PROTEOMIC SINGLE-SYNAPSE ANALYSIS WITH ARRAY …tq652pp7498/thesis-augmented.pdfIntroduction Array tomography (AT) is a high-resolution proteomic imaging method that exploits a combination

PROTEOMIC SINGLE-SYNAPSE ANALYSIS WITH ARRAY

TOMOGRAPHY

A DISSERTATION

SUBMITTED TO THE PROGRAM IN BIOPHYSICS

AND THE COMMITTEE ON GRADUATE STUDIES

OF STANFORD UNIVERSITY

IN PARTIAL FULFILLMENT OF THE REQUIREMENTS

FOR THE DEGREE OF

DOCTOR OF PHILOSOPHY

Brad Busse

August 2011

http://creativecommons.org/licenses/by-nc/3.0/us/

This dissertation is online at: http://purl.stanford.edu/tq652pp7498

© 2011 by Brad Lee Busse. All Rights Reserved.

Re-distributed by Stanford University under license with the author.

This work is licensed under a Creative Commons Attribution-Noncommercial 3.0 United States License.

ii



http://purl.stanford.edu/tq652pp7498

I certify that I have read this dissertation and that, in my opinion, it is fully adequatein scope and quality as a dissertation for the degree of Doctor of Philosophy.

Stephen Smith, Primary Adviser


Shaul Hestrin


Richard Lewis


Liqun Luo

Approved for the Stanford University Committee on Graduate Studies.

Patricia J. Gumport, Vice Provost Graduate Education

This signature page was generated electronically upon submission of this dissertation in electronic format. An original signed hard copy of the signature page is on file inUniversity Archives.

iii

Acknowledgements

This thesis would not have been possible without the contributions of pretty much

everyone I have ever known.

First off I want to thank my parents, Dale and Diane Busse, for raising me right.

My grandparents: Buck, Luella, John, and Aloise, for being such august role

models to emulate.

My undergraduate advisors Bruce McCormick and Yoonsuck Choe, for kindling

my interest in academia and neuroscience.

My graduate advisor, Stephen Smith, for his guidance, encouragement, numerous

introductions, and occasional whip-cracking.

My reading committee: Shaul Hestrin, Liqun Luo and Rich Lewis, for their valu-

able aid and prompt feedback in preparing my dissertation.

Gordon Wang, for being a good postdoc and a better friend.

Kristina Micheva and Nancy O’Rourke, for their patient proofreading and unpar-

alleled abilities at the lab bench and the microscope.

JoAnn Buchanan and Nafisa Ghouri, for their excellent technical support and

dependable repurposing of party leftovers as lab snacks.

Also Nick, Todd, Rachel and Forrest, for their friendship and myriad roles in

making the lab a lively place to work in.

Finally my fiancee, Stephanie Hold, for making my graduate career such an overall

enjoyable one in so many ways.

iv

Contents

Acknowledgements iv

1 Introduction 1

1.1 Production Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.1.1 Section discovery . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.1.2 Multistackreg . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.2 Analysis Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

1.2.1 Colocalization Analysis . . . . . . . . . . . . . . . . . . . . . . 5

1.2.2 Synaptogram . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

1.3 Synapse Classification . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2 Array Tomography: High-Resolution Three-Dimensional Immunoflu-

orescence 10

2.0.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.1.1 Array Tomography Procedures . . . . . . . . . . . . . . . . . . 12

2.2 Protocol A: Rodent Brain Tissue Fixation and Embedding . . . . . . 16

2.2.1 Materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2.2.2 Experimental Method . . . . . . . . . . . . . . . . . . . . . . 18

2.2.3 Troublesbooting . . . . . . . . . . . . . . . . . . . . . . . . . . 20

v

2.3 Protocol B: Production of Arrays . . . . . . . . . . . . . . . . . . . . 21

2.3.1 Materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21


2.3.3 Troubleshooting . . . . . . . . . . . . . . . . . . . . . . . . . . 24

2.4 Protocol C: Immunostaining and Antibody Elution . . . . . . . . . . 25

2.4.1 Materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25



2.5 Protocol D: Imaging Stained Arrays . . . . . . . . . . . . . . . . . . 30

2.5.1 Materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30



2.6 Protocol E: Semiautomated Image Alignment . . . . . . . . . . . . . 37

2.6.1 Materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38



2.7 Conclusion and Future Directions . . . . . . . . . . . . . . . . . . . . 39

2.7.1 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . 40

2.8 Recipes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

3 Classical MHCI molecules regulate retinogeniculate refinement and

limit ocular dominance plasticity 53

3.0.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

3.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

3.2.1 Enhanced ocular dominance plasticity in KbDb−/− mice . . . . 55

vi

3.2.2 Expanded thalamocortical projections to layer 4 of KbDb−/−

mice following ME . . . . . . . . . . . . . . . . . . . . . . . . 58

3.2.3 Abnormal retinogeniculate patterning in KbDb−/− mice . . . . 58

3.2.4 Abnormal segregation of eye-specific inputs in dLGN of KbDb−/− 59

3.2.5 MHCI Immunostaining is associated with LGN synapses and C1q 60

3.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

3.3.1 MHCI function during developmental refinement of the retino-

geniculate projection . . . . . . . . . . . . . . . . . . . . . . . 62

3.3.2 H2-Kb and H2-Db may function with PirB to limit OD Plasticity

in Visual Cortex . . . . . . . . . . . . . . . . . . . . . . . . . 63

3.4 Experimental Procedures . . . . . . . . . . . . . . . . . . . . . . . . . 65

3.4.1 Animals and Genotyping of mouse lines . . . . . . . . . . . . . 65

3.4.2 Mouse surgery and OD plasticity experiments . . . . . . . . . 65

3.4.3 Arc induction . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

3.4.4 Densitometric scans of Arc induction in specific cortical layers 66

3.4.5 Transneuronal labeling . . . . . . . . . . . . . . . . . . . . . . 67

3.4.6 Anterograde labeling of retinal ganglion axons and multiple

threshold analysis . . . . . . . . . . . . . . . . . . . . . . . . . 67

3.4.7 Array Tomography . . . . . . . . . . . . . . . . . . . . . . . . 68

3.4.8 Array Tomography Cross-Correlation Analysis of synaptic mark-

ers, MHCI, and C1q . . . . . . . . . . . . . . . . . . . . . . . 69

3.4.9 Statistical analyses . . . . . . . . . . . . . . . . . . . . . . . . 70

3.4.10 Supplemental Data . . . . . . . . . . . . . . . . . . . . . . . . 70

3.4.11 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . 70

4 Single-Synapse Analysis of a Diverse Synapse Population: Proteomic

Imaging Methods and Markers 79

vii

4.0.12 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

4.2.1 AT Resolves Individual Puncta of Multiple Synaptic Proteins

in Mouse Cortex . . . . . . . . . . . . . . . . . . . . . . . . . 83

4.2.2 Synaptic Protein Distributions Imaged by AT Correlate as Ex-

pected from Synapse Structure . . . . . . . . . . . . . . . . . 85

4.2.3 AT Immunofluorescence of Synapsin Is Highly Reliable as Synapse

Marker . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

4.2.4 Synapsin Is Detectable at Virtually All Dendritic Spines . . . 88

4.2.5 EM Analysis Supports the Identification of Synapses with Synapsin

Immunoreactivity . . . . . . . . . . . . . . . . . . . . . . . . . 88

4.2.6 Multiple Synaptic Proteins Can Be Visualized Volumetrically

as a Synaptogram Mosaic . . . . . . . . . . . . . . . . . . . . 90

4.2.7 AT Imaging Discriminates Multiple Glutamatergic and GABAer-

gic Synapse Subtypes . . . . . . . . . . . . . . . . . . . . . . . 91

4.2.8 AMPA and NMDA Receptors Distributions Vary at Different

Glutamatergic Synapses . . . . . . . . . . . . . . . . . . . . . 93

4.2.9 Synapsin Is Present at All Glutamatergic and GABAergic Synapses,

but in Varying Amounts . . . . . . . . . . . . . . . . . . . . . 93

4.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

4.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99

4.5 Experimental Procedures . . . . . . . . . . . . . . . . . . . . . . . . . 100

4.5.1 Tissue Preparation . . . . . . . . . . . . . . . . . . . . . . . . 100

4.5.2 LRWhite Sections . . . . . . . . . . . . . . . . . . . . . . . . . 100

4.5.3 ImmunoEM Staining . . . . . . . . . . . . . . . . . . . . . . . 101

4.5.4 Colocalization Analysis . . . . . . . . . . . . . . . . . . . . . . 103

viii

4.5.5 Transmission and Scanning Electron Microscopy . . . . . . . . 103

4.6 Supplemental Information . . . . . . . . . . . . . . . . . . . . . . . . 103

4.7 Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

5 Single-synapse analysis of a diverse synapse population: synapse

discovery and classification 122

5.0.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

5.0.2 Author Summary . . . . . . . . . . . . . . . . . . . . . . . . . 123

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124

5.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126

5.2.1 Identifying Putative Synaptic Loci . . . . . . . . . . . . . . . 126

5.2.2 Manual Classification . . . . . . . . . . . . . . . . . . . . . . . 128

5.2.3 Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . 129

5.2.4 Unsupervised Clustering . . . . . . . . . . . . . . . . . . . . . 131

5.2.5 Supervised Classification . . . . . . . . . . . . . . . . . . . . . 132

5.2.6 Post-Classification Analysis . . . . . . . . . . . . . . . . . . . 137

5.2.7 Classification Application . . . . . . . . . . . . . . . . . . . . 139

5.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

5.3.1 Limitations and Future Work . . . . . . . . . . . . . . . . . . 142

5.4 Materials and Methods . . . . . . . . . . . . . . . . . . . . . . . . . . 143

5.4.1 Acquisition of array tomographic volume . . . . . . . . . . . . 143

5.4.2 Normalization and background subtraction of volumetric data 145

5.4.3 Extraction of synaptic loci . . . . . . . . . . . . . . . . . . . . 146

5.4.4 PCA image treatment . . . . . . . . . . . . . . . . . . . . . . 146

5.4.5 Normalization of pairwise channel data . . . . . . . . . . . . . 146

5.4.6 Perpendicularization of cortical data . . . . . . . . . . . . . . 147

5.4.7 Software packages used . . . . . . . . . . . . . . . . . . . . . . 148

ix

6 Future Directions and Conclusions 162

Bibliography 165

x

List of Tables

2.1 Recipe: Alternative Antibody Dilution Solution with NGS (1 mL) . . . 41

2.2 Recipe: Alternative Blocking Solution with NGS (1 mL) . . . . . . . . 41

2.3 Recipe: Blocking Solution with BSA (1 mL) . . . . . . . . . . . . . . 41

2.4 Recipe: Elution Solution (10 mL) . . . . . . . . . . . . . . . . . . . . 42

2.5 Recipe: Fixative (4 mL) . . . . . . . . . . . . . . . . . . . . . . . . . 42

2.6 Recipe: Subbing Solution (300 mL) . . . . . . . . . . . . . . . . . . . 42

2.7 Recipe: Wash Buffer (50 mL) . . . . . . . . . . . . . . . . . . . . . . 43

2.8 Primary antibodies used with array tomography . . . . . . . . . . . . 44

4.1 Synaptic Antibodies Used in This Study . . . . . . . . . . . . . . . . 104

4.2 Proportion of Synapses from Different Synaptic Subtypes . . . . . . . 105

5.1 Machine Learning Algorithm Comparison . . . . . . . . . . . . . . . . 149

5.2 Estimated error rates . . . . . . . . . . . . . . . . . . . . . . . . . . . 149

xi

List of Figures

2.1 The sequence of steps for a basic immunofluorescence array tomogra-

phy process. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

2.2 Array tomographic images of layer 5 neuropil, barrel cortex of YFP-H

Thy-1 transgenic mouse . . . . . . . . . . . . . . . . . . . . . . . . . 47

2.3 Multiplexed staining for seven synaptic proteins in mouse cerebral cortex 49

2.4 A single iteration of the position-finding algorithm. . . . . . . . . . . 51

3.1 Enhanced ocular dominance plasticity in visual cortex of KbDb−/− mu-

tant mice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

3.2 Enhanced thalamocortical plasticity in KbDb−/− mutant mice . . . . 73

3.3 Incomplete segregation of RGC inputs to dLGN in KbDb−/− mutant

mice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

3.4 MHCI localization in relation to synaptic proteins during period of

retinogeniculate refinement . . . . . . . . . . . . . . . . . . . . . . . . 77

4.1 Array tomographic synapsin I immunofluorescence in the cerebral cor-

tex of an adult YFP-H mouse is punctate and consistent with synapse

identity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

4.2 Proteomic immunofluorescence AT of mouse somatosensory cortex yields

staining patterns consistent with synaptic protein distributions . . . . 108

xii

4.3 Multiple synaptic proteins are colocalized in a fashion consistent with

synaptic identity and glutamatergic and GABAergic synapse subtype 110

4.4 Dendritic spines in mouse cerebral cortex are contacted by synapsin

puncta and colocalize with other pre- and postsynaptic markers . . . 112

4.5 Ultrastructurally identified synapses are labeled with the synapsin an-

tibody . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114

4.6 Synaptograms are useful for viewing proteomic information from seri-

ally sectioned single synapses . . . . . . . . . . . . . . . . . . . . . . 116

4.7 Proteomic imaging with AT reveals the diversity of cortical synapses . 118

4.8 Double innervated spines receive both a glutamatergic VGluT1 and

GABAergic synapse . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120

5.1 The synaptogram as a tool for high-dimensional proteomic visualization150

5.2 Comparison of human rating to machine learning . . . . . . . . . . . 152

5.3 Unsupervised clustering of synapsin I imaged with array tomography 154

5.4 Feature importance for different molecular labels . . . . . . . . . . . . 156

5.5 Density and size of synapse classes as a function of depth through the

cortex . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158

5.6 Positive and negative pairwise channel copresence . . . . . . . . . . . 160

xiii

Chapter 1

Introduction

Array tomography (AT) is a high-resolution proteomic imaging method that exploits

a combination of light and EM techniques to resolve fine details at the synapse level

across large fields of view spanning entire circuits [1, 2]. This allows us to address a

long-standing, basic problem of large-scale synapse quantification. At the genomic

level, neurons display a staggering amount of diversity in the number of their cell

types and the variety of their spatial distributions [3]. The synapses of those neurons,

each comprising hundreds of distinct protein species [4–6], have enough proteomic

complexity to potentially display even more systemic variation. For instance, it is now

clear that within each neurotransmitter category (e.g., glutamatergic, GABAergic,

cholinergic) there is substantial diversity in the expression of many intrinsic synaptic

proteins, including neurotransmitter transporters and receptors [7–16]. Yet, to date

our ability to study diverse populations of synapses in situ has been somewhat lacking.

The resolution of AT, along with its proteomic multiplexing capabilities, is well suited

to rectify that.

Much of my graduate work has centered on helping to shape AT into an imag-

ing method fit for practical use, by developing computational algorithms which take

1

advantage of AT’s particular qualities to automate its operation. The material com-

ponents of an array tomographic pipeline are an amalgamation of those already used

in fluorescence and electron microscopy [17]; indeed the physical technique has ex-

isted for some time [18]. What has enabled its rebirth as AT is the proliferation of

fluorescence immunohistochemistry and the availability of software tools needed to

image, align and analyze AT volumes. Those tools required extensive development

to graduate AT from proof of concept to working technology.

1.1 Production Methods

1.1.1 Section discovery

The automated tools created to assist array tomographic imaging needed to address

a number of interesting computational problems, beginning with data acquisition

itself. Owing to the ribbon-like structure formed by consecutive sections adjoining

each other, tissue samples prepared for AT have a rather interesting layout that defies

most imaging applications. AT sections are laid consecutively on a slide in a fashion

which is fairly predictable for low-resolution (10x) imaging, but ribbon geometry

usually includes too much variation for pure extrapolation to track them adequately.

Small lateral shifts (within a section) correspond to lateral shifts in image space, while

larger shifts (to different sections) include a z-shift component and different lateral

shifting, almost as if the field of view “wrapped around” the volume and appeared

on the other side again. Further, depending on block treatment there may be little

to no distinguishing features which might tell us when sections change.

On the other hand, our ribbons universally come with a fluorescent nuclear stain

which makes a very good fiducial marker: DAPI, which intercalates into DNA and

thus labels cell nuclei. In the mammalian cortex, cell bodies are spaced tightly enough

2

that any given field of view will image a dozen or so nuclei. With roughly 10 µm

nuclear sizes and section thicknesses in the nanometer range, the view of the corre-

sponding point on the next consecutive section will remain almost entirely unchanged.

Therefore, while determining the break between sections may be a difficult problem,

finding the same position in the next section, barring tissue damage, and assuming

you have a reasonable idea where it is to start searching, is not.

In an effort to map AT ribbons by exploiting this observation, I developed a

method which uses short extrapolations refined by cross-correlation searches. The

exact mechanics, implemented as an Axiovision plugin, are presented later in this

chapter, but the basic procedure is: a human presents, as input, two points corre-

sponding to the same lateral position on two adjacent sections. The plugin uses the

vector between them to predict the stage coordinates of the next consecutive position.

Owing to AT’s linear ribbons of regularly-spaced sections, this estimate is accurate

enough that the true point is usually somewhere in the area. To find it, we use a

cross-correlation search comparing the areas around the second and third point to lo-

cate the best match. The addition of a Kalman filter [19] (such that the tissue patch

we attempt to locate is actually the weighted running average of recent patches) aids

recovery in the face of tissue damage.

The resulting plugin worked well enough to have become a staple tool for use in

AT applications for several years. Future improvements will likely focus on improving

the user interface of the plugin, such that it becomes easier to pause, change and

correct position lists to suit the particular ribbon under analysis. Support for the

use of fiducial markers would also be highly useful, and would probably be a vital

prerequisite for any automated imaging solution.

3

1.1.2 Multistackreg

Once an array tomographic data set has been imaged, it is not immediately ready for

analysis. Ribbons curve and warp during sectioning and plating, and removal/reinsertion

invariably introduces some rotation and offset. This necessitates an alignment step

to ensure that the tissue on each section is in the same positional reference frame as

the the tissue on the next.

While some alignment utilities existed at the time of AT’s creation, they were

inadequate for our purposes. We were using the best of the ImageJ plugins, Stack-

reg [20], but it had a number of areas which needed addressing. Stackreg could at

most align standard color (RGB) images, while at the time we were already producing

volumes with up to 5 different labels in a single imaging session. Although all of these

labels began in the same reference frame, it would be incorrect to consider multiple

independent alignments to still inhabit in the same frame, and more importantly,

almost impossible to correct if there was an error. Additionally, a few of the labels

(DAPI) made for excellent alignment between sections, and the mostly punctate rest

did not. We wanted to align with the signal from DAPI, without the noise from the

other channels. My solution was to modify Stackreg to include a save/load func-

tionality. This allowed us to treat all channels in an imaging session as independent

grayscale images, align with the best, and apply to the rest.

Multiple imaging sessions also proved to be a problem, as it is difficult for a human

to insert a standard glass slide into a slide holder with nanometer precision. I further

modified Stackreg (now renamed MultiStackreg and distributed independently, at

the request of Stackreg’s author) to be capable of aligning one stack to another, each

section of one stack aligned to its corresponding index in the other stack. This allowed

us to register multiple imaging sessions into the same reference frame, at which point

the generated transformation file could be used to align all of the labels at once.

The resulting plugin, with a revamped GUI and an API to make it accessible to

4

automation, proved sufficient for our alignment needs for some time. It remains the

first step in our alignment process, though we currently use another plugin which

boasts nonlinear dewarping (bUnwarpJ) for final alignment. Future development of

alignment methods will likely include nonlinear alignment algorithms which work

robustly enough to remove the need to use MultiStackReg for rough alignment.

1.2 Analysis Methods

Once an AT volume has been imaged and aligned, it is then amenable to analysis. Due

to the size of typical array tomographic volumes, on the order of thousands to millions

of cubic microns of high resolution proteomic data, these analysis methods must

necessarily be either automated or stereological. Both have particular advantages:

automated methods survey the entire data set and are not prone to confirmation bias,

while manual methods pass data more directly to human experts for interpretation.

1.2.1 Colocalization Analysis

Most of our early concerns with AT stemmed from uncertainty in the staining process.

A given label might look specific, and upon manual inspection it might correctly asso-

ciate with other labels we’d expect to find, but large scale surveys of each individual

label just to remove confirmation bias would be very inefficient. Cross-correlation

algorithms like Pearson’s correlation coefficient could give absolute measurements of

channel overlap, but those measurements would be more dependent on label density

than label proximity - a million weakly-correlated puncta would score higher than a

few completely overlapping examples. What we needed was a way to measure relative

correlation between the channels as imaged and deliberately uncorrelated channels

with the same image properties. Comparing the two correlation measures would give

an idea of the actual relationship between the channels of interest.

5

To put a relative correlation metric to use, I decided to implement and modify

Van Steensel’s algorithm for shifting colocalization [21]. By measuring Pearson’s

correlation with one channel undergoing a variety of lateral offsets, you can study

the spatial relationship between the channels. If your data is entirely punctate, as

synaptic labels in AT are, after some amount of shifting all local relationships will

be broken and you will be measuring baseline correlation, which can be compared

to the no offset case for the relative measure. In between the offset extremes, the

falloff of the correlation allows you to infer something of the relationship between the

channels. Sharper falloff shoulders imply either smaller puncta or puncta which do

not fully overlap.

To utilize several aspects specific to array tomographic volumes, I modified Van

Steensel’s algorithm in three ways. For ease of analysis and since AT features isotropic

x- and y- resolution, I modified the algorithm to be two-dimensional, returning a more

readily-analyzed heat map of correlations with x and y shifts. AT volumes are stacks,

so my implementation used the z-axis to repeat the x-y analysis on multiple z-levels,

resulting in measurements with lower variance than the original, often to the point

where the measurements were significant even using small patches of tissue. Finally,

I added a means by which the user could easily select a smaller window of tissue

for analysis. This had the effect of improving analysis speed without (due to the

preceding modification) losing statistical significance, and ensuring that only soma-

and blood vessel-less neuropil might be analyzed if desired (though the shifts were

small enough that they did not greatly impact the measurement regardless).

The resulting tool became a useful, if niche, part of the AT analysis kit. Its best

use is in the examination of antibody labeled images as a whole, without needing to

restrict the analysis volume to areas where the label is expected, e.g. the synapse.

It remains one of the best ways to examine the specificity and reliability of novel

labels, particularly in the case where multiple antibodies to the same protein are

6

being tested, as they can then be compared with each other was well as with a third

marker. For later analysis steps, especially those focused on synapse quantification,

additional tools are required.

1.2.2 Synaptogram

The manual side of our early analysis centered around visual inspection of AT data.

The user had to first “hunt through” the data to find suitable synapse candidates,

then iterate through all of the imaged channels, building a mental map of the lo-

cal geometry of that synapse. Due to the large amount of data and high proetomic

multiplexing, this process was very slow. However, it is still the best and most reli-

able method of synapse identification available to us. It relies on the perception and

expertise of the human viewer to apply the visual segmentation which defines the

presence of necessary synaptic components and verifies that they are in the correct

orientation relative to each other. This task incorporates a great deal of a priori

knowledge concerning the stearic and functional relationships between the different

molecular labels, the variance in labeling of each particular antibody, and the partic-

ular conditions under which that tissue had been fixed, embedded, labeled, imaged,

relabeled, etc.

In order to facilitate the manual effort required to find and identify synapses, we

devised the synaptogram as a way of displaying all of the requisite proteomic data at

once. It relies on the fact that synapses are small [112]: for many tissues, all of the

information needed to identify the synapse can be found in a volume no larger than

half a micron from the middle of the synapse. In AT resolution, this comprises a cube

11 pixels to a side. By splaying out the high dimensional data into a much larger

two dimensional space, we could present the information to the users more directly.

Synaptograms, and the human classifications performed with such, have crept into

virtually all of the tools I have developed since.

7

1.3 Synapse Classification

To utilize AT in synaptic quantification requires the development of new, automated

synapse detection and classification capabilities. Manual analysis using synaptograms

is acceptable for analyzing fragmentary subsets of a few hundred synapses, but it does

not scale beyond that. The use of synaptograms eases the difficulty of per-synapse

manual classification such that the effort of classifying a set of few hundred synapses

is no longer excessive, but no matter how convenient they are to analyze individually,

the sheer number of synapses makes manual analysis of the entire data set effectively

impractical.

The development of an automated synapse quantification algorithm, which essen-

tially automates the process of creating and analyzing synaptograms, can be broken

down into a few important pieces. To identify the sites of putative synapse locations I

decided to use Synapsin I puncta, for its synaptic ubiquity and robust labeling [23,24].

I used a local maximum filter to identify the peaks of Synapsin I staining, each of

which I considered a putative synapse location, or synaptic locus.

Determining which synaptic loci were actual synapses involved implementing a

supervised learning application. In this process, humans classify a small training

set of synapse examples on the basis of the presence of absence of relevant synaptic

labels. To facilitate this, I developed a browser-based active learning scheme which

aids training by pre-classifying the training examples, such that the humans only

need to correct the errors it makes. Once the training is complete, I extract the

salient features of the example loci’s proteomic distribution and extrapolate them,

using a random forest ensemble classifier, to the rest of the unknown synapses in

order to predict how the humans would classify those as well. To establish synaptic

identity, the different channel classifications are recombined into known or suspected

combinations corresponding to synapse classes. This allows us, in addition to accurate

8

quantification of expected synapse populations, to test for novel synapse types without

having to classify them before the fact, and successfully use it to identify a few

unexpected synaptic subpopulations.

Since AT data sets are likely to increase in size as our technique becomes more

consistent and our questions weightier, automated classification will only become a

more necessary piece of the analysis tool chain. A vital addition to this classification

process, when or if AT methods are usefully adapted to superresolution techniques

like STORM or STED, will be a move away from rotation-invariant features into a

feature set which incorporates the synapse geometry resolvable in those contexts.

9

Chapter 2

Array Tomography: High-Resolution

Three-Dimensional Immunofluorescence

Micheva KD, ORourke N, Busse B, Smith SJ. 2010a. Array tomography: High-

resolution three-dimensional immunofluorescence. Cold Spring Harbor Protocols doi:

10.1101/pdb.top89.

2.0.1 Abstract

Array tomography is a volumetric microscopy method based on physical serial sec-

tioning. Ultrathin sections of a plastic-embedded tissue specimen are cut using an

ultramicrotome, bonded in ordered array to a glass coverslip, stained as desired, and

then imaged. The resulting two-dimensional image tiles can then be computationally

reconstructed into three-dimensional volume images for visualization and quantita-

tive analysis. The minimal thickness of individual sections provides for high-quality,

rapid staining and imaging, whereas the array format provides for reliable and con-

venient section handling, staining, and automated imaging. In addition, the arrays

physical stability permits the acquisition and registration of images from repeated cy-

cles of staining, imaging, and stain elution and from imaging by multiple modalities

10

(e.g., fluorescence and electron microscopy). Array tomography offers high resolution,

depth invariance, and molecular discrimination, which justify the relatively difficult

tomography array fabrication procedures. With array tomography it is possible to

visualize and quantify previously inaccessible features of tissue structure and molecu-

lar architecture. This chapter will describe one simple implementation of fluorescence

array tomography and provide protocols for array tomography specimen preparation,

image acquisition, and image reconstruction.

2.1 Introduction

Our understanding of tissue function is constrained by incomplete knowledge of tis-

sue structure and molecular architecture. Genetics, physiology, and cell biology make

it overwhelmingly clear that all cell and tissue function depends critically on the

composition and precise three-dimensional configuration of subcellular organelles and

supramolecular complexes, and that such structures may consist of very large numbers

of distinct molecular species. Unfortunately, the intricacies of tissue molecular archi-

tecture badly outstrip the analytical capability of all presently known tissue imaging

methods.

Array tomography is a new high-resolution, three-dimensional microscopy method

based on constructing and imaging two-dimensional arrays of ultrathin (70–200 nm

thickness) specimen sections on solid substrates. (The word tomography derives from

the Greek words tomos, to cut or section, and graphein, to write: The moniker array

tomography thus simply connotes the writing of a volume image from an array of

slices.) Array tomography allows immunofluorescence imaging of tissue samples with

resolution, quantitative reliability, and antibody multiplexing capacity that is greatly

superior to previous tissue immunofluorescence methods [1]. Array tomography was

developed with neuroscience applications in mind (e.g. [61,69,70]), and the following

11

description will be illustrated with examples from neuroscience and particularly from

studies of synapses and circuits in rodent brain.

