Grouping and Classifying Electrophysiologically-Defined Classes of Neocortical Neurons by Single Cell, Whole-Genome Expression Profiling

The diversity of neuronal cell types and how to classify them are perennial questions in neuroscience. The advent of global gene expression analysis raised the possibility that comprehensive transcription profiling will resolve neuronal cell types into groups that reflect some or all aspects of their phenotype. This approach has been successfully used to compare gene expression between groups of neurons defined by a common property. Here we extend this approach to ask whether single neuron gene expression profiling can prospectively resolve neuronal subtypes into groups, independent of any phenotypic information, and whether those groups reflect meaningful biological properties of those neurons. We applied methods we have developed to compare gene expression among single neural stem cells to study global gene expression in 18 randomly picked neurons from layer II/III of the early postnatal mouse neocortex. Cells were selected by morphology and by firing characteristics and electrical properties, enabling the definition of each cell as either fast- or regular-spiking, corresponding to a class of inhibitory interneurons or excitatory pyramidal cells. Unsupervised clustering of young neurons by global gene expression resolved the cells into two groups and those broadly corresponded with the two groups of fast- and regular-spiking neurons. Clustering of the entire, diverse group of 18 neurons of different developmental stages also successfully grouped neurons in accordance with the electrophysiological phenotypes, but with more cells misassigned among groups. Genes specifically enriched in regular spiking neurons were identified from the young neuron expression dataset. These results provide a proof of principle that single-cell gene expression profiling may be used to group and classify neurons in a manner reflecting their known biological properties and may be used to identify cell-specific transcripts.


INTRODUCTION
The question of the diversity of neuron types is a challenging contemporary problem in neuroscience . It is generally held that a defi nitive parts list is essential to the accurate description of neuronal circuits and their potential functions (Bota and Swanson, 2007). However, the degree of neuronal diversity and its functional importance are areas of much debate, with a range of views on the degree to which neurons differ and the extent to which they should be grouped (see for example, Masland and Raviola, 2000). Much of the diffi culty in defi ning neuronal cell types has stemmed from the absence of agreement over which neuronal properties can and should be used to defi ne a neuronal cell type. Current and previous strategies for classifying neurons in any region of the central nervous system (CNS) rely on the use of one or more neuronal property, such as morphology, neurotransmitters, passive and active electrical properties, and expression of small numbers of molecular markers, typically proteins (Ascoli et al., 2008).
The use of those approaches is exemplifi ed by the concerted effort by many researchers to defi ne the diversity of GABAergic interneurons in the cerebral cortex (Ascoli et al., 2008). Despite labelling from targets was used to isolate two classes of layer V neurons, enabling the identifi cation of cell-type specifi c genes by expression profi ling (Arlotta et al., 2005). In a complementary approach, twelve mouse forebrain neuronal populations were isolated by cell-specifi c transgene expression and expression profi led to identify cell-specifi c gene expression combinations . A taxonomy of cell types based solely on the gene expression data correctly refl ected the expected relationships between the different populations, confi rming that global expression profi ling can correctly identify neuronal subtypes .
Here we extend this approach to ask whether single neuron gene expression profi ling can prospectively resolve neuronal subtypes into groups, independent of any phenotypic information, and whether those groups refl ect meaningful biological properties of those neurons. To do so, we applied methods we developed to compare gene expression among single neural stem cells Subkhankulova et al., 2008) to study global gene expression between single neurons from layer II/III of mouse neocortex (Yano et al., 2006).
In proof of principle experiments, we fi nd that global expression profi ling of single cortical neurons can be used prospectively to group and classify neurons in a manner refl ecting their physiological properties. The potential advantage of this approach is that it suggests that deep sampling of complex neuronal populations by single cell expression profi ling may be a useful approach for defi ning neuronal complexity in the CNS. However, in this case we used electrophysiology and morphology to restrict the diversity of the population analysed. It will be of interest in future studies to carry out single cell expression studies of unselected neuronal populations to assess whether it is possible to use such data to generate biologically meaningful groupings of neuronal types.
Single cortical neurons were identifi ed in layer II/III of the slices using an infrared differential interference contrast microscope (Olympus BX50WI) and their electrophysiological properties were obtained by whole cell patch clamping as described (Tateno et al., 2005), with the addition of recombinant RNase inhibitors to the pipette solution. Somatic patch-pipette recordings were made with a Multiclamp 700-A amplifi er (Axon Instruments, Foster City, CA, USA) in current-clamp mode, correcting for prenulled liquid junction potential. Whole cell recording pipettes (Clark GC150T-7.5) of resistance 3.9-4.3 MΩ were fi lled with the following intracellular solution, based on that of Monyer and Jonas (1995): 140 mM KCl, 5 mM EGTA, 3 mM MgCl2, 5 mM HEPES, pH adjusted to 7.3 with KOH; in this case 200 units/ml of SUPERase-In RNase inhibitor (Ambion, Austin Texas, USA) were added and KCl concentration reduced to 125-130 mM to achieve the same fi nal osmolarity as the Monyer and Jonas solution. Neurons for RNA harvesting were selected by morphology and on their fi ring properties in order to restrict the range of cell types sampled for gene expression profi ling. Recordings were carried out at room temperature to reduce RNase activity. Signals were low-pass fi ltered at 5 kHz (4-pole Bessel fi lter) and sampled at 20 kHz with 12-bit resolution, using a National Instruments analog card, and custom software (MatDAQ, Hugh Robinson, 1995-2010, written in Matlab and C. Stimulus generation was also at a 20 kHz rate, with 12-bit resolution. All analysis was performed in Matlab (Mathworks, Natick, MA, USA). Classifi cation of cells as regular spiking (RS) or fast spiking (FS) was based quantitatively on their clearly different spike shapes, and agreed with other features such as maximum spike rate and the qualitative pattern of spike fi ring at threshold (Tateno et al., 2004), and the pyramidal (RS) and nonpyramidal (FS) morphology under infrared differential interference contrast optics. Action potentials from responses just above threshold were convolved with the fi rst derivative of a Gaussian function: G'(t) = −t exp[−t 2 /(2σ 2 )], where σ = 1.3 ms, to produce a smoothed fi rst-derivative of the action potential voltage trajectory. This was then aligned by the fi rst negative-going zero-crossing (corresponding to the AP peak), and the interval to the second, positive-going zero-crossing (corresponding to the interspike membrane potential minimum) was measured. FS cells are defi ned as showing intervals less than 10 ms, while RS cells had intervals of 20 ms or greater.

