Professional Documents
Culture Documents
2015 Markram PDF
2015 Markram PDF
Correspondence
henry.markram@epfl.ch
In Brief
A digital reconstruction and simulation of
the anatomy and physiology of
neocortical microcircuitry reproduces an
array of in vitro and in vivo experiments
without parameter tuning and suggests
that cellular and synaptic mechanisms
can dynamically reconfigure the state of
the network to support diverse
information processing strategies.
Highlights
d The Blue Brain Project digitally reconstructs and simulates a
part of neocortex
*Correspondence: henry.markram@epfl.ch
http://dx.doi.org/10.1016/j.cell.2015.09.029
a b a b
A N A T O M Y
P H Y S I O L O G Y
D Electrical diversity of neurons: E Synaptic diversity: F Reconstructing virtual tissue volumes for
e-types s-types in silico experimentation
are sufficient single-cell gene and protein expression data to sub-plate pyramidal cells (L6_HPC and L6_SPC) were present
systematically identify cells, it can be extended to include molec- in the dataset and have also been reported in the literature
ular subtypes. The abbreviations used for each m-type are (Ghosh and Shatz, 1993; Hevner et al., 2001), the quality of the
provided in Figure 2. A mapping between the nomenclature stains was not sufficient for reliable reconstruction. These mor-
used in this study and alternative names present in the literature phologies are not represented in the first draft.
is provided in Table 1. Aggregating morphological reconstructions and reports in the
literature, we distinguished 55 m-types (65 if layers 2/3 are
Morphological Diversity of Neocortical Neurons considered separately and 67 if L6_HPC and L6_SPC are also
We recorded and labeled >14,000 neurons from all six layers in considered; Figure 2). Inhibitory types are mostly distinguished
the somatosensory cortex of P14 male Wistar (Han) rats, using by axonal features and excitatory types by dendritic features
patch-clamp electrodes in in vitro slices. Of these neurons, (for reviews, see Markram et al., 2004; Ramaswamy and Mark-
2,052 were sufficiently well stained to allow expert classification ram, 2015; Spruston, 2008). Figure S1 shows overlays of multiple
into m-types, based on well-established characteristic features exemplars of each of the 55 major m-types, and Figures S2A and
of their dendritic and axonal arbors, a procedure initiated by early S2B illustrate the objective classification. While in some cases,
neuroanatomists and still in use today (Fairén et al., 1984; Kara- it might have been possible to introduce a finer separation be-
giannis et al., 2009; Karube et al., 2004; Kawaguchi and Kubota, tween m-types, this would have limited the size of the samples
1997; Kisvárday et al., 1985; Larkman, 1991a; Perrenoud et al., for individual types, reducing the reliability of the classification.
2013; Peters and Kaiserman-Abramof, 1970; Ramón y Cajal, The same inhibitory types were present in all layers except
1909, 1911; Somogyi et al., 1982, 1998; Wang et al., 2004; Yuste, layer 1, which contained a unique set of inhibitory neuron types.
2005). We were able to digitally reconstruct a subset of 1,009 of Pyramidal cell morphologies varied across layers (Figure 2, right)
these neurons. This allowed validation of the expert classifica- and also with depth within layer, as illustrated by the diversity of
tion using an objective method (see below) based on clustering L23_PCs (Figure 2, upper-right). The number of pyramidal cell
of characteristic features and provided the initial pool of digital types, as defined by their local morphology, increased from
neuron models needed to reconstruct the microcircuitry. In a upper to lower layers. Several types of interneurons (e.g., LBC
few cases, we had no morphological reconstructions for rare and DBC) had axonal arbors that tended to descend to deeper
m-types known to be present in the microcircuitry (L5_BP, layers when they were in upper layers and to ascend to upper
L5_ChC, L6_NGC; Oláh et al., 2007; Szabadics et al., 2006). layers when they were in deeper layers. Consistent with this
These were represented using exemplars of the same mor- trend, one type of pyramidal cell (L6_IPC) also had inverted
phology from neighboring layers. Although L6 horizontal and axonal arbors.
Using multiple exemplars obtained from different animals for m-type. The height of the neocortex and heights of each layer
each m-type, we developed a repair process to recover arbors were measured experimentally in six animals, yielding an average
cut during the slicing process, which was validated using in vivo overall height of 2,082 ± 80 microns (mean ± SD; n = 6; Figure 3A).
reconstructed neurons (see Experimental Procedures; Anwar Layer thicknesses were determined experimentally by measuring
et al., 2009). To generate an even larger pool of unique morphol- the location of transitions in cell densities and soma sizes in
ogies, we cloned multiple exemplars of each m-type (Figures NeuN-stained tissue blocks (see Experimental Procedures). Frac-
S2C–S2F), jittering branch angles, and section lengths in the tions of excitatory and inhibitory neurons per layer (E-I fractions)
clones (see Experimental Procedures). The morphometric were established by counting cells stained for DAPI (all cells),
properties of the resulting population were validated against NeuN (all neurons), and GABA (all inhibitory neurons) in tissue
distributions of features obtained from reconstructed neurons blocks (Figure 3B; see Experimental Procedures). Overall, excit-
(see Experimental Procedures). This approach allowed us to atory and inhibitory neurons represented 87% ± 1% and 13% ±
establish a dataset of neuronal morphologies (see Movie S1A) 1% of the population, respectively, with a trend toward higher
that respects biological variability. Software applications for fractions of excitatory neurons in deeper layers (Figure 3B).
repairing and cloning in vitro neuron morphologies and for auto- The m-type composition for all excitatory and all inhibitory
mated classification of neurons into the 55 m-types are available neurons in each layer was obtained from the relative frequencies
through the NMC Portal. of each m-type in the experimental dataset of 2,052 classified
neurons mentioned earlier (Figure 3C; see Experimental Proce-
Reconstructing Neuron Densities, Ratios, and dures). It is not possible to exclude sampling bias in this dataset.