2.1.1 Array Tomography Procedures

A sequence of eight steps for a very basic array tomography protocol is illustrated

in Figure 2.1. Array tomography begins with (Step 1) the chemical fixation of the

specimen, followed by (Step 2) dissection and embedding in resin (LR White). Resin-

embedded specimen blocks are then (Step 3) mounted in an ultramicrotome chuck,

trimmed, and prepared for ultrathin sectioning. Block preparation includes careful

trimming of the block edges and application of a tacky adhesive to the top and bottom

block edges. As shown in the magnified detail of Step 3, this adhesive causes the

spontaneous formation of a stable splice between successive serial sections as they are

cut by the ultramicrotomes diamond knife blade. The automated cycling of a standard

ultramicrotome produces automatically a ribbon up to 45 mm in length, which may

consist of more than 100 serial sections held on a water surface. Ribbons are then

manually transferred to the surface of a specially coated glass coverslip (Step 4). The

resulting array can be stained using antibodies or any other desired reagents (Step 5).

After immunostaining, arrays can be imaged using fluorescence microscopy (Step 6).

The minimal thickness of array sections promotes very rapid and excellent staining

and imaging, whereas the array format promotes convenient and reliable handling

of large numbers of serial sections. The individual two-dimensional section images

are then computationally stitched and aligned into volumetric image stacks (Step 7)

to provide for three-dimensional image visualization and analysis (Figure 2.2). The

volumetric image stacks are stored electronically for analysis and archiving (Step

8). Although array tomography procedures are at present relatively complex and

demanding in comparison to many other imaging methods, each of the steps lends

itself potentially to automated and highly parallel implementations, and for many

12

applications the advantages outlined below can easily justify this extra effort.

Resolution

The volumetric resolution of fluorescence array tomography compares very favorably

with the best optical sectioning microscopy methods. The axial resolution limit for

array tomography is simply the physical section thickness (typically 70 nm). For a

confocal microscope, the z-axis resolution is limited by diffraction to e700 nm. The

confocals limiting z-axis resolution is usually worsened, however, by spherical aber-

ration when a high-numerical-aperture (high-NA) objective is focused more deeply

than a few micrometers into any tissue specimen. Array tomography physical sec-

tioning thus improves on ideal confocal optical sectioning by at least an order of

magnitude. Spherical aberrations also adversely impact the lateral resolution of con-

focal microscopes as they are focused into a tissue depth. Array tomography avoids

this problem, because the high-NA objective is always used at its design condition

(immediate contact between specimen and coverslip), with no chance of focus depth

aberration. The degradation of lateral resolution that occurs at focus depths of just a

few micrometers can easily exceed a factor of 2 (see http://www.microscopy.fsu.edu/),

so a very conservative approximation would imply that array tomography using or-

dinary high-NA, diffraction-limited optics would improve volumetric resolution (the

product of improvements in x-, y-, and z-axes) by a factor of 40 (= 2 x 2 x 10). The

improved volumetric resolution realized by array tomography can be very significant.

For instance, individual synapses in situ within mammalian cortex generally cannot

be resolved optically from their nearest neighbors by confocal microscopy but can be

resolved quite reliably by array tomography [1].

13

Depth Invariance

The major limitation to quantitative interpretation of whole-mount tissue immunoflu-

orescence images arises from reductions in both immunostaining and imaging efficien-

cies as focal plane depth increases. Diffusion and binding regimes typically limit the

penetration of labeling antibodies to the first few micrometers below the surface

of a tissue, even after multiday incubations. Imaging efficiency likewise decreases

with depth, as increasing spherical aberration and light scattering reduce signals

profoundly with focal plane depths of just a few micrometers. These staining and

imaging efficiency gradients make any quantitative comparison of specimen features

at different depths with whole-mount (e.g., confocal) volume microscopy difficult and

unreliable. Array tomography completely circumvents depth dependence issues, be-

cause each specimen volume element is stained identically owing to minimal section

thickness, and imaged identically because every section is bonded directly to the

coverslip surface.

Multiplexity

Traditional multicolor immunofluorescence techniques have provided compelling evi-

dence for the localization of multiple molecular species at individual subcellular com-

plexes. For example, because there is a very large number and a great diversity of

distinct molecules at individual synapses, there is a pressing need for imaging tech-

niques that can simultaneously discriminate many more than the three or four species

that can be distinguished by standard multicolor immunofluorescence. Attempts have

been made in the past to improve the multiplexity of immunofluorescence microscopy

by repeated cycles of staining, imaging, and stain elution, but the results have been

disappointing owing to the tendency of antibody elution treatments to destroy sam-

ples. In array tomography, specimens are stabilized by the embedding resin matrix

14

and by tight attachment to the coverslip substrate. An example of multiplexed stain-

ing with array tomography is shown in Figure 2.3. We have shown as many as nine

cycles of staining, imaging, and elution thus far [1]. With four fluorescence colors per

cycle, this would mean that 36 or more antigens could be probed in one specimen.

We now routinely acquire four colors in each of three cycles for a total of 12 marker

channels. Although 12–36 markers may still fall short of the degree of multiplex-

ing needed to fully probe the many and diverse molecules composing a synapse, it

is a substantial advance in comparison to traditional multicolor immunofluorescence

methods.

Volume Field of View

In principle, array tomography offers unique potential for the acquisition of high-

resolution volume images that extend seamlessly over very large tissue volumes. The

depth invariance of array tomography noted above eliminates any fundamental limit

to imaging in depth, whereas the availability of excellent automated image mosaic ac-

quisition, alignment, and stitching algorithms allows tiling over arbitrarily large array

areas. Ultimate limits to the continuous arrayable volume will be imposed by diffi-

culties in tissue fixation, processing, and embedding (owing to diffusion limitations)

as thicker volumes are encountered, and by mechanical issues of ultramicrotome and

diamond knife engineering as block face dimensions increase. Successful array tomog-

raphy has already been shown for volumes with millimeter minimum dimensions, and

it seems likely that volumes with minimum dimensions of several millimeters (e.g.,

an entire mouse brain) may be manageable eventually.

In practice, the size of seamless array tomography volumes is limited by the re-

quirement that numerous steps in the fabrication, staining, and imaging of arrays be

performed through many iterations without failure. At present, the most error-prone

steps are those involved in array fabrication, whereas the most time-consuming are

15

those involved in image acquisition. Ongoing engineering of array fabrication mate-

rials and processes will advance present limits to the error-free production of large

arrays, whereas image acquisition times will be readily reducible by dividing large ar-

rays across multiple substrates and imaging those subarrays on multiple microscopes.

The following protocols describe one simple implementation of immunofluores-

cence array tomography suitable for any laboratory with standard equipment and

some expertise in basic fluorescence microscopy and ultrathin sectioning. In addi-

tion, algorithms designed to fully automate the acquisition of array images are de-

scribed for the benefit of any laboratory having or planning to acquire the appropriate

automated fluorescence microscopy hardware and software.

2.2 Protocol A: Rodent Brain Tissue Fixation and

Embedding

Careful preparation of the tissue is essential for successful array tomography. These

steps take time to complete and require some practice to perfect.

2.2.1 Materials

CAUTION: See full Cold Spring Harbor citation, Appendix 6 for proper handling

of materials marked with <!>. See the end of the chapter for recipes for reagents

marked with <R>.

Reagents

• Ethanol <!>, 4◦C

• Fixative <R>

16

• Isoflurane <!> (VWR International)

• LR White resin <!> (medium grade, SPI Supplies 2646 or Electron Microscopy

Sciences 14381)

• Mice

• Wash buffer <R>, 4◦C

Equipment

• Capsule mold (Electron Microscopy Sciences 70160)

• Dissection instruments: handling forceps, small scissors, bone rongeur, forceps

#5, small spatula, scalpel

• Gelatin capsules, size 00 (Electron Microscopy Sciences 70100)

• Guillotine

• Microscope, dissection

• Microwave tissue processor system (PELCO with a ColdSpot set at 12◦C; Ted

Pella, Inc.) (optional)

• Oven (set at 51◦–53◦C)

• Paintbrush, fine

• Petri dishes, 35-mm

• Scintillation vials, glass, 20-mL

17

2.2.2 Experimental Method

Dissecting and Fixing Tissue

1. Anesthetize the rodent with isoflurane.

2. Remove head using the guillotine.

3. In a hood, using the dissection tools quickly remove the brain and plunge it into

a 35-mm Petri dish filled with fixative (room temperature). Remove the tissue

region of interest.

4. Transfer tissue to a scintillation vial with fixative solution. Use e1 mL of fixative

per vial, or just enough to cover the tissue; excessive liquid volume will cause

overheating in the microwave.

5. Microwave the tissue in the fixative using a cycle of 1 min on/1 min off/1 min

on at 100–150 W. After this and each subsequent cycle feel the glass vial to

check for overheating. If solutions are getting too warm (>37◦C), decrease the

amount of liquid added.

6. Microwave using a cycle of 20 sec on/20 sec off/20 sec on at 350–400 W. Repeat

three times.

7. Leave the tissue at room temperature for e1 h.

If a microwave is unavailable, fix the samples at room temperature for up to 3

h or overnight at 4◦C. Tissue can also be fixed by perfusion.

8. Prepare ethanol dilutions: 50%, 70%, 95%, and 100% in ultrapure H2O. Keep

at 4◦C.

9. Wash the tissue in wash buffer (4◦C) twice for 5 min each.

18

10. Transfer the tissue to a 100-mm Petri dish, cover with wash buffer, and under a

dissecting microscope dissect the tissue into smaller pieces (<1 mm in at least

one dimension).

11. Return the samples to scintillation vials and rinse them twice with wash buffer

for 15 min each at 4◦C.

12. Change to 50% ethanol (4◦C) and microwave the samples for 30 sec at 350 W.

Use just enough liquid to cover the tissue; excessive liquid volume will cause

overheating.

If a microwave processor is unavailable, Steps 12–20 can be performed for 5 min

per step on the bench.

13. Change to 70% ethanol (4◦C) and microwave the samples for 30 sec at 350 W.

Processing Samples that Contain Fluorescent Proteins

If processing samples with fluorescent proteins, then complete Steps 14–16. If

samples do not contain fluorescent proteins, then skip Steps 14–16, and instead

continue with Step 17.

14. Change one more time to 70% ethanol and microwave for 30 sec at 350 W.

15. Change to a mixture of 70% ethanol and LR White (1:3; if it turns cloudy add

1–2 extra drops of LR White) and microwave for 30 sec at 350 W.

16. Go to Step 20.

Processing Samples that Do Not Contain Fluorescent Proteins

17. Change to 95% ethanol (4◦C) and microwave for 30 sec at 350 W.

19

18. Change to 100% ethanol (4◦C) and microwave for 30 sec at 350 W. Repeat once.

19. Change to 100% ethanol and LR White resin (1:1 mixture, 4◦C) and microwave

for 30 sec at 350 W.

Embedding Brain Tissue

20. Change to 100% LR White (4◦C) for 30 sec at 350 W. Repeat two more times.

21. Change to fresh LR White (4◦C) and leave either overnight at 4◦C or 3 h at

room temperature.

22. Using a fine paintbrush, place the tissue pieces at the bottom of gelatin capsules

(paper labels can also be added inside the capsule) and fill to the rim with LR

White.

See Troubleshooting.

23. Close the capsules well and put in the capsule mold.

Gelatin capsules are used because they exclude air that inhibits LR White

polymerization. The little bubble of air that will remain at the top of the

capsule will not interfere with the polymerization.

24. Put the mold with capsules in the oven set at 51◦–53◦C. Leave overnight (e18–24

hours).

2.2.3 Troublesbooting

Problem (Step 22): It is difficult to orient the tissue.

Solution: If tissue orientation is important, it should be dissected in a shape that

will make it naturally sink in the resin the desired wayfor example, for mouse cerebral

20

cortex, a 300-µm coronal slice can be cut and trimmed to a rectangle, e1 x 2 mm,

that includes all of the cortical layers. Alternately, if the tissue is elongated and has

to be cut perpendicular to the long axis, the capsules can be positioned on the side,

instead of standing up in the mold.

2.3 Protocol B: Production of Arrays

Once the tissue has been embedded, the arrays are prepared. This protocol requires

familiarity with ultramicrotome sectioning for electron microscopy.

2.3.1 Materials

CAUTION: See full Cold Spring Harbor citation, Appendix 6 for proper handling

of materials marked with <!>. See the end of the chapter for recipes for reagents

marked with <R>.

Reagents

• Borax

• Contact cement (DAP Weldwood)

• Subbing solution <R>

• Tissue, fixed and embedded as in Protocol A

• Toluidine blue

• Xylene <!>

21

Equipment

• Coverslips (for routine staining: VWR International Micro Cover Glasses, 24 x

60-mm, No.1.5, 48393-252; for quantitative comparison between different arrays:

Bioscience Tools High Precision Glass Coverslips CSHP-No1.5-24 x 60)

• Diamond knife (Cryotrim 45; Diatome) (optional)

• Diamond knife (Histo Jumbo; Diatome)

• Eyelash tool

• Marker

• Razor blades

• Paintbrush, fine

• Slide warmer set at 60◦C

• Staining rack (Pacific Southwest Lab Equipment, Inc. 37-4470 and 4456)

• Syringe

• Transfer pipettes, extra fine-tip polyethylene (Fisher Scientific 13-711-31)

• Ultramicrotome (e.g., Leica EM UC6)


1. Prepare subbed coverslips. They can be prepared in advance and stored in

coverslip boxes until needed.

i Put clean coverslips into the staining rack.

22

ii Immerse the rack in the subbing solution and remove bubbles formed at the

surface of the coverslips using a transfer pipette.

iii After 30–60 sec, lift out and drain off excess liquid. Leave the coverslips in

a dust-free place until they are dry.

2. Using a razor blade, trim the block around the tissue. A blockface e2 mm wide

and 0.5–1 mm high works best.

3. Using a glass knife or an old diamond knife cut semithin sections until you

reach the tissue. Mount a couple of the semithin sections on a glass slide and

stain with 1% toluidine blue in 0.5% borax. View the stained sections under a

microscope to determine whether they contain the region of interest and decide

how to trim the block.

4. Trim the block again, to ensure that the blockface is not too big and the leading

and trailing edges of the blockface are parallel. The Cryotrim 45 diamond knife

works well for this purpose.

5. Using a paintbrush, apply contact cement diluted with xylene (1:2) to the lead-

ing and trailing sides of the block pyramid. Blot the extra glue using a tissue.

6. Insert a subbed coverslip into the knife boat of the Histo Jumbo diamond knife.

You may need to push it down and wet it using the eyelash tool. Make sure

that the knife angle is set at 0◦.

7. Carefully align the block face with the edge of the diamond knife. If the block

starts cutting at an angle, the leading and trailing edge of the block face will

no longer be parallel.

8. Start cutting ribbons of serial sections (70–200 nm) with the diamond knife. In

general, thinner sections stick better to the glass.

23


9. When the desired length of the ribbon is achieved, carefully detach it from the

edge of the knife by running an eyelash along the outer edge of the knife. Then

use the eyelash to gently push the ribbon toward the coverslip, so that the edge

of the ribbon touches the coverslip at the interface of the glass and the water.

The edge of the ribbon will stick to the glass.

10. Using a syringe, slowly lower the water level in the knife boat until the entire

ribbon sticks to the glass.

11. Remove the coverslip from the water and label it on one edge. Also, mark

the position of the ribbon by circling it with a marker on the backside of the

coverslip.

This allows you to keep track of the samples and provides a way to tell which

side of the coverslip the ribbon is mounted on (without a label, after the ribbon

dries, it is not possible to tell which side it is on).

12. Let the ribbon dry at room temperature and place the coverslip on the slide

warmer (e60◦C) for 30 min. The slides can be stored at room temperature for

at least 6 mo.

2.3.3 Troubleshooting

Problem (Step 8): The ribbons curve.

Solution: Sometimes, even when the leading and trailing edges of the blockface are

parallel, the ribbons are curved. This can happen when there is more resin around the

tissue on one side of the block than the other. As the section comes in contact with

water it expands, however, the resin and tissue expand to different degrees, causing

24

curving of the ribbon. Thus, make sure that the extra resin is trimmed on either side

of the block.

Problem (Step 8): The ribbons break.

Solution: Trim the block using a very sharp razor blade or, even better, the

Cryotrim diamond knife. Make sure that the blockface is at least twice as wide as it

is high. Apply glue again and take care to align the block so the edge of the blockface

is parallel to the knife edge.

2.4 Protocol C: Immunostaining and Antibody

Elution

The tissue arrays are prepared for imaging by binding primary antibodies against

specific cellular targets followed by secondary fluorescent antibodies. Alternatively,

fluorescent proteins can be used that have been introduced into the tissue before

dissection.

2.4.1 Materials

CAUTION: See Appendix 6 for proper handling of materials marked with <!>. See

the end of the chapter for recipes for reagents marked with <R>.

Reagents

• Alternative antibody dilution solution with normal goat serum (NGS) <R>

• Alternative blocking solution with NGS <R>

• Blocking solution with bovine serum albumin (BSA) <R>

• Elution solution <R>

25

• Glycine

• Mounting medium: SlowFade Gold antifade reagent with DAPI <!> (Invitrogen

S36939) or without DAPI (Invitrogen S36937)

• Primary antibodies, see Table 2.8

A detailed list of antibodies that have been tested for array tomography is

available from www.smithlab.stanford.edu.

• Secondary antibodies: for example, the appropriate species of Alexa Fluor 488,

594, and 647, IgG (H+L), highly cross-adsorbed (Invitrogen)

• Tissue sectioned as in Protocol B

• Tris buffered saline tablets (Sigma-Aldrich T5030)

Equipment

• Microcentrifuge

• Microscope slides (precleaned Gold Seal Rite-On micro slides; Fisher Scientific

12-518-103)

• PAP pen (ImmEdge Pen, Vector Laboratories H-4000)

• Petri dishes, 100-mm diameter

• Slide warmer set at 60◦C

• Transfer pipettes, extra fine-tip polyethylene (Fisher Scientific 13-711-31)

26


1. Encircle the ribbon of sectioned tissue with a PAP pen.

2. Place the coverslip into a humidified 100-mm Petri dish and treat the sections

with 50 mM glycine in Tris buffer for 5 min.

3. Apply blocking solution with BSA for 5 min.

If there is a problem with high background staining, see the alternate blocking

and staining protocol beginning with Step 21.

4. Dilute the primary antibodies in blocking solution with BSA. Approximately

150 µL of solution will suffice to cover a 30-mm-long ribbon.

5. Centrifuge the antibody solution at 13,000 revolutions per minute (rpm) for 2

min before applying it to the sections.

6. Incubate the sections in primary antibodies either overnight at 4◦C or for 2 h

at room temperature.

Primary antibodies are diluted to 10 µg/mL, although the best concentration

will need to be determined for each antibody solution.

7. Rinse the sections three to four times with Tris buffer for a total of e20 min.

Wash the sections using a manual perfusion method, simultaneously adding Tris

buffer on one end and removing if from another with plastic transfer pipettes.

8. Dilute the appropriate secondary antibodies in blocking solution with BSA

(1:150 for Alexa secondaries).

9. Centrifuge secondary antibody solution at 13,000 rpm for 2 min.

10. Incubate the sections in secondary antibodies for 30 min at room temperature

in the dark.

27

11. Rinse the sections three to four times with Tris buffer for e5 min each.

12. Wash the coverslip thoroughly with filtered ultrapure H2O to remove any dust

or debris, leaving some H2O on the sections so that they do not dry out.

13. Mount the sections on a clean, dust-free microscope slide with SlowFade Gold

Antifade containing DAPI.

14. Image the sections as soon as possible after immunostaining, or at least the same

day. If you are planning to restain the sections with additional antibodies, elute

the antibodies (Steps 15–19) as soon as possible after imaging.

Elute Antibodies Before Restaining

15. Add filtered ultrapure H2O around the edge of the coverslip to help slide it off

the microscope slide.

Wash the coverslip gently with filtered ultrapure H2O to rinse off the mounting

medium.

16. Apply elution solution for 20 min.

17. Gently rinse the coverslips twice with Tris, allowing them to sit for 10 min with

each rinse.

18. Rinse the coverslips with filtered ultrapure H2O and let them air dry completely.

19. Bake the coverslip on a slide warmer set to 60◦C for 30 min.

Staining the Sections Multiple Times

20. Restain using the Steps 2–13 above or store array at room temperature until

needed.

28


Alternative Staining Method to Reduce Background

21. Proceed through Steps 1 and 2 of the staining protocol above.

22. Incubate the sections for 30 min with alternative blocking solution with NGS.

If secondary antibodies are made in donkey, use normal donkey serum; if sec-

ondary antibodies are made in horse, use normal horse serum, etc. This protocol

can only be used if all of the secondary antibodies are made in the same animal.

23. Dilute the primary and secondary antibodies in alternative antibody dilution

solution with NGS.

24. Follow the rest of the staining protocol above, using the solutions with NGS.


Problem (Step 20): There is incomplete elution of antibodies.

Solution: To check for incomplete elution, which could interfere with subsequent

antibody staining, perform the following control experiment. Stain with the antibody

of interest and image a region that you can relocate later. Elute and apply the

secondary antibody again. Image the same region as before, using the same exposure

time; this will give an estimate of how much primary antibody was left after the

elution. Increase the exposure time to determine if longer exposure times reveal the

initial pattern of antibody staining. If the first antibody was not eluted sufficiently, try

longer elution times. Some antibodies elute poorly (e.g., rabbit synapsin or tubulin)

and, if followed by a weaker antibody, may still be detectable after the elution. In

such cases, begin the experiment with the weaker antibodies.

29

2.5 Protocol D: Imaging Stained Arrays

Tissue arrays are imaged using a conventional wide-field fluorescence microscopy.

Images can be captured manually or, with the appropriate software and hardware,

the process can be automated.

2.5.1 Materials

Reagents

• Immunostained brain sections prepared as in Protocol C

Equipment

• Digital camera (Axiocam HR, Carl Zeiss)

• Fluorescence filters sets (all from Semrock) YFP, 2427A; GFP, 3035B; CFP,

2432A; Texas Red, 4040B; DAPI, 1160A; FITC, 3540B; and Cy5, 4040A

• Illuminator series 120 (X-Cite)

• Objective (Zeiss Immersol 514 F Fluorescence Immersion Oil)

• Piezo Automated Stage (Zeiss)

• 10x Plan-Apochromat 0.45 NA

• 63x Plan-Apochromat 1.4 NA oil objective

• Software (e.g., Zeiss Axiovision with Interactive Measurement Module, Au-

tomeasure Plus Module and Array Tomography Toolbar; the toolbar can be

downloaded from http://www.stanford.edu/ebbusse/work/downloads.html)

• Upright microscope (Zeiss Axio Imager.Z1)

30


Manual Image Acquisition

1. Focus on your sample using the 10x objective. Find the ribbon by focusing on

the DAPI label or another bright label that is not prone to bleaching. Once you

have found the right general area of the sample, switch to the 63x objective.


2. Find the exact area of the sample that you want to image. Choose a landmark

that you can use to find the same spot in the next section. A useful landmark

should not change dramatically from one section to the next (e.g., a DAPI-

stained nucleus or blood vessel). Because the sections are 70–200 nm thick we

can often follow the same nucleus through the entire length of a long array. Line

up your landmark with a crosshair in the middle of the field.

3. Set the correct exposure for each of your fluorescence channels.

4. Beginning with the first section, collect an image of your area of interest.

5. Manually, move to the same area of the next section. The glue on the edge

of each section is autofluorescent, so you can tell when you have moved to the

next section. Align your landmark carefully in each section to assure that your

image alignment will run smoothly.


6. Continue to the end of the ribbon, collecting an image from each section. Align

your stack of images using Protocol E.

31

Automated Image Acquisition

Although we have developed our automated tools to work with Zeiss Axiovision

software, any microscopy software suite (such as Micro-Manager) controlling

an automated stage should be adaptable to this approach. Some steps may be

altered or eliminated, depending on your framework and implementation.

7. With the 10x objective, find the ribbon by focusing on the DAPI label or another

bright label that is not prone to bleaching.


8. Acquire a mosaic image of the entire ribbon with the MosaicX Axiovision mod-

ule, using a bright label that does not vary much between sections, such as

DAPI.

9. Find the top left and bottom right corners of the ribbon and use them to define

the limits of the mosaic in the Mosaic Setup dialog.

10. Set three to four focus positions along the length of the ribbon and enable focus

correction.

11. Collect the mosaic image. Convert the mosaic to a single image with the Convert

Tile Images dialog, setting the Zoom factor to 1 so that the resulting image is

the same size.


12. Choose a point of interest to be imaged in the ribbon. Place a marker on

that point via Measure → Marker. Place another marker at the same spot in

the next consecutive section. Create a table of the x and y coordinates of the

markers, DataTable, via Measure → Create Table, with the list option. This

allows Axiovisions Visual Basic scripts to read the marker locations.

32


13. With the large, stitched image selected, call PrepImage and MarkLoop from

the Array Tomography toolbar.

14. The preceding step will create a file (csv) with a list of the coordinates for the

same position in each section, which will be automatically saved in the same

folder as the mosaic and with the same name as the stitched image. To load

the position list, go to Microscope → Mark and Find, click the New icon, and

then the Import Position list button. In the Mark and Find dialog, switch to

the Positions tab which will let you review or edit the calculated positions by

double-clicking on any position.

15. Collect one field of view at each point via Multidimensional Acquisition with the

position list checkbox set. We recommend using a bright label that is present

throughout the field as the first channel, setting it to autofocus at each position.

Review your images at the end to make sure they are all in focus.



Problem (Steps 1 and 7): Sections cannot be found under the microscope.

Solution: Use DAPI in the mounting mediumit will stain the nuclei brightly and

make it easy to find the sections with the 10x objective. Make sure the coverslip has

been mounted with the sections on the same side as the mounting medium and that

there are no bubbles in the immersion oil.

Problem (Steps 5 and 15): Sections are wrinkled.

Solution: Section wrinkling can occur at several steps in the procedure. First, it

can occur during array preparation if the coverslip is put on the slide warmer while

33

the ribbon is still wet. Make sure that the sections are dry before putting them on

the slide warmer. It can also occur if the blockface is too big (>1 x 2 mm) or sections

are too thick (>200 nm). Second, wrinkles can be caused by improper subbing of

the coverslips. The gelatin must be 300 Bloom (measure of stickiness, higher number

indicates stickier) and should not be heated above 60◦C during solution preparation.

Third, sections can wrinkle if the ribbon is stored with the mounting solution for >2

days. Finally, wrinkling can occur after antibody elution, especially with sections 200

nm thick. Make sure that the solutions are applied gently during the elution and the

array is completely dry before putting it on the slide warmer.

Problem (Steps 5 and 15): There is no staining or fluorescent signal.

Solution: Use a high-power, high-NA objectiveideally a 63x oil objective. Only

immunofluorescence with antibodies against abundant antigens (e.g., tubulin, neu-

rofilament) will be visible with a low-power objective. Also, check if there are two

coverslips stuck to each other; this will make it impossible to focus at higher magni-

fication.

Problem (Steps 5 and 15): Punctate staining is seen with a seemingly random

distribution.

Solution: Immunostaining with thin array sections (≤200 nm) looks different from

staining on thicker cryosections or vibratome sections. Because a very thin layer of

tissue is probed, many stains that appear continuous on thicker sections will appear

punctate with array tomography. A 3D reconstruction of a short ribbon (10–20 sec-

tions) can be helpful for comparison. You may also need to test antibody performance.

First, compare the antibody staining pattern to that of different antibodies against

the same antigen or a different antigen with a similar distribution. For example, a

presynaptic marker should be adjacent to a postsynaptic marker. Other common

controls for immunostaining can be used, such as omitting primary antibodies, stain-

ing a tissue that does not contain the antigen, etc. Second, specific controls for array

34

tomography include comparison of the antibody staining patterns from adjacent sec-

tions or from consecutive stains (i.e., stain → image → elute → stain with the same

antibody → image the same region → compare). Not all antibodies that work well

for other applications will work for array tomography.

Problem (Steps 5 and 15): There is high background fluorescence.

Solution: Background fluorescence can have many causes. Often, there is high

autofluorescence when using the low-power (but not high-power) objectives. If the

autofluorescence levels are high with the 63x objective, try the following. First,

check whether the immersion oil is designed to be used with fluorescence. Second,

labeling marks on the back of the coverslip can dissolve in the immersion oil causing

autofluorescencewipe labels off with ethanol before imaging. Third, use high-quality

fluorescence filter sets. Fourth, try a longer fluorescence quenching step (glycine

treatment in Protocol C, Step 2), the alternative staining method (Protocol C, Step

21), or introduce an additional quenching step with 1% sodium borohydride in Tris

buffer for 5 min.

Problem (Steps 5 and 15): Green fluorescent protein (GFP)/YFP fluorescence is

lost.

Solution: First, confirm that the tissue was dehydrated only to 70% ethanol (Pro-

tocol A, Step 14). Second, make sure you are using a high-power, high-NA objective.

To check for GFP fluorescence use a short array with ultrathin sections (<200 nm).