SINGLE CELL TOTAL mRNA AMPLIFICATION AND EXPRESSION PROFILING
The cytoplasmic contents of each neuron were harvested after recording by aspiration into the patch pipette and deposited in a total volume of 5 µl. Total RNA was isolated from the cytoplasmic contents by the acid phenol-guanidinum extraction method (Trizol, Sigma; (Chomczynski and Sacchi, 1987). Total RNA was precipitated in ethanol in the presence of 100 pg polyinosinic acid (Sigma) and 1 µl of linear polyacrylamide (Ambion) as carriers to reduce the loss of low copy number mRNAs. Total mRNA from each of 18 neuronal cells was globally amplifi ed using a PCR-based technique (Brady and Iscove, 1993;Brady et al., 1995;Iscove et al., 2002) with minor variations . PCR products were purifi ed with the CyScribe GFX Purifi cation kit (Amersham Bioscience) and labeled with Cy3/Cy5-modifi ed dCTP using Klenow DNA polymerase (BD Bioscience). The labeled samples were hybridized against a reference of amplifi ed cDNA pooled from all of the neurons studied. Gene expression was studied using mouse oligonucleotide arrays printed in-house (65-mer, 23232 oligonucleotides representing ∼20,000 genes; Compugen oligonucleotide library) in pairs of dye-swapped hybridisations, as previously described Subkhankulova et al., 2008). Images of hybridised arrays were gathered on a microarray scanner (Axon Instruments GenePix). Data from single microarrays were extracted using the GenePix analysis package. The mean intensity of each feature (spot) and the local background at 532 nm (Cy3) and 635 nm (Cy5) were extracted. The background-corrected 635 nm/532 nm ratio was calculated for each feature as the initial measurement of relative mRNA levels for each gene represented on the array.

