroscience methods, 122(1):43–57, 2002. URL http://ac. els-cdn.com/S0165027002002765/1-s2.0-S0165027002002765-main. pdf?_tid=09dd1008-d505-11e3-a268-00000aacb35e&acdnat=1399370516_ 3352d841e54ae9a59b3da43cb45a5d6b.

Wilfrid Rall. Electrophysiology of a dendritic neuron model. Biophys. J., 2 (2Pt2):145–167, 1962. URL http://www.biophysj.org.

J. M. Rebesco, I. H. Stevenson, K. P. Kording, S. A. Solla, and L. E. Miller. Rewiring neural interactions by micro-stimulation. Front Syst Neurosci, 4, 2010. doi: 10.3389/fnsys.2010.00039. URL http://www.ncbi.nlm.nih.gov/ pubmed/20838477. Frontiers in systems neuroscience.

M. Rizk, C. A. Bossetti, T. A. Jochum, S. H. Callender, M. A. Nicolelis, D. A. Turner, and P. D. Wolf. A fully implantable 96-channel neural data acquisition system. J Neural Eng, 6(2):026002, 2009. doi: 10.1088/1741-2560/6/2/026002. URL http://www.ncbi.nlm.nih.gov/pubmed/19255459. Journal of neural engineering.

D.A. Robinson. The electrical properties of metal microelectrodes. Proceedings of the IEEE, 56(6):1065–1071, June 1968. ISSN 0018-9219. doi: 10.1109/PROC. 1968.6458.

Jacob T Robinson, Marsela Jorgolli, Alex K Shalek, Myung-Han Yoon, Rona S Gertner, and Hongkun Park. Vertical nanowire electrode arrays as a scalable platform for intracellular interfacing to neuronal circuits. Nature nanotechnol- ogy, 7(3):180–184, 2012.

J. D. Rolston, D. A. Wagenaar, and S. M. Potter. Precisely timed spatiotemporal patterns of neural activity in dissociated cortical cultures. Neuroscience, 148 (1):294–303, 2007. doi: 10.1016/j.neuroscience.2007.05.025. URL http:// www.ncbi.nlm.nih.gov/pubmed/17614210. Neuroscience.

J. D. Rolston, R. E. Gross, and S. M. Potter. Closed-loop, open-source electrophysiology. Front Neurosci, 4, 2010. doi: 10.3389/fnins.2010.00031.

127 URL http://www.ncbi.nlm.nih.gov/pubmed/20859448. Frontiers in neuroscience.

Peter J Rousseeuw. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of computational and applied mathematics, 20:53–65, 1987.

Isabel Delgado Ruz and Simon R Schultz. Localising and classifying neurons from high density mea recordings. Journal of neuroscience methods, 2014.

Andreas Schierwagen, Michael Ohme, Luigi M Ricciardi, Aniello Buonocore, and Enrica Pirozzi. A model for the propagation of action potentials in nonuniform axons. In Aip Conference Proceedings, volume 1028, page 98, 2008.

R. Schreier, J. Silva, J. Steensgaard, and G.C. Temes. Design-oriented estimation of thermal noise in switched-capacitor circuits. IEEE Transactions on Circuits and Systems I: Regular Papers, 52(11):2358–2368, November 2005. ISSN 1549- 8328. doi: 10.1109/TCSI.2005.853909.

Karsten Seidl, Stanislav Herwik, Tom Torfs, Herc P Neves, Oliver Paul, and Patrick Ruther. Cmos-based high-density silicon microprobe arrays for electronic depth control in intracortical neural recording. Microelectromechanical Systems, Journal of, 20(6):1439–1448, 2011.

M. Shahram and P. Milanfar. Statistical and Information-Theoretic Analysis of Resolution in Imaging. IEEE Transactions on Information Theory, 52(8): 3411–3437, August 2006. ISSN 0018-9448. doi: 10.1109/TIT.2006.878180.

F. Shahrokhi, K. Abdelhalim, D. Serletis, P.L. Carlen, and R. Genov. The 128-Channel Fully Diﬀerential Digital Integrated Neural Recording and Stim- ulation Interface. IEEE Transactions on Biomedical Circuits and Systems, 4 (3):149–161, June 2010. ISSN 1932-4545. doi: 10.1109/TBCAS.2010.2041350.

P Jesper Sjöström, Ede A Rancz, Arnd Roth, and Michael Häusser. Dendritic excitability and synaptic plasticity. Physiological reviews, 88(2):769–840, 2008. URL http://physrev.physiology.org/content/physrev/88/2/769.full. pdf.

S. Song, K. D. Miller, and L. F. Abbott. Competitive hebbian learning through spike-timing-dependent synaptic plasticity. Nat Neurosci, 3(9):919– 26, 2000. doi: 10.1038/78829. URL http://www.ncbi.nlm.nih.gov/pubmed/ 10966623. Nature neuroscience.

M. E. Spira, D. Kamber, A. Dormann, A. Cohen, C. Bartic, G. Borghs, J. P. M. Langedijk, S. Yitzchaik, K. Shabthai, and J. Shappir. Improved neuronal adhesion to the surface of electronic device by engulfment of protruding micro- nails fabricated on the chip surface. TRANSDUCERS ’07 & Eurosensors

128 Chapter 7. Outlook

XXI. 2007 14th International Conference on Solid-State Sensors, Actuators and Microsystems, pages 1247–50, 2007. URL ://INSPEC:9828705.