Let it sit for 5–10 min or more with Tris-glycine (50 mM glycine in Tris), mount over

a glass slide and look with the 63x objective. GFP can bleach very fast, so work

quickly to find the region with GFP fluorescence. For acquiring images, select the

region of interest with another stain (e.g., Alexa 594) and focus. Do not use the DAPI

stain for this purpose, because it can cause DAPI to bleed into the GFP channel. In

cases of weak GFP fluorescence, GFP antibodies may help identify GFP-positive cell

bodies and large processes, but are generally not useful for thinner processes. GFP

35

antibodies for array tomography include Roche 11814460001 (mouse), MBL 70 (rab-

bit), Invitrogen A11122 (rabbit), NeuroMabs 75-131 (mouse), GeneTex GTX13970

(chicken). All of these antibodies should be used at 1:100 dilution.

Problem (Step 11): The Convert Tile Images step keeps downsampling the stitched

image.

Solution: In the Tools→ Options→ Acquisition menu, change the Mx. MosaicX

image size to the maximum allowed: 1000000000 pixel.

Problem (Step 12): The microscopy software is not designed for array tomography.

Solution: We have developed an algorithm that automates position finding in the

arrays by using simple extrapolation to estimate the neighborhood of an unknown

point and then refining the estimate with an autocorrelation search. Given two known

points Pn and Pn1, we find the next point Pn+1 such that Pn+1 = Pn + (Pn [Pn1])

(Figure 2.4). This does not take into account ribbon curvature or changes in section

width, but gives a rough approximation of the unknown points locale. Pn+1 becomes

the center of an autocorrelation search to find the points true position. The size

of the search varies with the width of the sections; larger sections will have larger

warping and curvature effects, and any miscalculation in the estimate of Pn+1 will

be magnified.

To conduct the search, the algorithm compares the area centered at Pn+1 with

a Kalman-filtered image of recently processed points. Although our fiducial labels

(DAPI and tubulin immunostaining) have minor variations from section to section, it

does not disrupt the accuracy of the correlation search. To make the Kalman-filtered

image at each iteration, use the area around the current Pn, newSample, to update

the image using the following pseudocode: image = 0.3 x image + 0.7 x newSample.

The purpose of using the Kalman filter, when newSample alone would do, is to add

a measure of robustness to the algorithm. If the ribbon is damaged or has aberrant

staining on a single section, using newSample alone may result in the algorithm going

36

off course. With a running average of previous iterations to compare with, a defect

in a single section has a good chance of being ignored. This process continues until

one end of the ribbon is reached, then starts in the other direction.

We developed an implementation of this algorithm in Visual Basic script for Zeiss

Axiovision, available from http://www.stanford.edu/ebbusse/work/downloads.html,

and would welcome any ports to other microscopy software.

Problem (Step 15): Autofocus does not work using Axiovision.

Solution: The autofocus does not work every time. Typically, e5% of the images

collected with autofocus may be out of focus. In that case, you can move to the

positions on the ribbon with bad focus, focus by hand, and collect individual images.

Replace the out-of-focus images with the newly focused ones in the stack before to

alignment. If 10% or more of the images are out of focus, you can try using the

autofocus with a different channel. Pick a channel with antibody staining that is

bright, and present throughout the field of view. Using a channel with dim or sparse

immunostaining will not work well.

Problem (Step 15): Autofocus is grayed out.

Solution: In the Tools → Options → Acquisition menu, check the box marked

either Use calibration-free Autofocus or Enable new Autofocus.

2.6 Protocol E: Semiautomated Image Alignment

Successful array tomography requires that the captured images be properly stacked

and aligned. Software to achieve these ends is freely available.

37

2.6.1 Materials

Software

• Fiji can be obtained at http://pacific.mpi-cbg.de/wiki/index.php/Main Page

• MultiStackReg is available at

http://www.stanford.edu/ebbusse/work/downloads.html


1. Load your images into Fiji. If using Axiovision, Fijis Bio-Formats Importer

plugin can read .zvi files directly.

2. Pick a channel that is relatively invariant from one section to the next (e.g.,

DAPI or tubulin), and select a slice near the middle of the ribbon.

3. Align the sections of that channel using affine in MultiStackReg (Fiji), but do

not save over the misaligned stack. Save the resulting transformation matrix.

This is the intrasession matrix.


4. Using MultiStackReg, apply that matrix to the other channels of the same

imaging session.

5. For each subsequent imaging session, choose the same channel. Align the new

(misaligned) channel to the old (misaligned) channel, saving the matrix. This

is the intersession matrix.

6. For each channel in that imaging session, first apply the intersession matrix

from Step 5 and then the intrasession matrix from Step 3.

7. Repeat until all imaging sessions have been registered.

38


Problem (Step 3): The alignment steps are not working properly.

Solution: Detailed instructions with graphical illustration, compiled by Andrew

Olson, are available at http://nisms.stanford.edu/UsingOurServices/Training.html.

If an affine transformation does not align the images well, try either the rigid body

then affine or try rigid body alone. For each registration step, save the transformation

matrix and apply it to the other channels in sequence.

MultiStackReg is an extension of the StackReg ImageJ plugin, which is depen-

dent on TurboReg [20]. TurboReg aligns a single pair of images using a pyramid

registration scheme. StackReg aligns an entire stack by calling TurboReg on each

pair of consecutive slices in the stack, propagating the alignment to later slices. The

two principle changes added by MultiStackReg are the ability (1) to load and save

transformation matrices and (2) to align one stack to another by registering each pair

of corresponding sections independently. MultiStackReg can process TurboReg align-

ment files in the same manner as the files it generates for itself, so if your alignment

is failing owing to a single section, it is possible to manually align that section in

TurboReg, apply that transform to a copy of the stack, and splice the two together.

2.7 Conclusion and Future Directions

One important application of array tomography in the field of neuroscience is the

analysis of synapse populations. With this method it is possible to resolve individ-

ual synapses in situ within brain tissue specimens. Because 10 or more antibodies

can be used on an individual sample, the molecular signature of each synapse can

be defined with unprecedented detail. The throughput of the technique is inherently

high, approaching the imaging of one million synapses per hour. Compared with

3D reconstruction at the electron microscopic level, array tomography can image

39

much larger volumes and provide information about the presence of a much larger

number of molecules, but cannot presently provide the fine ultrastructure of electron

microscopy. On the other hand, the amount of effort involved in array tomography

may not be warranted for all studies. If it is not considered critical to resolve indi-

vidual synapses, immunostaining of vibratome sections or cryosections and confocal

microscopy imaging may be sufficient.

Currently, we are focused on developing array tomography in three directions.

First, we are refining current staining and imaging approaches to image larger and

larger tissue volumes with more antibodies. Second, we are combining light and

electron microscopic imaging to visualize both immunofluorescence and ultrastructure

on the same tissue sections. Finally, we are applying advanced computational methods

for data analysis, in particular with the goal to both count and classify millions of

synapses on a routine basis.

2.7.1 Acknowledgements

We thank JoAnn Buchanan and Nafisa Ghori for their help in refining the methods.

This work was supported by grants from McKnight Endowment Fund for the Neu-

rosciences, the National Institutes of Health (NS 063210), The Gatsby Charitable

Foundation, and the Howard Hughes Medical Institute.

2.8 Recipes

CAUTION: See Appendix 6 for proper handling of materials marked with <!>.

Recipes for reagents marked with <R> are included in this list.

40

Table 2.1: Recipe: Alternative Antibody Dilution Solution with NGS (1 mL)

Reagent Quantity Final concentrationTween (1%) (make the stock solution usingTween-20 [Electron Microscopy Sciences25564])

100 µL 0.1%

NGS (Invitrogen PCN5000) 30 µL 3%Tris buffer 870 µL

Prepare on the same day it is used. NGS can be kept frozen in aliquots for severalmonths.

Table 2.2: Recipe: Alternative Blocking Solution with NGS (1 mL)


100 µL 0.1%

NGS (Invitrogen PCN5000) 100 µL 10%Tris buffer 800 µL

Prepare on the same day it is used. NGS can be kept frozen in aliquots for severalmonths.

Table 2.3: Recipe: Blocking Solution with BSA (1 mL)


50 µL 0.05%

BSA (10%) (AURION BSA C [acety-lated BSA], Electron Microscopy Sciences25557)

10 µL 0.1%

Tris buffer 940 µL

Prepare the same day. The 1% Tween stock (10 µL Tween in 1 mL of H2O) and the10% BSA stock can be kept at 4◦C for several months.

41

Table 2.4: Recipe: Elution Solution (10 mL)

Reagent Quantity Final concentrationNaOH <!>, 10 N 200 µL 0.2 N%SDS <!> (20%) 10 µL 0.02%Distilled H2O 10 mL

Can be prepared in advance and stored at room temperature for several months.

Table 2.5: Recipe: Fixative (4 mL)

Reagent Quantity Final concentrationParaformaldehyde <!> (8%, EM grade;Electron Microscopy Sciences 157-8)

2 mL 4%

PBS, 0.02 M (use PBS powder, pH 7.4[Sigma-Aldrich P3813])

2 mL 0.01 M

Sucrose 0.1 gm 2.5%

Prepare the same day as it will be used.

Table 2.6: Recipe: Subbing Solution (300 mL)

Reagent Quantity Final concentrationGelatin from porcine skin, 300 Bloom(Sigma-Aldrich G1890)

1.5 g 0.5%

Chromium potassium sulfate (Sigma-Aldrich 243361)

0.15 g 0.05%

Distilled H2O 300 mL

Prepare the same day. Dissolve the gelatin in 290 mL of distilled H2O by heating to<60◦C. Dissolve 0.15 gm of chromium potassium sulfate in 10 mL of H2O. Whenthe gelatin solution cools down to e37◦C, combine the two solutions, filter, and pourinto the staining tank. Use fresh.

42

Table 2.7: Recipe: Wash Buffer (50 mL)

Reagent Quantity Final concentrationGlycine 187.5 mg 50 mMSucrose 1.75 g 3.5%PBS, 0.02 M 25 mL 0.01 MDistilled H2O 25 mL

Can be prepared in advance and stored at 4◦C for up to 1 month; discard if itappears cloudy.

43

Table 2.8: Primary antibodies used with array tomography

Antibody Source Supplier DilutionSynapsin I Rabbit Millipore AB1543P 1:100PSD95 Mouse NeuroMabs 75-028 1:100VGluT1 Guinea pig Millipore AB5905 1:1000GAD Rabbit Millipore AB1511 1:300Gephyrin Mouse BD Biosciences 612632 1:100Tubulin Rabbit Abcam ab18251 1:200Tubulin Mouse Sigma-Aldrich T6793 1:200Neurofilament 200 Rabbit Sigma-Aldrich N4142 1:100

44

Figure 2.1: The sequence of steps for a basic immunofluorescence arraytomography process.

45

Figure 2.1

46

Figure 2.2: Array tomographic images of layer 5 neuropil, barrel cortex ofYFP-H Thy-1 transgenic mouse [71] Yellow fluorescent protein (YFP) expressionin a subset of pyramidal cells (green), Synapsin I immunostaining (white), PSD95(red), DAPI staining of nuclear DNA (blue). (A) Four-color fluorescence image ofa single, ultrathin section (200 nm). (B) Volume rendering of a stack of 30 sectionsafter computational alignment as described in this chapter.

47

Figure 2.2

48

Figure 2.3: Multiplexed staining for seven synaptic proteins in mouse cere-bral cortex (layer 2/3, barrel cortex) using five cycles of staining and elution. Thisvolume of 18 x 16 x 1.3 µm was reconstructed from 19 serial sections (70 nm each).Individual synapsin puncta 1, 2, and 3 colocalize with synaptophysin and VGlut1 andare closely apposed to PSD95 and thus appear to be excitatory synapses. Synapsinpuncta 4–7 colocalize with synaptophysin, but do not have adjacent PSD95 puncta.Puncta 6 and 7 also colocalize with GAD and VGAT and thus have the characteristicsof inhibitory synapses.

49

Figure 2.3

50

Figure 2.4: (Top) A fragment of an array tomography ribbon stained with DAPI.(Bottom) A closer view of two sections in the ribbon showing a single iteration ofthe position-finding algorithm. An established field (red x) is used to maintain areference patch (red square) for a correlation-based search (green square) to find thenext point (green circle).

51

Figure 2.4

52

Chapter 3

Classical MHCI molecules regulate

retinogeniculate refinement and

limit ocular dominance plasticity

Datwani A, McConnell MJ, Kanold PO, Micheva KD, Busse B, Shamloo M, Smith

SJ, Shatz CJ. Classical MHCI molecules regulate retinogeniculate refinement and

limit ocular dominance plasticity. Neuron. 2009 Nov 25;64(4):463-70.

3.0.1 Abstract

Major histocompatibility complex Class I (MHCI) genes were discovered unexpect-

edly in healthy CNS neurons in a screen for genes regulated by neural activity. In

mice lacking just 2 of the 50+ MHCI genes H2-Kb and H2-Db, ocular dominance (OD)

plasticity is enhanced. Mice lacking PirB, an MHCI receptor, have a similar pheno-

type. H2-Kb and H2-Db are expressed not only in visual cortex, but also in lateral

geniculate nucleus (LGN) where protein localization correlates strongly with synaptic

markers and complement protein C1q. In KbDb−/− mice developmental refinement of

53

retinogeniculate projections is impaired, similar to C1q−/− mice. These phenotypes

in KbDb−/− mice are strikingly similar to those in β2m−/−TAP1−/− mice, which lack

cell surface expression of all MHCIs, implying that H2-Kb and H2-Db can account

for observed changes in synapse plasticity. H2-Kb and H2-Db ligands, signaling via

neuronal MHCI receptors, may enable activity-dependent remodeling of brain circuits

during developmental critical periods.

3.1 Introduction

Brain circuits are refined by early spontaneous activity and later on by sensory ex-

perience [25, 26]. In the developing mammalian visual system, spontaneous activ-

ity generated in retinal ganglion cells (RGCs) and relayed to the lateral geniculate

nucleus (LGN) drive initially intermixed retinogeniculate connections to refine into

non-overlapping eye-specific layers; blockade or perturbation of waves prevents refine-

ment [27, 28]. How neural activity ultimately leads to refinement is not well under-

stood, but it is known that synapse elimination, as well as synapse strengthening and

stabilization, are involved [29–32].

To discover genes downstream of spontaneous retinal activity, we conducted an

unbiased screen in which activity was blocked by tetrodotoxin (TTX) during the

period of eye-specific segregation [33]. TTX not only prevents retinogeniculate re-

finement [34,35], but unexpectedly also downregulated the expression of MHC Class

I mRNA in neurons. MHCI (Major Histocompatibility Class I), a large and highly

polymorphic family of 50+ genes, some with well known roles in the immune system

and T-cell function [36], were not previously thought to be expressed by neurons in

the healthy brain, let alone regulated by neural activity.

Here, by examining mice lacking both H2-Kb and H2-Db (Histocompatibility Locus

2 [37], we show that these 2 genes are required for retinogeniculate refinement. In the

54

immune system H2-Kb and H2-Db bind and signal not only via the T-cell receptor, but

also via PirB (Paired immunoglobulin-like receptor B), an innate immune receptor [38]

also expressed in visual cortical neurons [39]. We report that KbDb−/− mice have

enhanced ocular dominance (OD) plasticity, similar to PirB mutant mice [39]. These

findings imply a novel role for these 2 specific MHCI molecules and cognate immune

receptors in activity-dependent plasticity during CNS development. Given the recent

association between human Histocompatibility Locus (HLA) and Schizophrenia [40–

42], our findings also reveal a novel way to explore potential links between immune

and nervous systems: via the disruption of normal MHCI regulation of plasticity at

neuronal synapses.

3.2 Results

3.2.1 Enhanced ocular dominance plasticity in KbDb−/− mice

In the immune system, H2-Kb and H2-Db bind and signal via a variety of recep-

tors, including PirB [38, 43]. Cortical neurons not only express PirB, but also mice

lacking PirB have enhanced OD plasticity following monocular deprivation or eye re-

moval [39]. Because cortical neurons express both H2-Kb and H2-Db [44], Figure S1,

we examined if OD plasticity is also perturbed in KbDb−/− mutant mice. To create

a large imbalance in visually driven inputs to neurons normally receiving binocular

inputs, one eye was deprived (MD) or enucleated (ME) at P22, just at the onset of

the critical period in visual cortex but after eye segregation in the LGN is adult-

like [45–47].

The spatial extent of input from the eyes to visual cortex was assessed functionally

at P31 by in situ hybridization for the immediate early gene Arc. After a 30-minute

monocular light exposure, Arc mRNA is upregulated (induced) rapidly [39, 47] in

55

visually stimulated cortical neurons (Figure 3.1A). Following ME or MD, this method

reports OD plasticity reliably, as measured by the pronounced expansion in width of

Arc mRNA signal in visual cortex ipsilateral to the stimulated (open) eye (Syken et

al., 2006; Tagawa et al., 2005). This expansion in width of Arc induction correlates

with the well-known strengthening of the open eye following monocular deprivation

assessed in physiological studies measuring single units [48], visually-evoked potentials

(VEPs) [45,49] or optical imaging [50,51] and faithfully detects OD plasticity with the

advantage here of laminar and cellular resolution. As expected [47], ME in wildtype

mice (WT: KbDb+/+) led to a 51% increase in width of Arc induction in layer 4 of

visual cortex ipsilateral to the remaining eye, as compared to normally reared WT

mice (Figure 3.1C).

By contrast, Arc induction in KbDb−/− mice following ME is even wider; indicat-

ing that OD plasticity following ME is greater in KbDb−/− than WT mice. However,

in KbDb−/− mice reared with normal vision, the width of Arc induction ipsilateral to

the stimulated eye is already about 22% greater than WT (cf. Figure 3.1B,C). This

initially expanded ipsilateral representation might account entirely for the greater

OD plasticity observed in KbDb−/− mice. Thus, to compare the extent of OD plas-

ticity between the different genotypes, we computed a plasticity index (=width of

Arc induction following ME/width following normal visual rearing; Figure 3.1D) that

factors in the different starting points for each genotype. The plasticity index for

KbDb−/− (2.16 ± 0.11) is 65% larger than for WT (1.51 ± 0.08, Figure 3.1D), in-

dicating that even after correcting for the expanded representation of the ipsilateral

eye in normally reared KbDb−/− mice, OD plasticity is significantly enhanced.

To exclude the possibility that enhanced OD plasticity after eye removal inKbDb−/−

mice is due to an injury response, rather than visual deprivation, experiments were re-

peated using MD at similar ages (Figure 3.1C). A more modest expansion in width of

Arc induction was detected than after ME [47]. Nevertheless following MD, the width

56

of Arc signal in KbDb−/− mice, as well as the plasticity index, is 17% greater than

that in WT mice (plasticity index: 1.48 ± 0.103 KbDb−/−; 1.31 ± 0.105 KbDb+/+; P

= 0.042; Figure 3.1C, D). These differences in OD plasticity in KbDb−/− vs. WT mice

are also not likely due to differences in visual ability. We examined this possibility

behaviorally in normally reared mice using a visual acuity water maze test [52, 53].

Acuity is similar in WT and KbDb−/− mice (Figure S2: A-F), suggesting that the

expanded representation of the ipsilateral eye within the BZ of visual cortex following

MD or ME is not from degraded vision in the mutant mice. Indeed, following ME

in mutant mice, visual acuity is, if anything, slightly better than in WT mice (Fig-

ure S2: G-I), possibly reflecting a behavioral consequence of enhanced OD plasticity.

Thus, like PirB [39], H2-Kb and/ or H2-Db appear to limit the extent of OD plasticity

following MD or ME during the critical period.

The vast majority of MHCI cell surface proteins can be abrogated by genetically

deleting β2m, a subunit needed for MHCIs, and/or the peptide loading-associated

molecule TAP1 [36,54]. In the brains of β2m−/−, or β2m−/−TAP1−/− mice, there are

changes in hippocampal synaptic plasticity, homeostatic scaling [44,55], motor learn-

ing [56], pheromone driven behavior [57], and motoneuron axon regeneration [58]. In

addition, following ME here, the width of Arc induction in visual cortex of β2m−/−

and β2m−/−TAP1−/− mice is similar to that in KbDb−/− mice (Figure 3.1C, 3.1D,

and data not shown: β2m−/−TAP−/−+ME: 2258.8 ± 120.2 µm, n = 38 mice, P

= 0.0032; KbDb−/−+ME: 2404.6 ± 105.1 µm, n = 25 mice, P = 0.0017) The plas-

ticity index is also 44% greater than for WT (β2m−/−TAP−/− = 2.08 ± 0.140;

β2m+/+TAP1+/+ = 1.64 ± 0.080; P = 0.0078), similar to that for KbDb−/− mice

(2.16 ± 0.11). These observations suggest that loss of just H2-Kb and H2-Db can

account for enhanced OD plasticity seen in visual cortex of mice lacking surface ex-

pression of the majority of MHCI proteins.

57

3.2.2 Expanded thalamocortical projections to layer 4 of KbDb−/−

mice following ME

The expansion of Arc signal in layer 4 following ME or MD could arise from at

least two non-mutually exclusive possibilities: (1) thalamocortical axons could have

a wider distribution and therefore contact more neurons in layer 4; and/or (2) the

expansion in Arc signal could result from changes in intracortical connections. To

explore these possibilities, transneuronal tracing, using 3H-proline [39, 48, 59] was

performed after ME in WT (n = 5) or KbDb−/− (n = 7) mice. Similar to Arc

induction, transneuronal tracing labels a wider patch in KbDb−/− than in WT mice

(Figure 3.2B, C). Even so, the width measured from transneuronal transport is about

70% smaller than that from Arc induction (cf. Figure 3.1C, 3.2C)), suggesting that

expansion of thalamocortical axons alone cannot account entirely for the increased

OD plasticity observed in KbDb−/−mice; changes in intracortical connections likely

also contribute.

3.2.3 Abnormal retinogeniculate patterning in KbDb−/− mice

Blockade of neural activity prevents developmental refinement of retinogeniculate pro-

jections [27] and down-regulates MHCI mRNA and protein [33,55]. KbDb−/− mutant

mice resemble the situation in which 2 MHCI molecules have been completely down-

regulated by activity blockade. To examine if H2-Kb and H2-Db could be involved

in refinement of the retinogeniculate projection, we first assessed expression in the

dLGN by immunostaining (Figure S1C). Protein can be detected in LGN neurons at

P9, during the period of activity-dependent refinement of the retinogeniculate projec-

tion. Expression has decreased by P34, after refinement is complete, consistent with

transient expression of MHCI mRNA in the LGN [44].

58

To assess the distribution of retinal ganglion cell (RGC) terminals, we made in-

traocular injections of different fluorophores into the each eye (Experimental Proce-

dures, [60]. The LGN territory occupied by RGC projections was assessed at P34, 3

weeks after projections normally have completely segregated (Figure 3.3A, Figure S3).

The percent LGN area occupied by projections from the ipsilateral eye in KbDb−/−

mice (24.99 ± 2.97%) is almost twice that of WT (12.75 ± 2.42%, Figure 3.3B), while

total LGN area is similar between genotypes.

3.2.4 Abnormal segregation of eye-specific inputs in dLGN

of KbDb−/−

An increase in area of the ipsilateral eye projection to the LGN could come at the

cost of contralateral eye territory, or if RGC inputs failed to segregate [60]. Using

fluorescent double labeling to examine projections from both eyes simultaneously at

P34, significantly increased overlap of RGC projections from the two eyes was ob-

served in KbDb−/− vs. WT mice (Figure 3.3; for both hemispheres see Figure S3B).

These changes in the retinogeniculate projection of KbDb−/− mice also almost exactly

phenocopy β2m−/−TAP1−/− mice (Figure S3B; [44]). Using multiple threshold anal-

ysis (Experimental procedures [60]), we find that overlap of RGC inputs, measured

by pixels common to both red and green channels, is significantly greater in all MHCI

mutant mouse lines over WT controls (Figure S3C, D). For example, at 60% maxi-

mal threshold, the percent of dLGN area with overlapping pixels is 19.10 ± 2.80% in

KbDb−/−; 15.50 ± 3.01% in β2m−/−TAP1−/−; but only 4.23 ± 1.77% in WT (Figure

S3D). On average across examined intensity thresholds there is approximately a 3–5

fold increase in overlap in mutant mice over WT (Figure S3D). These findings suggest

that normal retinogeniculate refinement may require just H2-Kb and H2-Db.

59

Given the enhanced OD plasticity noted in visual cortex of mutant mice, we won-

dered if RGC projections in mutant mice remain stable even after retinogeniculate

segregation is normally completed. If not, ME might cause further expansion of the

RGC projection within the LGN. Intraocular injections of CTB-AF488 were per-

formed to label retinal afferents subsequent to ME between P22-31 in WT (n = 6)

and KbDb−/− (n = 6) mice and the percent LGN area occupied by the ipsilateral

RGC projection was measured. There was no change observed in either genotype,

the developmentally-expanded ipsilateral eye inputs remained stable. This obser-

vation correlates with our finding that the expression of both H2-Kb and H2-Db is

developmentally downregulated by P34, and demonstrates that the enhanced OD

plasticity seen in visual cortex of mutant mice during the critical period is not due to

a reorganization of RGC projections following ME.

3.2.5 MHCI Immunostaining is associated with LGN synapses

and C1q

A direct role for H2-Kb and/or H2-Db in refinement of RGC projections would be sup-

ported by synaptic localization of these proteins. To determine if MHCI localizes near

synapses in vivo, Array Tomography (AT) [1] was used to examine immunostained

LGN sections at P7, at time when eye-specific layer formation is in progress and

LGN expression levels of C1q [61] and MHCI (Figure S1; [44]) are high. This method

permits immunostaining of the same ultrathin 70nm section repeatedly for known

synaptic markers, as well as for MHCI and other molecules of interest (Figure 3.4A).

Serial sections are then tomographically reassembled, rendering a 3-dimensional im-

age containing patterns of protein localization. In AT images, MHCI is localized

in a punctate pattern, often closely associated both with PSD-95 and/ or synapsin

puncta (Figure 3.4A), consistent with previous observations of hippocampal neurons

60

in vitro [55].

Higher resolution examination of 4 serial sections (Figure 3.4B) shows MHCI

puncta associated with excitatory and inhibitory synapses, as well as with C1q, a com-

plement protein recently also found to be required for RGC refinement and synapse

elimination [61]. To assess these relationships quantitatively, a cross-correlation pixel

analysis was performed (Figure 3.4C, D and Figure S4 online; Experimental proce-

dures). As expected, pre- (synapsin I) and post- (postsynaptic density-95) synaptic

markers are highly correlated with each other, while inhibitory (GAD) and excitatory

(vGluT2) markers are not correlated with each other since they are not associated

with similar synaptic types. Close association of MHCI molecules and C1q at synapses

is consistent with the observation here that the specific MHCI molecules H2-Kb and

H2-Db, like C1q, are needed for retinogeniculate synapse refinement.

3.3 Discussion

In mice lacking expression H2-Kb and H2-Db, retinal projections to the LGN fail to

refine completely, and within visual cortex, OD plasticity is enhanced. These changes

phenocopy those present in β2m−/−TAP1−/− mice, which lack stable cell-surface

expression of most of the 50+ MHCIs [54]. Because H2-Kb and H2-Db mRNA and

protein are present in neurons within LGN and visual cortex, we propose that these

specific classical MHCI family members are not only required for activity-dependent

refinement and plasticity in the visual system, but can account for the majority of

the abnormalities observed in the visual system of β2m−/−TAP1−/− mice.

61

3.3.1 MHCI function during developmental refinement of the

retinogeniculate projection

Many experiments have shown that when neural activity is blocked or altered, seg-

regation of RGC afferents from the two eyes within the dLGN is incomplete [28, 62].

Here we show that H2-Db and H2-Kb are expressed in the dLGN primarily during

the period of eye-specific segregation and that, in their absence, RGC afferents fail

to segregate fully. The Array Tomography data provides strong support for the pro-

posal that these MHCI proteins are located near or at synapses, but the resolution

of the method is still not sufficient to conclude that they are situated at the pre-

or the postsynaptic membrane. Here, we found a higher correlation between MHCI

and PSD-95 than between MHCI and synapsin, consistent with previous observa-

tions of immunostaining associated with neuronal dendrites in vivo [33, 56], as well

as with synapses and PDS-95 in vitro [55]. Together, these observations place MHCI

at the postsynaptic membrane, but this suggestion must await electron microscopy,

fix-insensitive MHCI antibodies and or biochemical fractionation for confirmation.