CLASSIFICATION OF NEURONS INTO RS OR FS CELL TYPES FOR EXPRESSION PROFILING
The goal of this study was to explore whether the global expression profi les of neuronal cells could be used to prospectively cluster neurons into groups that refl ected their neuronal fi ring properties. For this purpose, we chose to expression profi le neurons in two broad categories of morphologically and electrophysiologically-defi ned cortical neurons: regular-spiking cells of pyramidal morphology, corresponding to glutamatergic, excitatory pyramidal neurons (Connors and Gutnick, 1990); and fast-spiking, non-pyramidal cells, corresponding to GABAergic inhibitory interneurons of basket morphology that express parvalbumin (Kawaguchi and Kondo, 2002). To do so, we analysed 18 cells from layer II/III of mouse neocortex harvested between postnatal days 5 (P5) and P20: 10 young neurons (all P13 and older) and 8 immature neurons (aged between P5 and P10). The fi ring properties of this set of neurons were measured individually for each cell, following which the cytoplasm of each cell was harvested for global mRNA amplifi cation (Figure 1).
The set of all recorded cells contained 7 RS and 11 FS cells. RS cells showed a pyramidal morphology, while FS cells generally had a nonpyramidal, multipolar morphology under infrared differential interference contrast microscopy (Connors et al., 1982). As in a previous study (Tateno et al., 2004), regular-spiking neurons (Figure 2A) showed an ability to fi re repetitively at very low frequencies (<10 Hz), a relatively slow spike waveform and lacked a pronounced after hyperpolarization. Fast-spiking neurons ( Figure 2B), in contrast showed a very fast action potential waveform, a largeamplitude, sharp after hyperpolarization, stable periodic fi ring only at a relatively high frequency (>10 Hz), and usually a high maximal fi ring rate. Depolarization block was pronounced at higher current levels in some recordings in each type of cell, potentially as a consequence of the high levels of RNase inhibitor, which is not a normal component of whole-cell pipette solutions.
In addition to qualitatively matching the well-known RS and FS fi ring patterns, we distinguished the two cell types on a quantitative basis, using the smoothed fi rst-derivative of action potentials elicited just above threshold (see Materials and Methods for details). The interval between zero-crossings of this derivative, corresponding to the interval between the peak of the action potential and the deepest point of the after-hyperpolarization was tightly clustered (Figures 2C,D), below 10 ms for FS cells (blue), while it ranged from 20 to 60 ms for RS cells (red).
Neurons studied here demonstrated some diversity in fi ring patterns, however all could be clearly defi ned as FS or RS spiking types (Figure 3; see Table S1 in Supplementary Material for details). Particularly at young ages, depolarization block is pronounced and can obscure the typical adult fi ring pattern, but the spike shape and the pattern of spiking within a certain window of stimulus currents leading to periodic fi ring, enabled a quite robust classifi cation of such cells.
After analysis and recording of the electrophysiological properties of each neuron, the cytoplasmic contents of each cell were harvested. Total RNA was purifi ed from each cell, reverse transcribed to cDNA and globally amplifi ed (see Materials and Methods for details). Typically, 32-34 cycles of PCR produced 15-20 µg of amplifi ed cDNA. Prior to global expression profi ling, we further