A. Stett, U. Egert, E. Guenther, F. Hofmann, T. Meyer, W. Nisch, and H. Haem- merle. Biological application of microelectrode arrays in drug discovery and basic research. Analytical and Bioanalytical Chemistry, V377:486–495, 2003b.

Ian H Stevenson and Konrad Koerding. Inferring spike-timing-dependent plasticity from spike train data. In NIPS, pages 2582–2590, 2011.

G. Stuart, N. Spruston, B. Sakmann, and M. Hausser. Action potential initia- tion and backpropagation in neurons of the mammalian cns. Trends in Neuro- sciences, 20:125–131, 1997. URL http://www.sciencedirect.com/science/ article/B6T0V-3RSNK05-9/2/1c1ac67919206a9bf9c6b8e7819bd80e.

Gina G Turrigiano and Sacha B Nelson. Hebb and homeostasis in neuronal plasticity. Current opinion in neurobiology, 10(3):358–364, 2000. URL http: //www.sciencedirect.com/science/article/pii/S095943880000091X.

Gert Van Dijck, Karsten Seidl, Oliver Paul, Patrick Ruther, Marc M Van Hulle, and Reinoud Maex. Enhancing the yield of high-density electrode arrays through automated electrode selection. International journal of neural systems, 22(01):1–19, 2012.

S. Venkatraman, K. Elkabany, J. D. Long, Y. Yao, and J. M. Carmena. A system for neural recording and closed-loop intracortical microstimulation in awake rodents. IEEE Trans Biomed Eng, 56(1):15–22, 2009a. doi: 10. 1109/TBME.2008.2005944. URL http://www.ncbi.nlm.nih.gov/pubmed/ 19224714. IEEE transactions on bio-medical engineering.

S. Venkatraman, C. Patten, and J.M. Carmena. Exploiting the 1/f structure of neural signals for the design of integrated neural ampliﬁers. In Annual International Conference of the IEEE Engineering in Medicine and Biology Society, 2009. EMBC 2009, pages 2050–2053, September 2009b. doi: 10.1109/ IEMBS.2009.5334436.

R. Vollgraf and K. Obermayer. Improved optimal linear ﬁlters for the discrimination of multichannel waveform templates for spike-sorting applications. Signal Processing Letters, IEEE, 13(3):121–124, 2006.

O. Vysotska, B. Frank, I. Ulbert, O. Paul, P. Ruther, C. Stachniss, and W. Bur- gard. Automatic channel selection and neural signal estimation across channels of neural probes. In 2014 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2014), pages 1453–1459, September 2014. doi: 10.1109/IROS.2014.6942748.

129 D. A. Wagenaar, J. Pine, and S. M. Potter. Eﬀective parameters for stimulation of dissociated cultures using multi-electrode arrays. J Neurosci Methods, 138 (1-2):27–37, 2004a.

A. Wallach, D. Eytan, A. Gal, C. Zrenner, and S. Marom. Neuronal response clamp. Front Neuroeng, 4:3, 2011. doi: 10.3389/fneng.2011.00003. URL http: //www.ncbi.nlm.nih.gov/pubmed/21519391. Frontiers in neuroengineering.

W. Wattanapanitch and R. Sarpeshkar. A Low-Power 32-Channel Digitally Programmable Neural Recording Integrated Circuit. IEEE Transactions on Biomedical Circuits and Systems, 5(6):592–602, December 2011. ISSN 1932- 4545. doi: 10.1109/TBCAS.2011.2163404.

Heinz W¨assle. Parallel processing in the mammalian retina. Nature Reviews Neu- roscience, 5(10):747–757, 2004. URL http://www.nature.com/nrn/journal/ v5/n10/pdf/nrn1497.pdf.

Sitong Yuan, L.G. Johnson, C.C. Liu, C. Hutchens, and R.L. Rennaker. Current biased pseudo-resistor for implantable neural signal recording applications. In 51st Midwest Symposium on Circuits and Systems, 2008. MWSCAS 2008, pages 658–661, August 2008. doi: 10.1109/MWSCAS.2008.4616885.

S. Zanos, A. G. Richardson, L. Shupe, F. P. Miles, and E. E. Fetz. The neurochip- 2: an autonomous head-ﬁxed computer for recording and stimulating in freely behaving monkeys. IEEE Trans Neural Syst Rehabil Eng, 19(4):427–35, 2011. doi: 10.1109/TNSRE.2011.2158007. URL http://www.ncbi.nlm.nih.gov/ pubmed/21632309. IEEE transactions on neural systems and rehabilitation engineering : a publication of the IEEE Engineering in Medicine and Biology Society.

G. Zeck, A. Lambacher, and P. Fromherz. Axonal transmission in the retina introduces a small dispersion of relative timing in the ganglion cell population response. PLoS ONE, 6(6):e20810, 2011a. doi: 10.1371/journal.pone.0020810PONE-D-11-04080[pii]. URL http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve& db=PubMed&dopt=Citation&list_uids=21674067.

Wei Zhang and David J Linden. The other side of the engram: experience-driven changes in neuronal intrinsic excitability. Nature Reviews Neuroscience, 4(11): 885–900, 2003.