Composition However, since E-I fractions were obtained in an unbiased
Reconstruction began by specifying the dimensions of the micro- manner, any bias is restricted to the proportions of m-types
circuit, the fractions of excitatory and inhibitory neurons, the within the excitatory and inhibitory neurons and does not affect
proportions of each m-type, and the number of neurons of each the overall E-I balance.
m-type
L4N BC
L4L C
B
L4N C
B
L4S hC
L4C C1
TP
bNAC cIR L5T PC2
T
(burst Non-accommodating) (continuous Irregular) L5T TPC
L5U PC
T
40
L5S MC
L5
TC
L5B C
B
L5D BP
L5
GC
L5N BC
L5L C
B
L5N C
dNAC bIR L5S C
B
(delayed Non-accommodating) (burst Irregular) h
L5C L1
PC 20
L6TPCL4
L6T TPC
U
L6 IPC
L6 C
P
L6B C
L6MC
T
L6B C
B
L6D P
25% 1.5% 1.5% L6B
cAD N
L6 BC
GC
(continuous Adapting)
L6L C
cAC bNAC cIR L6N C
B
B
L6S hC
20 mV L6C 0
R
R
cN C
bN C
AC
cS AC
D
C
bS UT
dS UT
T
bA
cI
bI
A
cA
cA
TU
dN
T
T
200 ms
e-type
1997; McGarry et al., 2010; O’Connor et al., 2009; Packer and classification scheme, we identified 11 e-types (10 inhibitory
Yuste, 2011; Santana et al., 2013; see also NMC portal). The e-types and 1 excitatory e-type) (Figure 4A; see Experimental
observed correspondence would be unlikely in the presence of Procedures). Objective clustering of the same features produced
major errors in neuron densities, m-type composition, or posi- a similar classification, validating the original classification
tioning of reconstructed neurons. However, the biological data scheme (Druckmann et al., 2013). The fact that the e-types iden-
are highly variable, and the validation of the inhibitory m-type tified in this way have characteristic ion channel profiles provides
composition used only a small proportion of markers reported further evidence for their distinctive identity (Khazen et al., 2012;
in the literature. The reconstruction should thus be considered Toledo-Rodriguez et al., 2004).
as a first draft, to be refined as it is challenged with additional Most inhibitory m-types expressed multiple e-types (Fig-
markers. ure 4B), consistent with previous observations (Ascoli et al.,
2008; Cauli et al., 2000; Nelson, 2002; Toledo-Rodriguez et al.,
Morpho-Electrical Composition 2005). Combining m- and e-types yielded 207 morpho-electrical
We applied a standardized battery of stimulation protocols types (me-types), providing an integrated view of the morpho-
(Le Bé et al., 2007; Wang et al., 2002, 2004) to >3,900 neurons electrical diversity of the microcircuit (Figure 4C). A dataset of
from all layers, recording and analyzing their responses. The 511 morphologically and electrically classified inhibitory neurons
neurons were classified using quantified features of the neuronal was used to determine the relative proportion of e-types for each
response to step current pulses, according to the criteria estab- inhibitory m-type (in a layer-dependent manner for m-types with
lished by the Petilla convention (Ascoli et al., 2008; Figure 4A, sufficient samples and otherwise in a layer-independent manner;
top), with the exception of stuttering cells, which were consid- Figure 4C, color map; see Experimental Procedures). The rela-
ered as a separate class (see Druckmann et al., 2013). tive proportions were combined with neuron densities to calcu-
Since no significant bursting behavior was observed in excit- late the number of neurons for each me-type in each layer. The
atory m-types from animals of the age used in this study, all resulting diversity and spatial distribution of inhibitory e-types
excitatory m-types were classified as continuous adapting is illustrated in Figure 5A. This integrated view of the micro-
(cAD) neurons (Figure 4A, bottom). Using this feature-based circuitry reveals that, at this age, the most common inhibitory
mature and reach higher layers. Pooling all excitatory and inhib- ascending projections, and that intralaminar inhibition is weakest
itory cells in each layer reveals that recurrent excitation increases in layer 4 (Figure S6C).
with cortical depth while recurrent inhibition is weak in all layers, The seven statistical instantiations of the mean microcircuit
that descending interlaminar projections are stronger than (BioM) yield 636 ± 10 million appositions and 36.5 ± 0.5 million
L1NGC-DA
L1NGC-SA
L23ChC
L4PC
L1DLAC
L4SP
L1HAC
C
L1DAC
L4SS
L6ChC
L1SLA
L6SBC
L4MC
L6NB
L4BTC
Presynaptic m-type
L6L
L4DBC
15
L6N BP
L4BP
C
C
BC
L4NGC
L6
3P
L6 TC
L4LBC
GC
L4NBC
L6
DB
L2
L4SBC
B
L4ChC
C
L6
L5TTPC1
M
L5TTPC2
C
C
L5UTPC
3M
L5STPC
L5MC 10 TC
L2
L6
L5BTC
BP 3B C
L2 3DB
L5DBC
L5BP C
L5NGC
L5LBC L 3BP
2
L5NBC L2 NGC
L5SBC 3
L5ChC L2
L6TPCL1 BC
3L
L6TPCL4
L6UTPC 5 L2 C
L6IPC
L6IP 2 3 NB
L6BPC L
L6MC C SBC
L6BTC
L6DBC
L23 hC
L6BP L23C
L6NGC
L6LBC
L6NBC
L6SBC
L6ChC 0
L1DAC
L1NGC-DA
L1NGC-SA
L1HAC
L1LAC
L1SAC
L23PC
L23MC