SINGLE CELL qRT-PCR
Single cell qRT-PCR was carried out using the globally amplifi ed single cell cDNA as a template. Each real time PCR mix contained was carried out in a 10 µl fi nal volume, containing 0.2 µl of each primer (10 mM) and 5 µl 2 × Master mix (DyNAmo Capillary SYBR Green qPCR Kit, Finnzymes). Intron-spanning primers were designed to Emx1, a pyramidal neuron-specifi c transcript (Chan et al., 2001), and GAD65/GAD2, a GABAergic interneuron-specifi c transcript (Katarova et al., 2000), using Primer3 software 1 , based on gene transcript information from the Ensembl genome browser 2 . Primer sequences were: Emx2, GAGCGAGCCTTTGAGAAGAA and CTGGAACCACACCTT CACCT; GAD65/GAD2, AGATCGCCCCTGTATTTGTG and GCATGGCATACATGTTGGAG. Real-time PCR was performed in a LightCycler (Roche Diagnostics) according to DyNAmo Capillary SYBR Green qPCR protocol. Ct values were determined using the maximum second derivative function in the LightCycler software (Roche Diagnostics). Generation of PCR products was confi rmed by melting curve analysis and gel electrophoresis. qPCR results were classifi ed as absent (−) if Ct values were higher than 35 and no single peak was observed from the melting curve, detectable above background (+/−) if a single peak was observed from the melting curve and Ct values were lower than 34, and expressed signifi cantly above background (+) if a single melting curve peak was present and Ct values were lower than 30.

ANALYSIS OF MICROARRAY DATA
All microarray data were analysed using the Acuity platform (Axon Instruments/Molecular Dynamics). Expression data from each array were normalized by locally-weighted regression (lowess) analysis. For data analysis of each group of arrays, data were fi ltered to remove low intensity features, features reporting gene expression in less than two thirds of the arrays in the set and features that did not show evidence for enriched expression in at least three cells in a set (empirically set at a 1.5-fold enrichment). Genes and arrays were pair-wise, hierarchically clustered using Pearson's correlation coeffi cients as similarity metrics. Similar results were obtained using alternative hierarchical clustering metrics (Euclidean distance, for example), with minor variations in the assignment of cells within clusters. Hierarchical cluster dendrograms were built by iterative, pair-wise similarity comparisons, using Pearson's correlation coeffi cients.
To identify genes with robust differential enrichment between young fast-and regular-spiking neurons, Student's t-testing was used to identify genes differentially expressed between the two groups (p < 0.05). Expression ratios of genes identifi ed as significantly differing in expression between the groups were averaged to rank genes for enrichment in RS or FS cells, and those groups were visualized by hierarchical clustering. In situ hybridization data for adult expression of genes identifi ed as enriched in each cell class were extracted from the Allen Brain Atlas' mouse brain database (mouse.brain-map.org) (Lein et al., 2007). two dye-swap hybridisations. Statistical analysis found that the expression of the abundant transcript GAPDH, printed on the arrays more then 200 times, did not vary signifi cantly among the cells (data not shown), confi rming the overall quality of the combined procedures of RNA harvesting, amplifi cation and microarray hybridisation.
To test the ability of global gene expression profi ling to group and distinguish well-characterised types of neurons, we used the set of 10 young cells (P13-22) to examine if the classifi cation of neurons by global expression analysis corresponded to the physiological classifi cation of RS and FS cells. Cells were assigned to groups by unsupervised clustering according to the degree of similarity between pairs of cells (as calculated by Pearson correlation coeffi cients; Figure 4). The cells formed two distinct groups of four and six cells each, as defi ned by gene expression (Figure 4). Those two groups corresponded to the two electrophysiologically-defi ned groups, with the three RS cells forming one branch of the overall dendrogram, with one misplaced FS neuron, and a set of fi ve FS cells plus one RS cell forming the other. Within the FS sub-tree there was a high degree of diversity of expression, although there were some cells that were highly similar in gene expression (cells 14, 15 and 19, Figure 4).