C. Zrenner, D. Eytan, A. Wallach, P. Thier, and S. Marom. A generic framework for real-time multi-channel neuronal signal analysis, telemetry control, and sub-millisecond latency feedback generation. Front Neurosci, 4:173, 2010. doi: 10.3389/fnins.2010.00173. URL http://www.ncbi.nlm.nih.gov/ pubmed/21060803. Frontiers in neuroscience.

130 Appendix A

Glossary

Abbreviations

ADC analog-to-digital converter

AFE analog frontend

AP action potential

API application programming interface

APS active-pixel sensor

ASIC application-speciﬁc integrated circuit

CAD computer-aided design

CMFB common-mode feedback

CMOS complementary metal oxide semiconductor

CMRR common-mode rejection ratio

CPU central processing unit

CRC cyclic redundancy check

DAC ditigal-to-analog converter

DAQ data acquisition

DDA differential-difference amplifier

DIV days in vitro

DMA direct memory access

DNL diﬀerential nonlinearity

131 Appendix

DRAM dynamic random access memory

EAP extracellular action potential

EPSC excitatory postsynaptic currents

EPSP excitatory postsynaptic potentials

FIM ﬁsher information matrix

FPGA ﬁeld programmable gate array. Reprogrammable logic device

GUI graphical user interface

HD-MEA high-density microelectrode array

HPF high pass ﬁlter

ICA independent component analysis

ILP integer linear program

INL integral nonlinearity

LDA linear discriminant analysis

LNA low noise ampliﬁer

LPF low pass ﬁlter

LSB least signiﬁcant bit

LTD long term depression

LTP long term potentiation

LVDS low-voltage diﬀerential signaling

MEA microelectrode array

MOS metal oxide semiconductor

OPAMP operational ampliﬁer

OTA operational transconductance ampliﬁer

PBS phosphate buﬀered saline

PC personal computer

PCA principal component analysis

PCB printed circuit board]

132 Appendix

PECVD plasma-enhanced chemical vapor deposition

PROM programmable read only memory

PSD power spectral density

PSNR peak signal-to-noise ratio

PSRR power supply rejection ratio

RGC retinal ganglion cells

RIE reactive-ion etching q Pn 2 RMS root mean square 1/n xi

SAR successive-approximation (ADC)

SEM scanning electron microscope

SFDR spurious-free dynamic range

SNDR signal-to-noise and distortion ratio

SNR signal-to-noise ratio

SPI serial programming interface

SRAM static random access memory

STA spike-triggered average

STDP spike-timing-dependent plasticity

TCP/IP transmission control protocol/internet protocol

THD total harmonic distortion

UDP user datagram protocol

USB universal serial bus

VGA variable gain ampliﬁer

VHDL VHSIC hardware description language

VHSIC very high speed integrated circuit

VLSI very-large-scale integration

133 Appendix

134 Appendix B

Publications

Peer Reviewed Journal Publications

1. Jan Muller¨ , Felix Franke, Michele Fiscella, Urs Frey, Douglas J. Bakkum and Andreas Hierlemann (2015) “Selection of best recording sites for optimizing spike-sorting yield”, in preparation

2. Jan Muller¨ , Marco Ballini, Paolo Livi, Yihui Chen, Milos Radivoje- vic, Amir Shadmani, Ian L. Jones, Michele Fiscella, Roland Diggelmann, Alexander Stettler, Urs Frey, Douglas J. Bakkum and Andreas Hierlemann (2015) “High-resolution CMOS MEA platform to study neurons at subcellular, cellular and network levels”, Lab Chip, 2015, 15, pp. 2767-2780

3. Jan Muller¨ , Douglas Bakkum and Andreas Hierlemann (2013) “Sub- millisecond closed-loop feedback stimulation between arbitrary sets of individual neurons”, Frontiers in Neural Circuits 2013, 6:121

Co-Authored Peer Reviewed Journal Publications

1. Keisuke Yonehara, Michele Fiscella, Antonia Drinnenberg, Federico Es- posti, Stuart Trenholm, Jacek Krol, Felix Franke, Brigitte Gross Scherf, Akos Kusnyerik, Jan Muller¨ , Arnold Szabo, Josephine Juttner,¨ Fran- cisco Cordoba, Ashrithpal Police Reddy, János Németh, Zoltán Zsolt Nagy, Francis Munier, Andreas Hierlemann, Botond Roska (2015) “Congenital nystagmus gene FRMD7 is necessary for establishing a neuronal circuit asymmetry for direction selectivity”, Neuron, 2015

2. Michele Fiscella, Felix Franke, Karl Farrow, Jan Muller¨ , Botond Roska, Rava Azeredo da Silveira, Andreas Hierlemann (2015) “Visual Coding with a Population of Direction-Selective Neurons”, Journal of Neurophysiology, 2015, 114(4):2485-99

135 Appendix

3. Ian L. Jones, Thomas L. Russell, Karl Farrow, Michele Fiscella, Felix Franke, Jan Muller¨ , David J¨ackel, Andreas Hierlemann (2015)“A method for electrophysiological characterization of hamster retinal ganglion cells using a high-density CMOS microelectrode array”, Frontiers in Neuroscience 2015, 9:360

