Content uploaded by Idan Segev
Author content
All content in this area was uploaded by Idan Segev on Sep 11, 2014
Content may be subject to copyright.
A Hierarchical Structure of Cortical Interneuron Electrical Diversity Revealed by
Automated Statistical Analysis
Shaul Druckmann1, Sean Hill2, Felix Schürmann2, Henry Markram2and Idan Segev1
1
Interdisciplinary Center for Neural Computation, and Department of Neurobiology, Edmond and Lily Safra Center for Brain
Sciences, Institute of Life Sciences, Hebrew University of Jerusalem, Jerusalem, Israel and
2
Blue Brain Project, Brain Mind
Institute, Ecole Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland
Address correspondence to Idan Segev, Interdisciplinary Center for Neural Computation, Edmond and Lily Safra Center for Brain Sciences,
Department of Neurobiology, Institute of Life Sciences, Hebrew University of Jerusalem, Edmond J. Safra Campus, Givat Ram, Jerusalem 91904,
Israel. Email: idan@lobster.ls.huji.ac.il
Received 18 March 2012; revised 8 August 2012; accepted 21 August 2012
Although the diversity of cortical interneuron electrical properties is
well recognized, the number of distinct electrical types (e-types) is
still a matter of debate. Recently, descriptions of interneuron varia-
bility were standardized by multiple laboratories on the basis of a
subjective classification scheme as set out by the Petilla convention
(Petilla Interneuron Nomenclature Group, PING). Here, we present a
quantitative, statistical analysis of a database of nearly five hundred
neurons manually annotated according to the PING nomenclature.
For each cell, 38 features were extracted from responses to supra-
threshold current stimuli and statistically analyzed to examine
whether cortical interneurons subdivide into e-types. We showed
that the partitioning into different e-types is indeed the major com-
ponent of data variability. The analysis suggests refining the PING
e-type classification to be hierarchical, whereby most variability is
first captured within a coarse subpartition, and then subsequently
divided into finer subpartitions. The coarse partition matches the
well-known partitioning of interneurons into fast spiking and adapt-
ing cells. Finer subpartitions match the burst, continuous, and
delayed subtypes. Additionally, our analysis enabled the ranking of
features according to their ability to differentiate among e-types.
We showed that our quantitative e-type assignment is more than
90% accurate and manages to catch several human errors.
Keywords: Cell type, Clustering, Dimensionality reduction, GABA,
Supervised classification
Introduction
The diversity of cortical inhibitory interneurons is a well-
accepted experimental phenomenon. This variability can be
observed in various domains: morphological, that is, the ana-
tomical structure of dendrites and axons (Ramon y Cajal 1904;
Gibson et al. 1999;Cauli et al. 2000;McGarry et al. 2010),
electrophysiological, that is, the firing patterns of the cells
(McCormick et al. 1985;Connors and Gutnick 1990;Maccafer-
ri and Lacaille 2003;McGarry et al. 2010), or the molecular,
that is, the profile of gene expression. (Kawaguchi 1993;Cauli
et al. 2000;Toledo-Rodriguez et al. 2004;Sugino et al. 2006).
Different strategies have been employed in order to character-
ize this diversity, considering one or more of the aforemen-
tioned domain types (Gupta et al. 2000;Wang et al. 2002,
2004;Nowak et al. 2003;Toledo-Rodriguez et al. 2005;Hala-
bisky et al. 2006;Ma et al. 2006;Helmstaedter et al.
2009a,2009b;McGarry et al. 2010).
Recently, representatives from more than a dozen labora-
tories worldwide convened in Petilla de Aragon, Spain, for a
discussion on the diversity of inhibitory interneurons (the
“Petilla Interneuron Nomenclature Group”[PING]; Ascoli
et al. 2008). The result of this convention was a unified no-
menclature describing different forms of variability in
interneurons.
When characterizing the diversity of interneuron electrical
properties, many studies, including the PING, describe the di-
versity in terms of several distinct groups, or electrical type
(e-types; sometimes also referred to as cell “classes”). In other
words, a number of prototypical types of electrical properties
(e.g., spiking response elicited by step currents) can be found
from which each individual cell belonging to a class might
vary to some extent, but considerably less than it varies from
the prototype corresponding to a different e-type. Such a de-
scription is closely related to the notion of data clustering
(Jain 2010). If the cells are indeed clustered into several dis-
tinct e-types rather than spread across a single continuum, the
existence of distinct types can be demonstrated, and their
number and nature can be objectively defined.
The caveat of the clustering approach is that one must
define the “space”in which to check whether the data are
clustered. This is especially problematic for electrophysiologi-
cal data, as such a space is notoriously difficult to define. One
main issue is the high-dimensional nature of the raw data.
Spiking responses to particular stimuli typically require digi-
tization at several kilohertz. If this were taken as the natural
space, each sampling point would be a dimension in this
space, and consequently there would be several thousand di-
mensions for even 1 s of voltage response recorded from a
cell. This is clearly an unfeasible solution that is never per-
formed in practice.
Alternatively, a set of descriptors, or features, of the voltage
response to a particular stimulus can be extracted (e.g., the
frequency of action potentials [APs], time of first AP following
a particular current input, and AP width and height [Druck-
mann et al. 2007]). This set of features, together with cluster-
ing methods (Jain 2010), could then be used to explore
whether the electrophysiological responses of the different
cells fall into clusters corresponding to distinct types. The
limitation of this approach is that the choice of features is
somewhat arbitrary and that the transition from the raw
voltage trace to the set of feature values involves information
loss, as all the information not contained within the feature
values is discarded. On the other hand, the manual PING
classification may have its own limitations and biases. There-
fore, it is only natural to study systematically, on the basis of a
© The Author 2012. Published by Oxford University Press. All rights reserved. For Permissions, please e-mail: journals.permissions@oup.com
Cerebral Cortex
doi:10.1093/cercor/bhs290
Cerebral Cortex Advance Access published September 17, 2012
by guest on September 17, 2012http://cercor.oxfordjournals.org/Downloaded from
large set of features and quantitative statistical criteria, how
robust the PING classification is and whether one can find
additional structure within the data.
In this study, we extracted a large set of electrical features
from the suprathreshold voltage response of several hundreds
of cortical interneurons and analyzed this database of features
using quantitative statistical approaches. The database of cor-
tical interneurons was also manually annotated in accordance
with the partitioning of types defined by the PING. In this
way, we could both evaluate the explanatory power of the
PING nomenclature for classifying e-types of cortical inter-
neurons and examine which of the different features were
most useful for capturing this partitioning of e-types.
We showed that the distribution of the electrophysiological
feature values is clearly multimodal. These modes correspond
quite well to the different types defined by the PING. More-
over, this coarse partition corresponds to the dimension in
the data accounting for the highest variance. We showed that
small subsets of features are sufficient to distinguish among
these coarse types. Highly informative features are related
both to the pattern of AP discharge (which is the basis of the
naming convention in the PING) and to the shape of the indi-
vidual APs. However, our analysis suggests a refinement for
the PING classification, toward a more hierarchical view of
the partitioning of cells into types. Notably, the aggregation
of large sets of the data and its statistical analysis yielded a
partitioning of the feature space into different “decision
regions”by determining the e-type to which each point in
space (representing a particular set of feature values) was
most likely to belong. These decision regions can be used to
quantitatively attribute a class to newly recorded cells, as well
as to form the basis of comparisons with data from additional
laboratories and different cell types.