SINGLE-CELL EXPRESSION PROFILING PROSPECTIVELY CLASSIFIED RS AND FS NEURONS AT DIFFERING DEVELOPMENTAL STAGES
To study whether gene expression could group and classify a more complex population of neurons, we proceeded to use unsupervised cluster analysis to analyse the entire group of 18 layer II/III neurons ranging in age from P5 to P22. Based on fi ring properties, all neurons in this set could be designated as either RS or FS, but with some variation in their developmental stage, as refl ected in their electrical properties (Figure 3; Table S1 in Supplementary Material).
investigated the expression of cell-specifi c transcripts in the single cells by quantitative PCR (qPCR) analysis, using the amplifi ed cDNA as a template. Quantitative PCR for the pyramidal cellspecifi c transcription factor Emx1 and for the GABAergic neuronspecifi c enzyme glutamic acid decarboxylase-2 (GAD2/GAD65) was used to distinguish between the two cell types (Figure 3). For example, cells 9 and 13 were defi ned as RS by fi ring properties (Table S1 in Supplementary Material) and expressed signifi cant levels of Emx1 and little or no GAD65 mRNA (Figure 3). Similarly, cell 12 was classed as FS by fi ring properties and qPCR found that this cell was GAD65-positive and Emx1-low.
However, there are notable exceptions in which expression of these markers was not detected in cells with clear electrophysiological phenotypes. For example, Emx1 was not detected in the RS cell 7. This absence of PCR-based detection of markers probably refl ects drop-out or sampling of transcripts during the global cDNA amplifi cation step, as we have observed previously in neural stem cells (Subkhankulova et al., 2008). In contrast, the RS cell 1 was found to have both Emx1 and GAD65 mRNA by qPCR, but has pyramidal morphology. It is unlikely that this cell corresponds to a regular-spiking, non-pyramidal cell. As discussed below, cluster analysis based on gene expression indicates that this cell is a regularspiking pyramidal neuron. Therefore, it may be that the presence of GAD65 mRNA represents contamination or a false-positive result due to the use of globally-amplifi ed cDNA as the template for the subsequent qPCR.

INDIVIDUAL EXPRESSION PROFILING DISTINGUISHED TWO GROUPS OF YOUNG CELLS CORRESPONDING TO FS AND RS NEURONS
The experimental design here was one in which gene expression in each cell was compared to a reference sample of cDNA pooled from the entire set of 18 young and immature cells studied in Cluster analysis by gene expression of all 18 neurons, independent of their fi ring properties, identifi ed three major clusters (Figure 5). The fi rst group contained four regular-spiking neurons of a range of developmental stages (Figure 5). The second cluster of 6 cells contained four FS neurons, of which three were immature, together with two cells, 2 (P9) and 11 (P22), whose fi ring properties were those of typical RS cells. The third cluster contains 6 fast-spiking neurons and two regular-spiking neurons (one young and one immature). Thus unbiased clustering based on global gene expression grouped this complex population of single cortical neurons in a manner that broadly refl ects their fi ring properties, with a subset of neurons whose cluster assignment did not accurately refl ect their fi ring properties.

IDENTIFICATION OF TRANSCRIPTS ENRICHED IN REGULAR-SPIKING NEURONS COMPARED WITH FAST-SPIKING CELLS
We used the fi ring properties and qPCR data to group the ten young (>P13) neurons into RS-and FS-type cells for the purpose of identifying genes differentially expressed between the two groups. Genes showing the most reproducible differences in expression between the two groups were selected using Student's t-test (see Materials and Methods for details), and the gene expression values for each gene averaged within each group. Those genes were then clustered for visualization (Figure 6A). At the statistical cutoff applied (p < 0.05), 482 mRNA transcripts with differential expression between the two groups were identifi ed, 279 of which were enriched in RS cells and 203 enriched in FS cells ( Table S2 in Supplementary Material). The sets of genes enriched in each group were functionally heterogeneous, including transcription factors, intracellular enzyme and plasma membrane proteins ( Table S2 in Supplementary Material).
To explore whether the enriched expression of genes identifi ed by the array analysis as such in RS or FS cells, we made use of the public, large-scale in situ hybridization dataset for the adult mouse brain, the Allen Brain Atlas (Lein et al., 2007). Many of the genes showing statistically signifi cant differences in expression between the two groups of cells, but of low magnitude, were found to be pan-neuronally expressed (data not shown). However, for those genes that were signifi cantly enriched in RS cells, that is pyramidal cells of the neocortex, we found many examples of genes that were either cortex-specifi c or highly enriched in expression in cortical and, in many cases, hippocampal pyramidal cells, when the expression patterns were compared with that of Emx1 (Figure 6). Several of those genes are not just enriched in cortex, but also demonstrate specifi c expression in layers 2-4 of the cortex (see Figure 6B for examples), demonstrating the ability of the single cell profi ling approach to identify cell type-specifi c genes.