4. Marco Ballini, Jan Muller¨ , Paolo Livi, Yihui Chen, Urs Frey, Alexander Stettler, Amir Shadmani, Vijay Viswam, Ian L. Jones, David J¨ackel, Milos Radivojevic, Marta Lewandowska, Wei Gong, Michele Fiscella, Douglas J. Bakkum, Flavio Heer, and Andreas Hierlemann (2014) ”A 1024-Channel CMOS Microelectrode Array With 26,400 Electrodes for Recording and Stimulation of Electrogenic Cells In Vitro”, IEEE Journal of Solid-State Circuits, 2014, 49, pp. 2705–2719

5. Douglas J. Bakkum, Urs Frey, Milos Radivojevic, Thomas L. Russell, Jan Muller¨ , Michele Fiscella, Hirokazu Takahashi, Andreas Hierlemann (2013) “Tracking axonal action potential propagation on a high-density microelectrode array across hundreds of sites”. Nature Communications 2013, 4:2181

6. Felix Franke, David J¨ackel, Jelena Dragas, Jan Muller¨ , Milos Radivoje- vic, Douglas Bakkum and Andreas Hierlemann (2012) “High-density microelectrode array recordings and real-time spike sorting for closed-loop experiments: an emerging technology to study neural plasticity”, Frontiers in Neural Circuits 2012, 6:105

7. Michele Fiscella, Karl Farrow, Ian L. Jones, David J¨ackel, Jan Muller¨ , Urs Frey, Douglas J. Bakkum, Peter Hantz, Botond Roska, Andreas Hierle- mann (2012) “Recording from deﬁned populations of retinal ganglion cells using a high-density CMOS-integrated microelectrode array with real-time switchable electrode selection”, Journal of Neuroscience Methods, 2012, Volume 211, Issue 1, pp. 103–113

8. Urs Frey, Jan Sedivy, Flavio Heer, Rene Pedron, Marco Ballini., Jan Muller¨ , Douglas J. Bakkum, Sadik Haﬁzovic, Francesca Faraci, Frauke Greve, Kay-Uwe Kirstein, Andreas Hierlemann (2010)“Switch-Matrix-Based High-Density Microelectrode Array in CMOS Technology”, IEEE Journal of Solid-State Circuits, Vol. 45, no. 2, pp. 467–482, 2010

Selected Oral Presentations at International Conferences

1. Jan Muller¨ , Douglas J. Bakkum, Milos Radivojevic, Marco Ballini, Paolo Livi, Yihui Chen, Andreas Hierlemann (2014) “Flexible Feedback Stim- ulation Loop(s) between Arbitrary Sets of Individual Neurons on High- Resolution CMOS-Based Microelectrode Arrays”, 9th FENS Forum, 2014, Milan, Italy

136 Appendix

2. Jan Muller¨ , Marco Ballini, Yihui Chen, Paolo Livi, Milos Radivojevic, Urs Frey, Douglas J. Bakkum, Andreas Hierlemann (2014) “High-Density CMOS-based Microelectrode Array Platform for Large-Scale Neuronal Net- work Analysis”, 9th International Meeting on Substrate-Integrated Micro Electrode Arrays, 2014, Reutlingen, Germany

3. Jan Muller¨ , Marco Ballini, Paolo Livi, Yihui Chen, Amir Shadmani, Urs Frey, Ian L. Jones, Michele Fiscella, Milos Radivojevic, Douglas J. Bakkum, Alexander Stettler, Flavio Heer and Andreas Hierlemann (2013) “Confer- ring Flexibility and Reconﬁgurability to a 26’400 Microelectrode CMOS Array for High Throughput Neural Recordings”, 17th IEEE International Conference on Solid-State Sensors, Actuators & Microsystems, Transduc- ers, Barcelona, Spain

Conference Contributions

1. J. Muller¨ , D.J. Bakkum, A. Hierlemann. ”Connectivity analysis of whole- network recordings obtained by a high-spatiotemporalresolution microelectrode array”, Society for Neuroscience (SFN) Meeting, 2015, Chicago, USA, Program No. 266.01/BB31

2. W. Gong, J. Sencar, D. J¨ackel, J. Muller¨ , M. Fiscella, M. Radivoje- vic, D. Bakkum, A. Hierlemann, (2015) “Long-term, High-spatiotemporalresolution Recording from Cultured Organotypic Slices with High-density Microelectrode Arrays”, 18th International Conference on Solid-State Sen- sors, Actuators and Microsystems, Transducers, 2015, Anchorage, Alaska, USA

3. J. Muller¨ , M. Ballini, P. Livi, Y. Chen, D. J. Bakkum, M. Radivojevic, U. Frey, A. Stettler, and A. Hierlemann, “Large-scale Recordings from Axonal Arbors of Single Neurons with CMOS based High-density Microelectrode Arrays” Proceedings of the 18th International Conference on Miniaturized Systems for Chemistry and Life Sciences (MicroTAS), San Antonio, TX, USA, 2014, pp. 986–988. ISBN: 978-0-9798064-7-6.

4. V. Viswam, D. J¨ackel, I. Jones, M. Ballini, J. Muller¨ , A. Stettler, U. Frey, F. Franke, and A. Hierlemann, “Eﬀects of Sub-10 Electrode Sizes on Extracellular Recording of Neuronal Cells”, Proceedings of the 18th In- ternational Conference on Miniaturized Systems for Chemistry and Life Sciences (MicroTAS), San Antonio, TX, USA, 2014, pp. 980–982. ISBN: 978-0-9798064-7-6.