Materials and Methods
Database of Electrophysiological Recordings
Data from in vitro intracellular recordings were collected as described
in Toledo-Rodriguez et al. (2004,2005) and aggregated into a large
database of cortical neurons consisting of the response of 466 cortical
interneurons from juvenile rats (age p 12–16). Features were calcu-
lated from responses to step currents of varying amplitude and
length. In brief, different amplitude step currents were first employed
to find the threshold current. Subsequent stimuli were then scaled ac-
cording to this value. We analyze the responses to features extracted
mainly from 2 amplitudes of step current injections: The first of ampli-
tude equal to 150% threshold current and the second of amplitude
equal to 300% threshold current (which we shall refer to as “standard”
and “strong”current, respectively). Each step current had a duration
of 2 s. Approximately 50 traces were used for each neuron.
Electrical Features
We extracted a set of 38 features describing the cell’s electrical
response to step current pulse injection of varying amplitude. These
features were based on a set of previously described electrical de-
scriptors (Toledo-Rodriguez et al. 2004,2005). Table 1presents a
brief description of each feature, a more detailed description of the
features is found in the Supplementary Material. As features were
defined in different units, prior to data analysis all features were trans-
formed into z-scores, that is, the mean was subtracted and the feature
value was normalized by the standard deviation (SD).
Multivariate Analysis
Following the extraction of the electrical descriptors, each cell was
represented by a vector in m-dimensional space, where m= 38 is the
number of features extracted (Table 1). As the data consisted of re-
cordings from 466 cells, the data set was represented by an n×m
matrix (n= 466, m= 38). Principal component analysis (PCA; Duda
et al. 2001)wasfirst performed to determine the prominent com-
ponents of the variability in the data, by calculating the eigenvectors
of the covariance matrix.
A linear Fisher discriminant analysis (Fisher 1936) was performed
in order to determine the features that provide the greatest separation
between the different types. Specifically, the discriminant functions
(DFs) maximize the distance between the means of 2 groups, normal-
ized by the scatter of each group. Formally, DFs are the eigenvectors
of the following matrix: S1
wSb;where S
w
, the within-group scatter
matrix, was calculated by measuring the scatter of the feature values
of cells belonging to each group following the subtraction from each
feature of the group mean, and S
b
is the scatter of the group means.
Feature subset selection was performed either by exhaustive search
for up to 7 features or by a branch-and-bound algorithm (Narendra
and Fukunaga 1977). In brief, branch-and-bound algorithms allow ef-
ficient evaluation of subsets of features if the criterion, in this case
group separation, is monotonic in the features. Informally, monotonic
functions are functions for which discarding additional features will
only diminish the value of the criterion. These algorithms take advan-
tage of the fact that if the criterion has already dropped below a
certain value, throwing out more features is not going to improve it.
Thus, when proceeding by eliminating (not adding) features, if a
higher criterion value has been found for a different subset with a
smaller number of features, then the lower-criterion subset need not
be considered any more. Importantly, along with the lower-criterion
subset, one can also reject without further consideration all the
Table 1
Brief description of each feature
Feature # Feature description
1 Drop in AP amplitude (amp.) from first to second spike (mV)
2 AP amplitude change from first spike to steady-state (mV)
3 AP 1 amplitude (mV)
4 AP 1 width at half height (ms)
5 AP 1 peak to trough time (ms)
6 AP 1 peak to trough rate of change (mV/ms)
7 AP 1 Fast AHP depth (mV)
8 AP 2 amplitude (mV)
9 AP 2 width at half height (ms)
10 AP 2 peak to trough time (ms)
11 AP 2 peak to trough rate of change (mV/ms)
12 AP 2 Fast AHP depth (mV)
13 Percent change in AP amplitude, first to second spike (%)
14 Percent change in AP width at half height, first to second spike (%)
15 Percent change in AP peak to trough rate of change, first to second spike (%)
16 Percent change in AP fast AHP depth, first to second spike (%)
17 Input resistance for steady-state current (Ohm)
18 Average delay to AP 1 (ms)
19 SD of delay to AP 1 (ms)
20 Average delay to AP 2 (ms)
21 SD of delay to AP 2 (ms)
22 Average initial burst interval (ms)
23 SD of average initial burst interval (ms)
24 Average initial accommodation (%)
25 Average steady-state accommodation (%)
26 Rate of accommodation to steady-state (1/ms)
27 Average accommodation at steady-state (%)
28 Average rate of accommodation during steady-state
29 Average inter-spike interval (ISI) coefficient of variation (CV) (unit less)
30 Median of the distribution of ISIs (ms)
31 Average change in ISIs during a burst (%)
32 Average rate, strong stimulus (Hz)
33 Average delay to AP 1, strong stimulus (ms)
34 SD of delay to AP 1, strong stimulus (ms)
35 Average delay to AP 2, strong stimulus (ms)
36 SD of delay to AP 2, strong stimulus (ms)
37 Average initial burst ISI, strong stimulus (ms)
38 SD of average initial burst ISI, strong stimulus (ms)
2A Hierarchical Structure of Cortical Interneuron •Druckmann et al.
by guest on September 17, 2012http://cercor.oxfordjournals.org/Downloaded from
subsets derived from it by discarding additional features. Accordingly,
many feature subsets do not need to be evaluated and the search
becomes much more efficient than exhaustive search.
Decision Regions
In order to infer to which class a data point was most likely to belong,
we needed to make some assumptions about the structure of the data.
We assumed that each class could be treated as a multidimensional
Gaussian, with different mean vectors and covariance matrices for
each group. The parameters of each Gaussian (mean vector and
covariance matrix) were estimated in standard fashion by maximum
likelihood estimation (Duda et al. 2001):
^
m
¼1
nX
n
k¼1
xk;
^
S¼1
nX
n
k¼1
ðxk^
m
Þðxk^
m
ÞT:
Given the Gaussian models and the estimated mean vector and covari-
ance matrix, the probability for each point in space is defined by the
following equation:
pðxj
m
;SÞ¼ 1
ð2
p
Þd=2jSj1=2exp 1
2ðx
m
ÞTS1ðx
m
Þ
;
where |∑| indicates the determinant of the covariance matrix, sigma.
This probability is calculated separately for each of the different Gaus-
sians corresponding to the different classes. If all classes were equally
common, every point in space would be assigned a class according to
the Gaussian that it was most likely to arise from. When classes are
not equally common, one must also factor in the relative contribution
of each class to the population and each point is assigned a class ac-
cording to the highest posterior probability:
c¼arg maxcpðcÞpðxj
m
c;ScÞ:
The lines of equal posterior probability between groups are marked
as the decision boundaries in Figure 4.
Testing for Normality
The classification analysis in the section above modeled each class as
a 2-dimensional Gaussian. Accordingly, the closer the data match the
Gaussian distribution, the more accurate the expected results. We
tested for multivariate Gaussianity using Mardia’s test for skewness
and kurtosis (Mardia 1970).
Objective Nested Clustering Analysis
We performed a nested clustering analysis by repeatedly performing
k-means clustering. Conceptually, this process is the methodic appli-
cation of the process performed manually transitioning from
Figures 2and 3(i.e., separating data sets and performing PCA on indi-
vidual parts of the data set). Namely, beginning with the full data set,
we calculated the first 10 PCs (that together capture 80% of the var-
iance) and then performed k-means clustering to find 2 clusters
within the data. These clusters are then split. For each of the clusters,
we repeated the PCA with only the data from the individual cluster in
order to reduce sensitivity to electrical class induced feature corre-
lations. Then we ran separate k-means clustering on each of the clus-
ters and split the cluster for which the splitting reduced the average
Euclidean distance in PCA space between the group cluster center
and each data point by the largest amount. For each of the k-means
steps, we performed 500 repetitions of the k-means algorithm with
different random seeds and selected the best cluster. We employed
10-fold cross validation to estimate the reliability of the results, that is,
the process was repeated 10 times on 90% of the data each time. We
note that we opted for this approach over more standard forms of di-
visive hierarchical clustering, such as (Macnaughton-Smith et al. 1964;
Guénoche et al. 1991) due to the simplicity of k-means, its intuitive
logic, and its widespread use in neuroscience and beyond. The
average variance captured in the first 10 PCs at each of the splits is:
(0.811, 0.825, 0.866, 0.869, 0.873, 0.916) respectively.