L23BTC
L23DBC
L23BP
L23NGC
L23LBC
L23NBC
L23SBC
L23ChC
L4PC
L4SP
L4SS
L4MC
L4BTC
L4DBC
L4BP
L4NGC
L4LBC
L4NBC
L4SBC
L4ChC
L5TTPC1
L5TTPC2
L5UTPC
L5STPC
L5MC
L5BTC
L5DBC
L5BP
L5NGC
L5LBC
L5NBC
L5SBC
L5ChC
L6TPCL1
L6TPCL4
L6UTPC
L6IPC
L6BPC
L6MC
L6BTC
L6DBC
L6BP
L6NGC
L6LBC
L6NBC
L6SBC
L6ChC
L6UTPC
L4PC
Postsynaptic m-type
B L1DAC 30%
L6TPC_L4
L1NGC-DA
L1NGC-SA
L1HAC L4SP
L1LAC
L1SAC
L23PC
L23MC L4SS
L23BTC
L1
PC_
L23DBC L4M
L23BP
L23NGC L6T L4B C
L23LBC T
L23NBC L4D C
L23SBC hC L4 BC
L23ChC
L4PC L5C C B
SB L4 P
Connection probability
L4SP
L4SS 20 L5 BC NG
L4MC N L4 C
L5
Presynaptic m-type
C L4 LBC
L4BTC
L4DBC LB
L4BP L5 C L4 NB
NG P SB C
L4NGC
L4LBC
5
L L5B
L4
L4NBC C
BT C
C
L4SBC
DB
hC
C
L4ChC
L5
L5TTPC1
MC
L5TTPC2
L5
L5UTPC
PC
L5
L5STPC
L5T
C
L5MC
ST
TP
L5BTC
TP
L5DBC
L5
L5TTPC2
L5U
L5BP
10
C1
L5NGC
L5LBC
L5NBC
L5SBC
L5ChC
L6TPCL1
L6TPCL4
L6UTPC
L6IPC
L6BPC
L6MC
L6BTC
L6DBC
L6BP
L6NGC
L6LBC
L6NBC
L6SBC
L6ChC 0
L1DAC
L1NGC-DA
L1NGC-SA
L1HAC
L1LAC
L1SAC
L23PC
L23MC
L23BTC
L23DBC
L23BP
L23NGC
L23LBC
L23NBC
L23SBC
L23ChC
L4PC
L4SP
L4SS
L4MC
L4BTC
L4DBC
L4BP
L4NGC
L4LBC
L4NBC
L4SBC
L4ChC
L5TTPC1
L5TTPC2
L5UTPC
L5STPC
L5MC
L5BTC
L5DBC
L5BP
L5NGC
L5LBC
L5NBC
L5SBC
L5ChC
L6TPCL1
L6TPCL4
L6UTPC
L6IPC
L6BPC
L6MC
L6BTC
L6DBC
L6BP
L6NGC
L6LBC
L6NBC
L6SBC
L6ChC
Postsynaptic m-type
synapses (25.8 ± 0.4 million excitatory and 10.6 ± 0.2 million predicted number of synapses in the microcircuit is thus 184 ±
inhibitory; n = 7; Table S1 and Movie S1C). The lower variability 6 million (mean ± SD; n = 35), of which only 20% ± 2% of
of the statistical instantiations compared to the individual recon- synapses are formed by neurons belonging to the microcircuit
structions (Bio1–Bio5; Table S1) indicates that the variation (i.e., intrinsic synapses), consistent with previous estimates
across digital reconstructions falls well within the bounds of bio- in neocortex (Stepanyants et al., 2009). In a parallel electron mi-
logical variability. croscopy study in which we determined average synapse den-
From the space remaining on dendrites after accounting for sity (0.63 ± 0.1/mm3; mean ± SD; n = 25) and calculated the
predicted intrinsic connectivity (assuming 1.1 synapses/mm; number of synapses in a comparable volume of the neocortex,
Datwani et al., 2002; Kawaguchi et al., 2006; Larkman, we obtained 182 ± 6 million synapses. On the assumption that
1991b), we predict that afferent fibers from beyond the the average number of synapses/connection is the same for
microcircuit (extrinsic synapses) form a further 147 ± 4 million afferent fibers as for excitatory connections within the microcir-
synapses (mean ± SD; n = 35) (Figures S6D and S6E). The total cuit (3.6 ± 0.04 synapses/connection; n = 35), we predict that
2 bAC L1SAC
L23PC
L23MC
L23BTC
Na+ L23DBC
3 L23BP
cNAC L23NGC
1. Mean frequency (Hz) L23LBC
2. AP height (mV) L23NBC
3. AHP depth (mv) L23SBC
4 4. Voltage base (mV) L23ChC
5. First ISI length (ms) L4PC
Exemplar features bNAC L4SP
K+ ... 40 L4SS
Feature value
L4MC
0 L4BTC
L4DBC
L4BP
-40 dNAC L4NGC
L4LBC
-80
m-type
L4NBC
1 2 3 4 5 L4SBC
Features L4ChC
L5TTPC1
cSTUT L5TTPC2
C Quality assurance L5UTPC
L5STPC
Rejected 5 L5MC
L5BTC
L5DBC
bSTUT
L5BP
L5NGC
d
te
L5LBC
ec
L5NBC
ej
Normalized feature score (in SD)
dSTUT
R
L5SBC
L5ChC
L6TPCL1
L6TPCL4
3 L6UTPC
cIR L6IPC
L6BPC
ed
Accepted
Median z-score
L6MC
pt
2
ce
L6BTC
Ac
L6DBC
bIR L6BP
L6NGC
1 L6LBC
L6NBC
L6SBC
cAD L6ChC
0 0
cAC
bAC
cNAC
bNAC
dNAC
cSTUT
bSTUT
dSTUT
cIR
bIR
L23 PC
L4PC
L5PC
L6PC
sl o.
h
e
de t
P igh
pt
og t n
op
...
I l rs
AH he
IS Bu
AP
cAD
e-type
coefficient of variation (c.v.) of first PSPs against reported exper- in vitro PSPs (Figure 10A). The in silico PSPs were systematically
imental data (r = 0.8; Figure 9D; Gupta et al., 2000; Markram lower. Since the neuron models and the numbers and locations
et al., 1998; Wang et al., 2006). of synapses between pairs of m-types had been validated, we
We then applied unitary synaptic conductances obtained in hypothesized that the reported synaptic conductances had
previous experiments that also measured somatic postsynaptic been underestimated, because of inadequate compensation
potentials (PSP) between specific pairs of m-types and for space-clamp errors (Feldmeyer et al., 2002; Gupta et al.,
compared the resulting in silico PSPs with the corresponding 2000; Rinaldi et al., 2008). To quantify the underestimate,
0.2 mV
20 mV 0.4 mV
200 ms 200 ms
Pre
cAC NGC
BP MC BTC
PC
cAC cSTUTcNAC bNAC
Pre Vm bAC cAC cNAC bIR
I1 E2 I2 I3
Afferent synapse types to PCs
D c.v. of first PSP amplitude
1.6
Experimentally measured
E1
c.v. of first PSP amplitude
r = 0.8 Extrapolated
1.2
0.8 E1 E2 E3
I1
I3 NBC/LBC MC NBC/LBC DBC SS PC
0.4
E3 I2 E2 cAC cIR bAC bIR cNAC bSTUT dNAC bNAC cSTUT bAC cAC
L6 PC
0 BP DBC BTC SBC BP BTC NGC ChC L1
0 0.4 0.8 1.2 1.6
int.