5. M. Radivojevic, F. Franke, J. Muller¨ , A. Hierleman, and D. Bakkum, ”Method to non-invasively study variations in waveforms and velocities between single action potentials in mammalian axons”, in Society for Neu- roscience (SfN) Conference, Washington DC, USA, 2014, p. 370.17

137 Appendix

6. M. Fiscella, F. Franke, J. Muller¨ , I. L. Jones, A. Hierlemann, ”Decoding of Motion Directions by Direction-Selective Retinal Cells”, in Proceedings of the 9th International Meeting on Substrate-Integrated Micro Electrode Arrays, 2014, Reutlingen, Germany, pp. 98-99, ISSN 2199-1596.

7. K. Reinhard, M. Mutter, M. Fiscella, J. Muller¨ , F. Franke, A. Hierle- mann, Thomas Munch,¨ ”Novel Insights into Visual Information Process- ing of Human Retina”, in Proceedings of the 9th International Meeting on Substrate-Integrated Micro Electrode Arrays, 2014, Reutlingen, Germany, pp. 102-103, ISSN 2199-1596.

8. W. Gong, D. Jäckel, J. Muller¨ , M. Fiscella, M. Radivojevic, D. J. Bakkum, F. Franke, F. Knoflach, B. Gähwiler, B. Roscic, T. Russell, A. Hierlemann, ”Long-Term Cultivation and Recording from Organotypic Brain Slices on High-density Micro-electrode Arrays”, in Proceedings of the 9th Interna- tional Meeting on Substrate-Integrated Micro Electrode Arrays, 2014, Reut- lingen, Germany, pp. 335-336, ISSN 2199-1596.

9. I. L. Jones, T. Russell, M. Fiscella, F. Franke, J. Muller¨ , M. Radivo- jevic, A. Hierlemann, ”Characterization of Mammalian Retinal Ganglion Cell Response to Voltage Stimulus”, in Proceedings of the 9th International Meeting on Substrate-Integrated Micro Electrode Arrays, 2014, Reutlingen, Germany, pp. 75-76, ISSN 2199-1596.

10.D.J ¨ackel, J. Muller¨ , T. L. Russell, M. Radivojevic, F. Franke, U. Frey, D. J. Bakkum, and A. Hierlemann, ”Simultaneous Intra- and Extracellu- lar Recordings using a Combined High-Density Microelectrode Array and Patch-Clamp System”, in Proceedings of the 9th International Meeting on Substrate-Integrated Microelectrode Arrays, 2014, Reutlingen, Germany, pp. 153–155, ISSN 2199-1596.

11. V. Viswam, D. J¨ackel, M. Ballini, J. Muller¨ , M. Radivojevic, U. Frey, F. Franke, and A. Hierlemann, ”An Automated Method for Characterizing Electrode Properties of High-Density Microelectrode Arrays”, in Proceed- ings of the 9th International Meeting on Substrate-Integrated Microelectrode Arrays, 2014, Reutlingen, Germany, pp. 302–303, ISSN 2199-1596.

12. M. Radivojevic, D. J¨ackel, J. Muller¨ , V. Viswam, I.L. Jones, A. Hi- erlemann, D. Bakkum, ”Finding the most eﬀective site for extracellular neuronal stimulation”, in Proceedings of the 9th International Meeting on Substrate-Integrated Microelectrode Arrays, 2014, Reutlingen, Germany, pp. 60-61, ISSN 2199-1596.

13. A. Shadmani, J. Muller¨ , P. Livi, M. Ballini, Y. Chen, A. Hierlemann, ”Stimulation Artifact Suppression Techniques for High-Density Microelec- trode Arrays”, in Proceedings of the 9th International Meeting on Substrate-

138 Appendix

Integrated Microelectrode Arrays, 2014, Reutlingen, Germany, pp. 326–327, ISSN 2199-1596.

14. D. J. Bakkum, U. Frey, M. Radivojevic, T. L. Russel, J. Muller¨ , M. Fiscella, H. Takahashi, and A. Hierlemann, “Electrical imaging of axon function”, 3rd International Symposium Frontiers in Neurophotonics, Bor- deaux, France, 2013, p. 25

15. D. J. Bakkum, U. Frey, M. Radivojevic, J. Muller¨ , M. Fiscella, H. Taka- hashi, and A. Hierlemann, “Local diﬀerences in axonal action potential conduction velocity”, Society for Neuroscience (SfN) Conference, San Diego, USA, 2013, p. 520.04

16. M. Ballini, J. Muller¨ , P. Livi, Y.Chen, U. Frey, A. Shadmani, I. L. Jones, W. Gong, M. Fiscella, M. Radivojevic, D. J. Bakkum, A. Stettler, F. Heer and A. Hierlemann, ”A 1024-Channel CMOS Microelectrode-Array Sys- tem with 26’400 Electrodes for Recording and Stimulation of Electro-active Cells In-vitro”, Digest of Technical Papers of Symposium on VLSI Circuits, 2013, Kyoto, Japan, pp. C54-C55, ISBN: 978-4-86348-348-4.

17. M. Fiscella, F. Franke, K. Farrow, I. L. Jones, J. Muller¨ , B. Roska, A. Hierlemann, ”Decoding the Activity of ON-OFF Direction-Selective Retinal Ganglion Cells”, European Retina Meeting, 2013, Alicante, Spain.