To check the consistency of the approach, we employed the Rand
index (Rand 1971). In brief, the Rand index is an approach to measur-
ing consistency between 2 sets of assignments of elements into
classes. In our case, the 2 assignments are 2 repetitions of the nested
clustering procedure, the classes are the e-types and the elements are
the neurons. For every pair of neurons, one determines whether these
2 neurons are in the same class or not, for each of the 2 assignments.
If the neurons are in the same class both in the first assignment and
in the second, this pair is consistent across the assignments. Similarly,
if the neurons are each in a different class both in the first assignment
and in the second assignment, this pair is consistent as well.
However, if the 2 points are in the same class in one assignment but
not in the other, this is an inconsistency. Though very intuitive, the
Rand index has one major disadvantage: It has no clear scale.
Namely, though complete agreement between 2 assignments yields a
score of 1, the expected value of the agreement between random as-
signments does not yield a specific score (such as zero). Therefore,
we employed the adjusted Rand index that uses the generalized hy-
pergeometric distribution as the control random model (Hubert and
Arabie 1985).
Software
Data were queried from the database and imported into Matlab (Math-
works, Natick, MA, United States of America). Feature extraction was
performed on all traces by custom software programmed in Matlab.
Statistics of features and their normalization, as well as PCA were per-
formed using built-in functions in Matlab. Multivariate analysis, in-
cluding DF analysis and generation of decision regions, was
programmed in Matlab.
Results
Figure 1depicts the subjective classification of our database of
cortical interneurons to e-types. Out of n= 466 cells recorded
in our database, the majority (418) were subjectively classified
in accordance with the Petilla (PING) nomenclature (Ascoli
et al. 2008). For most of the e-types defined in this convention,
we have a reasonable number of examples within the data set
(Fig. 1). The PING convention delineates the responses into
the steady-state part of the response, and the transient part of
the response. In Figure 1, the steady-state responses comprise
the rows (indicated by the following capital letters: fast
spiking [FS], non-adapting [NA] non-fast spiking, adapting
[AD], and irregular spiking, [IS]). The transient responses com-
prise the columns (indicated by lower case letters: burst [b],
continuous [c], delayed [d], and stuttering [s]). In addition, we
have the pyramidal cells (indicated by: Pyr). For 3 of the
e-types (“burst-IS,”“continuous-IS,”and “delayed-non-adapt-
ing non-FS”) the database consists of less than 10 cells (gray
traces). Accordingly, we did not treat these cells as separate
types but rather labeled them collectively as the “other”cell
type. Intrinsically bursting neurons were not considered
because of the lack of reliable examples within the data con-
sidered for this study. A few cells (14) were marked as poss-
ibly belonging to 1 of 2 types and an unambiguous subjective
classification was not found for a number of cells (31). Accord-
ingly, these were treated as the “unspecified”type.
PCA of Feature Space
We began our unsupervised analysis by performing PCA based
on the 38 features that we used for characterization of the elec-
trical properties of each cell in our database (Table 1), using all
Cerebral Cortex 3
by guest on September 17, 2012http://cercor.oxfordjournals.org/Downloaded from
the 466 neurons. Figure 2ashows the distribution of cells for
different values of the first PC; the total distribution is depicted
by the dotted black line, while the distribution for each of the
subjectively classified e-types is marked by a corresponding
colored line (AD, red, FS, blue, NA, green, Pyr, black, unspeci-
fied, orange, other, cyan, see legend). The total distribution
Figure 2. Cell classification based on PCA. (a) Distribution of cells as a function of the first PC score based on the 38 features used to characterize the cell’s electrical
properties (Table 1). Overall cell distribution is depicted by the dashed black line. Different subjectively classified e-types, according to PING convention, are marked by colored
lines. Classes are marked with different colors as shown in inset (FS, fast spiking; AD, adapting; NA, non-adapting non-FS; Stut., stuttering; Pyr, pyramidal; Other, other; Unspec.,
unspecified). (b) Each cell is plotted as a point in the space of the first 2 PCs (PC1 and PC2). Point color is as in a, transient e-type is indicated by marker shape (c) Percentage
of variance captured by each PC, PCs 1–10. (d) The weight of the contribution of each of the 38 features to the first (top) and second (bottom) PC. Features related to AP shape
are marked in gray, whereas features related to AP firing pattern are marked in dark blue. Feature descriptions are provided in Table 1.
Figure 1. Electrical cell types—subjective classification based on the PING convention. Color traces depict a single voltage trace in response to a depolarizing step current
injection of intensity equal to 1.5 times the firing threshold intensity. One example is shown for all 10 interneuron electrical classes as well as one trace for pyramidal cells (black
trace). The number of cells within the data set is shown for each group under the voltage trace. Gray traces are for cases containing less than 10 cells in the database and were
thus pooled together into an “others”group.
4A Hierarchical Structure of Cortical Interneuron •Druckmann et al.
by guest on September 17, 2012http://cercor.oxfordjournals.org/Downloaded from
function is clearly multimodal. The first 2 peaks correspond
well to the FS cell type and to the AD e-type, respectively. The
third peak is shared by Pyr neurons and adapting cells, and the
remaining cell types are distributed fairly uniformly in this
1-dimensional projection upon the first PC.
In Figure 2bwe plotted the cells in the space of the first 2
PCs. Each cell is represented as a point corresponding to its
value in the first 2 PCs. Each point is colored according to the
cell’s subjective e-type classification, as in Figure 2a, and
marker shape indicates transient e-type (circle—continuous,
triangle—delayed, and square—bursting). As above, 2 visible
clusters can be seen for the FS and adapting cells. Pyr cells
occupy a diffuse region in the lower right part. Nearly all sub-
jectively classified FS cells (blue) had negative values in the
first PC, whereas adapting cells (red) had slightly negative or
near-zero values and Pyr cells (black) obtained more positive
values. Thus, the first PC is strongly related to the variability
found in the subjective evaluation of cell types according to
their steady-state electrical behavior (the rows of Fig. 1).
The fraction of the variability represented by each of the first
10 PCs is plotted in Figure 2c. Note that the first component is
considerably larger than the subsequent ones. Thus, we
confirm that the cell’s steady-state e-type as defined by the
PING nomenclature is indeed the major component of variabil-
ity in the data. The data appear to be of a truly high-
dimensional nature, with 10 PCs required to account for 80% of
the variability in the data. In other words, there are numerous,
separate factors that contribute to the electrical variability of
cortical interneurons. Note that only the first 2 PCs are shown
because of the limitations of plotting high-dimensional data.
The distribution of the weights of features associated with
the first 2 PCs is depicted in Figure 2d. Features related to AP
shape are marked in gray and features related to AP firing
pattern in dark blue. Notably, the distribution is broad, with
many of the features carrying substantial weight, both for fea-
tures related to AP shape and to AP firing pattern [e.g.,
feature #4, width of the first AP at half height, and #30
median value of inter-spike-intervals (ISIs), corresponding to
shape- and pattern-related features, respectively Table 1).