c.v. of first PSP amplitude cAC cAC cAC cAC cNAC bNAC bAC cAC cNAC bIR
In vitro
synaptic conductances were adjusted until in silico PSPs 2007). The results suggested that reported conductances are
matched experimental levels (Figure 10B and Table S2; Angulo about 3-fold too low for excitatory connections, and 2-fold too
et al., 1999; Le Bé et al., 2007; Feldmeyer et al., 2006; Feldmeyer low for inhibitory connections (Table S2; Gupta et al., 2000; Ri-
et al., 1999, 2002; Markram et al., 1997; Silberberg and Markram, naldi et al., 2008). Other recent studies also suggest that
Post Vm
B 2.5
Internal validation MC L23PC BP BTC DBC NGC LBC NBC L5UTPC/ L5TTPC
L5STPC
PSP amplitude (mV)
2 r = 0.9 10,11,12 13
In silico
1.5
7,8 9
6
1 ChC SBC
2 4,5
3
0.5
1
0
0 0.5 1 1.5 2 2.5
PSP amplitude (mV)
In vitro
C External validation E
2.5
Quantal conductance
Inhibitory Excitatory
1.5
8
PSP amplitude (mV)
2 r = 0.6
0.43
(nS)
In silico
1.5 7 0.83
1
0.75
16
0.5 5
3 4
2
0.25 nS
0
0 0.5 1 1.5 2 2.5 3 ms
PSP amplitude (mV)
In vitro
This implies that it is a highly reproducible phenomenon, robust blies compared to randomly sampled neurons (Figure S14).
to biological and statistical variations in parameters such as Near the transition, a fall in [Ca2+]o of just 0.15 mM (Figure S14)
layer thickness, cell density, and composition; specific synaptic led to a sharp decrease in correlated spiking, clearly demar-
connectivity; and the specific dimensions of the microcircuit cating a transition in the SA spectrum.
(see Movie S3C). The mechanism underlying this sharp transition is likely to
We observed that a change in [Ca2+]o of < 1 mM can lead to a involve the differential Ca2+ sensitivities of inhibitory and excit-
transition from the synchronous to the asynchronous state, atory synapse types. Indeed, we found that changing [Ca2+]o
revealing two distinct activity regimes (Figure S12). The level of from 2 mM to 1.3 mM alters the ratio between excitatory and
[Ca2+]o at the transition varied slightly across the different instan- inhibitory synaptic PSPs by a factor of 3.5, in favor of inhibition
tiations of the microcircuit (Bio1–Bio5; Figure S13). (Figure S11). This suggests the existence of a threshold level of
Since the reconstructed microcircuitry displays synaptically Ca2+ beyond which inhibition is insufficient to prevent a super-
coupled assemblies comparable to those found experimentally critical state (see below).
(Perin et al., 2011; Reimann et al., 2015), we also analyzed corre- The finding that differential sensitivity of s-types to Ca2+
lations in neuronal activity within these assemblies. Neuronal levels determines the position of the network along the SA
activity was found to be slightly more correlated within assem- spectrum suggests that other mechanisms that change the
excitatory-inhibitory balance may have similar effects. We there- ling the position of the microcircuit on the spectrum (Figure S15).
fore performed in silico knockout experiments to understand the We found that blocking activity in the upper layers tended to shift
roles of the different layers, neurons, and connections in control- the network toward the synchronous state, while blocking the
500 10
Count
8
Count
300 6
II/III
4
100 2
0
0 5 10 15 20 25 30 700 800 900 1000 1100
Synapses/connection No. of efferent neurons per thalamic fiber
D In vitro-like
IV 70 L4 exc.
L5 exc.
50
Count
30
V
1 mV
100 ms 10
0 2 4 6 8 10
PSP amplitude (mV)
E In vivo-like
45 L4 exc. in vivo
* L5 exc. in vivo
*
35
VI
Count
25
15
0 0.5 1 1.5 2
PSP amplitude (mV)
(ER), spontaneous regenerative (SR), evoked non-regenerative from different neocortical regions that together provide statis-
(EN), and spontaneous non-regenerative (SN) (Figure 15D). tical distributions for layer heights, neuron densities, cellular
composition, and morphological and electrophysiological diver-
Reproducibility of Emergent Properties sity within and across types of neuron and reflect the diversity
The reconstructed microcircuitry is based on biological data of synaptic anatomy and physiology observed in biological
from a large number of different animals and, in some cases, experiments. The reconstruction process stochastically creates
P(spike)
(A) Raster plot of the spiking activity of an exem-
Trial
...
Count
0
in response to simulated thalamocortical stimula-
200
L4_PC tion with 60 fibers. The first spike after the stimulus
1 1 in each of 200 trials is indicated in red, other spikes
40 100
P(spike)
Trial
0.5
the neuron fires a spike in a 5 ms bin.
200
0 0 40 80 0 40 80 (B) Histograms of the response delay (delay of the
tstim Delay to first spike Delay to first spike first spike after stimulus presentation) for L23PCs
50 ms
(left) and L4PCs (right). (Black) For 200 trials of 25
C D E randomly chosen neurons of the indicated type.
Excitatory m-types L4_SP IV (Red) For 200 trials of the neurons indicated in A.
[Ca2+]o = 1.25 [Ca2+]o normalized
40 (C) Standard deviation of the response delay of
200
neurons in different layers across trials. Red line
30 60 indicates the median of neurons, blue boxes the
25th and 75th percentiles, and whiskers the full data
100 20 40 spread. (Top) Excitatory neurons; (bottom) inhibi-
tory neurons.
SD of delay to first spike (ms)
2
3
4
5
1
2
3
4
5
II/III IV V VI
io
io
io
io
io
io
io
io
io
io
B
B
B
B
B
B
B
B
B
ns
ts
m s
1
2
3
4
5
1
2
3
4
5
or uit
io
io
io
io
io
ia
ui
io
io
io
io
io
ro
.)