The failure of RGC axons from the two eyes to segregate completely from each

other even in the presence of intact retinal wave activity [44] implies H2-Kb and/or

H2-Db may function as molecular read-outs for activity-dependent synapse weaken-

ing and elimination. Other immune system molecules have also been implicated in

retinogeniculate synapse elimination. Like MHCI, neuronal pentraxins (e.g., NP1

and NP2), proteins homologous to immune system pentraxins, are upregulated by

neuronal activity [63]. In addition, the complement proteins C1q and C3 are present

in the dLGN during retinogeniculate refinement [61]. NP1−/−NP2−/− as well as

C1q−/− and C3−/− mutant mice have dLGN phenotypes strikingly similar to those

reported here. The localization of MHCI and C1q is highly correlated, implying that

these two molecules could interact and function together in developmental remodeling

62

at excitatory as well as inhibitory synapses. Note that PirB cannot be detected in

the LGN during the period of retinogeniculate remodeling [39], but CD3ζ, a signaling

component for other immune receptors, is present. CD3ζ mutant mice also have de-

fects in retinogeniculate refinement [44], implying that H2-Db and H2-Kb in the LGN

may collaborate with a CD3ζcontaining receptor.

The phenotypes observed in complement-deficient and in MHCI mutant mice are

stable and persist into adulthood. In contrast, the defect in NP1−/−NP2−/− is

transient [63]. Thus it may be that a series of tightly developmentally regulated events

operate sequentially to establish connectivity, then to stabilize and cluster glutamate

receptors (a process involving neuronal pentraxins [63] and finally to remodel and

eliminate synaptic inputs in an MHCI-C1q activity-dependent manner. The precise

mechanisms of how synapses are tagged for elimination based on their activity is far

from understood, but MHCI appears well-suited to act in this process, given that

action potential blockade downregulates both mRNA and protein [33,55].

3.3.2 H2-Kb and H2-Db may function with PirB to limit OD

Plasticity in Visual Cortex

The presence of enhanced OD plasticity in the visual cortex of mutant mice studied

here is notable. First, it argues for a role for H2-Kb and H2-Db in limiting the extent of

strengthening of the open (remaining) eye following an imbalance in activity created

by eye removal or closure. Just how the open eye is able to gain so much functional

territory in the visual cortex of KbDb−/ mice following visual deprivation remains to

be fully explored. Following ME in the mutant mice, we have shown that there is

a large mismatch in extent of expansion of the ipsilateral thalamocortical projection

as assessed using transneuronal tracing compared with that assessed using induction

of Arc mRNA. The expansion, as measured from anatomical tracing, is far less than

63

that measured functionally from the response of cortical neurons to visual stimulation,

suggesting that the enhanced OD plasticity in KbDb−/ mice may reflect not only the

presence of a larger pool of ipsilateral geniculocortical axons, but also changes within

the cortex. Future experiments will be required to elucidate how loss of these 2

MHCI family members alters the details of connectivity, as well as rules of synaptic

plasticity, at the cellular level.

H2-Kb and H2-Db join a very limited number of other molecules whose loss of func-

tion also results in enhanced, rather than diminished, OD plasticity. This small group

includes Nogo signaling (NgR−/−, NogoA/B−/− mice; [64] and PirB [39], as well as

infusion of tPA [65] which alters the extracellular matrix in cortex. PirB is expressed

on certain cells of the innate immune system and is a known receptor for H2-Kb and

H2-Db [43]. PirB is expressed in subsets of neurons including many pyramidal neurons

of the cerebral cortex (but notably is not detected in LGN neurons) and PirB has

recently been shown to bind Nogo [66]. Mice lacking PirB have enhanced OD plastic-

ity [39] almost identical to that seen here in KbDb−/ mice. Together these observations

suggest that PirB functions as a neuronal receptor for H2-Kb and H2-Db, possibly even

in collaboration with NgR/Nogo. In cultured cortical neurons, PirB immunostaining

is associated with axonal growth cones and is located near synapses [39], while MHCI

immunostaining is present in neuronal dendrites and colocalizes with PSD-95 [55],

implying a model in which postsynaptic MHCI interacts across the synapse with

presynaptic PirB [67]. Thus, MHCI-PirB signaling may operate either in parallel or

in conjunction with these other molecules to regulate negatively the strength and

stability of synaptic connections in an activity-dependent manner.

64

3.4 Experimental Procedures

3.4.1 Animals and Genotyping of mouse lines

The Institutional Animal Care and Use Committees at Harvard Medical School and

Stanford University approved all protocols. Animals were raised in a pathogen-free

facility, all mutant mice were outwardly normal. β2m−/−TAP1−/− mice were ob-

tained from D. Raulet [68]. Wildtype (β2m+/+TAP1+/+) and singly mutant mice

were derived by backcrossing β2m−/−TAP1−/− to the same background strain of

C57BL/6 wildtype mice purchased from Charles River (Wilmington, MA) for more

than five generations to obtain mice carrying the wildtype alleles for both β2m and

TAP1 genes. To avoid genetic drift, homozygous parents in each of these four colonies

are obtained from a common breeding colony containing mixed heterozygote geno-

types. Genotyping of β2m and TAP1 alleles was performed by PCR as described

previously [44]. KbDb−/− mice on a C57BL/6 genetic background were obtained

from H. Ploegh [37] and maintained as a homozygous breeding colony. Age-matched

C57BL/6 controls were purchased from Charles River (Wilmington, MA).

3.4.2 Mouse surgery and OD plasticity experiments

For monocular enucleation (ME) experiments to assess OD plasticity mice were anes-

thetized at P22 with isofluorane, one eye was removed (if needed sterile gelfoam was

inserted in the orbit to minimize bleeding). Eyelids were trimmed and sutured with

6-0 sterile surgical silk. A drop of Vetbond (3M, St. Paul, MN) was put on sutured

eyelids to prevent reopening. For monocular deprivation (MD) experiments, mice

were anesthetized at P25 and the procedure was identical as described above except

the eye was not removed until the day prior to Arc induction.

65

3.4.3 Arc induction

At P31 (for ME) or at P34 (for MD) one eye was removed under anesthesia (unless ME

had already been performed at an earlier age, as in OD plasticity experiments); mice

were revived and put in total darkness for 8 - 12 hours to minimize basal levels of Arc

mRNA in visual cortex. Mice were returned to a lighted environment for 30 minutes

to induce Arc mRNA in the cortex driven by vision through the remaining eye.

After light exposure, mice were euthanized with Halothane (Halocarbon, River Edge

NJ), brains were removed, flash-frozen in M-1 embedding media (Thermo Scientific,

Waltham, MA) and 14µm thick coronal sections were processed for situ hybridization

with Arc antisense riboprobe [39, 47]. Arc plasmid was provided by Dr. P. Worley,

Johns Hopkins University, Baltimore, MD.

3.4.4 Densitometric scans of Arc induction in specific corti-

cal layers

Quantitative analysis of Arc induction by stimulation of the ipsilateral eye was per-

formed in MATLAB (Mathworks, Inc, Natick, MA) by line scans across layer 4 of

primary and secondary visual cortex as described previously [47]: for each animal the

width of Arc mRNA signal two standard deviations above background was measured

in 3 - 4 sections scanned in randomized order, blind to genotype and manipulation.

Slides from different animals and manipulations were interleaved and only reassembled

once all width measurements were computed. Between 7-28 animals of each genotype

were studied; average widths of Arc induction were computed for each animal and

displayed in cumulative histograms.

66

3.4.5 Transneuronal labeling

To visualize the pattern of geniculocortical projections to layer 4 of mouse visual

cortex 1-2 µl of L-[2,3,4,5-3H]-proline (GE Healthcare, Cat#TRK750; approx. 100

Ci/mmol) was injected intraocularly at a concentration of 150-200µCi/µl dissolved

in 0.1M PBS pH 7.4; a week later mice were euthanized with euthasol 50mg/kg,

brains were frozen in M-1 mounting medium (ThermoShandon) and 14 µm coronal

cryosections were prepared on glass superfrost Plus slides (Fisher scientific). Sections

were fixed in 4% paraformaldehyde in PBS, pH 7.4, washed twice in 0.1M PBS and

dehydrated through a graded ethanol series. Sections were coated with NTB-2 emul-

sion (Kodak, Inc.), dried in dark room, and then stored at 4◦C. After 2-3 months,

slides were developed and imaged using dark field microscopy. Width measurements

of transported radioactive signal were measured in MATLAB (Mathworks, Inc) by

making line scans across layer 4 of primary and secondary visual cortex, as described

for Arc in situ measurements above and as used previously [47].

3.4.6 Anterograde labeling of retinal ganglion axons and mul-

tiple threshold analysis

P31-34 mice were anesthetized with isofluorane (in the case of LGN plasticity ex-

periments ME was performed from P22-31) then 1 - 2µl of cholera toxin B (CTB)

subunit conjugated to AF488 was injected in the right eye and AF594 in the left

(1mg/mL dissolved in 0.2% DMSO in nuclease-free water; Invitrogen, Inc. Carlsbad,

CA). After 24 hours, animals were overdosed with euthasol (50mg/kg), brains fixed

by transcardial perfusion of 0.1M PBS then ice-cold 4% PFA in 0.1M PBS. Brains

were removed, postfixed overnight in 4% PFA in 0.1M PBS at 4◦C, then sectioned

coronally using a vibratome at 100µm. Sections were mounted with prolong antifade

67

gold media (Invitrogen, Inc. Carlsbad, CA) on glass sides, and coverslipped for imag-

ing on a Zeiss LSM 510 META confocal microscope (Carl Zeiss MicroImaging, Inc.

Thornwood, NY)

All analysis was performed blind to genotype. To minimize variability analysis

was performed on dLGN sections where the ipsilateral projection area was the largest,

typically at the middle of the rostral-caudal extent of the dLGN (Figure S3B online).

All images were acquired such that the peak intensity values were just below saturat-

ing and multiple threshold analysis was carried out for the series of signal thresholds,

described previously [60]. In brief, dLGN sections were imaged in the red and then

green channel, and by varying each ipsi and contra channel at each intensity threshold

of 20%, 40%, 60%, 80% and 100% of maximum. To obtain overlap measurements,

the amount of overlapping red and green pixels in LGNs in both hemispheres was

measured in Image J (NIH, Bethesda, MD) using the Colocalization plugin tool and

displayed as yellow pixels (Figure 3.3A lower panel, overlap). The total area of over-

lapping pixels was represented as a percentage of total dLGN area (Figure 3.3C).

3.4.7 Array Tomography

Two postnatal day (P)7 mice, were perfused intracardially with 0.1M PBS, followed

by 4% paraformaldehyde in 0.1M PBS and the tissue was processed for array to-

mography [1]. The LGN was dissected out, further fixed in the same fixative using

microwave irradiation (PELCO 3451 laboratory microwave system; Ted Pella), then

dehydrated in ethanol and embedded in LRWhite resin (medium grade, SPI). Serial

ultrathin sections (70 nm) were cut on an ultramicrotome (Leica), mounted on subbed

coverslips and immunostained using either of two MHCI antibodies that yielded sim-

ilar AT staining patterns (Ox18, 1:50 AbDSerotec Cat#MCA51G or ErHr52, 1:10

BMA Cat#T2105). For secondary antibodies, Alexa 488, Alexa 594 and Alexa 647

(Invitrogen, 1:150) from the appropriate species were used. Up to 3 antibodies from

68

different hosts were applied together and imaged, followed by an antibody elution.

Sections were then restained with a different set of antibodies and re-imaged. Other

antibodies included those against synaptic proteins: synapsin I (rabbit, Millipore

AB1543P, 1:100), PSD-95 (mouse, NeuroMabs 75-028, 1:100), GAD65/67 (rabbit,

Millipore AB1511, 1:300), vGluT2 (guinea pig, Millipore AB2251, 1:1000), as well as

a C1q antibody (goat, Quidel A301, 1:300; ref. 31). Sections were mounted using

SlowFade Gold antifade reagent with DAPI (Invitrogen). Imaging was done on a Zeiss

AxioImager.Z1 fluorescence microscope with AxioCam HRm CCD camera, using a

Zeiss 63x/1.4 NA Plan Apochromat objective. Images were aligned using ImageJ and

the Multistack Reg plugin.

3.4.8 Array Tomography Cross-Correlation Analysis of synap-

tic markers, MHCI, and C1q

To examine the spatial relationship between known synaptic markers and MHCI

and C1q, we used a cross-correlation analysis method similar to that described pre-

viously [21]. For each pair of channels analyzed, a cross-correlation score Si was

computed over a range of lateral offset distances for images in the two channels.

Differences in the mean brightness in different channels were corrected by repeating

the analysis with one channel transposed (and therefore uncorrelated) to obtain a

baseline score St. To ensure that only labeled molecules in the synaptic neuropil

were included, the DAPI channel was used as a mask to remove nuclear and somatic

staining from the analysis. Si/St = Cr, describing the correlation (Cr) between two

channels as a multiple of their baseline correlation. A Cr of 1 indicates no correla-

tion, Cr >> 1 indicates high correlation, and Cr < 1 indicates negative correlation.

This method of analysis enables one to directly compare the correlation of different

immunolabels in a channel-independent manner (Figure 3.4D).

69

3.4.9 Statistical analyses

Statistical comparisons were performed using Excel (Microsoft Corp., Redmond, WA).

Tests on independent groups were Mann Whitney U test for comparisons of cumula-

tive distributions (widths of Arc mRNA signal or radioactively labeled thalamocorti-

cal terminals in layer 4) with population samples of unequal size, two-tailed Students

t test for group comparisons of LGN areas, two-way ANOVA for LGN pixel overlap

across multiple thresholds, and one-way ANOVA for Plasticity Index comparisons.

3.4.10 Supplemental Data

Supplemental Data include four figures and Supplemental Experimental Procedures

and can be found with this article online at http://www.cell.com/neuron/S0896-

6273(09)00844-7.

3.4.11 Acknowledgements

We thank members of the Shatz lab for helpful suggestions and comments. For tech-

nical assistance we thank B. Printseva, M. Marcotrigiano, J. Neville-Golden, and P.

Kemper. β2m−/−TAP1−/− mice were a gift from D. Raulet (UC Berkeley, Berkeley

CA) and KbDb−/− mice from H. Ploegh (MIT, Cambridge MA). We thank Dr. Beth

Stevens (Harvard Medical School, Boston MA) for providing the C1q antibody and

helpful discussions on using Array Tomography. Thanks also to Matthew Priestley

and Nay Lui Saw of the Stanford Institute for Neuroinnovation and Translational

Neuroscience NeuroBehavior Core Faci lity. This work was supported by NIH R01

EY02858, NIH R01 MH071666, the G. Harold and Leila Y. Mathers Charitable Foun-

dation, the Dana Foundation (CJS & AD), and NIH T32CA09361 (MJM).

70

Figure 3.1: Enhanced ocular dominance plasticity in visual cortex of KbDb−/−

mutant mice (A) Schematic of mouse visual system. Binocular zone (BZ), locatedbetween primary (V1) and secondary (V2) visual cortex, receives input from botheyes. Arc mRNA induction was used to map cortical neurons driven by stimulationof one eye (Experimental procedures, [47]). Below: darkfield autoradiograph of Arc insitu hybridization in a coronal section from a P34 WT (KbDb+/+) mouse; 30 minutesof visual stimulation upregulates Arc mRNA in neurons of layers 2-4 and 6 of visualcortex (layer 5 neurons only express very low levels of Arc mRNA). Box indicatesregion of Arc mRNA upregulation driven by ipsilateral eye stimulation within the BZ;broad induction of Arc mRNA is present throughout V1 and V2 contralateral to thestimulated eye, Scale = 900µm. (B) Ipsilateral eye representation in cortex expandsmore in KbDb−/− than in WT (KbDb+/+) mice following monocular enucleation (ME)during the critical period (P22-31). Top: In situ hybridization for Arc mRNA inKbDb+/+ (upper) and KbDb−/− (lower) visual cortex ipsilateral to the remaining eye;arrows indicate borders of signal in layer 4. Below: cumulative histograms of meanwidth of Arc induction in layer 4 ± sem for KbDb+/+ (upper) and KbDb−/− (lower)mice reared with normal vision (open symbols) or mice that received ME (filledsymbols). Note increased width of Arc induction following ME in both genotypes.(KbDb+/++ME: 1426.3 ± 89.7 µm, n = 18 mice vs. KbDb+/+ normal vision: 945± 58.3 µm, n = 2 8 mice; P < 0.05; KbDb−/−+ME: 2404.6 ± 105.1 µm, n = 25mice; P = 0.0017). Note also that width of Arc induction in KbDb−/− mutant micereared with normal visual experience (open squares) is slightly larger than that ofnormally-reared KbDb+/+ mice (open circles): KbDb−/−: 1111 ± 50.8 µm, n = 25mice; KbDb+/+: 945 ± 58.3 µm n = 2 8 mice. Each symbol represents the average ofseveral scanned sections from a single animal ± sem. Scale = 400µm. (C) Averagewidth of Arc induction in layer 4 for normally reared KbDb+/+ or KbDb−/− mice (openbar) vs. mice receiving monocular visual deprivation from P25-34 (MD: gray bar, n= 7) or from P22-31 (ME: black bar, n = 25). (D) Plasticity Index (see text) revealsgreater OD plasticity in KbDb−/− than in KbDb+/+ visual cortex by MD and ME, (*)statistical significance determined by one-way ANOVA. Error bars = standard errorof mean (sem) in B and C, and root mean square error (RMSE) in D.

71

Figure 3.1

72

Figure 3.2: Enhanced thalamocortical plasticity in KbDb−/− mutant mice (A)Schematic of connections in mouse visual system: thalamocortical axon terminalsinnervate layer 4 of cortex. Below, transneuronal transport of H3proline followingintraocular injection reveals (age P34; Experimental procedures) a broad contralateralsignal (left), and smaller ipsilateral patch (right) in layer 4. Scale=1500µm. (B)Higher magnification of representative sections from KbDb+/+ (top) and KbDb−/−

(bottom) mice. Arrows indicate measurement borders used. Scale = 450µm. (C)Population averages for width ± sem of layer 4 label for each genotype after ME.P<0.05; (Mann-Whitney U test).

73

Figure 3.2

74

Figure 3.3: Incomplete segregation of RGC inputs to dLGN in KbDb−/− mu-tant mice (A) RGC projections labeled by intraocular injections of CTB AF594(red) into the contralateral (contra) eye and CTB AF488 (green) into the ipsilat-eral (ipsi) eye. Top: Merged fluorescent micrographs of dLGN from KbDb+/+ andKbDb−/− mice at P34. Middle (green pixels): ipsilateral eye projection pattern (in-tensity threshold = 60% of maximum). Bottom (yellow pixels): overlapping pixelsfrom ipsilateral and contralateral eye projections each at an intensity threshold 60%of maximum (Figure S3 and Experimental procedures). Note ectopic patches of ipsi-lateral eye projections not eliminated during development in KbDb−/− mice. Scale =150µm. (B, C) Mean % of dLGN area ± sem of ipsilateral eye projection and meanoverlapping pixels in both channels. More ipsilateral territory as well as overlappingpixels are present in mutant dLGN: KbDb+/+ = 5 mice, KbDb−/− = 5 mice, (*) P <0.05 (two-tailed t-test).

75

Figure 3.3

76

Figure 3.4: MHCI localization in relation to synaptic proteins during pe-riod of retinogeniculate refinement (A) AT micrographs reconstructed from 25serial ultrathin sections (each 70nm thick) of P7 LGN showing MHCI immunos-taining (green) in relation to known synaptic markers (Syn=synapsin (orange),PSD=postsynaptic density-95 (white), GAD=glutamic acid decarboxylase 65/67(cyan)), as well as to C1q=complement protein C1q (magenta). DAPI stain of nuclei,blue. Scale = 5µm. (B) Four serial sections of 2 different synapsin positive puncta:Left example is characteristic of an excitatory synapse (close apposition of presy-naptic marker synapsin with postsynaptic excitatory synapse marker PSD-95); Rightexample is characteristic of an inhibitory synapse (overlap of synapsin with GADand absence of PSD-95). MHCI and C1q are closely associated with both types ofsynapses. (C) Single ultrathin section, showing colocalization between MHCI punctaand PSD, Synapsin, GAD and C1q. Bottom right, zoomed in view of puncta num-bered on left, showing immunofluoresence signal in separate channels, Scale = 2µm.Note colocalization of signal for MHCI, PSD and C1q at puncta #1 and 2. (D) Crosscorrelations showing pairwise comparisons of degree of spatial overlap between punctaimmunostained for the various markers (Experimental procedures and Figure S4 on-line). Synapsin vs. PSD95 shows strongest correlation, and GAD vs. vGluT2 theweakest. MHCI is more correlated with C1q than with other markers. Si/St = Cr,the correlation ratio between two channels as a multiple of their baseline correlation.Cr = 1 indicates no correlation, Cr >> 1 indicates high correlation, Cr < 1 indicatesnegative correlation.

77

Figure 3.4

78

Chapter 4

Single-Synapse Analysis of a

Diverse Synapse Population:

Proteomic Imaging Methods and

Markers

Micheva KD, Busse B, Weiler NC, O’Rourke N, Smith SJ. Single-synapse analysis

of a diverse synapse population: proteomic imaging methods and markers. Neuron.

2010 Nov 18;68(4):639-53.

4.0.12 Abstract

A lack of methods for measuring the protein compositions of individual synapses in

situ has so far hindered the exploration and exploitation of synapse molecular diver-

sity. Here, we describe the use of array tomography, a new high-resolution proteomic

imaging method, to determine the composition of glutamate and GABA synapses in

somatosensory cortex of Line-H-YFP Thy-1 transgenic mice. We find that virtually

79

all synapses are recognized by antibodies to the presynaptic phosphoprotein synapsin

I, while antibodies to 16 other synaptic proteins discriminate among 4 subtypes of

glutamatergic synapses and GABAergic synapses. Cell-specific YFP expression in the

YFP-H mouse line allows synapses to be assigned to specific presynaptic and postsy-

naptic partners and reveals that a subpopulation of spines on layer 5 pyramidal cells

receives both VGluT1-subtype glutamatergic and GABAergic synaptic inputs. These

results establish a means for the high-throughput acquisition of proteomic data from

individual cortical synapses in situ.

4.1 Introduction

Rapidly accumulating physiological and genetic evidence establishes that the molec-

ular diversity of synapses extends far beyond that envisioned by traditional classi-

fication schemes based solely on neurotransmitter identity. For instance, it is now

clear that within each neurotransmitter category (e.g., glutamatergic, GABAergic,

cholinergic) there is substantial diversity in the expression of many intrinsic synaptic

proteins, including neurotransmitter transporters and receptors [7–16]. Until synapse

molecular diversity is properly fathomed, it is likely to be a troublesome source of

variability in physiological and neurodevelopmental experimentation. Conversely, a

systematic understanding of synapse diversity (i.e., the synaptome) is likely to provide

valuable new perspectives on the organization of synaptic circuitry (i.e., the connec-

tome), its development, plasticity and disorders. It is easy to envision, for instance,

that a potential catalog of molecular synapse types [4] would help explorations of

the synaptic basis of specific memory or disease processes to focus more fruitfully on

specific synapse subpopulations.

To place a possible molecular catalog of synapse types on a firm footing, two

broad experimental challenges remain. First, it is essential that synapse populations

80

be explored at the single-synapse level. Until recently, the only way to reliably resolve

and characterize individual synapses was by way of electron microscopy (EM). While

traditionally a time-consuming and very volume-limited method, recent advances in

EM [72–76] have greatly improved its throughput, even offering the possibility of

detailed neuronal circuit reconstruction. Nonetheless, EM still provides only very

limited proteomic discrimination (although Anderson et al. [75] describe a very pow-

erful new approach to integrating small-molecule discrimination with EM). Second,

synapse diversity must be explored in situ, in ways that retain full fidelity to the

intact tissue setting and allow for the acquisition of as much information as possible

about circuit context and cellular morphology.

Array tomography (AT) is a high-resolution proteomic imaging method [1, 17]

that exploits a combination of light and EM approaches to resolve fine details at the

level of synapses across large fields of view spanning entire circuits. Of prime sig-

nificance to the present application, AT allows the immunofluorescence resolution of

single synapses within cortical neuropil, where such resolution is highly problematic

for other optical methods. Additionally, AT can acquire many more dimensions of

immunofluorescence information about single synapses than previous methods (up

to 17 in the present work, as compared to the standard immunofluorescence limit of

three or four). AT also benefits from greatly improved quantitative reliability, since

both staining and imaging are completely independent of depth within a tissue sam-

ple. Finally, AT delivers very high experimental throughput: our present automated

methods acquire image data at a rate of approximately one million synaptic protein

puncta per hour. Such throughput will help advance the analysis of synaptic diversity

from the anecdote to the realm of solid bioinformatics. AT thus seems uniquely suited

to meet the challenges of exploring the molecular diversity of cortical synapses.

Here, we describe array tomographic immunofluorescence methods for the single-

synapse analysis of mouse cortex, focusing on the discrimination and analysis of

81

glutamatergic and GABAergic synapses. Toward a goal of identifying every single

cortical synapse as unambiguously as possible, we evaluated antibody markers to

presynaptic proteins likely to be common to all synapses, such as synaptophysin,

bassoon, and synapsin. We find that antibodies to the presynaptic phosphoprotein

synapsin I [23, 24] are particularly robust and useful, labeling the vast majority of

cortical synapses with a minimum of labeling at nonsynaptic loci. For increased

confidence in synapse identification, we also develop here a basis for conjoint use of

multiple synaptic markers. We argue that antibodies to the glutamatergic synaptic

proteins VGluT1, VGluT2, PSD-95, GluR2, NMDAR1, and the GABAergic synaptic

proteins GAD, VGAT, and gephyrin can be used both to distinguish reliably between

glutamatergic and GABAergic synapses and begin the work of searching for finer

synapse molecular subtypes within these broad categories.

4.2 Results

A Note about Color Use

We have adopted a colorblind-friendly scheme in as many figures as possible. In

figures with only two immunofluorescence channels (Figures 4.3, 4.6, 4.7, 4.8C–D),

we use magenta and green as additive colors, such that regions of overlap display as

white. When three or more channels need to be displayed (Figures 4.1, 4.2, 4.4, 4.8A–

B), we represent each channel with a nontransparent color plane. In this case, colors

are nonadditive, for example, white in such figures is not the result of magenta and

green overlap but rather a distinct color representing a distinct immunofluorescence

channel.

82

4.2.1 AT Resolves Individual Puncta of Multiple Synaptic

Proteins in Mouse Cortex

Figure 4.1 offers a panoramic view of a volume of somatosensory cortex from a YFP-

H Thy-1 transgenic mouse [71] representative of the specimens used in the present

work. The volume image rendered here was acquired by automated AT imaging

of a mosaic of 52 fields per section over 60 serial sections (200 nm each) in three

fluorescence spectral channels, comprising a total of 9360 individual image tiles. The

tiles were stitched and aligned in three dimensions and rendered as described in

Experimental Procedures. This volume is 12 µm thick, 0.5 mm wide and extends a

distance of 1.4 mm from the pial surface of the cortex through all cortical layers past

the subcortical white matter and into a portion of the underlying striatum. The three

fluorescence channels represented are YFP fluorescence (green), anti-tubulin (blue),

and anti-synapsin I (magenta) immunofluorescence. The vast information content

of the volume presented in Figure 4.1 is better appreciated from dynamic volume

renderings as in Movie S1 in the Supplemental Information available online.

YFP fluorescence in the cortex of line H mouse represents a soluble YFP marker

transgenically expressed in a large subset of layer 5 pyramidal neurons. This mouse

line was used because the YFP-expressing neurons provide a useful anatomical frame-

work, with their apical dendrites extending all the way to layer 1 and their axons

forming conspicuous bundles in the white matter. However, YFP fluorescence is not

necessary for the subsequent single synapse analysis, which can be performed also

in a wild-type mouse. In addition to the YFP labeled neurons, the apical dendrites

of pyramidal cells not expressing YFP are evident from tubulin immunostaining of

their core microtubule bundles (Figure 4.1D). Finally, the presence of aggregates

of synapsin I protein in the neuropil is apparent from the magenta puncta, which

can be individually discerned (Figure 4.1C–H). The spatial distribution of synapsin

83

immunoreactivity is consistent with that expected of neocortical synapses. For ex-

ample, cell bodies appear as circular spaces largely devoid of synapsin puncta. The

few puncta observed within such voids are actually situated in front or behind the

cell bodies, as the depth of this volume is larger than the diameters of most cortical

cell bodies. There is no staining in blood vessels and very few puncta in the white

matter (Figure 4.1H). All of the data that follow were collected from arrays similar

to that represented in Figure 4.1, but image acquisition was, for expediency’s sake,

carried out in single fields of view in layers 4 and 5, corresponding to the areas in

Figure 4.1E and 4.1F, respectively. Thinner, 70 nm sections were used to increase z

resolution and the sampling of each synapse.