Some features have only small weight [e.g., features #12 and
#34 correspond to fast after-hyperpolariztion (AHP) of second
AP, and SD of the delay to first AP in a strong stimulus]
indicating that they do not contribute strongly to the main
source of the variability in the data and consequently to the
discrimination between the different steady-state neuron
types.
We note that the different transient e-types do not stand
out as separate clusters in Figure 2b. This could occur for 2
reasons: Either these different types are not in fact different
types in the full high-dimensional feature space, or the par-
ticular projection chosen by the first 2 PCs causes these types
to appear mixed. Why would the latter option be true?
Judging by eye, the variability between the steady-state types
is clearly large and immediately evident, whereas the differ-
ences between the transient behaviors are almost by defi-
nition more subtle. Moreover, the vast majority of cells come
from the “continuous”transient behavior (see numbers in
Fig. 1), further biasing the variance in the data toward differ-
ences between the steady-state e-types rather than those
between transient types. Thus, PCA favors those directions in
the high-dimensional space that correspond to separation
between steady-state e-types. If these directions in feature
space are different from those that separate between transient
e-types, then the latter separation may be masked by the PCA
(Supplementary Figure 1). We note that this is not a limitation
of the PCA technique per se, but rather of using PCA (whose
purpose is to find the directions in space that capture
the most variance) to explore a different question, one of
discrimination between e-types.
To test whether the differences in transient electrical
behavior are masked by the differences in the steady-state
electrical behaviors, we repeated PCA on data from a single
steady-state e-type—the adapting interneurons, 158 neurons.
The distribution of cells across the new first PC appears multi-
modal, though the 2 peaks overlap to some degree (Fig. 3a).
The 2 peaks mostly correspond to the 2 transient electrical be-
haviors (burst transient and continuous) found for the adapt-
ing cells.
To understand why the difference between the transient
types was not immediately apparent in the original PCA, we
compared the weights assigned to the different features by
PCA performed on all cells with the weight assigned by PCA
performed separately on the adapting interneurons. The
weights of the PCs were quite similar (Fig. 3b), yet the relative
Figure 3. PCA on adapting interneurons alone. (a) Spread of adapting interneuron feature values across the first PC. Black solid trace shows all adapting interneurons, dotted red
trace shows cAD interneurons, and red solid line bAD interneurons. Note that the solid trace was artificially offset upwards to allow clearer viewing of all lines. (b) Weights
assigned to different features composing the first PC. Features related to AP shape are shown in pink and features related to AP discharge pattern in dark red. Note that larger
weights were assigned on average to the latter features (and compare also to Fig. 2dtop). (c) Relative weight of the firing pattern features in different PCs. Weight is defined as
the sum of the absolute value of feature weights, and is normalized by the different number of firing pattern versus AP shape features. Left and middle shows original PCs 1 and
2 (Fig. 2d) and right shows the PC plotted in Figure 3b.
Cerebral Cortex 5
by guest on September 17, 2012http://cercor.oxfordjournals.org/Downloaded from
contribution of features related to AP discharge pattern, as
opposed to AP shape, was larger in the separate PCA for
adapting interneurons (Fig. 3c). This is important since the
differences between burst-adapting (bAD) and continuous-
adapting (cAD) are much more distinct in the AP discharge
pattern. A one-way analysis of variance (ANOVA) shows
highly significant group differences when based on the dis-
charge features (n= 158, f= 23.61, P<1e–5), yet the shape
features show a much weaker, borderline non significant
effect (n= 158, f= 3.68, P= 0.06). In summary, the high
weights put on features related to AP shape in the original PC
analysis (due to the distinct difference in shape between FS
and AD interneurons and the large number of such neurons)
masks the mostly discharge related differences between bAD
and cAD interneurons.
DF Analysis
Having shown that the neurons’e-type is a major component
of the variability in the data, we next explored the partitioning
of neurons into different e-types by performing DF analysis
(DFA, see Materials and Methods). In brief, DFA is a super-
vised data analysis method that finds the features that lead to
the sharpest distinction between different classes. We stress
that this method requires stating beforehand which neurons
belong to which e-type and is in this sense biased. As noted
previously, manual classification was performed for most cells
by an expert based on the PING subjective criteria.
We performed DFA using the supervised manual classifi-
cation. We use the 8 e-types that had a sufficient number of
neurons (cFS, dFS, sFS, bNA, cNA, bAD, cAD, Pyr, 8 e-types,
387 neurons). In Figure 4a, we plotted the cells in the space
of the first 2 DFs. The FS cells (blue, top right) are well separ-
ated from the adapting (red) cells, and even more distant
from the Pyr (black) cells. Most adapting cells are well separ-
ated from the Pyr cells, but a portion are mixed among the
Pyr cells (upper left portion of plot, red circles mixed with
black circles) and appear to be more of a continuation of the
Pyr cells than part of the main adapting e-type. The difference
between the 3 e-types is highly significant, P<1e–5 for all
pairwise differences. The non-adapting non-FS (green) e-type
is almost in full overlap with the adapting cells. The separ-
ation between the transient behavior (burst/continuous/
delayed) cannot be distinguished at this resolution.
Figure 4bshows the contribution of different features to
the separation between the e-types. In general, features that
Figure 4. DF analysis of electrical classes. (a) Each cell is plotted as a point in the space of the first 2 DFs. Marker color corresponds to subjectively characterized e-type, see
legend. (b) Each feature in the discriminant analysis shown in a is plotted as a circle in the x,y location corresponding to the CV (mean over SD) and ratio of the within-group
variance to the total variance, respectively (i.e., higher y-values indicate better separability of groups using this particular feature). Features related to AP shape are marked in
gray and features related to AP firing pattern in black. For feature definition, see Table 1.(c) DF is performed on the fast-spiking and stuttering cells alone (the “blue”classes).
Each fast-spiking cell is represented as a single point. Transient class is marked by different symbols (triangle—delayed, circle—continuous). (d) Each feature is plotted as in (b),
but now with CV and variance ratio calculated for fast-spiking cells alone. Note the different y-axis scale in band d.
6A Hierarchical Structure of Cortical Interneuron •Druckmann et al.
by guest on September 17, 2012http://cercor.oxfordjournals.org/Downloaded from
are well suited to separating any pair of 2 groups are those
that take very different values if the neuron belongs to one
group or another, but are relatively constant within each
group. Formally, this is expressed as the ratio of the
between-group variance to the within-group variance (see
Materials and Methods). A second desirable property is that
the variable not be very noisy, as expressed, for example, by
its coefficient of variation (CV). Accordingly, each feature is
plotted as a circle in the x,y position determined by the
feature CV (x-axis) and the ratio of the between-group var-
iance to the total variance (y-axis). Thus, features that offer
good separability between the different types are found on
the upper left part of the plot, corresponding to high
between-to-within group variance and low CV. Indeed, a
group of 8 features stands out on the top part of the plot.
These are (in descending order of the ratio between group
variance and total variance): (1) AP 1 width at half height; (2)
average rate of strong stimulus; (3) AP 1 fast AHP depth; (4)
average initial burst interval; (5) median of the distribution of
ISIs; (6) AP 1 peak to trough rate of change; (7) SD of average
initial burst interval; and (8) average ISI CV. Notably, the 2
main features that have been traditionally used to characterize
FS cells—spike width and AP firing rate—are represented
within this set.