Tr
irc
irc
B
B
B
B
B
B
B
B
B
B
eu
C
N
(n
B1
B2 B3
designed to produce this phenomenon, it nonetheless generated shifts toward the synchronous state, the correlation is strong
excitatory conductances in single neurons that were highly but broad, resulting in a temporally imprecise increase in
correlated but effectively cancelled out by anti-correlated inhib- spiking more suitable for a rate code. When it shifts toward
itory conductances (Figure 17B). the asynchronous regime, the correlation is sharp but too
Deeper investigation revealed that spiking is correlated with weak to effectively drive spiking, a regime more suitable for a
momentary imbalances between excitatory and inhibitory con- population code based on a high degree of correlated activity
ductances lasting <10 ms and that the timing of spikes can (Figure S18).
be predicted from the difference in the E and I conductances Temporally Sequential Structure during Spontaneous
(Figure 17B). We also found that the precision with which Activity of L5 Neurons
these imbalances drive spiking falls dramatically as the network The search for precise temporal structures in brain activity, such
state shifts away from the transition in either direction. When it as synfire chains, motifs, repeated spike patterns, etc., has a
PSTH [Hz]
A1 A2 B1 40
Neuron t2 – t1 ! 150 73 20 5s
0
1 stPR
t2 – t1 (ms)
25
2 20
15
10
3 5
0
t3 – t1
-150 0 35 shuffled sTPR
Shuffling in time 10 0 1 2 3
t2 – t1; t3 – t1 (ms)
5
0
−400 −200 0 200 400
Time (ms)
B3
long history. These patterns are thought to reflect ‘‘stereotypical presynaptic neurons becomes highly synchronized, for example,
organized sequential spread of activation through local cortical by external input.
networks,’’ as demonstrated recently (Luczak et al., 2007). Luc- Soloists versus Choristers
zak et al. (2007) found a temporally sequential structure during A recent study showed that some neurons in a network display
spontaneous activity of L5 neurons in vivo in the somatosensory spiking activity that is tightly correlated with the average activity
cortex (Luczak et al., 2007). In particular, they found that, after of the population of neurons in the network (choristers), while
the onset of an UP state, trios of neurons generated spike motifs others display a diversity of spiking patterns whose correlation
(triplets) with a precisely defined temporal relationship between with that of the population is smaller than expected by chance
spikes that could not be explained by random correlations during (soloists), suggesting that they actively avoid correlating with
high-frequency spiking (see their Figure 5). A similar analysis of the rest of the population (Okun et al., 2015). We simulated the
the evoked response to thalamic stimulation of L5 neurons in spontaneous activity of a single microcircuit in the asynchronous
the digital reconstruction found the same repeating triplet struc- state but close to the transition to synchronous state for 800 s
tures as observed in vivo (Figure 18A). A second in silico exper- (Figure 18B1; see also Figure S19D) and analyzed the spiking
iment further into the asynchronous regime (i.e., at lower Ca2+ activity of every individual neuron in L5 and L6 with respect to
levels; 1.0 mM) showed no evidence of triplet structures (Figures the spiking of all others. Replication of the analysis in Okun
S19A–S19C), supporting our prediction that, in the highly asyn- et al. (2015) yielded comparable results, although the proportion
chronous regime, it is difficult for single neurons to track fine of choristers appears to be somewhat higher in the digital recon-
temporal structure in network activity unless the population of struction (Figure 18B2 in silico versus Figures 1E and 1G in vivo).
C1 C2
E1 E2
size of network required to reproduce key functional properties biological principles. For instance, additional cell type markers
of the microcircuit shows that it is roughly equivalent to the vol- would improve the accuracy of the morphological composition,
ume of neocortical tissue used as the basis for the reconstruc- saturated EM reconstructions could be used to further validate
tion. This is evidence that a network of this size is the minimum the derived connectivity, more experiments reporting combined
functional unit required for neocortical information processing. voltage and current measurements for synaptic responses will
strengthen conclusions on quantal conductances and connec-
Validity of the Digital Reconstruction tion-specific synaptic dynamics, and further characterization
The reconstruction certainly includes errors due to mistakes and of the sensitivity of different synapses to [Ca2+]o may allow
gaps in experimental datasets and incomplete understanding of more accurate demarcation of the transitions between different
[Ca2+]o (mM)
are shown for a selection of circuit sizes and cal-
1.6 cium concentrations (100% depolarization).
(B) Overview of a broad range of circuit sizes and
300
1.4 calcium concentrations as in A. Red crosses and
I
green dots indicate regenerative and non-regen-
II/III
IV erative circuit behavior, respectively, as assessed
V
1.2 by visual inspection. Black curve depicts interpo-
VI
1000 30
lated transition between regenerative and non-
Hz
~
~
0 regenerative regimes.
0 1 2 30 1 2 30 1 2 3 0 50 100 150 200 250 300 1000
Time (s) Circuit Size (minicolumns) (C) Spatial profile of instantaneous firing rates for
circuits of increasing size. Mean instantaneous
C Circuit Size (minicolumns) D firing rates were estimated for contiguous group-
50 300 1000 t1 t2 ings (clusters) of approximately ten minicolumns
Circuit Size (minicolumns)
6.0
using a K-means algorithm. Six spatial firing rate
50
activity states and may allow a more precise determination of its underlying anatomy and physiology. For instance, major
the role played by each neuron and synapse type in maintaining errors in cell morphology, densities, composition, and con-
and shifting regimes. For example, a recent EM study found nectivity would make it difficult to reproduce the types of
evidence for a higher number of synapses per connection than neuronal assemblies discovered in 12-patch experiments, the
predicted by a naive interpretation of Peters’ rule (Kasthuri numbers of GABAergic synapses on pyramidal somata, protein
et al., 2015). Applying the same analysis to the digital reconstruc- staining patterns, layer-wise synapse densities, connection
tion produced comparable findings (Figure S20 in silico versus probabilities, bouton densities, and distributions, etc. (see
Figures 7D, 7F, and S6B in Kasthuri et al., 2015). These proper- Table S3). These properties lie well within experimentally re-
ties emerge in the digital reconstruction as a consequence of ported ranges. The reproducibility of observations and pre-
preferential pruning of connections with low numbers of synap- dictions in multiple reconstructions using data from different
ses (Reimann et al., 2015). animals and incorporating statistical variations provide evidence
The validation tests conducted at multiple stages of the recon- that they are robust.
struction process reduce the risk that errors could lead to major Although the reconstruction is, to our knowledge, the most
inaccuracies in the reconstruction or in simulations of its emer- detailed to date, it omits many important details of microcircuit
gent behavior. For example, validation of electrical neuron structure and function, such as gap junctions, receptors, glia,
models against independent data insulates the emergent vasculature, neuromodulation, plasticity, and homeostasis.