While the distribution of synapsin puncta resembles that of cortical synapses, it

is not clear whether all synapsin puncta represent synapses and whether all synapses

are immunoreactive for synapsin. Synapsin I is highly concentrated in presynap-

tic boutons [23] and has been used extensively as a general synaptic marker, but

a one-to-one relationship between synapsin puncta and synapses has not yet been

demonstrated. Therefore, it cannot be assumed that synapsin immunofluorescence

data alone are sufficient for synapse identification. Also, there are other proteins, for

example synaptophysin and bassoon, which are highly concentrated at presynaptic

boutons and could be useful as general markers for synapses. To evaluate candidate

cortical synapse markers, we developed a panel of antibodies that label a variety

of pre- and postsynaptic proteins (Table 4.1; see Supplemental Experimental Proce-

dures for antibody characterization). Because of the proteomic capability of AT to

immunostain sections multiple times with different sets of antibodies, we were able to

measure numerous pre- and postsynaptic markers at every putative synaptic locus.

An example of multiple antibody labeling is shown in Figure 4.2, which represents

a volume rendering from layer 4 of the somatosensory cortex of a YFP-H mouse

84

immunostained with 10 different antibodies against synaptic proteins (synapsin, bas-

soon, VGluT1, VGluT2, PSD-95, GluR2, NMDAR1, GAD, VGAT, gephyrin) and

one against tubulin. The sequence of antibody application is presented in Table S1

(data set KDM-SYN-090416). In addition, two other fluorescent labels (YFP and

DAPI) were imaged, making a total of 13 fluorescent channels collected from each

section.

4.2.2 Synaptic Protein Distributions Imaged by AT Corre-

late as Expected from Synapse Structure

Some of the antibody markers in Table 4.1 are expected to be present at all synapses

in cortical neuropil, while others are specific for particular synapse subtypes. For ex-

ample, as universal presynaptic proteins, synapsin and synaptophysin puncta overlap

(Figure 4.3A). The majority of synapsin puncta also overlap with VGluT1, known to

be present in most cortical glutamatergic synapses. In addition, synapsin puncta are

closely apposed with PSD-95 puncta as would be anticipated from imaging pre- and

postsynaptic proteins at single synapses. GAD puncta, expected to label inhibitory

GABAergic synapses, overlap with a small subset of synapsin puncta, and synapsin

levels are generally lower in these synapses.

To quantify globally the extent of spatial correlation among various synaptic

marker candidates, we designed a correlation matrix test based on the van Steensel

method [21]; see Experimental Procedures for full description. The basic idea is to

test for the effect of very small relative displacements between pairs of marker images

on a measurement of image overlap. Because of the abundance of many synaptic

markers, overlapping spatial distributions might occur by chance. If the association

between two channels is real, however, then any shift of one channel relative to the

other will decrease the observed degree of colocalization. On the other hand, if two

85

channels tend to be mutually exclusive, then a shift will increase the degree of colo-

calization. Finally, if the association between two channels is occurring by chance,

then a shift will not substantially affect the degree of colocalization. Using a 20 x

20 x 6.3 µm3 volume of neuropil from data set KDM-SYN-091207 (Table S1), we

computed a cross-correlation score for pairs of channels over a range of lateral offset

distances. From the 17 antibodies used in this dataset, we focused on the general

presynaptic markers synapsin, synaptophysin and bassoon, as well as several specific

markers for glutamatergic (VGluT1, VGluT2, PSD-95, and GluR2) and GABAergic

synapses (GAD and VGAT).

The cross-correlation score is represented in Figure 4.3B as a grid of false col-

ored squares with centers corresponding to the score at 0 offset and each pixel shift

equal to 0.1 µm offset. To visualize the data, different channel pairs are also shown

as immunofluorescent images from a small area of a single section of the same data

set. As can be seen in the correlation matrix, both synapsin and synaptophysin,

and to a lesser extent bassoon, colocalize with all other synaptic markers, includ-

ing those of smaller subsets of synapses that contain VGluT2 or GAD. All synaptic

markers are anticorrelated with tubulin, which labels microtubules within dendrites

and cell bodies. VGluT1 and VGluT2, found in cortical glutamatergic synapses, do

not colocalize with the GABAergic markers. PSD-95 and GluR2, both present at

the postsynaptic side of glutamatergic synapses, correlate strongly with each other

and more weakly with the presynaptic glutamatergic markers. GAD and VGAT,

presynaptic markers for GABAergic synapses, show strong correlation. An interest-

ing distinction can be made between the presynaptic markers with respect to their

colocalization with postsynaptic markers. Presynaptic markers that are associated

with synaptic vesicles (e.g., synapsin, synaptophysin, VGluTs) show high colocaliza-

tion among themselves, while their colocalization with postsynaptic markers such as

PSD-95 and GluR2 is weaker. On the other hand, the presynaptic marker bassoon,

86

which labels the presynaptic active zone, shows similar colocalization with both pre-

and postsynaptic markers. This is due to the fact that the synaptic vesicle cluster is

situated far enough from the postsynaptic density to be resolved by AT. On the other

hand, the presynaptic active zone is only one synaptic cleft (around 20 nm) away

from the postsynaptic density which is below the resolution capabilities of AT. For

example, in single section images in Figure 4.3B, synapsin puncta are seen next to

PSD-95 and GluR2 puncta, while bassoon overlaps with these postsynaptic markers.

4.2.3 AT Immunofluorescence of Synapsin Is Highly Reliable

as Synapse Marker

A single marker protein detectable at all synapses and only at synapses would be

very useful for many purposes, but thus far there has been no conclusive demonstra-

tion of any such marker. While numerous markers, e.g., intrinsic proteins of synaptic

vesicles, might be localized at every chemical synapse, the usefulness of any such

antibody marker would be diminished if it were found at nonsynaptic loci as well.

From the colocalization matrix of Figure 4.3B, it is evident that both synapsin and

synaptophysin colocalize strongly with all other synaptic markers and thus might be

useful as general markers for synapses. Further examination of the immunofluores-

cence images revealed, however, that synaptophysin immunoreactivity is also fairly

often detectable at obviously extrasynaptic sites, e.g., in cell body and dendritic cy-

toplasm and nuclei (Figure 4.3A). Synaptophysin puncta moreover tend to be smaller

and less continuous than synapsin puncta. For these reasons, the synapsin I antibody

appeared to be the stronger candidate as a reliable synaptic marker and was subjected

to further evaluation.

87

4.2.4 Synapsin Is Detectable at Virtually All Dendritic Spines

Almost all dendritic spines in adult cortex receive synapses and therefore a general

synaptic marker should be present at these sites. To determine the distribution of

synapsin puncta at spines, we reconstructed the apical dendrites of YFP-positive

layer 5 pyramidal cells extending through layer 4 in tissue that was immunostained

for both pre- and postsynaptic proteins (Figure 4.4). Immunofluorescence reveals

PSD-95 puncta within spine heads that are closely associated with both synapsin

and bassoon puncta. Two dendritic segments from data set KDM-SYN-091207 were

used to quantify the number of spines contacted by synapsin puncta. Only synaptic

marker immunofluorescence within 0.5 µm of the YFP dendrites was considered for

this analysis. One of the dendritic segments was 45 µm in length, 2 µm in width

and had 131 spines (2.9 spines/µm). From the 116 spines included in their entirety

within the imaged volume, 114 had at least one synapsin punctum associated with

them. The other dendritic segment was 93 µm long, 1.7 µm wide, and had 117 spines

(1.3 spines/µm), from which 110 were completely included in the volume. All of

these spines were associated with a synapsin punctum. Thus, more than 99% of the

dendritic spines on layer 5 pyramidal neurons were in the immediate vicinity of a

synapsin punctum. Moreover, other pre- and postsynaptic proteins colocalized with

the synapsin puncta at these dendritic spines (100% of synapsin puncta with pre- and

98% with postsynaptic markers).

4.2.5 EM Analysis Supports the Identification of Synapses

with Synapsin Immunoreactivity

The use of the synapsin antibody as a general synaptic marker was also assessed at

the EM level. Postembedding immunoEM using synapsin and a secondary antibody

conjugated to colloidal gold (15 nm) labeled presynaptic boutons, as identified by

88

the presence of synaptic vesicles and adjacent postsynaptic densities (Figure 4.5C).

The relatively low density of immunogold labeling is likely due to the addition of

0.1% glutaraldehyde and 0.1% OsO4 during tissue preparation, which is necessary

for ultrastructural preservation but significantly impairs immunogenicity. Indeed,

synapsin immunofluorescence on sections from tissue treated this way is much weaker

than on our conventional sections used for the rest of this study (Figure S2).

We then compared synapsin immunofluorescence with the corresponding ultra-

structure to assess what proportion of synapses identified at the EM level are fluores-

cently labeled. Because the tissue preparation for EM significantly reduces synapsin

immunolabeling this analysis will result in an underestimate of the presence of synapsin

at synapses. Serial sections from tissue prepared for EM observation and mounted

on coverslips were first immunofluorescently labeled with the synapsin antibody and

imaged with the fluorescent microscope. The sections were then post-stained with

uranyl acetate and lead citrate and viewed in the SEM using the backscattered elec-

tron detector. The fluorescent and SEM images were aligned using the DAPI signal

and the nuclei as viewed in the SEM. The bright DAPI-stained puncta in the nuclei

correspond to the electron dense heterochromatin masses (Figure 4.5A). A compar-

ison of the ultrastructurally identified synapses and synapsin immunofluorescence

revealed that 91% of synapses (279 out of 305) were synapsin positive on at least

one section. The intensity of synapsin immunofluorescence did not correlate with the

size of the synapse as seen in the SEM. For example, some big presynaptic boutons

(asterisk on Figure 4.5B) were very weakly labeled. Thus, despite the reduced im-

munoreactivity with conjugate immunofluorescence-SEM imaging, 91% of synapses

were synapsin positive, which is consistent with synapsin being a reliable marker for

immunofluorescent imaging of cortical synapses.

89

4.2.6 Multiple Synaptic Proteins Can Be Visualized Volu-

metrically as a Synaptogram Mosaic

The large number of immunofluorescence stains used with AT presents a challenge for

visualization. The traditional color coding cannot be used for so many channels, and

volume reconstruction along any single axis can obscure weak labels or show false colo-

calization of markers. Therefore, we devised a representation for multichannel volu-

metric image data, called a synaptogram (Figure 4.6), that is useful for single-synapse

analysis. All possible synaptic loci are identified with synapsin immunostaining and

represented by single channel serial sections in a matrix where each section occupies

a column and each channel a row. Sections represent a 1 x 1 µm2 area centered on

the centroid of the synapsin punctum. With synaptograms many antibody labels

can be visualized simultaneously and spatial relationships among labeled structures

can be examined with precision and relative ease. For example, the synaptograms in

Figure 4.6 use 18 different fluorescent signals: the general synaptic markers synapsin,

synaptophysin, and bassoon (two different antibodies), eight glutamatergic markers,

four GABAergic markers, and two structural markers (data set KDM-SYN-091207;

Table S1). The glutamatergic synapse on the left has a distinct synaptogram appear-

ance compared to the GABAergic synapse on the right. Both synapses contain the

general synaptic markers synapsin, synaptophysin, and bassoon. The glutamatergic

synapse contains presynaptic VGluT1 as well as a number of postsynaptic scaffold

and receptor markers (PSD-95, MAGUK, GluR2, NMDA receptor subunits). The

GABAergic synapse is distinguished by the presence of presynaptic GAD and VGAT

as well as the postsynaptic scaffold protein gephyrin and GABAA receptor subunit.

Both of the synapses are adjacent to a YFP-positive process and it appears that the

glutamatergic synapse is making a contact with this process (postsynaptic markers

overlap with the YFP signal), while the GABAergic synapse is not (postsynaptic

90

markers away from the YFP process).

The synaptogram makes it easy to check for the continuity of a given marker

punctum from one serial section to the next, and the 3D colocalization of multiple

markers that would be expected at a true synapse. The synaptogram can also be

useful in identifying fluorescence signals that are clearly not synaptic structures, such

as staining artifacts, and excluding them from the analysis. For example, the presence

of a synaptic vesicle immunofluorescence signal in just one isolated section is unlikely

to originate from an actual synapse, because synaptic vesicle clusters almost always

have a minimum extent greater than the 70 nm thickness.

4.2.7 AT Imaging Discriminates Multiple Glutamatergic and

GABAergic Synapse Subtypes

To begin characterizing the diversity of cortical synapses we focused on a panel of 10

antibodies. Synapsin and bassoon were used as general synaptic markers. VGluT1,

VGluT2, PSD-95, GluR2, and NMDAR1 were used as markers for glutamatergic

synapses. The vesicular glutamate transporters VGluT1 and VGluT2 were included

because their expression reportedly varies depending on the intracortical or subcorti-

cal origins of the synapses. In particular it is believed that VGluT2 is predominantly

expressed in thalamocortical synapses [77–79]. GAD, VGAT, and gephyrin were used

as markers for GABAergic synapses. This combination of antibodies allowed the iden-

tification of two general types of synapses: glutamatergic and GABAergic, with the

glutamatergic synapses further subdivided into subtypes containing VGluT1, both

VGluT1 and 2, VGluT2, and other (lacking VGluT1 and 2) (Figure 4.7A).

Synapses formed by axons of layer 5 pyramidal neurons belonged to the glu-

tamatergic VGluT1 subtype (Figure 4.7D), as evidenced by examination of YFP-

positive presynaptic boutons. From 96 YFP synapses in layer 4 and 110 YFP synapses

91

in layer 5, all had associated VGluT1, not VGluT2, immunofluorescence.

To further evaluate the reliability of our approach to single-synapse analysis, we

calculated the proportion of synapses falling into each synaptic subtype from two

experiments performed on tissue sections from the same region of the same animal

but with a different order of antibody application (Table S1; data sets KDM-SYN-

090416 and KDM-SYN-091207). A cluster of glutamatergic markers was considered

to be a synapse only if both pre- and postsynaptic markers were present. The re-

quirement for the presence of pre- and postsynaptic markers was not applied to the

GABAergic synapses, because it is not yet known whether the postsynaptic scaffold

gephyrin is present at all synapses of this type [80]. Instead, we adopted the less strin-

gent criterion of colocalization of several presynaptic markers belonging to different

compartments (e.g., cytoplasmic, presynaptic active zone, and vesicular). Thus, in

addition to the presence of a ubiquitous synaptic marker (e.g., synapsin or bassoon)

and the cytoplasmic GAD, we required the presence of the vesicular marker VGAT,

which has been shown to specifically localize to presynaptic boutons [81]. The results

obtained from these two experiments were very similar (Table 4.2). Approximately

84% synapses were glutamatergic and 16% GABAergic. There were almost twice as

many VGluT2 containing synapses (VGluT1 + 2 subtype and VGluT2 only subtype)

in layer 4 compared to layer 5 (21.4% versus 12.9%). On average, only 4% of synapsin

puncta were not associated with other synaptic proteins and therefore most likely do

not represent synapses. The proportion of glutamatergic and GABAergic synapses,

as well as the preference of VGlut2 synapses for layer 4 are consistent with previous

studies [77, 79, 82]. These results, reproducible between two separate experiments,

confirm the reliability of single-synapse analysis with AT. It should be noted that

only one animal was analyzed quantitatively, and the results are aimed at evaluating

the technique, not providing reference proportions of synapse types in the mouse so-

matosensory cortex. This will, undoubtedly, require a larger number of animals and

92

is beyond the scope of the present study.

4.2.8 AMPA and NMDA Receptors Distributions Vary at

Different Glutamatergic Synapses

There is great variability in the expression levels, subunit composition, and local-

ization of AMPA and NMDA receptors at synapses, which significantly affects their

functional properties [11]. To further characterize cortical synapses based on the type

of postsynaptic receptors present (Figure 4.7C), 110 synapses were randomly selected

in each of layers 4 and 5 using synapsin immunostaining. Inhibitory synapses were

excluded from the analysis. The distribution of glutamatergic receptors was very

similar in both cortical layers. The great majority of excitatory synapses contained

both AMPA and NMDA receptors, as identified with antibodies against GluR2 and

NMDAR1 respectively (85.4% in layer 4 and 84.5% in layer 5). In 11.8% of synapses

in both layers only AMPA receptors could be detected and in 2.7% of synapses in

layer 4 and 3.6% in layer 5only NMDA receptors. GluR2 and, to a lesser extent,

NMDA labels often extended through more sections than PSD-95, or were observed

in sections adjacent to PSD labeling.

4.2.9 Synapsin Is Present at All Glutamatergic and GABAer-

gic Synapses, but in Varying Amounts

In addition to enabling the study of synaptic diversity, the establishment of markers

for synapse subtypes also allowed us to revisit the question of whether synapsin is

expressed in all cortical synapses. From the conjugate immunofluorescence-SEM anal-

ysis, it was observed that 91% of ultrastructurally identified synapses were labeled for

synapsin, but, as mentioned before, those were suboptimal conditions for immunos-

taining. To better understand what proportion of synapses are labeled with synapsin

93

in our tissue prepared for immunofluorescence, synapses were identified using different

combinations of pre- and postsynaptic markers, excluding synapsin. Glutamatergic

VGluT1 synapses were identified by the combination of VGluT1-PSD-95 antibodies,

VGluT2 synapses with VGluT2-PSD-95 antibodies and GABAergic synapses with

VGAT and gephyrin antibodies. One hundred synapses from each group were chosen

randomly and the synapsin immunofluorescence associated with them was measured

on all sections through the synapse. Synapsin immunofluorescence was above back-

ground levels in all analyzed synapses (Figure 4.7B). VGluT1 synapses contained the

highest average synapsin levels (151 ± 7 arbitrary units) compared to VGluT2 (111

± 7 a.u.) and VGAT synapses (81 ± 4 a.u.). A significant proportion of VGluT2

(17%) and VGAT synapses (12%) contained low levels of synapsin immunofluores-

cence (below 40 a.u.) compared to only 2% of the VGluT1 synapses.

These data further confirm that synapsin I can be used as a general synaptic

marker because it appears to be present in all mouse glutamatergic and GABAergic

cortical synapses. However, synapsin content as detected with immunofluorescence

varies depending on synapse type with some VGluT2 and VGAT synapses exhibiting

low levels of synapsin. Thus, applying a simple intensity threshold of synapsin im-

munofluorescence should be avoided because this can lead to underestimation of the

synapse subtypes exhibiting low synapsin levels. Double Innervated Spines on Layer

5 Pyramidal Neurons Are Contacted by a VGluT1 and a GABAergic Synapse

Double innervated dendritic spines are an intriguing synaptic arrangement in-

volved in cortical plasticity and are found in a variety of species and cortical areas.

Until now they have been observed only by EM [82–85]. Very little is known about

the identity of either the input or the target in this arrangement [86, 87] especially

in the mouse somatosensory cortex. AT can resolve neighboring synapses, as can

be seen on Figure 4.5B, where adjacent synapses imaged in SEM are represented

by separate immunofluorescent synapsin puncta. We therefore used AT to further

94

characterize dually innervated spines. Examples of dendritic spines receiving two

synaptic inputs, one glutamatergic and one GABAergic, are presented in Figure 4.8.

These spines emerge from the apical dendrites of YFP-positive layer five pyramidal

neurons. From two data sets (KDM-SYN-090416 and KDM-SYN-091207), 22 un-

equivocal double-innervated spines were identified, i.e., both pre- and postsynaptic

markers were present for the two inputs. Of those spines, 82% received a glutamater-

gic VGluT1 input and a GABAergic input. The remaining 18% received a gluta-

matergic VGluT2-containing (VGluT2 only, 9%, and VGluT1 and 2, 9%) as well as a

GABAergic input. The great majority of the glutamatergic synapses contained both

AMPA and NMDA receptors (91%), one spine was seen with NMDA receptors only,

and one spine had neither AMPA nor NMDA receptor markers present. Thus, layer

5 pyramidal neurons have apical dendrites that contain spines innervated by both

excitatory and inhibitory synapses, and the glutamatergic input to these spines is

predominantly VGluT1 positive.

4.3 Discussion

In the present study, we identify a synapsin I antibody as a reliable marker for cor-

tical synapses. Synapsin I associates exclusively with small synaptic vesicles and is

concentrated in presynaptic boutons [23]. The great majority of synapses are thought

to contain synapsin and this protein has been used extensively as a general synap-

tic marker. However, the list of possible exceptions to this rule has grown recently

and now includes ribbon synapses in the retina [88], reticulogeniculate synapses [89],

and some GABAergic- and VGluT2-containing synapses in the cerebral cortex [90].

AT, because of its increased sensitivity of immunofluorescence detection, ability to

analyze multiple antibody stains and use immunofluorescence in conjunction with

EM, allowed us to examine in detail the relationship between synapsin puncta and

95

synapses in the mouse cerebral cortex.

For the synapsin antibody to be relied upon as a universal synapse marker, the

following conditions had to be confirmed: (1) synapsin labels the great majority

of synapses; (2) synapsin background labeling is minimal and can be differentiated

from the labeling of synapses; and (3) synapsin imaging has the resolution for dis-

cerning adjacent synapses. This study presents multiple lines of evidence that a

rabbit polyclonal synapsin I antibody labels the vast majority of synapses in the

mouse somatosensory cortex. For example, conjugate SEM of sections stained and

imaged for synapsin immunofluorescence revealed that 91% of ultrastructurally iden-

tified synapses are immunofluorescently labeled on at least one section. Because the

preparation of the tissue for EM sharply decreases antigenicity, this is a conserva-

tive estimate of the number of synapsin positive synapses. In tissue prepared for

immunofluorescence, more than 99% of dendritic spines have an adjacent synapsin

punctum. Also, all glutamatergic and GABAergic synapses identified by different

combinations of pre- and postsynaptic markers contain synapsin label, albeit some

at low levels. The background synapsin immunofluorescence is very low and only

occasional puncta can be seen in cell body or large dendrite cytoplasm, or nuclei.

Very few synapsin puncta (around 4%) exist alone, away from other synaptic mark-

ers. Synapsin immunofluorescence colocalizes with other presynaptic markers such as

bassoon, VGluT1, VGluT2, VGAT, and GAD and is also found immediately adjacent

to postsynaptic markers such as PSD-95, GluR2, NMDAR1, and gephyrin. These re-

lationships were confirmed both at the synapse population level using a normalized

cross-correlation analysis and for individual synapses using synaptograms. Finally,

synapsin labeling with AT allows for the resolution of juxtaposed synapses as can be

seen from conjugate EM. At the light level, when multiple immunofluorescent labels

are used, adjacent synaptic puncta can be observed that colocalize with different sets

of pre- and postsynaptic antibodies and thus clearly belong to two different synapses.

96

The reliability of synapsin as a universal synaptic marker can be strengthened with

the concomitant use of multiple strategically chosen synaptic markers. This not only

helps the unequivocal identification of synapses but also allows the positive identifi-

cation of glutamatergic and GABAergic synapses within cortical neuropil and reveals

the existence of several synaptic subtypes within those broad categories. Based on

the presence of vesicular glutamate transporters, the glutamatergic synapses could be

divided into those containing only VGluT1, only VGluT2, both VGluT1 and 2, and

neither VGlut1 or 2. Synapsin immunofluorescence was detected in all synapses but

varied in intensity depending on the synapse type. It was highest in VGluT1 con-

taining synapses, followed by VGluT2 synapses and GABAergic synapses. Previous

studies [90] have noted this heterogeneity of synapsin content in cortical synapses of

rats. Using confocal microscopy in Vibratome sections, synapsin was observed in e90%

of VGluT1 puncta and only 30%–50% of VGluT2 and VGAT puncta. This discrep-

ancy with our results is probably due to the inability to detect small synapsin puncta

using confocal microscopy. Previously, we observed fewer small synapsin puncta with

confocal microscopy of Vibratome sections compared to AT on LR White sections pre-

pared from the same animal [1]. The varying synapsin content in different synapse

types is probably related to the different functional properties of the synapses. For

example, release probability is low at VGluT1 synapses and high at VGluT2 [77] and

many VGAT synapses (e.g., synapses made by parvalbumin-containing fast-spiking

interneurons). Interestingly, there is evidence that synaptophysin is also expressed at

a lower level in GABAergic synapses [91].

The proportions of different synaptic types was found to be very similar in layers

4 and 5, with the exception of VGluT2 containing synapses which, as observed in

previous studies [77], were more prominently represented in layer 4. In addition, the

existence of a sizable population of glutamatergic synapses that contain both VGluT1

and 2 was detected. There were several-fold more synapses containing both VGluT1

97

and 2 (15% in layer 4 and 10% in layer 5) than purely VGluT2 synapses (6% in layer

4 and 2.5% in layer 5). It was previously thought that the expression of VGluT1 and

2 in synapses in the adult animal is mostly complementary [77], but later studies have

revealed the existence of both VGluTs in the same cortical synapses [92], particularly,

the thalamocortical terminals in layer 4 [79]. VGluT1, VGluT1 and 2, and VGluT2

containing synapses appear to have distinct intracortical and subcortical origins, but

the exact details of their identities are still being explored [77–79]. Interestingly, the

expression of the two vesicular glutamate transporters can be regulated by activity

in opposite directions [78]. Thus, determining the VGluT1 and 2 content of synapses

may provide information about their synaptic activity as well. Including more molec-

ular markers in the single-synapse analysis is expected to reveal additional synaptic

categories and contribute further to our understanding of synaptic diversity.

AT also allowed us to observe double innervated spines at the light level and to be-

gin characterizing their input. The YFP fluorescence expressed in pyramidal neurons

in the YFP-H mouse line conveniently outlines dendritic spines, but similar analysis

can also be performed in wild-type mice on neurons labeled by intracellular microin-

jections of fluorescent tracers or by in utero electroporation. Between 5% and 30%

of cortical dendritic spines in a variety of species are thought to receive two synaptic

inputs, one excitatory and one inhibitory [82, 83, 93]. This intriguing synaptic ar-

rangement is involved in cortical plasticity and is an example of a specific synapse

subtype, namely inhibitory synapses on spines in layer 4, that changes in response to

modifications in sensory experience and/or learning [82, 84, 85]. Sensory deprivation

induced by whisker removal results in a selective reduction of inhibitory synapses

on spines, while increased sensory stimulation or classical conditioning involving the

whiskers results in a selective increase in inhibitory synapses on spines. No direct

evidence existed about the identity of either the dendritic spines or their inputs in

98

mouse somatosensory cortex. A recent study in the frontal cortex of young rats sug-

gests that double-innervated spines are preferentially targeted by VGluT2-containing

thalamocortical afferents, while the inhibitory input is from all subtypes of cortical

interneurons [87]. Using AT, we were able to observe the double innervated spines of

layer 5 pyramidal neurons in mouse somatosensory cortex at the light level and con-

firm that they receive one glutamatergic and one GABAergic synapse. We also show

for the first time that the glutamatergic input is predominantly VGluT1-containing

and that both AMPA and NMDA receptors are present at the postsynaptic site. The

addition of synaptic markers such as neuropeptide markers could provide information

about the identity of the inhibitory input.

4.4 Conclusions

Here, we demonstrate the usefulness of AT for the proteomic examination of indi-

vidual synapses in natural brain tissue with full preservation of neuroanatomical and

circuit context information. As efficient automated analysis strategies are developed

to complement the inherently high throughput of array tomographic image acquisi-

tion, this tool should open new doors to the large-scale bioinformatic exploration of

the molecular diversity and architecture of synapses. One likely consequence of such

exploration could be the development of new schemes for the differentiation and cata-

loging of molecular synapse types. By isolating specific subsets of synapses, a synapse

catalog could help enormously in pinpointing the specific synapse changes involved in

particular neurological disorders [94] or forms of neural plasticity [82,84,95–97]. AT’s

unique abilities to extract simultaneously rich proteomic and fine-scale structural in-

formation also suggests that the method may substantially advance ongoing efforts

to integrate the structural and molecular views of neuronal microcircuit function.

99

4.5 Experimental Procedures

4.5.1 Tissue Preparation

All procedures related to the care and treatment of animals were approved by the

Administrative Panel on Laboratory Animal Care at Stanford University. Four adult

mice: three C57BL/6J and one YFP-H [71], were used for this study. The ani-

mals were anesthetized by halothane inhalation and their brains quickly removed and

placed in 4% formaldehyde and 2.5% sucrose in phosphate-buffered saline (PBS) at

room temperature. Each cerebral hemisphere was sliced coronally into three pieces

and fixed and embedded using rapid microwave irradiation (PELCO 3451 laboratory

microwave system with ColdSpot; Ted Pella, Redding CA) as described in [17]. To

preserve YFP fluorescence in the YFP-H mouse, the tissue was dehydrated only up

to 70% ethanol.

For EM, the tissue was processed as above except that the fixative also contained

0.1% glutaraldehyde, and a postfixation step was added with osmium tetroxide (0.1%)

and potassium ferricyanide (1.5%) with rapid microwave irradiation, 3 cycles of 1 min

on-1 min off-1 min on at 100W, followed by 30 min at room temperature.