FIGURE 3 | Typical fi ring patterns of the 18 layer II/III neurons analysed and their corresponding expression of pyramidal cell (Emx1) and interneuron (GAD65) markers.
Expression levels: −, not detected; +/−, detected at low level; +, robustly detected. Cell numbers refer to the cell numbering system reported in Table S1 in Supplementary Material and used throughout the text.
Single cortical neuron expression profi ling

DISCUSSION
In this paper we have addressed the question of whether single neuron gene expression profi ling can prospectively resolve neuronal subtypes into groups, independent of any phenotypic information, and whether those groups refl ect meaningful biological properties of those neurons. To do so, we applied methods we have developed to compare gene expression among single neural stem cells to study global gene expression in 18 randomly picked neurons from layer II/III of mouse neocortex. The morphology, fi ring characteristics and electrical properties of the neurons enabled the classifi cation of each cell as either fast-or regular-spiking, corresponding to inhibitory interneurons or excitatory pyramidal cells (Connors and Gutnick, 1990). Unsupervised clustering of neurons by global gene expression resolved the cells into three groups and those corresponded broadly with the groups of fastand regular-spiking neurons. Genes specifi cally enriched in RS neurons were identifi ed from the expression dataset and confi rmed as such in the public mouse brain in situ hybridisation databases. These results provide an initial proof of principle that single-cell gene expression profi ling has the potential to be used to group and classify neurons in a manner refl ecting their known biological properties and can be used to identify novel, cell type-specifi c transcripts.

GLOBAL EXPRESSION PROFILING CAN GROUP SINGLE NEURONS INTO CLASSES THAT REFLECT MEANINGFUL BIOLOGICAL PROPERTIES
Previous expression profi ling studies at the resolution of defi ned neuronal cell types have either focused on a small number of functionally important transcripts, such as ion channels or neurotransmitter systems (Cauli et al., 1997(Cauli et al., , 2000van Hooft et al., 2000), in single neurons or studied groups of neurons defi ned by a shared property: synaptic target (Arlotta et al., 2005) or common expression of a transgene . While very powerful, those studies did not address the question of the ability of global expression profi ling at the single cell level to prospectively group and classify neuronal cell types. In addition, the use of a single property to defi ne group identity, by its very nature, will give an averaged expression profi le of that population with no indication of underlying cellular heterogeneity. Given the diffi culties encountered in classifying cortical interneurons using combinations of morphology, electrical properties and marker protein or gene expression (Ascoli et al., 2008), we wished to investigate whether an unbiased, global assay of gene expression could identify groups of neurons independent of any other cellular property. We generated global expression data from a relatively small sample of 18 morphologically and electrophysiologically characterized neurons and used those data to group the neurons by hierarchical cluster analysis. It is noteworthy that the data were not fi ltered to fi rst derive a subset of transcripts that displayed variation in expression among the cells that would then be used to group the cells. In this case, the clustering based on gene expression was carried out independent of any phenotypic information. We found that this approach can group and classify cells that would be considered quite different in their cellular properties, including neurotransmitter type, morphology and developmental origins. However, we did fi nd errors in assigning cells to clusters within this small and complex population of neurons. Secondly, we found that this approach also enabled the identifi cation of novel cell type-enriched transcripts. It is not clear currently if single cell profi ling is capable of distinguishing related, but distinct, classes of neurons, such as pyramidal cells from different cortical layers.
However, there are inherent limitations to this approach at present. Single cell expression profi ling presents specifi c technical and biological challenges. Technically, current expression profi ling technologies require amplifi cation of the population of single cell mRNA, a process that can and does introduce noise, resulting in an altered transcript distribution. Importantly, this can lead to low abundance transcripts being called as absent in a random manner, due to ineffi ciencies in the early steps of reverse transcription and amplifi cation (Subkhankulova et al., 2008). In the data reported here, for example, we did not identify some genes that are known to be pyramidal or interneuron-specifi c, such as Emx1 or GAD65 (Katarova et al., 2000;Chan et al., 2001). From our previous work in neural stem cells, it is likely that this is due to a combination of technical issues, including ineffi ciencies in the global amplifi cation method, the sensitivity of the array technique and the use of the pooled cells as a common reference. This last aspect of the experimental design is useful as it ensures that all genes expressed in the population are represented in the reference sample. However, the drawback of this approach is that, depending on the abundance of cell-specifi c transcripts in the pooled reference, some cell-enriched transcripts may not appear enriched when compared directly to the reference, rather they will appear depleted in the cells that do not express those genes.
A second potentially confounding biological problem is the noise inherent in gene expression in any cell and the consequent variation in absolute transcript numbers among phenotypically identical cells. We have previously found that there are signifi cant differences in transcript abundance between phenotypically identical neural stem cells and that technical noise did not account for those differences (Subkhankulova et al., 2008). Therefore, it is likely that it is necessary to sample the population deeply to ensure suffi cient representation of each cell type and thereby compensate for the