behavior of the network from the impact of our limited knowledge Furthermore, it represents a snapshot of just one brain region,
of ion channel kinetics and distributions. More generally, in one strain of male rat, at a young age. This limits the gener-
the reconstruction passed multiple tests broadly validating ality of the conclusions that can be drawn. For instance, in
animals of the age used for the study, dendritic morphologies a spectrum of activity states ranging from synchronous to asyn-
have already matured to adult levels (Larkman, 1991a; Romand chronous behavior. Varying the Ca2+ level profoundly changes
et al., 2011), but the ascending axons may not be fully repre- the overall E-I balance and hence the position of the network
sented and are certainly not completely mature (Romand along the spectrum, leading to a sharp transition between activ-
et al., 2011). However, studies at a greater level of biological ity regimes. These in silico predictions were verified by new
detail (e.g., including glia, receptors, and signaling pathways) in vitro experiments.
and investigations of different brain regions in animals of Further simulations showed that the level of Ca2+, where the
different ages, gender, and species, as well as in disease transition occurred, varies across digital reconstructions that
models can use the reconstruction as a reference point. Find- use data from different animals and that this accounts for a
ings consistent with the reconstruction would indicate the significant proportion of the variance in neuronal spiking and
sufficiency of the principles of organization used in the recon- the spatial resolution of the network. We also found that a
struction process; discrepancies may point to new principles. small adjustment in Ca2+ levels (0.05 mM) in individual recon-
For example, if application of the connectivity algorithm to structions significantly reduces their physiological variability.
another brain region or to animals at a different age or belong- These simulations provide an example of how variations in in-
ing to a different species failed to yield results consistent with dividual neuroanatomy may lead to functional differences.
experimental findings, this would point to specific variations Inspired by this finding, we performed further simulations,
in the connectivity rules. which demonstrated that activating or inhibiting specific layers,
Failure in validation could also indicate errors in experi- neurons, and synaptic connections also shifts the network
mental data. For instance, the reconstruction indicated that along the spectrum. While it is well known from previous
cell densities from a dozen previous studies were all too low theoretical findings that changing E-I balance changes the state
to account for spine and synapse densities, suggesting new of the network (Brunel, 2000; van Vreeswijk and Sompolinsky,
experiments, which verified this prediction. The reconstruction 1996), the simulations further suggest that any mechanism
also revealed that many experiments underestimate synaptic that differentially changes the synaptic dynamics of different
conductances and suggests that in vitro experiments that do types of synapses (e.g., through neuromodulation; for reviews,
not account for calcium level in the bath may misinterpret see Lee and Dan, 2012; Zagha and McCormick, 2014) could
the relevance of their findings for in vivo conditions. These ex- alter the boundaries between activity regimes in complex
amples illustrate how the reconstruction process does not take ways. We speculate that other emergent properties, such as
experimental data at face value but uses complementary, UP and DOWN states with two meta-stable fixed points,
related datasets to constrain the use as parameters, wherever as observed in vivo (Steriade et al., 1993), which are not
possible. reproduced by the digital reconstruction, may require thalamo-
cortical interactions (Hughes et al., 2002), cortico-cortical inter-
Functional Implications actions (Timofeev et al., 2000), intrinsic oscillators (Lo } rincz
Simulations of the spontaneous and evoked activity that ac- et al., 2015; Sanchez-Vives and McCormick, 2000), or neuro-
counted for the differential sensitivity to Ca2+ of different types modulation (Constantinople and Bruno, 2011; Lo } rincz et al.,
of synapses and that explored changes in Ca2+ levels revealed 2015; Sigalas et al., 2015). Modulation of cellular or synaptic
Hernando, J., Schürmann, F., and Pastor, L. (2012). Towards real-time Kawaguchi, Y., Karube, F., and Kubota, Y. (2006). Dendritic branch typing and
visualization of detailed neural tissue models: View frustum culling for parallel spine expression patterns in cortical nonpyramidal cells. Cereb. Cortex 16,
rendering. BioVis, 25–32. 696–711.
Hestrin, S., and Armstrong, W.E. (1996). Morphology and physiology of cortical Keller, A., and Carlson, G.C. (1999). Neonatal whisker clipping alters intracort-
neurons in layer I. J. Neurosci. 16, 5290–5300. ical, but not thalamocortical projections, in rat barrel cortex. J. Comp. Neurol.
412, 83–94.
Hevner, R.F., Shi, L., Justice, N., Hsueh, Y., Sheng, M., Smiga, S., Bulfone, A.,
Goffinet, A.M., Campagnoni, A.T., and Rubenstein, J.L.R. (2001). Tbr1 regu- Khazen, G., Hill, S.L., Schürmann, F., and Markram, H. (2012). Combinatorial
lates differentiation of the preplate and layer 6. Neuron 29, 353–366. expression rules of ion channel genes in juvenile rat (Rattus norvegicus)
neocortical neurons. PLoS ONE 7, e34786.
Hill, S.L., Wang, Y., Riachi, I., Schürmann, F., and Markram, H. (2012). Statis-
Khazipov, R., Congar, P., and Ben-Ari, Y. (1995). Hippocampal CA1 lacuno-
tical connectivity provides a sufficient foundation for specific functional con-
sum-moleculare interneurons: modulation of monosynaptic GABAergic IPSCs
nectivity in neocortical neural microcircuits. Proc. Natl. Acad. Sci. USA 109,
by presynaptic GABAB receptors. J. Neurophysiol. 74, 2126–2137.
E2885–E2894.
Killackey, H.P., Gould, H.J., 3rd, Cusick, C.G., Pons, T.P., and Kaas, J.H.
Hines, M.L., and Carnevale, N.T. (1997). The NEURON simulation environ-
(1983). The relation of corpus callosum connections to architectonic fields
ment. Neural Comput. 9, 1179–1209.
and body surface maps in sensorimotor cortex of new and old world monkeys.
Hines, M.L., Markram, H., and Schürmann, F. (2008a). Fully implicit parallel J. Comp. Neurol. 219, 384–419.
simulation of single neurons. J. Comput. Neurosci. 25, 439–448.