4.5.2 LRWhite Sections

Ribbons of serial ultrathin (70 nm) sections were cut with an ultramicrotome (EM

UC6, Leica Microsystems, Wetzlar, Germany) as described in [17]. The ribbons were

mounted on subbed coverslips (coated with 0.5% gelatin and 0.05% chromium potas-

sium sulfate) and placed on a hot plate ( e60◦C) for 30 min. For SEM imaging, the

subbed coverslips were also carbon coated using a Denton Bench Top Turbo Carbon

Evaporator (Denton Vacuum, Moorestown, NJ). Subbed and carbon coated cover-

slips were also prepared for mounting ribbons of sections to be used for multiple

100

immunostaining rounds (>6). For transmission electron microscope (TEM) the sec-

tions were collected on formvar-coated nickel grids. Immunofluorescence Staining and

Antibodies

Staining was performed as described in [17]. The coverslips with sections were

mounted using SlowFade Gold antifade with DAPI (Invitrogen, Carlsbad CA). To

elute the applied antibodies, the mounting medium was washed away with dH2O and

a solution of 0.2 M NaOH and 0.02% SDS in distilled water was applied for 20 min.

After an extensive wash with Tris buffer and distilled water, the coverslips were dried

and placed on a hot plate (60◦C) for 30 min.

The primary antibodies and their dilutions are listed in Table 4.1. Only well

characterized commercial antibodies were used and they were evaluated specifically for

AT as described in Supplemental Experimental Procedures. For immunofluorescence,

Alexa Fluor 488, 594, and 647 secondary antibodies of the appropriate species, highly

preadsorbed (Invitrogen, Carlsbad CA) were used at a dilution 1:150. The sequence

of antibody application in the multiround staining is presented in Table S1.

4.5.3 ImmunoEM Staining

The staining protocol was similar to the immunofluorescence staining with the addi-

tion of two steps in the beginning: treatment for 1 min with saturated sodium meta-

periodate solution in dH2O to remove osmium and 5 min with 1% sodium borohydride

in Tris buffer to reduce free aldehydes resulting from the presence of glutaraldehyde

in the fixative. A 15 nm gold labeled goat anti-rabbit IgG secondary antibody (SPI

Supplies, West Chester, PA) was used at 1:25 for 1 hr. After washing off the sec-

ondary antibody, the sections were treated with 1% glutaraldehyde for 1 min to fix

the antibodies in place and the sections were post-stained with 5% uranyl acetate for

30 min and lead citrate for 1 min. Fluorescence Microscopy and Image Processing

Sections were imaged on a Zeiss Axio Imager.Z1 Upright Fluorescence Microscope

101

with motorized stage and Axiocam HR Digital Camera as described in [17]. Briefly,

a tiled image of the entire ribbon of sections on a coverslip was obtained using a

10x objective and the MosaiX feature of the software. The region of interest was

then identified on each section with custom-made software and imaged at a higher

magnification with a Zeiss 63x/1.4 NA Plan Apochromat objective, using the image-

based automatic focus capability of the software. The resulting stack of images was

exported to ImageJ, aligned using the MultiStackReg plugin and imported back into

the Axiovision software to generate a volume rendering. When a ribbon was stained

and imaged multiple times, the MultiStackReg plugin was used to align the stacks

generated from each successive imaging session with the first session stacks based on

the DAPI channel, then a second within-stack alignment was applied to all the stacks.

To reconstruct large volumes of tissue (Figure 4.1), we first used Zeiss Axiovision

software to stitch together the individual high-magnification image tiles and produce

a single mosaic image of each antibody stain for each serial section in the ribbon. We

created a z stack of mosaic images for each fluorescence channel, and then grossly

aligned the stacks using the MultiStackReg plugin. Finally, to remove non-linear

physical warping introduced into the ribbons by the sectioning process, we used a sec-

ond ImageJ plugin, autobUnwarpJ (available at http://www.stanford.edu/enweiler),

which adapts an algorithm for elastic image registration using vector-spline regular-

ization [98].

For the figures, images representing single sections were upscaled using bicu-

bic interpolation. No other image processing was used except for adjustment of

brightness/contrast in some of the channels (NMDA receptor subunits and gephyrin).

Synapsin immunofluorescence was not adjusted. For volume renderings, meant only

for visual appreciation, more extensive image processing was used to adequately illus-

trate the spatial distribution and relationship between different markers. No quan-

tifications or substantive comparisons were based on these images.

102

4.5.4 Colocalization Analysis

To examine the spatial relationships between synaptic markers, we developed a colo-

calization detection function similar to the van Steensel method [21]. Using a 20 x

20 x 6.3 µm3 volume of neuropil, for each pair of channels we computed the three-

dimensional normalized cross-correlogram [99] for a range of lateral offsets approx-

imating the size of a synapse, using Eaton’s extension of the MATLAB function

normxcorr2 (http://www.cs.ubc.ca/edeaton/tut/normxcorr3.html). Pairs of labeled

channels with nonrandom associations (either positive or negative) should demon-

strate a nonzero correlation effect which asymptotically approaches 0 at offset ranges

exceeding their scale of interaction.

4.5.5 Transmission and Scanning Electron Microscopy

The immunostained TEM grids were imaged using a JEOL TEM1230 equipped with

a Gatan 967 slow-scan, cooled CCD camera. For SEM following fluorescence imaging,

the immunostained arrays were washed with dH2O to remove the mounting medium

and post-stained with 5% uranyl acetate in H2O for 30 min and lead citrate for 1 min.

The arrays were imaged on a Zeiss Sigma scanning electron microscope equipped with

field emission gun using the backscattered electron detector at 10 kV. SEM images

were aligned with the corresponding immunofluorescent images in ImageJ using the

TurboReg plugin [20]. The nuclei as viewed with DAPI fluorescence, and the SEM

defined the identical regions in the two imaging modes.

4.6 Supplemental Information

Supplemental Information includes three figures, one table, and one movie and can

be found with this article online at doi:10.1016/j.neuron.2010.09.024.

103

4.7 Acknowledgments

This work was supported by grants from the National Institutes of Health (NS063210),

Gatsby Charitable Trust, Howard Hughes Medical Institute (Collaborative Innovation

Award #43667), by funds from the Stanford’s BioX Program, Stanford’s Departments

of Neurosurgery and Neurology and Neurological Science, and by a gift from Dr.

Lubert Stryer. We thank Nafisa Ghori for her expert technical help and JoAnn

Buchanan and Gordon Wang for help and advice. We thank Profs. Liqun Luo,

Miriam Goodman, and Thomas Clandinin (Stanford University) and Bradley Hyman

(Harvard Medical School) for their very helpful comments on the manuscript.

Table 4.1: Synaptic Antibodies Used in This StudyAntibody Localization Species Source Cat. No. Dilution

All synapses Synapsin I presynaptic Rabbit Millipore AB1543P 1:100Bassoon presynaptic Mouse Abcam ab13249 1:100Bassoon presynaptic Rabbit Synaptic Systems 141003 1:100Synaptophysin presynaptic Mouse Abcam ab8049 1:10Synaptophysin presynaptic Rabbit Abcam ab68851 1:100

Glutamatergic VGluT1 presynaptic Mouse NeuroMab 75-066 1:100VGluT1 presynaptic Guinea Pig Millipore AB5905 1:1000VGluT2 presynaptic Guinea Pig Millipore AB2251 1:1000PSD-95 postsynaptic Mouse NeuroMab 75-028 1:100panMAGUK postsynaptic Mouse NeuroMab 75-029 1:100GluR2 postsynaptic Mouse Millipore MAB397 1:50GluR2/3 postsynaptic Rabbit Millipore AB1506 1:100NMDAR1 postsynaptic Mouse Millipore MAB363 1:200NMDAR2A postsynaptic Mouse Millipore MAB5216 1:25NMDAR2B postsynaptic Mouse NeuroMab 75-101 1:500

GABAergic GAD presynaptic Rabbit Millipore AB1511 1:300VGAT presynaptic Mouse Synaptic Systems 131 011 1:100Gephyrin postsynaptic Mouse BD Biosciences 612632 1:100GABAAR1 postsynaptic Mouse NeuroMab 75-136 1:100

See also Figure S3.

104

Table 4.2: Proportion of Synapses from Different Synaptic SubtypesSynaptic Subtype Layer 4 Layer 5

Exp. 1 Exp. 2 Exp. 1 Exp. 2VGluT1 60.1% (149) 57.5% (100) 66.7% (156) 67.3% (113)VGluT1+2 15.3% (38) 15.5% (27) 9.0% (21) 11.9% (20)VGluT2 5.6% (14) 6.3% (11) 2.6% (6) 2.4% (4)Other glutamatergic 3.6% (9) 4.6% (8) 5.6% (13) 2.4% (4)GABAergic 15.3% (38) 16.1% (28) 16.2% (38) 16.1% (27)All synapses 100% (248) 100% (174) 100% (234) 100% (168)All synapsin puncta 262 183 242 172Not synapse 5.3% (14) 4.9% (9) 3.3% (8) 2.3% (4)

105

Figure 4.1: Array tomographic synapsin I immunofluorescence in the cere-bral cortex of an adult YFP-H mouse is punctate and consistent withsynapse identity.(A) A volume rendering of 60 serial sections (200 nm each) through the entire corti-cal depth, including portions of the striatum. While all subsequent experiments andanalysis were performed on thinner, 70 nm sections, the thicker sections in this casehave allowed us to collect a larger volume and to better visualize the extensive den-drites of pyramidal neurons. Synapsin (magenta), tubulin (blue), and YFP (green).Scale bar, 50 µm.(B) A close up of layer 5 pyramidal neurons labeled with YFP.(C–H) Zoomed-in view of layers 1 (C), 2/3 (D), 4 (E), 5a (F), 6a (G), and whitematter and striatum (H).Scale bar for (B)–(H), 10 µm. See also Movie S1 for a more revealing rendering of thesame image volume and Figure S1 for comparison of different synapsin antibodies.

106

Figure 4.1

107

Figure 4.2: Proteomic immunofluorescence AT of mouse somatosensory cor-tex yields staining patterns consistent with synaptic protein distributions.Volume rendering from 20 sections, 70 nm each, from an array stained with 11 anti-bodies (Table S1, data set KDM-SYN-090416).(A) Tubulin (blue), synapsin (magenta), YFP (green), and DAPI (gray) fluorescence.(B–D) The boxed area in (A). DAPI (gray) and YFP (green). (B) Distribution of allpresynaptic boutons as labeled with synapsin (magenta). (C) Distribution of VGluT1(red), VGluT2 (yellow), and VGAT (cyan) presynaptic boutons. (D) Postsynapticlabels: GluR2 (blue), NMDAR1 (white), and gephyrin (orange) next to synapsin(magenta).Scale bar 10 µm. See also Table S1 for sequence of antibody application.

108

Figure 4.2

109

Figure 4.3: Multiple synaptic proteins are colocalized in a fashion consistentwith synaptic identity and glutamatergic and GABAergic synapse subtype.(A) Volume rendering of 20 sections (70 nm) from the mouse somatosensory cor-tex immunostained for synapsin (magenta) and synaptophysin, VGluT1, PSD-95, orGAD (green). Colocalization of the magenta and green channels is displayed as white.DAPI, blue. These volume renderings are from an array stained with 17 antibodies(Table S1, data set KDM-SYN-091207). Scale bar, 5 µm.(B) Colocalization matrix of nine synaptic markers and tubulin (left) and correspond-ing pairwise representation of the channels on a small area (4 x 4 µm) of a single sec-tion (right). For each pair of channels we computed a cross-correlation score over arange of lateral offset distances for images in the two channels. The cross-correlationscore is represented as a grid of false colored squares with their center representingthe score at 0 offset and each pixel equal to 0.1 µm offset.(C) For a subset of channel comparisons, the cross-correlation score is plotted as afunction of the lateral offset. Each trace is obtained by averaging 16 equally spacedradii. Left, with no lateral shift the normalized cross-correlation is equal to thePearson correlation coefficient and at shifts beyond the rough size of a synapse thecorrelation drops to e0 for all channels. Right, the same is normalized such thateach curve is 1.0 in the no-shift case. Pre-presynaptic and post-postsynaptic channelcomparisons drop off sharply, while pre-postsynaptic do not.

110

Figure 4.3

111

Figure 4.4: Dendritic spines in mouse cerebral cortex are contacted bysynapsin puncta and colocalize with other pre- and postsynaptic mark-ers.Volume rendering of 45 sections from dataset NAOR-081118 (Table S1). To bettervisualize the synaptic markers associated with dendritic spines, only immunofluores-cence within 0.5 µm of the YFP dendrite was displayed.(A) In the left panel, a 20 µm long segment from a spiny dendrite of a layer 5 pyramidalcell (green) is shown as it traverses layer 4. In each subsequent panel the labeling of asynaptic protein is added. PSD-95 (blue), bassoon (yellow), and synapsin (magenta).The postsynaptic protein PSD-95 is found within spine heads and closely apposed tothe presynaptic proteins bassoon and synapsin (arrow). Scale bar, 2 µm.(B) The opposite side of the spine marked with an arrowhead in (A) at higher mag-nification. Scale bar, 0.5 µm.See also Figure S2.

112

Figure 4.4

113

Figure 4.5: Ultrastructurally identified synapses are labeled with thesynapsin antibody.(A and B) Conjugate synapsin immunofluorescence and SEM of the adult mousecerebral cortex. Synapsin (magenta) and DAPI (blue) signal as obtained with thefluorescence microscope are overlaid on the SEM image from the same section. (B)Four serial sections through the boxed region in (A). Boxed region is section #2 inthe series. The majority of presynaptic boutons are consistently labeled from sectionto section (arrows), but some are labeled only on few sections with a weak signal(asterisk). Scale bar, 0.5 µm.(C) A TEM image of postembedding gold immuno-EM for synapsin. The 15 nm goldparticles label presynaptic terminals as identified by the presence of synaptic vesiclesand postsynaptic density. Scale bar, 0.5 µm.See also Figure S2 for effect of tissue processing on synapsin immunostaining.

114

Figure 4.5

115

Figure 4.6: Synaptograms are useful for viewing proteomic information fromserially sectioned single synapses.A glutamatergic (left) and a GABAergic synapse (right) are shown. Each squarerepresents an area of 1 x 1 µm from a single 70 nm section. Each section through thesynapsin punctum occupies a column and each antibody label a row. See also TableS1 for sequence of antibody application.

116

Figure 4.6

117

Figure 4.7: Proteomic imaging with AT reveals the diversity of corticalsynapses.(A) Examples of synaptograms representing the main synapse subtypes observed inmouse somatosensory cortex with the current antibody panel.(B) Synapsin content of different synaptic subtypes. For each subtype, 100 synapseswere randomly selected using the VGluT1-PSD-95, VGluT2-PSD-95, and VGAT-gephyrin channels and synapsin immunofluorescence was measured on each sectionthrough the synapse. Top panel, histograms of synapsin immunofluorescence in thethree synapse subtypes. Lower panel, scatterplot of synapsin intensity versus therespective vesicular transporter immunofluorescence for each synapse.(C) Examples of glutamatergic synapses with different postsynaptic receptor combi-nations.(D) Example of a synapse made by the axon of a YFP-positive layer 5 pyramidalneuron.

118

Figure 4.7

119

Figure 4.8: Double innervated spines receive both a glutamatergic VGluT1and GABAergic synapse.(A and B) Volume rendering of dendritic spines from YFP-positive pyramidal celldendrites (green), each receiving 2 synaptic inputs on the head. The glutamater-gic synapses are represented by postsynaptic PSD-95 label (blue) and presynap-tic synapsin (magenta). The GABAergic synapses are represented by postsynapticgephyrin (orange) and presynaptic GAD (cyan). The labels are added consecutivelyfrom left to right. Additional synapses not contacting the spines are also observedwithin the reconstructed volume.(C and D), Single sections through double innervated spines labeled with multipleantibodies. For each spine, the two adjacent sections where most of the markers werepresent was chosen. Each panel shows the spine (green) and one synaptic marker(magenta). Direct overlap of the two labels is seen as white. The punctuated lineseparates adjacent sections. (C’) and (D’) show a volume rendering of the spines in(C) and (D) with the plane of the single sections represented in gray. Scale bar, 1µm.

120

Figure 4.8

121

Chapter 5

Single-synapse analysis of a diverse

synapse population: synapse

discovery and classification

5.0.1 Abstract

Synapses of the mammalian central nervous system are recognized today as being

highly diverse in function and in molecular composition. Indeed, synapse diversity

per se is likely to be critical to brain function, since the abilities of synaptic circuits to

store and retrieve memories, and to adapt homeostatically to developmental and en-

vironmental change are thought to be rooted primarily in activity-dependent plastic

changes in specific subsets of individual synapses. Unfortunately, the measurement of

synapse diversity has been restricted by limitations of methods capable of measuring

synapse properties at the level of individual synapses. Array tomography is a new

high-resolution, high-throughput proteomic imaging method that has the potential

to very substantially advance the measurement of unit-level synapse diversity across

large and diverse synapse populations. Here we introduce and compare automated

122

feature extraction and classification algorithms designed to discriminate, classify, and

quantify synapses from high-dimensional array tomographic data much too volumi-

nous for manual analysis. We identify a random forest classifier, trained from exam-

ples classified by human experts, as being particularly suitable to high-throughput

synapse classification. We demonstrate the use of this method to quantify laminar

distributions synapses in mouse somatosensory cortex and validate the classification

process by detecting the presence of known but uncommon proteomic profiles. We

suggest that such classification and quantification is likely to be highly useful in iden-

tifying specific subsets of synapses exhibiting plasticity in response to perturbations

of the environment or sensory periphery.

5.0.2 Author Summary

Synaptic connections are fundamental to every aspect of brain function. There is

growing recognition of the individual synapse as the key sites of the functional plastic-

ity that allows brain circuits to store and retrieve memories and to adapt to changing

demands and environments. There is also growing recognition that many neurologi-

cal, psychiatric, neurodevelopmental and neurodegenerative disorders must be under-

stood at he level of individual synapses and proteomically-defined synapse subsets.

Here, we introduce and validate computational analysis tools designed to complement

array tomography, a new high-resolution proteomic imaging method, to enable the

analysis of diverse synapse populations of unprecedentedly large size at the single-

synapse level. We expect these new single-synapse classification and analysis tools

to substantially advance the search for the specific physical traces, or engrams, of

specific memories in the brains synaptic circuits. We also expect these same tools

to be useful for identifying the specific subsets of synapses that are impacted by the

various synaptically-rooted afflictions of the brain.

123

5.1 Introduction

Synapses are fundamental to every aspect of brain function. They are recognized

today as being highly complex structures and highly diverse in both function and

molecular composition. At the structural level, individual synapses of the mam-

malian central nervous system are thought to comprise hundreds of distinct protein

species [4–6], and genomic and gene expression data available implies very strongly

that there are multiple isoforms of many of these proteins and that their expression

is differentially patterned across the brains many different neuron types [3]. It thus

seems inescapable that synapses of the brain, even within traditional transmitter-

defined synapse categories (e.g., glutamatergic, GABAergic, cholinergic, etc.), must

be highly diverse in protein composition. This conclusion is consistent with the avail-

able functional data, where physiological studies report wide differences in synaptic

transmission as different brain regions and pathways are examined (again, even when

results are compared only within traditional neurotransmitter categories). Moreover,

the well-known functional plasticity of both synapse structure and synapse function

in response to electrical activity implies directly that even an otherwise homogeneous

synapse population must become heterogeneous or diverse after individual synapses

experience differential activity. In this light, it seems likely that synapse diversity

per se may be critical to the proper function of neural circuitry. For instance, there

is now widely believed that the plasticity (and therefore resulting diversity) of indi-

vidual synapses is fundamental to memory storage and retrieval and to many other

aspects of neural circuit adaptation to environmental change [100,101].

Unfortunately, the measurement of synapse diversity has been restricted by limi-

tations of methods capable of measuring synapse properties at the level of individual

124

synapses. Array tomography (AT) is a new high-resolution, high-throughput pro-

teomic imaging method that has the potential to very substantially advance the mea-

surement of unit-level synapse diversity across large and diverse synapse populations.

AT uses multiple cycles of immunohistochemical labeling on thin sections of resin-

embedded tissue to image the proteomic composition of synapse-sized structures in

a depth-invariant manner. We have applied AT to freshly-fixed mouse cerebral cor-

tex, where our volumes have typical sizes of thousands to millions of µm3 of tissue,

contain millions of individually-resolved synapses, and label over a dozen multiplexed

proteomic markers.

With proper analysis, the informational density of array tomographic volumes has

numerous potential applications. Synapse-level resolution of large volumes of tissue

is an ideal tool for addressing interesting hypotheses concerning principles like synap-

tic scaling [100], structural arrangement [102], and novel synapse types [103, 104].

Combined with connectomic data [105, 106], genetic models [107, 108] or dye filling

techniques [109, 110], array tomography can also address questions regarding pro-

teomic distributions in specific subsets of cells. We are interested in investigations

of this nature and others in the mouse cerebral cortex, where the anatomical distri-

bution of synapses, aside from cortical layer cytoarchitectonics, is currently largely

unexplored.

Developing a Method of Synapse Quantification

To utilize array tomography to its fullest extent requires the development of new

synapse detection and classification capabilities. Simple analysis, using repeated hu-

man observation of a fraction of the channels available in the full volume, may be

acceptable for analyzing fragmentary subsets of a few hundred synapses but cannot

scale beyond that. We have developed tools and methods to assist in handling the

high proteomic dimensionality of array tomographic volumes (Figure 5.1), principally

125

the synaptogram [2], which splays out a small 3d volume surrounding a single synapse

into a larger 2d image. This eases the difficulty of per-synapse manual classification

such that the effort of classifying a set of few hundred synapses is no longer excessive,

but no matter how convenient they are to analyze individually, the sheer number of

synapses makes manual analysis of the entire data set effectively impractical.

Given that just a few hundred analyzed examples can be obtained with a reason-

able expenditure of effort, there are two approaches to consider. The first is to use

those examples as a representative sample, in a manner similar to stereology. That

may work well for some questions, but not others. Rare or novel synapse types and

cortical laminar distributions would be especially difficult to study. An alternative,

which this paper will present, is to take that sample of accurately classified synapses

and extrapolate its decision-making information to the much larger population of

unclassified individuals.

5.2 Results

5.2.1 Identifying Putative Synaptic Loci

The first necessary step in our classification process is to locate the sites which

may contain synapses. Despite their appreciable proteomic diversity [111], cortical

synapses are small: from the ostensible midpoint of the synapse, all relevant synaptic

protein labeling can fit within a 500 nanometer radius for mouse cortex [112]. Given a

reliable method of locating synapses, all information needed to verify and type those

synapses can be had from the local volume surrounding them, greatly reducing the

spatial analysis needed per synapse. To avoid confusion with actual synapses, we

refer to these sorts of putative synapse locations as ”synaptic loci.” They are specific

places which might be synaptic.

126

In order to find putative synapses to help limit the necessary search space, we are

using an antibody targeting Synapsin I. Synapsin is a scaffolding protein reportedly

found in all cortical synapses [113], and labeled antibodies targeting synapsin have

previously been used on their own to estimate synapse counts [114]. A Millipore

Rabbit anti-Synapsin I antibody (Millipore AB1543P) demonstrates robust and re-

liable labeling, and is likely to be colocalized with all relevant synaptic markers [2].

For these reasons the core of our analysis uses Synapsin I labeling to derive a list of

locations likely to contain synapses from which to begin small volumetric searches for

confirmation. Our approach is to use the brightest point of each Synapsin I punc-

tum as the site of a possible synapse to designate a local volume for further analysis,

without attempting to explicitly determine the punctum boundaries.

We prefer our local maxima-based approach over thresholding-based segmentation

because the latter has a number of issues arising from AT’s largely anisotropic reso-

lution (e200nm x e200nm x 70nm). This anisotropy, combined with (often unknown)

epitope density and labeling variance means that any segmented punctum boundary

is at best an estimate. An approach using local maxima, paired with a voxel-based

rotation-invariant feature set, is not affected by the exact boundaries of the puncta

of interest, but by the puncta themselves.

While our approach to synapse discovery sidesteps segmentation, it does so at

the cost of introducing potential false positives: background local maxima which

segmentation would have discarded, but whose peak brightness rises over our low

threshold for consideration. However, it is possible to filter those out in later clas-

sification. Conversely, this method is ideal for teasing apart “clumps” of synaptic

labeling, where multiple synapses exist in close proximity but can be resolved by the

Rayleigh criterion and thus having separate local maxima.

127

5.2.2 Manual Classification

Using Human Experts

Humans can visually identify the synaptic category of a given locus via the use of

synaptograms (Figure 5.1), using the spatial juxtaposition of a number of relevant

synaptic molecules for classification [2]. Glutamatergic synapses, for example, will

by definition have at least one vesicular glutamate transport protein and at least one

post synaptic density scaffolding protein present. Similarly, GABAergic synapses can

be identified by the presence of glutamic acid decarboxylase (GAD) and a vesicular

GABA transport protein.

This process of human synapse identification is the best and most reliable method

of synapse identification available to us. It relies on the perception and expertise of

the human viewer to apply the visual segmentation which defines the “presence” of

necessary labels. This task incorporates a great deal of a priori knowledge concerning

the stearic and functional relationships between the different molecular labels, the

variance in labeling of each particular antibody, and the particular conditions under

which that tissue had been fixed, embedded, labeled, imaged, relabeled, etc.

Although manual classification of fluorescence data is orders of magnitude faster

than EM stereology, it is still orders of magnitude slower than needed to keep up

with the synaptic output rate of AT volumes. For that, we decided to use human-

generated classifications as training data, then liberally applied a number of clustering

and supervised learning methods to quantitatively mimic the human decision making

process.

Human Rater Agreement

In order to gauge the reliability of any single human expert’s rating, we performed a

qualitative test of the consistency of human classification. We presented a set of one

128

hundred randomly-selected synaptograms to a group of six human raters who were

familiar with the task of interpreting synaptograms, and instructed them to classify

the set based on whether or not the synaptogram was centered on a glutamatergic

synapse. Once collated, we considered the true classification of a given synaptogram

to be that of a simple majority vote. When we compared each rater’s performance

relative to the average, we found an average accuracy rate of 77.7%, with a standard

deviation of 10.1% (Figure 5.2). The largest source of variance arose from the self-

reported stringency of the raters, in how much ambiguity they found acceptable when

classifying a locus as positive.

5.2.3 Machine Learning

Machine learning methods come in two broad categories. Supervised learning al-

gorithms, trained using a sufficient number of human rated synapses, are capable

of producing numerical descriptions of human judgment as it is applied to synapse

classification, as well as extrapolating that judgment to the hundreds of thousands

of synapses which comprise an average data set. Unsupervised clustering, on the

other hand, when applied to raw synaptic loci or already classified synapses is a great

approach to the discovery of marginal classes or subtle subtypes.

Feature Extraction

The first step in constructing a computational framework for either form of synapse

classification is to find a set of explicit measurements which span the feature space

that human raters implicitly search. We are using a small set of ad hoc, channel-

independent, rotationally invariant features to measure the spatial distribution of

each channel’s fluorescence about the synaptic locus. These features are calculated

per voxel, without relying on segmentation, combinatorial information or a priori

129

geometrical information, in keeping with the rationale behind finding the loci in a

similarly parameter-independent manner. The equations used to calculate the four

features are given below.

For every voxel i in the local 11x11x11 voxel window with brightness b and pixel-

wise distance from the synaptic locus d:

IntegratedBrightness = B =∑i∈V

bi (5.1)

LocalBrightness =∑i∈V

bi

di2 (5.2)

CenterofMass =

∑i∈V bidi

B(5.3)

MomentofInertia =

∑i∈V bidi

2

B(5.4)

Of these features, the Integrated Brightness is the simplest to describe, as it is

the sum of all the pixel values within 5 pixels. Local Brightness is also the sum of all

values within 5 pixels, but the contribution of each pixel is reduced by the square of

its distance from the locus. It can be used as a metric for estimating the volume of the

punctum without segmentation because nearby pixels (more likely to be part of the

punctum) contribute much more heavily than distant ones (more likely to be noise

or neighbors). The remaining features, Center of Mass and Moment of Inertia, treat

the puncta brightness as if it is a mass distribution in a synaptogram-sized object,

and respectively compute the distance to the center of that object and its angular

inertia for a rotation about the locus. Collectively, these four features do a good job

of describing the fluorescence distribution in a synaptogram.

The result of this feature extraction, when performed on a multidimensional image

130

of c channels, is a 4c-long numerical vector of proteomic measurements describing the

putative synapse. This analysis is repeated for each of p synaptic loci in the data

set, giving us a p x 4c matrix of measurements to be further analyzed. To enhance

consistency between data sets, which may well have different imaging conditions, we

normalize each of the extracted features by dividing by the population’s mean score.

5.2.4 Unsupervised Clustering

Although visual analysis is the traditional and preferred method of examining biolog-

ical data, long strings of numbers such as our feature vectors are difficult for humans

to visualize. In response, high-dimensional numerical measurements have often been

approached using some form of dimensionality reduction as a first step in numerical

analysis. Simply put, reducing a long string of numbers to a short string of numbers

makes them easier graphically display and understand. Principal Component Anal-

ysis (PCA) is a venerable method of dimensionality reduction which has seen use in

similar applications [115,116], and has proven useful in ours as well.