FIGURE 6 | (A) Clustering of transcripts identifi ed by t-test as enriched in either RS or FS cells. Expression is represented by a scale from green (enriched in FS cells) through red (enriched in RS cells). Each row represents expression for a single gene. (B)
In situ hybridization patterns for the pyramidal neuron-specifi c transcript Emx1 and the GABAergic interneuron-specifi c transcript Gad65. Low power images are shown of mRNA in situ hybridizations of a set of differentially expressed genes on parasagittal sections of the adult mouse brain, from the Allen Brain Atlas (mouse.brain-map.org), along with a high power image of the expression in the somatosensory cortex. In all images the olfactory bulbs are to the left and the cerebellum to the right. For each non-radioactive ISH image, the corresponding Allen Brain Atlas expression analysis image is shown to clarify the relative expression of each gene according to scale of blue (low) to red (high) expression, as illustrated in the attached diagram. intrinsic variability of single cell transcript levels. Alternatively, reducing expression levels from continuous measurements to discrete bins or, in the extreme, to simple binary presence or absence calls, would potentially simplify the classifi cation of cells based on expression of thousands of transcripts.

CAN GLOBAL EXPRESSION PROFILES BE USED TO PREDICT BIOLOGICAL PROPERTIES OF NEURONS?
One long-term goal of single cell profi ling is to use the combinatorial expression data to predict the biological properties of the neuron, essentially by constructing cellular models from single cell transcriptome data. Previous studies, based on single cell PCR data of expression of sets of ion channels, indicated that it is possible to correlate electrical properties with the expression of particular combinations of channels . However, scaling this approach from tens to thousands of transcripts presents a considerable challenge, particularly if the relative levels of those transcripts, and not simply presence or absence of a given transcript, are included in the cellular modeling. Analyses of gene expression in single neurons have found that the levels of mRNAs for specifi c genes can vary considerably among cells of the same type (Schulz et al., 2006(Schulz et al., , 2007Tobin et al., 2009). It is likely that models of neuronal phenotype and electrical properties will require data on the levels of the corresponding proteins and also the possible protein-protein interactions and their functional signifi cance. Current technologies for protein analysis have not achieved single cell resolution, but large-scale proteomic analyses of neurons have begun to provide data that will be useful in this context (Emes et al., 2008).

TOWARDS A CATALOGUE OF CORTICAL CELL TYPES
Previous global expression studies of cortical neurons have produced a taxonomy of forebrain cell types  and identifi ed many novel cell-specifi c transcripts in pyramidal cells (Arlotta et al., 2005). Those studies have demonstrated that there are many examples of transcripts that can be used singly or in combination as discriminators of cortical cell types. The data presented here indicate that the global expression profi ling approach has the potential to be extended to the single cell level, independent of any cellular property of cortical neurons, to sample the population of neurons in a cortical region in order to assess cellular diversity at the genetic level, group those cells and, ultimately, classify them by gene expression. As such, this approach provides a bottom-up approach that complements the top-down approach of identifying cell-specifi c expression patterns based on some group-defi ning property of a set of neurons. Together, these complementary approaches may enable the generation of a catalogue of cortical cell types, based on gene expression, which can be one element of a holistic classifi cation of cortical neurons.