18. D. J. Bakkum, U. Frey, J. Muller¨ , M. Fiscella, B. Roscic, H. Takahashi, A. Hierlemann, ”Capability of an 11,011-electrode CMOS array to study action potential propagation plasticity”, Annual Meeting of the Society for Neuroscience (SfN), 2012, New Orleans, USA 322.07.

19.D.J ¨ackel, D. J. Bakkum, M. Radivojevic, J. Muller¨ , M. Fiscella, U. Frey, F. Franke, A. Hierlemann, ”Using extracellular high-resolution microelectrode array recordings to estimate intracellular features of cultured neurons”, Annual Meeting of the Society for Neuroscience (SfN), 2012, New Orleans, USA.

20. M. Ballini, J. Muller¨ , P. Livi, Y. Chen, U. Frey, F. Heer, A. Stettler, A. Hi- erlemann, ”A 1024-channel 26k-electrode, low noise, CMOS microelectrode array for in-vitro recording and stimulation of electrogenic cells at high resolution”, in Proceedings of the 8th International Meeting on Substrate- Integrated Micro Electrode Arrays, 2012, Reutlingen, Germany, pp. 272- 273, ISSN 2194-5519.

21. V. Viswam, M. Ballini, V. Jagannadh, R. Branka, J. Muller¨ , J. David, A. Hierlemann, ”High Resolution Bio Impedance Imaging and spectroscopy with CMOS based Micro Electrode Arrays”, in Proceedings of the 8th In- ternational Meeting on Substrate-Integrated Micro Electrode Arrays, 2012, Reutlingen, Germany, pp. 284-285, ISSN 2194-5519.

139 Appendix

22. M. Fiscella, K. Farrow, I. L. Jones, D. J¨ackel, J. Muller¨ , U. Frey, D. J. Bakkum, B. Roska, A. Hierlemann, ”Targeting Deﬁned Populations of Retinal Ganglion Cells with CMOS Microelectrode Arrays”, in Proceedings of the 8th International Meeting on Substrate-Integrated Micro Electrode Arrays, 2012, Reutlingen, Germany, pp. 114-116, ISSN 2194-5519.

23. E. Masi, M. Fiscella, J. Muller¨ , U. Frey, S. Mancuso, A. Hierlemann, ”In- vestigating Electrical Network Dynamics in Plant Root Cells by Means of a High-Density CMOS-based Microelectrode Array”, in Proceedings of the 8th International Meeting on Substrate-Integrated Micro Electrode Arrays, 2012, Reutlingen, Germany, pp. 246-247, ISSN 2194-5519.

24. M, Radivojevic, D. J¨ackel, J. Muller¨ , M. Fiscella, U. Frey, B. Roscic, A. Hierlemann, D. J. Bakkum, ”Methods for Long-term High-resolution Characterization of In Vitro Developing Neuronal Networks Grown over High-density CMOS-based Microelectrode Arrays”, in Proceedings of the 8th International Meeting on Substrate-Integrated Micro Electrode Arrays, 2012, Reutlingen, Germany, pp. 68-69, ISSN 2194-5519.

25. K. Deligkaris, J. Kaneko, M. Fiscella, J. Muller¨ , I. L. Jones, A. Hier- lemann, M. Takahashi, U. Frey, ”Light Response Patterns in Silenced rd1 Mice Retinal Ganglion Cells”, in Proceedings of the 8th International Meet- ing on Substrate-Integrated Micro Electrode Arrays, 2012, Reutlingen, Ger- many, pp. 126-127, ISSN 2194-5519.

26. D. J. Bakkum, U. Frey, J. Muller¨ , M. Fiscella, B. Roscic, H. Takahashi, A. Hierlemann, ”Capabilities of a High-Density CMOS Microelectrode Ar- ray to Identify, Record, and Stimulate Individual Neurons in Cultured Networks”, in Proceedings of the 8th International Meeting on Substrate- Integrated Micro Electrode Arrays, 2012, Reutlingen, Germany, pp. 206- 207, ISSN 2194-5519.

27. J. Muller¨ , D. J. Bakkum, B. Roscic, A. Hierlemann, ”A ﬂexible system to provide sub-millisecond feedback stimulation loops between multiple sets of individually identiﬁable neurons”, Proc. of the 8th International Meeting on Substrate-Integrated Micro Electrode Arrays, 2012, Reutlingen, Germany, pp. 76-77, ISSN 2194-5519.

28. D. J. Bakkum, U. Frey, J. Muller¨ , M. Fiscella, H. Takahashi, A. Hi- erlemann, ”Towards an ideal interface to neural cultures: recording and stimulation via high-density CMOS electrode arrays”, Medical Physics and Biomedical Engineering (MPBE) World Congress, Beijing, China, 2012, May 26th: TH.16/01.1-1.

29. Douglas J. Bakkum, U. Frey, J. Muller¨ , M. Fiscella, H. Takahashi, A. Hierlemann, ”Uncovering neuronal cellular and network function using a

140 Appendix

high-density 11,011-microelectrode CMOS array”, Swiss Society for Neu- roscience Annual Meeting, February 2012, Zurich; Symposium: ”Advances in neuromorphic systems”.