Kisvárday, Z.F., Martin, K.A., Whitteridge, D., and Somogyi, P. (1985). Synaptic
Hines, M.L., Eichner, H., and Schürmann, F. (2008b). Neuron splitting in connections of intracellularly filled clutch cells: a type of small basket cell in the
compute-bound parallel network simulations enables runtime scaling with visual cortex of the cat. J. Comp. Neurol. 241, 111–137.
twice as many processors. J. Comput. Neurosci. 25, 203–210.
Klausberger, T., Magill, P.J., Márton, L.F., Roberts, J.D.B., Cobden, P.M., Buz-
Hines, M., Kumar, S., and Schürmann, F. (2011). Comparison of neuronal spike sáki, G., and Somogyi, P. (2003). Brain-state- and cell-type-specific firing of
exchange methods on a Blue Gene/P supercomputer. Front. Comput. Neuro- hippocampal interneurons in vivo. Nature 421, 844–848.
sci. 5, 49.
Le Bé, J.-V., Silberberg, G., Wang, Y., and Markram, H. (2007). Morphological,
Holmgren, C., Harkany, T., Svennenfors, B., and Zilberter, Y. (2003). Pyramidal electrophysiological, and synaptic properties of corticocallosal pyramidal cells
cell communication within local networks in layer 2/3 of rat neocortex. in the neonatal rat neocortex. Cereb. Cortex 17, 2204–2213.
J. Physiol. 551, 139–153.
Krimer, L.S., Zaitsev, A.V., Czanner, G., Kröner, S., González-Burgos, G., Po-
Honey, C.J., Kötter, R., Breakspear, M., and Sporns, O. (2007). Network struc- vysheva, N.V., Iyengar, S., Barrionuevo, G., and Lewis, D.A. (2005). Cluster
ture of cerebral cortex shapes functional connectivity on multiple time scales. analysis-based physiological classification and morphological properties of
Proc. Natl. Acad. Sci. USA 104, 10240–10245. inhibitory neurons in layers 2-3 of monkey dorsolateral prefrontal cortex.
Horton, J.C., and Adams, D.L. (2005). The cortical column: a structure without J. Neurophysiol. 94, 3009–3022.
a function. Philos. Trans. R. Soc. Lond. B. Biol. Sci. 360, 837–862. Larkman, A.U. (1991a). Dendritic morphology of pyramidal neurones of the vi-
Hughes, S.W., Cope, D.W., Blethyn, K.L., and Crunelli, V. (2002). Cellular sual cortex of the rat: I. Branching patterns. J. Comp. Neurol. 306, 307–319.
mechanisms of the slow (<1 Hz) oscillation in thalamocortical neurons Larkman, A.U. (1991b). Dendritic morphology of pyramidal neurones of the
in vitro. Neuron 33, 947–958. visual cortex of the rat: III. Spine distributions. J. Comp. Neurol. 306, 332–343.
Jones, H.C., and Keep, R.F. (1988). Brain fluid calcium concentration and Larkum, M.E., Zhu, J.J., and Sakmann, B. (2001). Dendritic mechanisms un-
response to acute hypercalcaemia during development in the rat. J. Physiol. derlying the coupling of the dendritic with the axonal action potential initiation
402, 579–593. zone of adult rat layer 5 pyramidal neurons. J. Physiol. 533, 447–466.
Reimann, M.W., Muller, E.B., Ramaswamy, S., and Markram, H. (2015). Silberberg, G., and Markram, H. (2007). Disynaptic inhibition between neocor-
An algorithm to predict the connectome of neural microcircuits. Front. Com- tical pyramidal cells mediated by Martinotti cells. Neuron 53, 735–746.
put. Neurosci. 9, 28. Silberberg, G., Wu, C., and Markram, H. (2004). Synaptic dynamics control the
Renart, A., de la Rocha, J., Bartho, P., Hollender, L., Parga, N., Reyes, A., and timing of neuronal excitation in the activated neocortical microcircuit.
Harris, K.D. (2010). The asynchronous state in cortical circuits. Science 327, J. Physiol. 556, 19–27.
587–590. Silva, L.R., Amitai, Y., and Connors, B.W. (1991). Intrinsic oscillations of
Reyes, A., and Sakmann, B. (1999). Developmental switch in the short-term neocortex generated by layer 5 pyramidal neurons. Science 251, 432–435.
modification of unitary EPSPs evoked in layer 2/3 and layer 5 pyramidal neu- Silver, R.A., Lubke, J., Sakmann, B., and Feldmeyer, D. (2003). High-probabil-
rons of rat neocortex. J. Neurosci. 19, 3827–3835. ity uniquantal transmission at excitatory synapses in barrel cortex. Science
302, 1981–1984.
Reyes, A., Lujan, R., Rozov, A., Burnashev, N., Somogyi, P., and Sakmann, B.
(1998). Target-cell-specific facilitation and depression in neocortical circuits. Simkus, C.R.L., and Stricker, C. (2002). Properties of mEPSCs recorded in
Nat. Neurosci. 1, 279–285. layer II neurones of rat barrel cortex. J. Physiol. 545, 509–520.
Reyes-Puerta, V., Sun, J.-J., Kim, S., Kilb, W., and Luhmann, H.J. (2015). Singer, W. (1993). Synchronization of cortical activity and its putative role in
Laminar and Columnar Structure of Sensory-Evoked Multineuronal Spike Se- information processing and learning. Annu. Rev. Physiol. 55, 349–374.
quences in Adult Rat Barrel Cortex In Vivo. Cereb. Cortex 25, 2001–2021. Somogyi, P., Freund, T.F., and Cowey, A. (1982). The axo-axonic interneuron
Rinaldi, T., Silberberg, G., and Markram, H. (2008). Hyperconnectivity of local in the cerebral cortex of the rat, cat and monkey. Neuroscience 7, 2577–2607.
neocortical microcircuitry induced by prenatal exposure to valproic acid. Somogyi, P., Tamás, G., Lujan, R., and Buhl, E.H. (1998). Salient features of
Cereb. Cortex 18, 763–770. synaptic organisation in the cerebral cortex. Brain Res. Brain Res. Rev. 26,
Romand, S., Wang, Y., Toledo-Rodriguez, M., and Markram, H. (2011). 113–135.