Our PCA result, illustrated in Figure 5.3, identifies some synaptic populations

but does not separate them sufficiently for classification. The loci tend to aggregate

in clusters which correspond to a few of the broader synaptic categorizations, namely

GABAergic and two common subtypes of glutamatergic synapses. We identified the

clusters using multivariate regression, that is, taking a few of the more distant exam-

ples and inferring the contribution of channels which brought them from the mean.

We had hoped when applying PCA to find discrete, easily-separable clusters corre-

sponding to each class, but in the reduced dimensionality of PCA, simple thresholds

are insufficient for proper class discrimination.

We had hoped the dimensionality reduction accomplished by the above methods

would have proven amenable to simple thresholding. If that where the case, multi-

variate regression might have led to identification and, combined with a measure of

131

the statistical significance of cluster separation, classification of unknown synapses

based solely on where they fell in the unsupervised plot. Since our clusters were not

so cleanly separable, we resorted to a more subtle stratagem involving supervised

learning.

5.2.5 Supervised Classification

The “supervision” of supervised learning refers to the supervised training set, a ran-

dom or semi-random collection of human-rated examples from which the machine

learning algorithm (MLA) infers the rules for classification to extrapolate onto novel

synapses. To generate each item of the set, we presented a synaptogram to a human

trainer, who rated the synaptogram in one or more binary categories representing

the presence or absence of channels relevant to synapse classes of interest. We could

then associate those categorizations with the already-derived feature vectors of those

examples, compiling them into a library of “correct” classifications for training.

MLA Selection

Another necessary choice in supervised learning is that of the MLA used as a classi-

fier. In an early training experiment, we created a training set of 200 examples clas-

sified into glutamatergic/non-glutamatergic and GABAergic/non-GABAergic cate-

gories. We fed these results into an assortment of MLAs available in MATLAB with

minimal parameter optimization. The error rates of the various MLAs are presented

in Table 5.1. Although many of these algorithms performed well, the random forest

ensemble [117] slightly edged out the competition and earned its place as our classifier

of choice.

132

Global Feature Importance

An additional point in favor of the random forest ensemble was the useful poste-

rior metrics made readily available in MATLAB’s random forest implementation (the

TreeBagger class). Posterior metrics are methods of analyzing the process of clas-

sification after classification. Their primary purpose is to relate information about

why a given locus was classified one way or another, and meta-information such as

the relationship between classes and the features which proved more important than

others during classification.

Each decision tree in a random forest is a series of optimal feature threshold

branches with decisions for leaves. By keeping track of which feature was used for

each branch point, along with the confidence that branch point engenders, we could

gauge the importance of the various features relative to each other. Overall, our

local brightness feature proved most useful, with the rest decreasing in performance.

Normalized to the local brightness importance, feature values were 0.76, 1.00, 0.59

and 0.55 for the integrated brightness, local brightness, center of mass and moment of

inertia features, respectively. Though the local brightness feature may have outshone

the others, all proved useful in classification.

Channel-based Classification

In order to facilitate the discovery of novel synapse populations, we decided to classify

loci based on channel presence, rather than synapse identity. For the MLA choice

analysis and the human agreement test our raters classified synaptograms based on

full synapse classes: whether or not the synaptogram included all the requisite markers

for a synapse of the class in question. While this strategy was effective at classifi-

cation of known synapse types, for the full analysis we wished to be able to extend

the analysis to detect synapse configurations which the raters did not have a priori

133

knowledge of. Such novel synapse types might be discarded as being the product of

noise or aberrancies in the data, if they were accessible at all. For example, suppose

there was a small population of glutamatergic synapses which also expressed VGAT.

If all you have is glutamatergic/non-glutamatergic, there’d be no way of discovering

it even though the data is there.

We chose a classification system which addresses these concerns by using multiple

MLAs per synapse type, each trained to detect a single proteomic marker. A given

locus can be said to be a synapse of a certain class if all requisite markers of that class

are present at that locus. To avoid confusion, we elected to highlight the difference

by identifying synapses which contain a given marker as “marker synapses,” and loci

which contain a given marker (but may not be synaptic), “marker loci” or “marker

positive loci.” For example, rather than using a glutamatergic synapse classifier to

detect glutamatergic synapses, we use individual classifiers for the relevant channels

(VGluT1, VGluT2, PSD95), and then use their outputs in the same logical way (

(VGluT1 ∪ VGluT2) ∩ PSD95) to identify glutamatergic synapses. Using the previ-

ous example of VGAT-positive glutamatergic synapses, it would be straightforward to

add a ∩ VGAT to the equation, and see if the resulting population occurs significantly

above chance.

Active Learning and Rare Classes

In most supervised learning models, training set examples are sampled entirely at

random in order for the training set to have the same statistical properties of the full

data set. This can be inefficient for us in the of case of uncommon channels. The less

common a given channel is, the more negative results a human has to sort through

before reaching a usable number of positive results. For example, VGluT3 positive

loci can be identified in much the same manner as VGluT1 or VGluT2 loci, but due

to their paucity in the cortex (we see roughly 1.2 VGluT3+ loci per one thousand

134

negative loci), human raters would have to classify excessive numbers of negative loci

for each positive locus in the training set.

In order to address this possibility, our classification process is a two-phased non-

random selection of training examples. The first phase is to “prime” the training set

data for rare classes by choosing one of each class’s requisite presynaptic channels and

randomly sampling a subset from the loci for which the channel’s local brightness is

more than two standard deviations above the mean. A number of class subsets gen-

erated in this manner are collated, each class contributing to the negative examples

of the rest. The second phase is an “active” training process in which a human rater

and the MLA being trained work in tandem to speed training, a technique known

as active learning [118]. At each step, the half-trained classifier selects a few exam-

ples, half of which it thinks are positive and half negative, to present to the rater for

verification and feedback.

135

In pseudocode, the training proceeds according to the following algorithm:

while Human wishes to train do

Load training synaptogram population, P

Human selects a synaptic category

Train RFE using partially classified training set T , display predicted error rate

while Human wishes to add training examples do

Randomly choose c, where c ∈ [True, False]

Randomly choose a synaptogram s from subpopulation Pc, the elements of P

classified as c

Display s and c to human for verification

Add/Update s in T to reflect human input

end while

end while

The application of active learning in this instance was inspired by a similar effort

by Dr. Badrinath Roysam (personal correspondence) for use in the classification of

cell types in confocal volumes.

The net effect of the training modification is to focus the human role more on

verification and correction than strict classification. Aside from accomplishing the

goal of efficiently training classifiers for rare classes, we find that the active version

seems to be much less of a strain on human patience than de novo training, even

that aided by synaptograms. It also reduces the necessary training set size to roughly

twice the number of requisite positive synapses in the training set, despite the rarity

of the class in question.

Once the human raters are satisfied with their training sets, we pass the entire

data volume through the classifiers for identification, and collate the results into a

combinatorial set of vectors.

136

5.2.6 Post-Classification Analysis

After classification, the predicted presence of each channel for a given locus can be

derived from the percentage of decision trees in the random forest ensemble which

attest to its presence. This effectively serves as a confidence metric for the entire

ensemble, and is generally referred to as the “posterior probability.” An instance

with a posterior probability of 1.0 is unequivocally positive for the class in question,

one of 0.0 is undeniably negative. In this manner, we reduce the 4c-long numeric

feature vector to a c1 -long numeric posterior vector, representing the presence or

absence of all c1 relevant channels. We can then use these vectors in a combinatorial

fashion to recreate synaptic classes. Glutamatergic VGluT1-expressing synapses, for

example, should at a minimum be positive (posterior probability ≥ 0.5) for VGluT1

and PSD95.

Per-Channel Feature Importance

Since our labeled channels occupy a number of spatial niches in the canonical synapse,

we were interested in determining which features contributed most to which channel

classifier, in case that reflected the differential distribution. The results are shown

in Figure 5.4. The channels which differ from the norm (Figure 5.4-A) in selecting

the center of mass or moment of inertia features as their most important included

VGluT2, VGluT3 and VAChT. These channels are all presynaptic, which eliminates

spatial differentiation as a cause, but interestingly they are all uncommon to rare.

TH, also rare, did not display this behavior, and also differs from the rest in that

”neighboring” puncta were deemed acceptable for positive classification. This may

suggest that for rare classes where neighbor discrimination is important, determining

whether a discovered punctum is part of the synapse in question or a close neighbor

plays a bigger role in the accuracy rate than discovering the punctum in the first

137

place.

OOB Error

The training process of the random forest classification itself provides a reliable ap-

proximation of its error rate. During training, each tree in a random forest excludes

a random fraction of examples from its construction, which can later be used in the

manner of cross-validation testing to gauge the accuracy of that tree. More precisely,

each training example can function as withheld data for a sub-random forest ensemble

composed of the fraction of decision trees to have excluded it during training, and,

taken in aggregate, are an estimate of the performance of the full forest. This is called

the “out-of-bag error” [117]. OOB performance for the classes we are interested in

can be found in Table 5.2. The OOB error can be interpreted as a self-estimation of

the classifier’s true error rate.

Comparison to Human Rating

To quantitatively examine this system’s performance when applied to real synapse

classification, we ran our human accuracy test set through the VGluT1 and PSD95

classifiers, then compared the combined output (VGluT1 ∩ PSD95) loci with that

given by humans. Although these two channels had the worst OOB performance, the

intersection of the two was about as accurate as the best human raters. We performed

a receiver operating characteristics analysis to describe the classifier performance in a

more detailed fashion; it is shown in Figure 5.2B. The fact that the worst OOB error

is still equal to the agreement of human raters implies the output of the classifiers

should be usable with the same degree of confidence as that of human raters.

138

5.2.7 Classification Application

Synapse Class Definition

The use of a channel-based classification process allows us somewhat greater flexibility

in the definition of synaptic classes. Our lab has years of experience in recognizing

VGluT1-glutamatergic, VGluT2-glutamatergic and GABAergic synapses [2], which

compose the majority of synapses in the cortex, and all are defined by the presence

of at least two specific markers, in addition to Synapsin I. For this paper we have

also included a number of labels targeting synaptic populations for which we haven’t

found a robust label for a ubiquitous second protein. This includes VGluT3-positive

synapses, cholinergic (vesicular acetylcholine transporter [VAChT]) and dopaminer-

gic/noradrenergic (tyrosine hydroxylase [TH] positive) synapses. It is our intention

to find such corroborating labels before these channels are used in a full experiment.

Additionally, dopaminergic synapses have been reported not to express much of the

Synapsin I/II isoforms, if they express them at all [119]. Since we are using a Synapsin

I marker to discover putative synapse loci, those which are positive for TH may ac-

tually be identifying simple synaptic complexes - dopaminergic synapses adjoining

those of another class.

Cortical Depth Analysis

One straightforward application of synapse-classified array tomography can be had

via cytoarchitectonics, as seen in Figure 5.5. We first segregated the data into a

number of synaptic classes, then subdivided those into 10 µm bins stretching from

the pial surface of the cortex to the striatum. We calculated the density of each bin’s

population, and averaged the Synapsin I local brightness feature to estimate the mean

synapse size. Overall, the synaptic densities were nearly twice as high as expected

in the literature [120], but tissue shrinkage during LR White embedding [121] can

139

potentially account for some or all of that.

Although a sample size of one precludes statistical certainty and excessive hypoth-

esizing about function, there are three interesting effects observed by this analysis.

First, there is an increase in VGluT2-positive synapse density in layer IV, which we

expected given the laminar characterization of VGluT2-expressing synapses [2]. Sec-

ond, we notice a decrease in the density of parvalbumin-positive GABAergic synapses

in layers I and VI, similar to [122]. Finally, we find that VGluT1-positive synapses in

layer 5a, though not more dense than elsewhere in the cortex, are somewhat larger.

Pairwise Proteomic Analysis

Another promising possibility is the use of data sets classified in our per-channel

fashion to search for unexpected proteomic combinations which may correspond to

novel synaptic subsets, particularly of rare classes. In any volume, some background

noise is to be expected: given the spatial distribution of synapses, it is inevitable that

some synapses will have asynaptic puncta, or those belonging to nearby synapses,

expressed in the region of analysis. Assuming that two classified markers have inde-

pendent distributions, the expected number of loci in a volume which will be classified

positive for both is the product of their probabilities, Eij = Pi ∗Pj. We can compare

this with Fij, the number of colocalized loci actually found in the data set, and use

a two-tailed binomial test to check for significance and reject stochastic noise as an

explanation.

For example, VGluT3 has previously been intimated to be present in a very small

subset of cortical GABAergic synapses [123]. Since we have labeled both GABAergic

synapses and VGluT3 puncta in the course of classifying their respective categories,

we can simply retain those GABAergic synapses which classed VGluT3-positive. A

two-tailed binomial test can tell if the overlap we observe (82 synapses) is significantly

different from that we would expect by multiplying the two class probabilities together

140

(43 synapses). Those are small numbers in a data set of nearly a million classified

synapses, but the difference between them is significant (p < 0.001).

Using the nine classified channels in our present analysis, we ran binomial tests to

calculate the normalized pairwise relationship between each of them. Our results are

presented in Figure 5.6. The significant results match our expected relationships for

the most part - GAD, VGAT, parvalbumin all colocalize, as do VGluT1/PSD95 and

VGluT2/PSD95, and all three categories are mutually exclusionary. There are a few

points of interest - as mentioned, VGluT3 colocalizes with all GABAergic channels

and excludes itself from PSD95, corroborating the literature’s suggestion of VGluT3

as a supporting neurotransmitter and not a primary glutamatergic synapse class on

its own [124]. Additionally, TH generally avoids both VGluT1 and VAChT, but

shows positive copresence with VGluT2 (though this relationship disappears in the

striatum).

5.3 Discussion

Synapse Class Discovery

When we began the class discovery process as shown in Figure 5.6, we expected re-

lationships based on our preconceived notions of a few synapse classes: that GAD,

VGAT and parvalbumin should all be copresent to some extent [122], and that VG-

luT1 and VGluT2 should each colocalize with PSD95 (but not with each other) [125].

The other channels, VGluT3, VAChT and TH, had fewer performance expectations.

Of them, VGluT3 had the most interestingly unexpected behavior, avoiding the com-

mon glutamatergic markers and colocalizing instead with GABAergic synapses. That

the literature corroborates these as possible roles of VGluT3 in the cortex [124] lends

a degree of confidence that our analysis returns usable results. Another interesting re-

sult is the cortical-only localization of VGluT2 and TH. Dopaminergic (TH) neurons

141

have been reported to express VGluT2 in rat cultures [126], midbrain and hypotha-

lamus [127]. With current volume sizes we only find a dozen of these appositions,

however, so it would be problematic to assert certain confirmation.

Human Consensus Formation

The variance of human raters raises a few interesting questions to look into in the

future. Two of the six raters (#2 and #6) self-reported using a stricter standard

of classification than the rest: when an example was at all doubtful, they classified

it as being negative. Effectively, these raters elected to position themselves on the

left side of the ROC curve, trading an increase in false negatives for reduced false

positives. Depending on the application, stricter classification may be preferable.

Based on our experiences, we would recommend taking time to discuss questionable

examples and reasons for rating them one way or another. Such conversations are

rather illuminating and very effective at getting everyone to agree on a common

standard of classification.

5.3.1 Limitations and Future Work

There are two significant limitations to the questions which can be asked using this

method. The first and strictest: an array tomography volume is a decidedly terminal

snapshot of a piece of tissue. This precludes experiments which would watch a par-

ticular cell or dendrite change over time, or in response to learning [128], except in

animal models which are stereotyped enough for different animals to have equivalent

nervous systems, namely C. Elegans [129] and Drosophila [130]. Synapse populations

are assumed to be fairly invariant between individual mice (and presumably humans),

however, which allows us to study changes to synaptic classes as a whole in response

to plasticity or disease.

142

The second limitation is more easily rectified. Our analysis partially depends on

limiting the scope of the problem to that required to identify synapses at locations al-

ready suspected to contain a synapse. For common synapse classes this is easy. They

all express Synapsin I, so wherever we find our Synapsin I marker, there may be a

synapse. As mentioned, we have already begun to abut the usefulness of Synapsin I,

which may not be expressed in dopaminergic synapses. Using a pan-Synapsin anti-

body would be a straightforward solution to catching all dopaminergic synapses, but

it is fully possible that other, more exotic synapse types would not express Synapsin

at all.

Establishing a robust system for synapse classification in array tomographic vol-

umes opens up a number of avenues for addressing biological questions. It allows us

to conduct single-synapse analyses in large regions of tissue, which lets us study rare

or spatially-segregated populations. It helps us discover new synaptic populations

and novel variations on known synapse types, and gives us an unprecedented level of

control over the proteomic complexity we can bring to bear.

5.4 Materials and Methods

5.4.1 Acquisition of array tomographic volume

All procedures related to the care and treatment of animals were approved by the

Administrative Panel on Laboratory Animal Care at Stanford University. All volumes

were acquired from mouse cortex, line C57BL/6J, using the methodology given in [2].

One adult mouse was used for this study. The animal was anesthetized by

halothane inhalation and its brain quickly removed and placed in 4% formaldehyde

143

and 2.5% sucrose in phosphate-buffered saline (PBS) at room temperature. Its cere-

bral hemisphere was sliced coronally into three pieces and fixed and embedded us-

ing rapid microwave irradiation (PELCO 3451 laboratory microwave system with

ColdSpot; Ted Pella, Redding CA) as described in [17]. The tissue was dehydrated

up to 70% ethanol.

Ribbons of serial ultrathin (70 nm) sections were cut with an ultramicrotome

(EM UC6, Leica Microsystems, Wetzlar, Germany) as described in [17]. The ribbons

were mounted on subbed coverslips (coated with 0.5% gelatin and 0.05% chromium

potassium sulfate) and placed on a hot plate (60 C) for 30 min. For SEM imaging,

the subbed coverslips were also carbon coated using a Denton Bench Top Turbo

Carbon Evaporator (Denton Vacuum, Moorestown, NJ). Subbed and carbon coated

coverslips were also prepared for mounting ribbons of sections to be used for multiple

immunostaining rounds (>6).

Staining was performed as described in [17]. The coverslips with sections were

mounted using SlowFade Gold antifade with DAPI (Invitrogen, Carlsbad CA). To

elute the applied antibodies, the mounting medium was washed away with dH2O and

a solution of 0.2 M NaOH and 0.02% SDS in distilled water was applied for 20 min.

After an extensive wash with Tris buffer and distilled water, the coverslips were dried

and placed on a hot plate (60C) for 30 min.

The primary antibodies and their dilutions are listed in [2], Table 1. Only well

characterized commercial antibodies were used and they were evaluated specifically for

AT as described in Supplemental Experimental Procedures. For immunofluorescence,

Alexa Fluor 488, 594, and 647 secondary antibodies of the appropriate species, highly

preadsorbed (Invitrogen, Carlsbad CA) were used at a dilution 1:150. The sequence

of antibody application in the multiround staining is presented in [2], Table S1.

Sections were imaged on a Zeiss Axio Imager.Z1 Upright Fluorescence Microscope

with motorized stage and Axiocam HR Digital Camera as described in [17]. Briefly,

144

a tiled image of the entire ribbon of sections on a coverslip was obtained using a

10 objective and the MosaiX feature of the software. The region of interest was

then identified on each section with custom-made software and imaged at a higher

magnification with a Zeiss 63/1.4 NA Plan Apochromat objective, using the image-

based automatic focus capability of the software. The resulting stack of images was

exported to ImageJ, aligned using the MultiStackReg plugin and imported back into

the Axiovision software to generate a volume rendering. When a ribbon was stained

and imaged multiple times, the MultiStackReg plugin was used to align the stacks

generated from each successive imaging session with the first session stacks based on

the DAPI channel, then a second within-stack alignment was applied to all the stacks.

To reconstruct large volumes of tissue, we first used Zeiss Axiovision software to

stitch together the individual high-magnification image tiles and produce a single mo-

saic image of each antibody stain for each serial section in the ribbon. We created a

z stack of mosaic images for each fluorescence channel, and then grossly aligned the

stacks using the MultiStackReg plugin. Finally, to remove non-linear physical warping

introduced into the ribbons by the sectioning process, we used a second ImageJ plu-

gin, autobUnwarpJ (available at http://www.stanford.edu/enweiler), which adapts

an algorithm for elastic image registration using vector-spline regularization [98].

5.4.2 Normalization and background subtraction of volumet-

ric data

Before analyzing imaged volumes, we subtracted the background from each fluorescent

channel using a 10x10 pixel (1 µm2) rolling ball filter to remove systematic apunctal

background fluorescence, then normalized each slice of the stack without saturating

any pixels, such that the brightness histogram of each section was stretched as much

as possible without loss of information. No other image processing, including removal

145

of fluorescence due to foreign material, nonspecific staining, etc, was performed before

analysis.

5.4.3 Extraction of synaptic loci

To extract a list of putative synapse locations from raw volume data, we first identi-

fied individual synapsin puncta by convolving the synapsin channel with a 3x3x3 local

maxima filter; retaining all voxels with a brightness ≥ those of its 26-voxel neighbor-

hood. Then, we passed the synapsin maxima through a connected component filter

to reduce peak voxel clumps (caused by discretization of the fluorescence data) to

centroids, and discarded those below a deliberately low threshold (10% of the total

brightness range) as being too dim to represent a real synapse. What remained was

a list of putative synapse locations, or “synaptic loci,” so named for their central role

in later classification steps.

5.4.4 PCA image treatment

The color of each point in the PCA figure was determined by taking the extreme

outliers of the three clusters, determining their feature composition via multivariate

regression, taking the dot product of the feature weight vectors with the feature vector

of each locus, and assigning that to red, green or blue for the VGluT1, VGluT2, and

GABA clusters respectively. Colors were manually normalized to be of approximately

equal intensity, and synaptic loci not strongly represented in any of the three colors

were removed to better visualize cluster relationship.

5.4.5 Normalization of pairwise channel data

To produce the pairwise channel copresence map, for each marker pair (i, j) we cal-

culated the probability of co-occurrence Eij = Fi/N ∗ Fj/N , where Fi is the number

146

of loci found to be positive for i, and N is the number of total loci in the population.

Multiplying by N gives us the expected population, Eij. We compared this num-

ber with the observed population Fij using difference over sum normalization to find

the normalized pairwise relationship Rij = (Fij − Eij)/(Fij + Eij). These relation-

ships made pairwise comparisons easy to interpret, with one minor counter-intuitive

exception: markers which comprised a substantial proportion of the synaptic loci

population (VGluT1 and PSD95) had reduced values, even with themselves, owing

to their high Eij. To bring those into the same reference frame as the rest, we nor-

malized again using the reciprocal of the sum of the relationship identity reciprocals,

that is, Nij = Rij ∗ (1/Rii + 1/Rjj)/2. Finally, since the previous setup disrupted

negative relationship scaling such that the most negative pairs (VGluT1 vs GABAer-

gic markers) reached nearly -3.0, we multiplied the positive ratings by 3 to match

once more.

5.4.6 Perpendicularization of cortical data

To simplify the calculation of the cortical depth-dependent metrics used in Figure 5.5,

such that any given Y-value represented tissue at the same cortical depth, we needed

to correct a minor slant in the raw volume. We measured the degree of tissue slope

using the pial surface and the white matter/striatum boundary, and imposed an affine

transformation on the loci, linearly interpolating them to be level. The underlying

data and the features used to classify the loci were not changed as a result of this

process.

147

5.4.7 Software packages used

Image normalization, locus discovery and feature extraction were implemented and

performed using Fiji (http://pacific.mpi-cbg.de/). Training set generation was im-

plemented as a browser-based application, coded in Python, to permit our experts to

work at their leisure. We used R for interactive classification for its ease of Python

integration, but the final random forest classifiers, trained on the complete training

set alone, used MATLAB (the TreeBagger class). Imaris was used to render the data

for visualization of Figure 5.1.

All implemented code used in this analysis is available at http://stanford.edu/ebbusse/.

148

Table 5.1: Machine Learning Algorithm Comparison

LDA QDA NB NBkd RFE kNN SVMGlut 0.128 0.114 0.110 0.104 0.084 0.176 0.222

GABA 0.044 0.036 0.062 0.052 0.036 0.070 0.178

Comparison of various supervised machine learning algorithms. A smalltraining set was used to compare the error rates of multiple MLAs when classifyingglutamatergic and GABAergic synapses in an early data set. From left to right:Linear Discriminant Analysis (LDA); Quadratic Discriminant Analysis (QDA);Naive Bayesian filter, gaussian distribution assumption (NB); Naive Bayesian filter,normalized kernel distribution assumption (NBkd); Random Forest Ensemble(RFE); k-means clustering (kNN); Support Vector Machine (SVN). k-meansclustering, an unsupervised clustering method, was included for comparison’s sake.

Table 5.2: Estimated error rates

Channel OOB ErrorGAD 0.0400VGAT 0.0767PV 0.0814VGluT3 0.1146VGluT2 0.0716VGluT1 0.0690PSD95 0.1215VAChT 0.1394TH 0.0333

Out-of-bag (OOB) error rate estimates for various classified markers.Order of markers is the same as in Figure 5.6. Each classifier had a minimumtraining set size of one hundred examples.

149

Figure 5.1: The synaptogram as a tool for high-dimensional proteomic visu-alization. (A) A maximum projected volume of Synapsin I labeling. 41 slices, 70nmper slice, total thickness of 2.87 µm. (B) Randomly-colored segmentation of individ-ual synapsin puncta. (C) Rendering of a single punctum from the volume showingsynapsin (white), imaged together with VGluT1 (red), PSD95 (green), GluR2 (blue),GAD (magenta) and VGAT (magenta). From top to bottom: all proteomic mark-ers, glutamatergic presynaptic labels, glutamatergic postsynaptic labels, GABAergiclabels. This appears to be a glutamatergic synapse. (D) The synaptogram derivedfrom the same synapse. Synapsin, top row, is repeated in red for the rest to providespatial context. Not shown, sixteen other colors and two redundant labels (synapsinand VGluT1). Scale bar: 5 um, size of synaptogram/render volume, 1100 nm x 1100nm x 630 nm

150

Figure 5.1

151

Figure 5.2: Comparison of human rating to machine learning. (A) Accuracyrates. i-vi - When compared against the average decisions of their peers in a VGluT1synapse discrimination task, humans performed at different accuracy levels basedon their stringency of classification. vii - The random forest ensemble, (VGluT1∩ PSD95), trained by human rater 1, performed comparably to the humans. (B)Receiver operating characteristics (ROC) curve, for VGluT1 and PSD95 classifica-tions on human-rated data. The ROC curve describes the tradeoff between reducingfalse positives (left side of the curve) and maximizing true positives (right side of thecurve). The displayed diagonal line represents chance, with better classifiers occupy-ing large areas between the diagonal and their own curves. A perfect classifier wouldhave no rounded corner; there would be no need to compromise. In this case, ourclassifiers perform well.

152

Figure 5.2

153

Figure 5.3: Unsupervised clustering of synapsin I imaged with array tomog-raphy. When the first and third principle components of the local brightness featureeq 1 are plotted against each other, they form clusters identifiable as known synapticsubtypes.

154

Figure 5.3

155

Figure 5.4: Feature importance for different molecular labels. (A) When allclasses were averaged, our local brightness feature (ii) saw the most use, followed byintegrated brightness (i), center of mass (iii) and moment of inertia (iv). (B-J) GAD,VGAT, PV, VGluT3, VGluT2, VGluT1, PSD95, VAChT, and TH (respectively)each make slightly different use of the feature set. VGluT3, VGluT2, and VAChTare notable in that they rely most heavily on features other than local brightness.

156

Figure 5.4

157

Figure 5.5: Density and size of synapse classes as a function of depth throughthe cortex. (A) Synapse density through the cortex. ** - VGluT2 synapses peak inlayer IV. PV-positive GABAergic synapse density is slightly decreased in layer I, andsignificantly lacking in layer VI. (B) Synapse size estimated using the synapsin localbrightness measurement. * - VGluT1 size peaks in layer Va (p < 0.05).

158

Figure 5.5

159

Figure 5.6: Positive and negative pairwise channel copresence. Symbols de-note interesting comparisons with statistical significance of p < 0.001. Red squaresrepresent label pairs which are copresent more than expected, blue squares less thanexpected by chance. * - GABAergic markers are copresent with each other, but avoidglutamatergic and TH markers. ˆ - VGluT1/2 are copresent with PSD95, but notwith each other. # - VGluT3 is present with all three GABAergic markers, butavoids VGluT1 and PSD95. & - VGluT2 shows some presence with TH. e - TH tendsto avoid VAChT.

160

Figure 5.6

161

Chapter 6

Future Directions and Conclusions

For the past few years the image processing tools required by AT have been sub-

stantially refined through the efforts of our lab and others. The algorithms I have

presented here, though initially developed in completely different programming en-

vironments, compose the seed around which a unified image processing pipeline is

precipitating. This is largely thanks to a strong emphasis on open source software,

which encourages the quick adoption and modification of useful third party code.