30. D. J. Bakkum, U. Frey, J. Muller¨ , M. Fiscella, H. Takahashi, A. Hier- lemann, ”Advancing neuronal cellular and network analysis using a high- density 11,011-microelectrode CMOS array”, International Brain Research Organization (IBRO), 2011, Florence, Italy.

31. M. Fiscella. I. L. Jones, D. J¨ackel, J. Muller¨ , U. Frey, K. Farrow, B. Roska and A. Hierlemann, ”Recording of Light Induced Neural Activity of Mouse Retinal Ganglion Cells on a CMOS-Integrated High-Density Mi- croelectrode Array”, Proceedings of the European Retina Meeting (ERM), 2011, Amsterdam, The Netherlands, pp. 92.

32.D.J ¨ackel, J. Muller¨ , M.U. Khalid, U. Frey, D.J. Bakkum, and A. Hierle- mann ”High-Density Microelectrode Array System and Optimal Filtering for Closed-Loop Experiments”, Proceedings of the 16th IEEE International Conference on Solid-State Sensors, Actuators & Microsystems, Transduc- ers, 2011, Beijing, China, pp. 1200-1203. ISBN: 978-1-4577-0157-3.

33. I. L. Jones, M. Fiscella, U. Frey, D. J¨ackel, J. Muller¨ , B. Roscic, R. Stre- ichan, A. Hierlemann, ”Recording of Neural Activity of Mouse Retinal Gan- glion Cells by Means of an Integrated High-density Microelectrode Array”, Proceedings of the 16th IEEE International Conference on Solid-State Sen- sors, Actuators & Microsystems, Transducers, 2011, Beijing, China, pp. 186-189. ISBN: 978-1-4577-0157-3.

34. M. Fiscella, I.L. Jones, D. J¨ackel, J. Muller¨ , U. Frey, K. Farrow, B. Roska, A. Hierlemann, ”Recording of Light Induced Neural Activity of Mouse Reti- nal Ganglion Cells on an Integrated High-density Microelectrode Array”, Abstract Book 1st International SystemsX.ch Conference on Systems Bi- ology, Oct 24th – 26th, 2011, Basel, Switzerland, p.142. ISBN: 978-3- 909386-21-5. Received ’Best Student Poster’ Award.

35. M. Fiscella, U. Frey, D. J¨ackel, J. Muller¨ , R. Streichan, I. L. Jones, B. Roscic, K. Farrow, B. Roska, A. Hierlemann, ”Recording of Neural Activity of Mouse Retinal Ganglion Cells by Means of an Integrated High-Density Microelectrode Array”, in FENS Abstr., vol.5, 019.10, 2010, Amsterdam, The Netherlands

36. M. Fiscella, U. Frey, D. J¨ackel, J. Muller¨ , R. Streichan, I. Jones, B. Roscic, C. Farrow, B. Roska, A. Hierlemann, ”Recording of Neural Activity of Mouse Retinal Ganglion Cells by Means of an Integrated High Density Microelectrode Array”, Proc. of the 7th International meeting on substrate- integrated micro electrode arrays, 2010, Reutlingen, Germany, p. 106-107, ISBN 3-938345-08-5.

141 Appendix

37.D.J ¨ackel, U. Frey, J. Muller¨ , I. Jones, U. Khalid, J. Sedivy, A. Hierle- mann, ”Online Spike Extraction for Bidirectional High-Density Microelec- trode Arrays using Optimal Filters”, Proc. of the 7th International meeting on substrate-integrated micro electrode arrays, 2010, Reutlingen, Germany, p. 201-202, ISBN 3-938345-08-5.

38. D. Bakkum, U. Frey, J. Muller¨ , M. Fiscella, H. Takahashi, A. Hierle- mann, ”Novel neuronal cellular and network measurements enabled by a high-density 11’011-Microelectrode CMOS Array”, Proc. of the 7th In- ternational meeting on substrate-integrated micro electrode arrays, 2010, Reutlingen, Germany, p. 324-326, ISBN 3-938345-08-5.

39. D. Bakkum, U. Frey, T. Mita, J. Muller¨ , M. Fiscella, H. Takahashi, A. Hierlemann, ”Neuronal cellular and network analysis using a high-density 11’011-electrode CMOS Array”, 33rd Annual meeting of the Japan Neuro- science Society, 2010, Kobe, Japan, oral contribution O2-10-2-4.

40. D. J. Bakkum, U. Frey, J. Muller¨ , M. Fiscella, A. Hierlemann, and H. Takahashi, ”Accessing neuronal network activity with an 11,011 electrode CMOS array”, Proc. of the 24th Symposium on Biological and Physiological Engineering, Sendai, Japan, pp. 27–28, 2009. Received ’Young Investiga- tors’ Award (Society of Instrument and Control Engineers).

41. D. J. Bakkum, U. Frey, J. Muller¨ , M. Fiscella, A. Hierlemann, and H. Takahashi, ”Subcellular-resolution electronic recording and stimulation of cultured cortical networks using an 11,011 electrode CMOS array”, in So- ciety for Neuroscience Conference, 2009, Chicago, USA, p. 390.23/HH31, 2009.