We next attempted to distinguish between the transient
electrical behaviors by DFA. As these can be obscured by the
large differences between the steady-state electrical cell types
as shown earlier, we performed discriminant analysis anew
separately for each steady-state type. Figure 4cshows this
analysis for the FS cells (the “blue”class—cFS, dFS, sFS; 116
neurons). Each FS cell is plotted as a point corresponding to
the 2 first newly calculated DFs. Now the different transient
types can be fairly well separated into their 3 subtypes.
Figure 4dshows which features best discriminate between
the different subtypes in this case. The important features in
this case include: AP 2 peak to trough rate of change, AP 1
amplitude, average ISI CV, SD of delay from first to second
spike, and AP 1 peak to trough time. These are mostly differ-
ent from the features that separated between the steady-state
firing types. Note, however, that even for the best feature, the
between-group variance is only approximately 15% of the
total variance of that feature (Fig. 4d,y-value of top circle), as
opposed to the best feature for the steady-state electrical be-
havior discrimination shown in Figure 3b, where 50% of its
variance was explained by the between-group variance
(Fig. 4b,y-value of top circle). This indicates that the separ-
ation between the transient e-types is much weaker than the
separation between the steady-state e-types. Similar analysis
was performed for the difference between the different AD
subtypes (Supplementary Figure 2).
Reliability of Discriminant Analysis
In order to test how well the e-types are separated, we em-
ployed multiple class separation in the space of the first 2
DFs. Namely, the mean and SD of a 2-dimensional Gaussian
was estimated for each class separately. In addition, each
point in space was assigned to the class whose Gaussian was
most likely to generate it. Figure 5ashows the partition of the
space into the points assigned to the 3 major types found
above (black—Pyr, red—AD, blue—FS) the decision bound-
ary is marked by respectively colored lines. The cells from
these 3 types (335 neurons, encompassing all transient types)
are shown atop this grid as full large dots. Testing for the
hypothesis that each class arises from a 2-dimensional Gaus-
sian (see Materials and Methods), we found that the null-
hypothesis is not rejected (P> 0.05) for the FS and Pyr
classes, whereas for the AD cells the hypothesis is rejected
(P< 0.01) likely because of the existence of a number of
extreme values.
Although some cells lie deep within the region belonging
to a single e-type, they were subjectively classified as belong-
ing to a different electric type. Four FS cells (blue circles) are
found deep within the adapting cell (red) region and 5 adapt-
ing cells are found deep within the FS cell region. Figure 5c
shows the firing response of 6 cells that were classified as be-
longing to a different e-type. As can be seen by the firing
pattern, each of these cells actually belongs to the opposite
e-type, as indeed they were assigned by the statistical classifi-
cation of the decision boundaries. We note that such manual
misclassification of 9 cells out of more than 300 corresponds
only to a small percentage of the total number of cells and is
an expected part of any human process applied to hundreds
of examples, e.g., neuron tracing (Helmstaedter et al. 2011).
Most importantly, the fact that mistakes in classifications can
be caught alleviates at least in part the concern of using a su-
pervised method like DFA, since it directly shows that the
method does not simply recapitulate any arbitrary human la-
beling. Rather, by using the statistical power afforded by
classification of a large number of cells, reliable measures of
separation between the different e-types were determined
despite conflicting examples (i.e., examples that reside in the
portion of space occupied by an e-type different from that
originally labeled).
Once such a decision region is found, it can be used to
define the e-type of newly measured cells without the need
for further subjective classification. In order to assess the
accuracy of such an approach, we randomly excluded a
portion of the data and used the remaining data to generate
the decision region. We then checked the validity of the stat-
istical classification by ten-fold cross validation (see Materials
and Methods). Namely, we set aside data that were not used
to construct the decision region, and estimated error rates by
comparing the manual and statistical assignment on this vali-
dation set. The average rate of error was 6%, [mean error 5.9%
± 2.3% (SD), 10-fold cross validation], meaning that we expect
to correctly classify with no further human intervention about
94% of cells that will be measured in the future. We note that
although FS interneurons are commonly distinguished in
experiments from Pyr cells by the properties of their electrical
response, adapting cells are considered to be difficult to dis-
tinguish from Pyr cells on the basis of their response to elec-
trical stimuli alone. However, we find that this separation,
although slightly more difficult, can still be achieved using
our objective method with an 89% accuracy (mean error
89.1% ± 3.5%, 10-fold cross validation).
Between-Group and Within-Group Correlations
The importance of recognizing and describing e-types is
especially relevant to examining the correlations between
different biophysical properties. Correlations in the value of
biophysical properties are often treated as suggestive of a tra-
deoff or balance between these different properties. However,
Cerebral Cortex 7
by guest on September 17, 2012http://cercor.oxfordjournals.org/Downloaded from
one should be careful to distinguish between correlations that
exist because of differences between group means and other,
more straightforward correlations, which exist regardless of
whether or not neurons belong to the same or a different
class. For instance, we considered the relationship between
the firing rate of a cell and the width of its AP for FS and AD
cells (a total of 273 neurons). When pooling over the 2 cell
types, we found a highly significant linear relationship for the
Figure 5. Classification by decision regions. The space of the first 2 DFs is divided into regions corresponding to the e-type that the data point would be most likely to have
come from (see Materials and Methods): Adapting (red fine dots), FS (blue fine dots), or Pyr cells (black fine dots). Position of cells is shown by circles, colored according to
e-type. Note that several circles lie far outside their decision area. (b) An example voltage response to a step current of 1.5 times firing threshold intensity is shown for the
adapting (red, left) and FS (blue, right) electrical classes. (c). Three examples of cells manually classified as 1 e-type, yet located deep within the decision area of another
e-type, are shown. The left column shows cells manually classified as adapting cells (red dots) yet located within the selection region of (and thus statistically classified as) FS
cells. The right column shows cells manually classified as FS and statistically classified as adapting. Comparing the voltage responses to those found in b, the statistical analysis
clearly gave the correct classification and the manual classification is erroneous.
8A Hierarchical Structure of Cortical Interneuron •Druckmann et al.
by guest on September 17, 2012http://cercor.oxfordjournals.org/Downloaded from
data (Fig. 6,r
2
= 0.422, P< 0.0001, confidence interval [CI] of
slope [−72.1, −50.4]). Accordingly, one may be tempted to
think that there is some biophysical mechanism that allows an
increase in firing rate but causes spike width to be narrower.
However, when we considered each of the 2 e-types separ-
ately, we found that this correlation was not significant, with
the CI of the slope containing both negative and positive
values. Thus, the correlation is almost exclusively due to the
between-group effects, that is, the difference in the mean of
the groups. Indeed, within each class, neurons with differ-
ences of almost half a millisecond in spike width (which is
nearly the difference between the group means) exhibited
similar firing rates. Thus, the nature and interpretation of the
correlation between firing rate and AP width is more
complex. In other words, an analysis blind to the e-type
might have mistakenly assumed that the correlation exists
within each group and not just due to differences in mean
group values, a point that should guide its biophysical study.
Relative Importance of Features
Until now, we described analyses that had access to all the
features. However, using a large number of features may com-
plicate visualization and interpretation of data. Therefore, one
may also consider using only smaller sets of features, thus
raising the question: which are the most important features?
To answer this question, we attempted to maximize the dis-
tinction between the types of electrical behavior as before,
but here we used a more limited subset, each time of a differ-
ent size. For small subsets one can search through all possible
combinations of features, while for larger subsets, the large
number of feature combinations requires more sophisticated
searching, such as the use of branch-and-bound algorithms
(Narendra and Fukunaga 1977, see Materials and Methods).