Morphological development of thick-tufted layer v pyramidal cells in the rat so- Sporns, O., and Kötter, R. (2004). Motifs in brain networks. PLoS Biol. 2, e369.
matosensory cortex. Front. Neuroanat. 5, 5. Spruston, N. (2008). Pyramidal neurons: dendritic structure and synaptic inte-
Rozov, A., Burnashev, N., Sakmann, B., and Neher, E. (2001). Transmitter gration. Nat. Rev. Neurosci. 9, 206–221.
release modulation by intracellular Ca2+ buffers in facilitating and depressing Stepanyants, A., Martinez, L.M., Ferecskó, A.S., and Kisvárday, Z.F. (2009).
nerve terminals of pyramidal cells in layer 2/3 of the rat neocortex indicates The fractions of short- and long-range connections in the visual cortex.
a target cell-specific difference in presynaptic calcium dynamics. J. Physiol. Proc. Natl. Acad. Sci. USA 106, 3555–3560.
531, 807–826.
Steriade, M., Nuñez, A., and Amzica, F. (1993). A novel slow (< 1 Hz) oscillation
Rudy, B., Fishell, G., Lee, S., and Hjerling-Leffler, J. (2011). Three groups of in- of neocortical neurons in vivo: depolarizing and hyperpolarizing components.
terneurons account for nearly 100% of neocortical GABAergic neurons. Dev. J. Neurosci. 13, 3252–3265.
Neurobiol. 71, 45–61.
Svoboda, K., Denk, W., Kleinfeld, D., and Tank, D.W. (1997). In vivo dendritic
Sakata, S., and Harris, K.D. (2009). Laminar structure of spontaneous and sen- calcium dynamics in neocortical pyramidal neurons. Nature 385, 161–165.
sory-evoked population activity in auditory cortex. Neuron 64, 404–418. Szabadics, J., Varga, C., Molnár, G., Oláh, S., Barzó, P., and Tamás, G. (2006).
Salinas, E., and Sejnowski, T.J. (2001). Correlated neuronal activity and the Excitatory effect of GABAergic axo-axonic cells in cortical microcircuits.
flow of neural information. Nat. Rev. Neurosci. 2, 539–550. Science 311, 233–235.
Sancesario, G., Pisani, A., D’Angelo, V., Calabresi, P., and Bernardi, G. (1998). Tamás, G., Somogyi, P., and Buhl, E.H. (1998). Differentially interconnected
Morphological and functional study of dwarf neurons in the rat striatum. Eur. J. networks of GABAergic interneurons in the visual cortex of the cat.
Neurosci. 10, 3575–3583. J. Neurosci. 18, 4255–4270.
Sanchez-Vives, M.V., and McCormick, D.A. (2000). Cellular and network Tan, A.Y.Y., Chen, Y., Scholl, B., Seidemann, E., and Priebe, N.J. (2014). Sen-
mechanisms of rhythmic recurrent activity in neocortex. Nat. Neurosci. 3, sory stimulation shifts visual cortex from synchronous to asynchronous states.
1027–1034. Nature 509, 226–229.
Santana, R., McGarry, L.M., Bielza, C., Larrañaga, P., and Yuste, R. (2013). Thomson, A.M., and Bannister, A.P. (1998). Postsynaptic pyramidal target
Classification of neocortical interneurons using affinity propagation. Front. selection by descending layer III pyramidal axons: dual intracellular recordings
Neural Circuits 7, 185. and biocytin filling in slices of rat neocortex. Neuroscience 84, 669–683.
Sarid, L., Bruno, R., Sakmann, B., Segev, I., and Feldmeyer, D. (2007). Thomson, A.M., and Deuchars, J. (1997). Synaptic interactions in neocortical
Modeling a layer 4-to-layer 2/3 module of a single column in rat neocortex: local circuits: dual intracellular recordings in vitro. Cereb. Cortex 7, 510–522.
interweaving in vitro and in vivo experimental observations. Proc. Natl. Thomson, A.M., and Lamy, C. (2007). Functional maps of neocortical local
Acad. Sci. USA 104, 16353–16358. circuitry. Front. Neurosci 1, 19–42.
Schroeder, C.E., and Foxe, J.J. (2002). The timing and laminar profile of Thomson, A.M., Deuchars, J., and West, D.C. (1993). Large, deep layer
converging inputs to multisensory areas of the macaque neocortex. Brain pyramid-pyramid single axon EPSPs in slices of rat motor cortex display
Res. Cogn. Brain Res. 14, 187–198. paired pulse and frequency-dependent depression, mediated presynaptically
Wang, Z., and McCormick, D.A. (1993). Control of firing mode of corticotectal Yoshimura, Y., Kimura, F., and Tsumoto, T. (1999). Estimation of single chan-
and corticopontine layer V burst-generating neurons by norepinephrine, nel conductance underlying synaptic transmission between pyramidal cells in
acetylcholine, and 1S,3R-ACPD. J. Neurosci. 13, 2199–2216. the visual cortex. Neuroscience 88, 347–352.
Wang, Y., Gupta, A., Toledo-Rodriguez, M., Wu, C.Z., and Markram, H. (2002). Yuste, R. (2005). Origin and classification of neocortical interneurons. Neuron
Anatomical, physiological, molecular and circuit properties of nest basket cells 48, 524–527.
in the developing somatosensory cortex. Cereb. Cortex 12, 395–410. Yuste, R., Tank, D.W., and Kleinfeld, D. (1997). Functional study of the rat
Wang, Y., Toledo-Rodriguez, M., Gupta, A., Wu, C., Silberberg, G., Luo, J., and cortical microcircuitry with voltage-sensitive dye imaging of neocortical slices.
Markram, H. (2004). Anatomical, physiological and molecular properties of Cereb. Cortex 7, 546–558.
Martinotti cells in the somatosensory cortex of the juvenile rat. J. Physiol. Zagha, E., and McCormick, D.A. (2014). Neural control of brain state. Curr.
561, 65–90. Opin. Neurobiol. 29, 178–186.
Wang, Y., Markram, H., Goodman, P.H., Berger, T.K., Ma, J., and Goldman- Zhu, J.J. (2000). Maturation of layer 5 neocortical pyramidal neurons: ampli-
Rakic, P.S. (2006). Heterogeneity in the pyramidal network of the medial fying salient layer 1 and layer 4 inputs by Ca2+ action potentials in adult rat
prefrontal cortex. Nat. Neurosci. 9, 534–542. tuft dendrites. J. Physiol. 526, 571–587.