142 Appendix C

Acknowledgements

I would like to thank Prof. Andreas Hierlemann for giving me the opportunity to pursue my Ph.D. thesis at the Bio Engineering Laboratory (BEL). I’m grateful for his professional advice and guidance throughout the project. He created a unique and inspiring working environment at the confluence of engineering and neuroscience, provided me with a lot of freedom to pursue my ideas, and was of great help with proofreading my manuscripts. A special thank goes to Prof. Tobi Delbruck for co-examining this thesis and for the valuable input he provided. Special thanks go to Douglas Bakkum, who was not only a co-examiner for my thesis but also an invaluable source of scientific guidance throughout the project. His very scientific mindset, his patience, and his practical approach to problem- solving inspired me and helped to keep me focused on the sciences. I am very thankful to Yihui Chen for his advice and support on all aspects of circuit design. It is not an overstatement that there would have hardly been a single tape-out without his tireless and hard-working efforts. I thoroughly enjoyed designing circuits with all the designers from BEL. Thanks to Jörg Rothe for being a good listener and for his expert advice on circuits and systems design and for sharing with me his truly unparalleled hardware debugging skills. Thanks go to Marco Ballini for being a great office mate and teaching me the upsides of being meticulous and to Paolo Livi for his positive attitude. I would like to thank Urs Frey for introducing me to the Neuro-Chip project. I very much enjoyed the short times we were working together. His attitude and commitment to the project was and still is inspirational to me. Many people contributed in numerous ways to this thesis. I want to give special thanks to Michele Fiscella for the great times we spent together trying to overcome technical hurdles in the early days of our projects and for his infinite patience with me, particularly when I tried to do yet another software update in the middle of one of his experiments. Thanks to Ian Jones for his unparalleled humor (“I have to be frank”), positive attitude, and the fruitful collaboration we

143 Appendix had when exploring electrical stimulation. Thanks to David Jäckel for radiating a relaxed attitude and for letting me participate in his very exciting project, combining MEAs with a patch-clamp setup. Special thanks go to Milos Radi- vojevic for his many culturing efforts with my chips. I thank Amir Shadmani for spending countless late night hours helping me debug the various setups and sharing with me his expert circuit design intuition; and Vijay Viswam for the numerous discussions on impedance measurements. I would like to thank the IT department of D-BSSE for the outstanding infrastructure and service they provide. Especially John Ryan for his expert support and his patience with my outrageous storage needs. Furthermore, I thank the FIS and administrative staff for their motivated and professional support. Alexander Stettler is acknowledged for all the expert clean room work he did and together with Albert Martel for the very reliable and rapid die bonding service. I am thankful to my office friends, with whom I had the pleasure to share great times, beers and wines: Zhu Zhen, Felix Franke, Sebastian Burgel,¨ Jinyoung Kim, Patrick Misun, Roland Diggelmann, Niels Haandbaek, Marco Ballini, Sergey Sitnikov, Derk Wild, Carlos Escobedo and Klaus Mathwig. Furthermore, all current and past BEL members not mentioned so far are acknowledged for contributing to the unique and enjoyable working environment at BEL: Sydney Geissler, Wei Gong, Thomas Russell, Marta Lewandowska, Jelena Dragas, Ben van Lier, Ketki Chawla, Olivier Frey, Gregor Schmidt, YongHong Tao, Branka Roˇsˇcić,Nils Goedecke and Ralf Streichan. Throughout different projects, I had the great opportunity to work with many ex- ceptional researchers from around the globe. Riley Zeller-Townson from Georgia Tech, Benjamin Naecker from Stanford University, XueYing Chua from McGill University, Mita Takeshi and Yuichiro Yada from Tokio University, Katja Rein- hard from Tubingen,¨ Patrick Dini from IMTEK in Freiburg, Anne Tscherter from University of Bern, Kosmas Deligkaris from University of Twente and Asaf Gal from the Technion in Haifa. I am thankful to all of them for introducing me to their research and their perspectives on science. I would like to thank the students with whom I had the pleasure to work. Sun Yan, who developed an analog/digital interface to the patch-clamp setup; Us- man Khalid, who implemented the FPGA-based real-time spike sorter; Dzhihan I. Ahmedov, who characterized stimulation hardware; and Roland Diggelmann, who worked tirelessly to get the neuronal data visualization software into excel- lent shape. This work was financially supported by the ERC Advanced Grant “NeuroCMOS” under contract number AdG 267351. Finally, I want to thank my friends and family for their continuous support throughout my life. I am grateful to my parents, who always supported me wherever I went and to Corina, my love.

144 Curriculum Vitae

Jan Muller¨ Born January 7, 1984, Basel, Switzerland Swiss Citizen

March 2015 Defense of the dissertation entitled High-Density Microelec- trode Array Platform in CMOS Technology.

2009 - present Work on Ph.D. thesis and related topics at the Bio Engi- neering Laboratory, ETH Zurich, under the supervision of Prof. Dr. Andreas Hierlemann.

April 2009 M.Sc. Degree in Electrical Engineering from ETH Zurich, Switzerland.

2008 - 2009 Diploma Thesis at the Bio Engineering Laboratory, ETH Zurich, entitled High Resolution Measurements of Brain Tissue Conductivity.

2008 Internship at Dynatronics AG, Uster, Switzerland.

2005 - 2008 Linux System Administrator at Lehrentwicklung und - technologie, ETH Zurich. 2003 - 2009 Studies at the Department of Information Technology and Electrical Engineering, ETH Zurich, Switzerland.

2002 Swiss Federal Matura, Bern, Type C (Science).

145