Alternatively, there are numerous feature subset selection ap-
proaches that can be used also for large sets of features (for a
review see Saeys et al. 2007). We performed an exhaustive
search for feature subsets of up to size 4 and found that very
accurate classifications could be performed on the basis of
only a few features. This is perhaps to be expected since
many of the features measure similar properties (Table 1).
We quantified the separability of the 3 steady-state e-types
considered above (FS, AD, Pyr, 335 neurons) by the distance
of the group means over the SD of the distributions (standar-
dized mean difference, SMD). The larger SD of the 2 for each
group was chosen as the normalization. In the case of the full
feature sets, these differences were: FS-AD: 2.58 units of SD,
AD-Pyr: 2.98 units. Considering subsets composed of only a
single feature, we found that the most effective feature, AP 1
width at half-height, allows a rough classification (FS-AD
1.43, AD-Pyr 1.51, SMD). The best set of 3 features resulted in
nicely separated groups (Fig. 7, FS-AD 2.21, AD-Pyr 2.16,
SMD) and included the following features: AP 2 width at half-
height, firing rate, AP 1 fast AHP depth. Notably, this set
shares features that describe both the shape of the AP and the
pattern of AP discharge. We obtained a classification nearly as
accurate as that of the full feature set by using only 4 features:
width of AP 2 at half-height, rate of AP discharge, AP 1 fast
AHP depth, time to second AP from stimulus onset (FS-AD
2.33, AD-Pyr 2.51, SMD). Rerunning the analysis while allow-
ing only features from 1 of the 2 sets (AP shape, discharge
pattern) resulted in less accurate classification, which again
highlights the usefulness of incorporating an analysis of the
shape of the AP even though the electric types were originally
defined (and are still named) according to their AP discharge
pattern.
A Hierarchical Picture of Interneuron Electrical
Variability
Taken together, these results strongly suggest a hierarchical
picture for the classification of cortical interneuron electrical
variability. Namely, the steady-state interneuron e-types (FS
and AD cells) are easy to distinguish, account for a large
Figure 6. Electrical class and feature correlations. FS cells (blue) and adapting cells
(red) are each represented by a colored dot located according to the value of their
spike width (x-axis) and rate of AP discharge (y-axis). The overall correlation between
AP width and firing rate for all cells is highly significant (black regression line, r
2
-value
in black top right). However, when each group is considered separately, the
regression between the 2 variables is found to be not significant (blue and red
regression lines and r
2
-values in top right).
Figure 7. Feature selection for steady-state e-type discrimination. A feature
selection procedure was used to choose the features that best discriminate between
FS interneurons (blue), adapting interneurons (red) and Pyr neurons (black). Plot
shows the distribution of neuron feature values upon the first DF composed of the
one best feature (top) and 3 best features (bottom). Note the sharper separation of
groups in the bottom plot.
Cerebral Cortex 9
by guest on September 17, 2012http://cercor.oxfordjournals.org/Downloaded from
portion of the variability while the transient electrical behav-
ior types have smaller between-group differences (Figs 1–4
and the appropriate "Results" sections). To test this notion, we
calculated the between-group distances for all e-type pairs
(Fig. 8). For each of the e-types, we calculated the average of
each feature value. This average defines a single point in
feature space, that is, a space where each direction in is a
given feature, yielding dimensionality equal to the number of
features. Then, the distance between all pairs of points is
computed. Since many of the features are correlated, we use
the Mahalanobis distance (Mahalanobis 1936) rather than Eu-
clidean distance. The distances found between the different
e-types are indeed consistent with a hierarchical picture
(Fig. 8a). First, the steady-state e-types have large distances
between themselves (FS to AD). Second, the transient e-types
within one steady-state response have smaller distances
between themselves, that is, the cFS to dFS distance is smaller
than that between FS and AD. Third, the distance between the
same transient behavior but different steady-state behavior is
larger than the distance between different transient behavior
and same steady-state behavior, that is, the distance between
cFS and cAD is larger than that between cFS and dFS. We
note that the same properties also hold for distances calcu-
lated on the basis of groups directly taken from the original
subjective classification.
To further test this question, we employed an unsupervised
nested clustering approach that allows for changes in the cor-
relations between different features (see Materials and
Methods). In brief, considering the unambiguously labeled
interneurons (404 neurons), the first split caused the majority
of interneurons to break away from the Pyr neurons (Sup-
plementary Figure 3). The following split caused the majority
of FS cells to break away from the adapting cells. Next, the
adapting cells broke into 2 groups, one of which contained
most of the bAD interneurons. Next, the FS interneurons
broke into a group that contained most of the delayed-FS
type. The non-adapting non-FS cells tended to be distributed
evenly between the different groups. Results were consistent
within 10-fold cross validation (e.g., removing part of the data
and rerunning the analysis, see Materials and Methods). We
quantified consistency by the adjusted Rand index (Hubert
and Arabie 1985, see Materials and Methods). Briefly, the ad-
justed Rand index compares 2 assignments of data points to
partitions and yields an expected value of 1 if the 2 assign-
ments are identical and 0 if the assignments are random. The
adjusted Rand index found was highly significant (mean 0.68,
SD = 0.03, P< 0.0001). Groups were then named by examin-
ing the subjective classification of each cell belonging to the
group (which we stress was not used during the classification)
and selecting the e-type that had the largest contribution.
Though excluded from the full analysis because of the limited
number of examples within our data set, we note that the IS
neurons appear to branch from the adapting interneurons,
with the aforementioned caveat of the small sample size.
In summary, the emerging hierarchical picture of inter-
neuron e-types (Fig. 8b) appears well supported from the
electrophysiological variability standpoint.
Discussion
In this study, we found that by quantifying the nomenclature
described by the PING meeting (Ascoli et al. 2008) and apply-
ing several analysis approaches, a quantitative, statistical
evaluation of the e-types characterized in this meeting could
be readily obtained. Historically, the existence of multiple
types of cortical interneurons (as well as hippocampal inter-
neurons, e.g., Somogyi and Klausberger 2005) was suggested
very early on using morphological features (Ramon y Cajal
1904). Since that time, cell classification has generated much
interest, with different measures being used for neuron classi-
fication, including their firing properties, or e-types (McCor-
mick et al. 1985;Connors and Gutnick 1990;Maccaferri and
Lacaille 2003), their morphological structure (Ramon y Cajal
1904;Gibson et al. 1999;Cauli et al. 2000), and their genetic
profile (Kawaguchi 1993;Cauli et al. 2000;Toledo-Rodriguez
et al. 2004;Sugino et al. 2006). Regarding the classification
into e-types, different approaches to what defines a class
(Somogyi and Klausberger 2005;Ascoli et al. 2008;Helm-
staedter et al. 2009a;McGarry et al. 2010) have led to a con-
siderable amount of confusion, resulting in the often
expressed feeling that such classification is inherently subjec-
tive, depending on whether one is a “lumper”or a “splitter.”
Importantly, the translation of the seemingly straightfor-
ward question “how many classes succinctly describe a
certain dataset”into a quantitative framework is problematic.
This is the case since it implies 2 contradictory goals: Describ-
ing the data well and having a minimal number of classes.
The first, describing the data accurately, can be formalized as
reducing the average distance between each individual data
point and its associated group center (in our case, each data
Figure 8. Hierarchical view of electrical variability of cortical interneurons. (a) Color
plot of difference of class mean feature values between each pair of e-types. (b)
Proposed schematic of the hierarchical structure of cortical interneuron electrical
variability based on the hierarchical classification results. IS is marked by a question
mark due to small representation in data.