This process is still ongoing, but I have every confidence that in the near future we

will have a single piece of AT software capable of performing every image processing

step required to go from ribbon to volume.

One thing which will require additional attention is the future state of array

tomographic analysis. In this dissertation I have presented solutions to two analyt-

ical problems: one of label validation and one of synapse quantification. These are

powerful tools which can facilitate a wide array of biological experiments, but they

barely scratch the surface of the hypotheses AT can potentially address. For ex-

ample, without even leaving the context of proteomics, we have previously used the

close apposition of synapses and labeled dendrites as evidence of connections [2]. This

tells us something of the population of synapses onto that dendrite, but not with the

162

confidence necessary to positively identify the specific synapses in an specific circuit.

That would require a segmentation-based approach to synapse discovery, where we

segregate the pre- and postsynaptic sides of the cleft, and look for whatever the cir-

cuit has been labeled with in the pre- and postsynaptic boutons. Those circuits also

can be traced with AT’s resolution, barring some axonal processes, but suffer from a

similar problem as synapse identification: there’s too much material to be segmented

for manual methods to be efficient and no current automated solutions (most all of

which are closed source) perform acceptably. Both of these problems can probably

be solved with similar expenditures of time and effort as the synapse quantification I

have presented, but they too are but two specialized analytical methods.

I would argue that AT could stand to benefit more from a more general-purpose

approach to analytics, akin to that which ImageJ offers for image processing [131].

The strength of ImageJ (and its descendant, FIJI) lies in enabling the quick adap-

tation of image processing capabilities through its open source framework. Likewise,

a general open-source analytics package should be structured to encourage users to

extend its segmentation and classification functionality to address similar but subtly

different applications without forcing them to reimplement algorithms from scratch.

This should make it simpler to then apply associative feature measurements like Sholl

analysis, freeing up developer time for more substantive science.

As for the field of Neuroscience as a whole, the next decade is likely to see a

pressing need for exactly the kind of large-scale biological imaging AT excels at.

Computational modeling environments like the Blue Brain project [132] are already

capable of simulating hundreds of thousands of NEURON models with performance

sufficient to allow for in silico experimentation [133], and owing to the parallel archi-

tecture of the brain this capability will probably increase in the near future at about

the same exponential rate as general computing power. Given how much is currently

unknown about molecular physiology at the circuit level, projects like Blue Brain will

163

need to rely on gross and almost certainly inaccurate assumptions unless and until

circuit-level connectomic or proteomic imaging can experimentally test them [134].

Array tomography is presently well-suited to address these sorts of questions, and is

on a trajectory which should presently make it a powerful platform for exploratory

science in its own right.

164

Bibliography

1. Micheva K, Smith S (2007) Array tomography: a new tool for imaging the

molecular architecture and ultrastructure of neural circuits. Neuron 55: 25–

36.

2. Micheva K, Busse B, Weiler N, O’Rourke N, Smith S (2010) Single-synapse

analysis of a diverse synapse population: proteomic imaging methods and

markers. Neuron 68: 639–53.

3. Lein E, Hawrylycz M, Ao N, Ayres M, Bensinger A, et al. (2007) Genome-wide

atlas of gene expression in the adult mouse brain. Nature 445: 168–76.

4. Grant S (2007) Toward a molecular catalogue of synapses. Brain research

reviews 55: 445–9.

5. Sheng M, Hoogenraad C (2007) The postsynaptic architecture of excitatory

synapses: a more quantitative view. Annual review of biochemistry 76: 823–

47.

6. Sassoe-Pognetto M, Frola E, Pregno G, Briatore F, Patrizi A (2011) Under-

standing the molecular diversity of gabaergic synapses. Frontiers in cellular

neuroscience 5: 4.

165

7. Gupta A, Wang Y, Markram H (2000) Organizing principles for a diversity

of gabaergic interneurons and synapses in the neocortex. Science (New York,

NY) 287: 273–8.

8. Hausser M, Spruston N, Stuart G (2000) Diversity and dynamics of dendritic

signaling. Science (New York, NY) 290: 739–44.

9. Staple J, Morgenthaler F, Catsicas S (2000) Presynaptic heterogeneity: Vive

la difference. News in physiological sciences: an international journal of phys-

iology produced jointly by the International Union of Physiological Sciences

and the American Physiological Society 15: 45–49.

10. Cherubini E, Conti F (2001) Generating diversity at gabaergic synapses.

Trends in neurosciences 24: 155–62.

11. Craig A, Boudin H (2001) Molecular heterogeneity of central synapses: affer-

ent and target regulation. Nature neuroscience 4: 569–78.

12. Cull-Candy S, Brickley S, Farrant M (2001) Nmda receptor subunits: diver-

sity, development and disease. Current opinion in neurobiology 11: 327–35.

13. Grabowski P, Black D (2001) Alternative rna splicing in the nervous system.

Progress in neurobiology 65: 289–308.

14. Huang Y, Bergles D (2004) Glutamate transporters bring competition to the

synapse. Current opinion in neurobiology 14: 346–52.

15. Mody I, Pearce R (2004) Diversity of inhibitory neurotransmission through

gaba(a) receptors. Trends in neurosciences 27: 569–75.

16. Grant S (2006) The synapse proteome and phosphoproteome: a new paradigm

for synapse biology. Biochemical Society transactions 34: 59–63.

166

17. Micheva K, O’Rourke N, Busse B, Smith S (2010) Array tomography: high-

resolution three-dimensional immunofluorescence. Cold Spring Harbor proto-

cols 2010: pdb.top89.

18. Newman G, Jasani B, Williams E (1983) A simple post-embedding system

for the rapid demonstration of tissue antigens under the electron microscope.

The Histochemical journal 15: 543–55.

19. Kalman R (1960) A new approach to linear filtering and prediction problems.

Journal of Basic Engineering 82: 35–45.

20. Thevenaz P, Ruttimann U, Unser M (1998) A pyramid approach to subpixel

registration based on intensity. IEEE transactions on image processing : a

publication of the IEEE Signal Processing Society 7: 27–41.

21. van Steensel B, van Binnendijk E, Hornsby C, van der Voort H, Krozowski

Z, et al. (1996) Partial colocalization of glucocorticoid and mineralocorticoid

receptors in discrete compartments in nuclei of rat hippocampus neurons.

Journal of cell science 109 ( Pt 4): 787–92.

22. Arellano J, Benavides-Piccione R, Defelipe J, Yuste R (2007) Ultrastructure of

dendritic spines: correlation between synaptic and spine morphologies. Fron-

tiers in neuroscience 1: 131–43.

23. De Camilli P, Cameron R, Greengard P (1983) Synapsin i (protein i), a nerve

terminal-specific phosphoprotein. i. its general distribution in synapses of the

central and peripheral nervous system demonstrated by immunofluorescence

in frozen and plastic sections. The Journal of cell biology 96: 1337–54.

167

24. Hilfiker S, Pieribone V, Czernik A, Kao H, Augustine G, et al. (1999)

Synapsins as regulators of neurotransmitter release. Philosophical transactions

of the Royal Society of London Series B, Biological sciences 354: 269–79.

25. Katz L, Shatz C (1996) Synaptic activity and the construction of cortical

circuits. Science (New York, NY) 274: 1133–8.

26. Knudsen E (2004) Sensitive periods in the development of the brain and be-

havior. Journal of cognitive neuroscience 16: 1412–25.

27. Huberman A, Feller M, Chapman B (2008) Mechanisms underlying develop-

ment of visual maps and receptive fields. Annual review of neuroscience 31:

479–509.

28. Stellwagen D, Shatz C (2002) An instructive role for retinal waves in the

development of retinogeniculate connectivity. Neuron 33: 357–67.

29. Butts D, Kanold P, Shatz C (2007) A burst-based ”hebbian” learning rule at

retinogeniculate synapses links retinal waves to activity-dependent refinement.

PLoS biology 5: e61.

30. Chen C, Regehr W (2000) Developmental remodeling of the retinogeniculate

synapse. Neuron 28: 955–66.

31. Hooks B, Chen C (2006) Distinct roles for spontaneous and visual activity in

remodeling of the retinogeniculate synapse. Neuron 52: 281–91.

32. Mooney R, Madison D, Shatz C (1993) Enhancement of transmission at the

developing retinogeniculate synapse. Neuron 10: 815–25.

33. Corriveau R, Huh G, Shatz C (1998) Regulation of class i mhc gene expression

in the developing and mature cns by neural activity. Neuron 21: 505–20.

168

34. Shatz C, Stryker M (1988) Prenatal tetrodotoxin infusion blocks segregation

of retinogeniculate afferents. Science (New York, NY) 242: 87–9.

35. Sretavan D, Shatz C, Stryker M (1988) Modification of retinal ganglion cell

axon morphology by prenatal infusion of tetrodotoxin. Nature 336: 468–71.

36. Zinkernagel R, Doherty P (1979) Mhc-restricted cytotoxic t cells: studies on

the biological role of polymorphic major transplantation antigens determin-

ing t-cell restriction-specificity, function, and responsiveness. Advances in

immunology 27: 51–177.

37. Schott E, Bonasio R, Ploegh H (2003) Elimination in vivo of developing t cells

by natural killer cells. The Journal of experimental medicine 198: 1213–24.

38. Takai T (2005) Paired immunoglobulin-like receptors and their mhc class i

recognition. Immunology 115: 433–40.

39. Syken J, Grandpre T, Kanold P, Shatz C (2006) Pirb restricts ocular-

dominance plasticity in visual cortex. Science (New York, NY) 313: 1795–800.

40. Purcell S, Wray N, Stone J, Visscher P, O’Donovan M, et al. (2009) Common

polygenic variation contributes to risk of schizophrenia and bipolar disorder.

Nature 460: 748–52.

41. Shi J, Levinson D, Duan J, Sanders A, Zheng Y, et al. (2009) Common variants

on chromosome 6p22.1 are associated with schizophrenia. Nature 460: 753–7.

42. Stefansson H, Ophoff R, Steinberg S, Andreassen O, Cichon S, et al. (2009)

Common variants conferring risk of schizophrenia. Nature 460: 744–7.

43. Nakamura A, Kobayashi E, Takai T (2004) Exacerbated graft-versus-host dis-

ease in pirb-/- mice. Nature immunology 5: 623–9.

169

44. Huh G, Boulanger L, Du H, Riquelme P, Brotz T, et al. (2000) Functional

requirement for class i mhc in cns development and plasticity. Science (New

York, NY) 290: 2155–9.

45. Frenkel M, Bear M (2004) How monocular deprivation shifts ocular dominance

in visual cortex of young mice. Neuron 44: 917–23.

46. Rossi F, Pizzorusso T, Porciatti V, Marubio L, Maffei L, et al. (2001) Require-

ment of the nicotinic acetylcholine receptor beta 2 subunit for the anatomical

and functional development of the visual system. Proceedings of the National

Academy of Sciences of the United States of America 98: 6453–8.

47. Tagawa Y, Kanold P, Majdan M, Shatz C (2005) Multiple periods of functional

ocular dominance plasticity in mouse visual cortex. Nature neuroscience 8:

380–8.

48. Antonini A, Fagiolini M, Stryker M (1999) Anatomical correlates of functional

plasticity in mouse visual cortex. The Journal of neuroscience: the official

journal of the Society for Neuroscience 19: 4388–406.

49. Iny K, Heynen A, Sklar E, Bear M (2006) Bidirectional modifications of visual

acuity induced by monocular deprivation in juvenile and adult rats. The

Journal of neuroscience: the official journal of the Society for Neuroscience

26: 7368–74.

50. Cang J, Kalatsky V, Lowel S, Stryker M (2005) Optical imaging of the intrinsic

signal as a measure of cortical plasticity in the mouse. Vis Neurosci 22: 685–

691.

51. Hubener M (2003) Mouse visual cortex. Current opinion in neurobiology 13:

413–20.

170

52. Prusky G, West P, Douglas R (2000) Behavioral assessment of visual acuity

in mice and rats. Vision research 40: 2201–9.

53. Robinson L, Bridge H, Riedel G (2001) Visual discrimination learning in the

water maze: a novel test for visual acuity. Behavioural brain research 119:

77–84.

54. Ljunggren H, Van Kaer L, Sabatine M, Auchincloss H, Tonegawa S, et al.

(1995) Mhc class i expression and cd8+ t cell development in tap1/beta 2-

microglobulin double mutant mice. International immunology 7: 975–84.

55. Goddard C, Butts D, Shatz C (2007) Regulation of cns synapses by neuronal

mhc class i. Proceedings of the National Academy of Sciences of the United

States of America 104: 6828–33.

56. McConnell M, Huang Y, Datwani A, Shatz C (2009) H2-k(b) and h2-d(b)

regulate cerebellar long-term depression and limit motor learning. Proceedings

of the National Academy of Sciences of the United States of America 106:

6784–9.

57. Loconto J, Papes F, Chang E, Stowers L, Jones E, et al. (2003) Functional

expression of murine v2r pheromone receptors involves selective association

with the m10 and m1 families of mhc class ib molecules. Cell 112: 607–18.

58. Oliveira A, Thams S, Lidman O, Piehl F, Hokfelt T, et al. (2004) A role for

mhc class i molecules in synaptic plasticity and regeneration of neurons after

axotomy. Proceedings of the National Academy of Sciences of the United


171

59. LeVay S, Stryker M, Shatz C (1978) Ocular dominance columns and their

development in layer iv of the cat’s visual cortex: a quantitative study. The

Journal of comparative neurology 179: 223–44.

60. Torborg C, Feller M (2004) Unbiased analysis of bulk axonal segregation pat-

terns. Journal of neuroscience methods 135: 17–26.

61. Stevens B, Allen N, Vazquez L, Howell G, Christopherson K, et al. (2007)

The classical complement cascade mediates cns synapse elimination. Cell 131:

1164–78.

62. Penn A, Riquelme P, Feller M, Shatz C (1998) Competition in retinogeniculate

patterning driven by spontaneous activity. Science (New York, NY) 279:

2108–12.

63. Bjartmar L, Huberman A, Ullian E, Renterıa R, Liu X, et al. (2006) Neuronal

pentraxins mediate synaptic refinement in the developing visual system. The

Journal of neuroscience: the official journal of the Society for Neuroscience

26: 6269–81.

64. McGee A, Yang Y, Fischer Q, Daw N, Strittmatter S (2005) Experience-driven

plasticity of visual cortex limited by myelin and nogo receptor. Science (New

York, NY) 309: 2222–6.

65. Mataga N, Nagai N, Hensch T (2002) Permissive proteolytic activity for visual

cortical plasticity. Proceedings of the National Academy of Sciences of the

United States of America 99: 7717–21.

66. Atwal J, Pinkston-Gosse J, Syken J, Stawicki S, Wu Y, et al. (2008) Pirb is a

functional receptor for myelin inhibitors of axonal regeneration. Science (New

York, NY) 322: 967–70.

172

67. Shatz C (2009) Mhc class i: an unexpected role in neuronal plasticity. Neuron

64: 40–5.

68. Dorfman J, Zerrahn J, Coles M, Raulet D (1997) The basis for self-tolerance

of natural killer cells in beta2-microglobulin- and tap-1- mice. Journal of

immunology (Baltimore, Md: 1950) 159: 5219–25.

69. Smith S (2007) Circuit reconstruction tools today. Current opinion in neuro-

biology 17: 601–8.

70. Koffie R, Meyer-Luehmann M, Hashimoto T, Adams K, Mielke M, et al. (2009)

Oligomeric amyloid β associates with postsynaptic densities and correlates

with excitatory synapse loss near senile plaques. Proc Natl Acad Sci 106:

4012–4017.

71. Feng G, Mellor R, Bernstein M, Keller-Peck C, Nguyen Q, et al. (2000) Imag-

ing neuronal subsets in transgenic mice expressing multiple spectral variants

of gfp. Neuron 28: 41–51.

72. Denk W, Horstmann H (2004) Serial block-face scanning electron microscopy

to reconstruct three-dimensional tissue nanostructure. PLoS biology 2: e329.

73. Harris K, Perry E, Bourne J, Feinberg M, Ostroff L, et al. (2006) Uniform serial

sectioning for transmission electron microscopy. The Journal of neuroscience

: the official journal of the Society for Neuroscience 26: 12101–3.

74. Knott G, Marchman H, Wall D, Lich B (2008) Serial section scanning electron

microscopy of adult brain tissue using focused ion beam milling. The Journal

of neuroscience: the official journal of the Society for Neuroscience 28: 2959–

64.

173

75. Anderson J, Jones B, Yang J, Shaw M, Watt C, et al. (2009) A computational

framework for ultrastructural mapping of neural circuitry. PLoS biology 7:

e1000074.

76. Kasthuri N, Lichtman J (2010) Neurocartography. Neuropsychopharmacol-

ogy: official publication of the American College of Neuropsychopharmacology

35: 342–3.

77. Fremeau R, Troyer M, Pahner I, Nygaard G, Tran C, et al. (2001) The ex-

pression of vesicular glutamate transporters defines two classes of excitatory

synapse. Neuron 31: 247–60.

78. De Gois S, Schafer M, Defamie N, Chen C, Ricci A, et al. (2005) Homeostatic

scaling of vesicular glutamate and gaba transporter expression in rat neocor-

tical circuits. The Journal of neuroscience : the official journal of the Society

for Neuroscience 25: 7121–33.

79. Graziano A, Liu X, Murray K, Jones E (2008) Vesicular glutamate trans-

porters define two sets of glutamatergic afferents to the somatosensory tha-

lamus and two thalamocortical projections in the mouse. The Journal of

comparative neurology 507: 1258–76.

80. Fritschy J, Harvey R, Schwarz G (2008) Gephyrin: where do we stand, where

do we go? Trends in neurosciences 31: 257–64.

81. Chaudhry F, Reimer R, Bellocchio E, Danbolt N, Osen K, et al. (1998) The

vesicular gaba transporter, vgat, localizes to synaptic vesicles in sets of glycin-

ergic as well as gabaergic neurons. The Journal of neuroscience : the official

journal of the Society for Neuroscience 18: 9733–50.

174

82. Micheva K, Beaulieu C (1995) An anatomical substrate for experience-

dependent plasticity of the rat barrel field cortex. Proceedings of the National

Academy of Sciences of the United States of America 92: 11834–8.

83. Jones E, Powell T (1969) Morphological variations in the dendritic spines of

the neocortex. Journal of cell science 5: 509–29.

84. Knott G, Quairiaux C, Genoud C, Welker E (2002) Formation of dendritic

spines with gabaergic synapses induced by whisker stimulation in adult mice.

Neuron 34: 265–73.

85. Jasinska M, Siucinska E, Cybulska-Klosowicz A, Pyza E, Furness D, et al.

(2010) Rapid, learning-induced inhibitory synaptogenesis in murine barrel

field. The Journal of neuroscience: the official journal of the Society for

Neuroscience 30: 1176–84.

86. Dehay C, Douglas R, Martin K, Nelson C (1991) Excitation by geniculocortical

synapses is not ’vetoed’ at the level of dendritic spines in cat visual cortex.

The Journal of physiology 440: 723–34.

87. Kubota Y, Hatada S, Kondo S, Karube F, Kawaguchi Y (2007) Neocortical

inhibitory terminals innervate dendritic spines targeted by thalamocortical

afferents. The Journal of neuroscience: the official journal of the Society for


88. Mandell J, Townes-Anderson E, Czernik A, Cameron R, Greengard P, et al.

(1990) Synapsins in the vertebrate retina: absence from ribbon synapses and

heterogeneous distribution among conventional synapses. Neuron 5: 19–33.

175

89. Kielland A, Erisir A, Walaas S, Heggelund P (2006) Synapsin utilization differs

among functional classes of synapses on thalamocortical cells. The Journal of

neuroscience: the official journal of the Society for Neuroscience 26: 5786–93.

90. Bragina L, Candiracci C, Barbaresi P, Giovedı S, Benfenati F, et al. (2007)

Heterogeneity of glutamatergic and gabaergic release machinery in cerebral

cortex. Neuroscience 146: 1829–40.

91. Grønborg M, Pavlos N, Brunk I, Chua J, Munster-Wandowski A, et al. (2010)

Quantitative comparison of glutamatergic and gabaergic synaptic vesicles un-

veils selectivity for few proteins including mal2, a novel synaptic vesicle pro-

tein. The Journal of neuroscience : the official journal of the Society for


92. Nakamura K, Watakabe A, Hioki H, Fujiyama F, Tanaka Y, et al. (2007)

Transiently increased colocalization of vesicular glutamate transporters 1 and

2 at single axon terminals during postnatal development of mouse neocortex:

a quantitative analysis with correlation coefficient. The European journal of

neuroscience 26: 3054–67.

93. LeVay S, Gilbert C (1976) Laminar patterns of geniculocortical projection in

the cat. Brain research 113: 1–19.

94. Luscher C, Isaac J (2009) The synapse: center stage for many brain diseases.

The Journal of Physiology 587: 727–9.

95. Hofer S, Mrsic-Flogel T, Bonhoeffer T, Hubener M (2009) Experience leaves

a lasting structural trace in cortical circuits. Nature 457: 313–7.

176

96. Xu T, Yu X, Perlik A, Tobin W, Zweig J, et al. (2009) Rapid formation and

selective stabilization of synapses for enduring motor memories. Nature 462:

915–9.

97. Yang G, Pan F, Gan W (2009) Stably maintained dendritic spines are associ-

ated with lifelong memories. Nature 462: 920–4.

98. Arganda-Carreras I, Sorzano C, Marabini R, Carazo J, Ortiz-de Solorzano C,

et al. (2006) Consistent and elastic registration of histological sections using

vector-spline regularization. Springer, 85–95 pp.

99. Lewis J (1995) Fast template matching. Canadian Image Processing and

Pattern Recognition Society, pp. 120–123.

100. Turrigiano G, Nelson S (2004) Homeostatic plasticity in the developing ner-

vous system. Nature reviews Neuroscience 5: 97–107.

101. Citri A, Malenka R (2008) Synaptic plasticity: multiple forms, functions, and

mechanisms. Neuropsychopharmacology : official publication of the American

College of Neuropsychopharmacology 33: 18–41.

102. Garner C, Zhai R, Gundelfinger E, Ziv N (2002) Molecular mechanisms of cns

synaptogenesis. Trends in neurosciences 25: 243–51.

103. Hioki H, Fujiyama F, Nakamura K, Wu S, Matsuda W, et al. (2004) Chemi-

cally specific circuit composed of vesicular glutamate transporter 3- and pre-

protachykinin b-producing interneurons in the rat neocortex. Cerebral cortex

(New York, NY : 1991) 14: 1266–75.

104. Gutierrez R (2005) The dual glutamatergic-gabaergic phenotype of hippocam-

pal granule cells. Trends in neurosciences 28: 297–303.

177

105. Briggman K, Helmstaedter M, Denk W (2011) Wiring specificity in the

direction-selectivity circuit of the retina. Nature 471: 183–8.

106. Anderson J, Jones B, Watt C, Shaw M, Yang J, et al. (2011) Exploring the

retinal connectome. Molecular vision 17: 355–79.

107. Li L, Tasic B, Micheva K, Ivanov V, Spletter M, et al. (2010) Visualizing the

distribution of synapses from individual neurons in the mouse brain. PloS one

5: e11503.

108. Kuhlman S, Huang Z (2008) High-resolution labeling and functional manip-

ulation of specific neuron types in mouse brain by cre-activated viral gene

expression. PloS one 3: e2005.

109. Katz Y, Menon V, Nicholson D, Geinisman Y, Kath W, et al. (2009) Synapse

distribution suggests a two-stage model of dendritic integration in ca1 pyra-

midal neurons. Neuron 63: 171–7.

110. Davie J, Clark B, Husser M (2008) The origin of the complex spike in cerebellar

purkinje cells. The Journal of neuroscience : the official journal of the Society

for Neuroscience 28: 7599–609.

111. Bayes A, Grant S (2009) Neuroproteomics: understanding the molecular or-

ganization and complexity of the brain. Nature reviews Neuroscience 10:

635–46.

112. Dani A, Huang B, Bergan J, Dulac C, Zhuang X (2010) Superresolution imag-

ing of chemical synapses in the brain. Neuron 68: 843–56.

113. De Camilli P, Cameron R, Greengard P (1983) Synapsin 1 (protein 1), a nerve

terminal-specific phosphoprotein. Journal of Cell Biology 96: 1337–1354.

178

114. Evans N, Facci L, Owen D, Soden P, Burbidge S, et al. (2008) Abeta(1-42)

reduces synapse number and inhibits neurite outgrowth in primary cortical

and hippocampal neurons: a quantitative analysis. Journal of neuroscience

methods 175: 96–103.

115. Jackson JE (1991) A User’s Guide to Principal Components. John Wiley and

Sons.

116. Belkacem-Boussaid K, Pennell M, Lozanski G, Shana’ah A, Gurcan M (2010)

Computer-aided classification of centroblast cells in follicular lymphoma. An-

alytical and quantitative cytology and histology / the International Academy

of Cytology [and] American Society of Cytology 32: 254–260.

117. Breiman L (2001) Random forests. Machine Learning 45: 5-3-2.

118. Settles B (2009) Active learning literature survey. Computer Sciences Tech-

nical Report 1648, University of Wisconsin–Madison.

119. Bogen I, Boulland J, Mariussen E, Wright M, Fonnum F, et al. (2006) Absence

of synapsin i and ii is accompanied by decreases in vesicular transport of

specific neurotransmitters. Journal of neurochemistry 96: 1458–66.

120. Schuz A, Palm G (1989) Density of neurons and synapses in the cerebral

cortex of the mouse. The Journal of comparative neurology 286: 442–55.

121. Lawton D, Oswald W, McClure J (1995) The biological reality of the interlacu-

nar network in the embryonic, cartilaginous, skeleton: a thiazine dye/absolute

ethanol/lr white resin protocol for visualizing the network with minimal tissue

shrinkage. Journal of microscopy 178: 66–85.

179

122. Gonchar Y, Wang Q, Burkhalter A (2007) Multiple distinct subtypes of

gabaergic neurons in mouse visual cortex identified by triple immunostain-

ing. Frontiers in neuroanatomy 1: 3.

123. Fremeau R, Burman J, Qureshi T, Tran C, Proctor J, et al. (2002) The identi-

fication of vesicular glutamate transporter 3 suggests novel modes of signaling

by glutamate. Proceedings of the National Academy of Sciences of the United


124. Fremeau R, Voglmaier S, Seal R, Edwards R (2004) Vgluts define subsets of

excitatory neurons and suggest novel roles for glutamate. Trends in neuro-

sciences 27: 98–103.

125. Fremeau R, Kam K, Qureshi T, Johnson J, Copenhagen D, et al. (2004) Vesic-

ular glutamate transporters 1 and 2 target to functionally distinct synaptic

release sites. Science (New York, NY) 304: 1815–9.

126. Dal Bo G, St-Gelais F, Danik M, Williams S, Cotton M, et al. (2004)

Dopamine neurons in culture express vglut2 explaining their capacity to re-

lease glutamate at synapses in addition to dopamine. Journal of neurochem-

istry 88: 1398–405.

127. Kawano M, Kawasaki A, Sakata-Haga H, Fukui Y, Kawano H, et al. (2006)

Particular subpopulations of midbrain and hypothalamic dopamine neurons

express vesicular glutamate transporter 2 in the rat brain. The Journal of

comparative neurology 498: 581–92.

128. Fu M, Zuo Y (2011) Experience-dependent structural plasticity in the cortex.

Trends in neurosciences 34: 177–87.

180

129. White JG, Southgate E, Thomson JN, Brenner S (1986) The structure of

the nervous system of the nematode caenorhabditis elegans. Philosophical

Transactions of the Royal Society B: Biological Sciences 314: 1–340.

130. Cardona A, Saalfeld S, Arganda I, Pereanu W, Schindelin J, et al. (2010) Iden-

tifying neuronal lineages of drosophila by sequence analysis of axon tracts. The

Journal of neuroscience : the official journal of the Society for Neuroscience

30: 7538–53.

131. Abramoff M, Magelhaes P, Ram S (2004) Image processing with imagej. Bio-

photonics International 11: 36–42.

132. Markram H (2006) The blue brain project. Nature reviews Neuroscience 7:

153–60.

133. Migliore M, Cannia C, Lytton W, Markram H, Hines M (2006) Parallel net-

work simulations with neuron. Journal of computational neuroscience 21:

119–29.

134. Mishchenko Y, Hu T, Spacek J, Mendenhall J, Harris K, et al. (2010) Ultra-

structural analysis of hippocampal neuropil from the connectomics perspec-

tive. Neuron 67: 1009–20.

181

Documents

PROTEOMIC SINGLE-SYNAPSE ANALYSIS WITH ARRAY …tq652pp7498/thesis-augmented.pdfIntroduction Array tomography (AT) is a high-resolution proteomic imaging method that exploits a combination