10 A Hierarchical Structure of Cortical Interneuron •Druckmann et al.
by guest on September 17, 2012http://cercor.oxfordjournals.org/Downloaded from
point corresponds to a neuron and group center to e-type).
The optimal solution to this first goal is achieved by having
one group for each data point, resulting in zero group center
distance. The second goal, having a minimal amount of
classes, is maximized by having one group for all the data
points. Thus, these 2 aims are clearly in conflict and the
tension between “lumping”and “splitting”is not merely a
subjective one. We note that different approaches and heuris-
tics can be employed in the attempt to find reasonable
answers to the question of number of types (Duda et al. 2001)
but the problem cannot be unequivocally resolved unless the
exact generative model for the data is known, which is typi-
cally not the case.
We demonstrated in this study, which is based on a particu-
larly large experimental database, that a partition into differ-
ent major e-types is the foremost source of explainable
variability in the data and that this partitioning is extremely
salient statistically both in significance and in effect size.
Notably, this partitioning can be found and confirmed with
the simplest of statistical approaches (PCA, Gaussian decision
theory). Thus, although one can argue that the natural ten-
dency toward being a “lumper”or a “splitter”will affect one’s
view of the structure of interneuron electrical variability, the
steady-state electrical properties should most certainly not be
thought of as one continuum.
Supervised and Unsupervised Classification
Our simple statistical criteria were able to correctly predict the
steady-state e-type of a cell (FS or AD) with an ∼95% accuracy.
Like previous studies, we found that combining supervised
and unsupervised classification techniques is a powerful ap-
proach (Guerra et al. 2011). We used the unsupervised tech-
niques (PCA) to verify, in an exploratory unbiased way, that
the separation into different e-types is indeed a salient feature
of the data and not some preconceived notion that we have
imposed on it. Following that, we used supervised techniques
(DFA) to better understand the strongest features in terms of
separating the e-types and to construct decision regions for
classification. Although the criteria were learnt through analy-
sis of the subjective classification, they were robust enough to
uncover cases of manual misclassification (Fig. 5). Impor-
tantly, this means that we are now able to automatically classi-
fy a newly recorded cell without the need for further manual,
subjective intervention. In addition, the entire process of tran-
sition from voltage traces to electrical features to classification
is automated and requires very modest computational power;
it can thus be readily extended to additional large databases
of experimental recordings. Finally, we used the nested clus-
tering approach (chosen on the basis of our finding that
different features support different classifications) to confirm,
in an unsupervised manner, the partitioning into different
e-types. We note that the manual annotation used in this
paper was based on a single expert. More robust labeling by
taking the consensus opinion among multiple experts could
have led perhaps to even stronger results.
Features
One of the main sources of controversy regarding the classifi-
cation of interneuron e-types is the fact that there is no
“natural space”in which the electrical activity of the cells
should be described and their clustering (or lack thereof)
examined, since the raw voltage traces’data are too high di-
mensional (including as many dimensions as sampling
points). Thus, some form of dimensionality reduction is re-
quired. Typically, this takes the form of a set of features (e.g.,
firing rate, AP height, input resistance, etc.) that are used to
represent the full data (Druckmann et al. 2007), entailing a
loss of information.
The set of possible features can be quite large and any par-
ticular choice of features is partially arbitrary. Our choice in
the current paper was to include a large set of features related
to both the shape of individual APs (e.g., AP width) and the
firing pattern (e.g., firing rate). Individual features were
chosen with the aim of broadly characterizing these 2 types
of information, and with the goal of including features used
in previous studies. We note that large sets of features can be
used, since as we demonstrated in this study, when a well-
defined question is addressed (such as separation between
labeled e-types), appropriate statistical tools for feature selec-
tion are available to assist in choosing minimal sets of these
features in a principled fashion. Importantly, different fea-
tures are informative for different questions, such as the sep-
aration between steady-state e-types versus transient e-types.
One notable set of features that is likely to contain interesting
information but was not included in this study in detail is the
characterization of the firing rate of a neuron as a function of
input current (F-I curves).
For the classification of the 3 major types of cortical
neurons (FS interneurons, adapting interneurons, and Pyr
amidal cells), we found that a highly accurate classification
could be achieved even when only using a few features (in
our hands: width of AP at half-height, rate of AP discharge,
depth of AHP, time to second AP from stimulus onset). We
note that the set of features found to be effective might vary
for different data sets.
Use of Morphological, Electrophysiological, and
Molecular Descriptors
In this study, we considered the electrophysiological proper-
ties of interneurons as a basis for their analysis and classifi-
cation. Qualitatively different and equally important data
regarding interneuron identity should also be considered,
most notably morphological features of axons and dendrites,
connectivity patterns, and molecular properties. The combi-
nation of different types of descriptors is crucial and can
serve to both sharpen and validate the distinction between
different neurons (Helmstaedter et al. 2009a,2009b). We
believe that the appropriate strategy is to first consider each
type of descriptor separately and then combine them at the
final stage of the analysis. Independently analyzing different
types of descriptors allows their statistical power to be clearly
examined. For instance, we show that adapting interneurons
can be accurately distinguished from Pyr neurons based on
their electrophysiological response alone, even if these
responses appear quite similar to the naked eye. Notably, had
we included a morphological attribute, such as maximal den-
dritic branch length, this major difference between inter-
neurons and Pyr neurons might have masked the more subtle
differences in their electrophysiological characteristics. Ulti-
mately we believe the existence of neuronal types may be
most convincingly shown if descriptors of different types (e.
g., electrical, morphological, genetic) are clustered separately
Cerebral Cortex 11
by guest on September 17, 2012http://cercor.oxfordjournals.org/Downloaded from
and then combined, thereby demonstrating that the types
derived from the separate analyses can be related to each
other as well as their associations (one-to-one, one-to-many,
etc.).
A Hierarchy of Electrical Interneuron Types
In summary, we demonstrated that the PING nomenclature
could be evaluated and backed up by quantitative statistical
analysis. We found that the central aspects of the subjective
classification described by the PING mostly matched the
variability in our data and that it indeed can be parsimo-
niously explained as a partitioning of cortical cells into at
least 3 major steady-state electrical activity types. However,
the non-adapting non-FS class appears, at least in our data, to
not be a distinct class (e.g., Fig. 1). Within these types further
distinctions can be made according to transient properties.
These distinctions are finer in terms of the variability in the
data that they account for and are more difficult to quantify
because of sampling biases in favor of the more common be-
haviors. We thus propose modifying the previous description
of e-type partitioning through a 2-dimensional table
(steady-state and transient behavior as proposed by the
PING), in favor of a hierarchical view with the steady-state be-
havior being the more dominant factor (Fig. 8). Importantly,
the ability to mark subsets of cells and specifically target
them for data collection (Hempel et al. 2007) offers effective
tools to offset sampling biases inherent to the under-
representation of rare e-types and further examine these finer
partitions.
Could the approach described above be extended for cell
types beyond the neocortical interneurons considered here or
to different brain areas? Since the analysis is largely automatic,
adding different electrical properties to the feature set (e.g.,
those tied to different stimuli; Druckmann et al. 2011)to
adapt the analysis to future data sets is rather straightforward.
This is an important future direction that we will continue to
address. We hope that the strength of large experimental data-
bases shown in the framework provided above will provide
incentive for the standardization of data collection and exper-
imental procedures. Combining such data from different labs
will more robustly test the approach presented in this study
and mark new directions for extending the analysis.
Supplementary Material
Supplementary material can be found at: http://www.cercor.
oxfordjournals.org/.
Funding
This work was supported by the EPFL fund for the Blue Brain
Project, the Israel Science Foundation and by the Gatsby
Charitable Foundation.
Notes
Conflict of Interest: None declared.
References
Ascoli GA, Alonso-Nanclares L, Anderson SA, Barrionuevo G,
Benavides-Piccione R, Burkhalter A, Buzsaki G, Cauli B, Defelipe
J, Fairen A et al. 2008. Petilla terminology: nomenclature of fea-
tures of GABAergic interneurons of the cerebral cortex. Nat Rev
Neurosci. 9:557–568.
Cauli B, Porter JT, Tsuzuki K, Lambolez B, Rossier J, Quenet B,
Audinat E. 2000. Classification of fusiform neocortical inter-
neurons based on unsupervised clustering. Proc Natl Acad Sci
USA. 97:6144–6149.
Connors BW, Gutnick MJ. 1990. Intrinsic firing patterns of diverse
neocortical neurons. Trends Neurosci. 13:99–104.
Druckmann S, Banitt Y, Gidon A, Schürmann F, Markram H, Segev I.
2007. A novel multiple objective optimization framework for con-
straining conductance-based neuron models by experimental data.
Front Neurosci. 1:7–18.
Druckmann S, Berger TK, Schürmann F, Hill S, Markram H, Segev I.
2011. Effective stimuli for constructing reliable neuron models.
PLoS Comput Biol. 7:e1002133.
Duda RO, Hart PE, Stork DG. 2001. Pattern classification. New York:
Wiley.
Fisher RA. 1936. The use of multiple measurements in taxonomic pro-
blems. Ann Eugenic. 7:179–188.
Gibson JR, Beierlein M, Connors BW. 1999. Two networks of electri-
cally coupled inhibitory neurons in neocortex. Nature. 402:75–79.
Guénoche A, Hansen P, Jaumard B. 1991. Efficient algorithms for di-
visive hierarchical clustering with the diameter criterion. J Classif.
8:5–30.
Guerra L, McGarry LM, Robles V, Bielza C, Larranaga P, Yuste R.
2011. Comparison between supervised and unsupervised classifi-
cations of neuronal cell types: A case study. Dev Neurobiol.
71:71–82.
Gupta A, Wang Y, Markram H. 2000. Organizing principles for a di-
versity of GABAergic interneurons and synapses in the neocortex.
Science. 287:273–278.
Halabisky B, Shen F, Huguenard JR, Prince DA. 2006. Electrophysio-
logical classification of somatostatin-positive interneurons in
mouse sensorimotor cortex. J Neurophysiol. 96:834–845.
Helmstaedter M, Briggman KL, Denk W. 2011. High-accuracy neurite
reconstruction for high-throughput neuroanatomy. Nat Neurosci.
14:1081–1088.
Helmstaedter M, Sakmann B, Feldmeyer D. 2009a. L2/3 interneuron
groups defined by multiparameter analysis of axonal projection,
dendritic geometry, and electrical excitability. Cereb Cortex.
19:951–962.
Helmstaedter M, Sakmann B, Feldmeyer D. 2009b. The relation
between dendritic geometry, electrical excitability, and axonal pro-
jections of L2/3 interneurons in rat barrel cortex. Cereb Cortex.
19:938–950.
Hempel CM, Sugino K, Nelson SB. 2007. A manual method for the
purification of fluorescently labeled neurons from the mammalian
brain. Nat Protoc. 2:2924–2929.
Hubert L, Arabie P. 1985. Comparing partitions. J Classif. 2:193–218.
Jain AK. 2010. Data clustering: 50 years beyond K-means. Pattern
Recogn Lett. 31:651–666.
Kawaguchi Y. 1993. Physiological, morphological, and histochemical
characterization of three classes of interneurons in rat neostriatum.
J Neurosci. 13:4908–4923.
Ma Y, Hu H, Berrebi AS, Mathers PH, Agmon A. 2006. Distinct sub-
types of somatostatin-containing neocortical interneurons revealed
in transgenic mice. J Neurosci. 26:5069–5082.
Maccaferri G, Lacaille JC. 2003. Interneuron diversity series: Hippo-
campal interneuron classifications—making things as simple as
possible, not simpler. Trends Neurosci. 26:564–571.
Macnaughton-Smith P, Williams WT, Dale MB, Mockett LG. 1964. Dis-
similarity analysis: A new technic of hierarchical subdivision.
Nature. 202:1034–1035.
Mahalanobis PC. 1936. On the generalised distance in statistics. Proc
Natl Inst Sci. 2:49–55.
Mardia KV. 1970. Measures of multivariate skewness and kurtosis
with applications. Biometrika. 57:519.
McCormick DA, Connors BW, Lighthall JW, Prince DA. 1985. Com-
parative electrophysiology of pyramidal and sparsely spiny stellate
neurons of the neocortex. J Neurophysiol. 54:782–806.
12 A Hierarchical Structure of Cortical Interneuron •Druckmann et al.
by guest on September 17, 2012http://cercor.oxfordjournals.org/Downloaded from
McGarry LM, Packer AM, Fino E, Nikolenko V, Sippy T, Yuste R. 2010.
Quantitative classification of somatostatin-positive neocortical
interneurons identifies three interneuron subtypes. Front Neural
Circuits. 4:12.
Narendra PM, Fukunaga K. 1977. A branch and bound algorithm for
feature subset selection. IEEE Trans Comput. 100:917–922.
Nowak LG, Azouz R, Sanchez-Vives MV, Gray CM, McCormick DA.
2003. Electrophysiological classes of cat primary visual cortical
neurons in vivo as revealed by quantitative analyses. J Neurophy-
siol. 89:1541–1566.
Ramon y Cajal S. 1904. Textura del sistema nervioso del hombre y de
los vertebrados. Madrid: Imprenta N. Moya.
Rand WM. 1971. Objective criteria for evaluation of clustering
methods. J Am Stat Assoc. 66:846–853.
Saeys Y, Inza I, Larranaga P. 2007. A review of feature selection tech-
niques in bioinformatics. Bioinformatics. 23:2507–2517.
Somogyi P, Klausberger T. 2005. Defined types of cortical interneur-
one structure space and spike timing in the hippocampus. J
Physiol. 562:9–26.
Sugino K, Hempel CM, Miller MN, Hattox AM, Shapiro P, Wu C,
Huang ZJ, Nelson SB. 2006. Molecular taxonomy of major neur-
onal classes in the adult mouse forebrain. Nat Neurosci. 9:99–107.
Toledo-Rodriguez M, Blumenfeld B, Wu C, Luo J, Attali B, Goodman
P, Markram H. 2004. Correlation maps allow neuronal electrical
properties to be predicted from single-cell gene expression pro-
files in rat neocortex. Cerebral Cortex. 14:1310–1327.
Toledo-Rodriguez M, Goodman P, Illic M, Wu C, Markram H. 2005.
Neuropeptide and calcium-binding protein gene expression pro-
files predict neuronal anatomical type in the juvenile rat. J Physiol.
567:401–413.
Wang Y, Gupta A, Toledo-Rodriguez M, Wu CZ, Markram H. 2002.
Anatomical, physiological, molecular and circuit properties of nest
basket cells in the developing somatosensory cortex. Cereb
Cortex. 12:395–410.
Wang Y, Toledo-Rodriguez M, Gupta A, Wu C, Silberberg G, Luo J,
Markram H. 2004. Anatomical, physiological and molecular prop-
erties of Martinotti cells in the somatosensory cortex of the juven-
ile rat. J Physiol. 561:65–90.
Cerebral Cortex 13
by guest on September 17, 2012http://cercor.oxfordjournals.org/Downloaded from
