Genetically targeted photostimulation of interneurons

Focal 473-nm illumination of acute neocortical slices from mice carrying homozygous recombined R26::ChR2-EGFP loci in Gad2-positive cells elicited action potentials in interneurons but not in pyramidal cells (Fig. 2a,b). The comparatively low expression levels of ChR2 from the genomic cassette required longer stimulating light pulses than those used to activate virally transduced or transfected neurons^34,35; a 20-ms pulse at 2 mW represented a favorable trade-off between reliable action potential iniation (100 and 91.1 % in cell-attached and whole-cell mode, respectively; n=55 cells) and a ~1:1 ratio of spikes per optical pulse (single action potentials in 8/10 cells in cell-attached mode; occasional doublets in 2/10).

Open in a separate window

Figure 2

Genetically targeted photostimulation of GABAergic interneurons

(a, b) Responses to photostimulation in cell-attached (a) and whole-cell current-clamp recordings (b); native resting potentials are indicated in parentheses. Interneurons follow trains of 20-ms optical pulses at 10 Hz with action potentials; pyramidal neurons are unresponsive to light. (c) ChR2-mediated photocurrents desensitize during repeated optical stimulation at frequencies >1 Hz (left) but remain stable at stimulation frequencies ≤0.2 Hz (right). (d) Spiking probabilities as a function of time after stimulus onset were estimated by analyzing 29–198 light-evoked action potentials per cell (n=4 interneurons in cell-attached recordings; colored traces). Spike times are defined as the times at which the upstroke of an action potential reaches half-maximal amplitude. Average spike latencies (± s.d.) are indicated in matching colors. (e) Wide-field optical stimulation at 5 Hz (grey bars) evokes IPSCs in pyramidal cells voltage-clamped at 0 mV (grey traces). IPSCs are abolished after bath application of 10 μM gabazine (red trace, top) or 0.5 μM TTX (red trace, bottom).

Interneurons followed optical pulse trains with maximal frequencies of 0.5–40 Hz in cell-attached (n=10; Fig. 2a) or whole-cell recordings (n=41; Fig. 2b). Individual pulses caused peak photocurrents of −188±106 pA at holding potentials of −70 mV and −145±82 pA at −60 mV (means±s.d.; n=4 cells; Fig. 2c), which evoked action potentials with latencies of 15.8±3.7 ms from the onset of illumination (means±s.d., Fig. 2d). Interneurons in areas M1, S1, and V1 were equally responsive to light, and no significant interregional differences in our ability to control fast- or non-fast-spiking cells were detected (Supplementary Fig. 2).

Driving inhibitory neurons periodically by wide-field illumination caused stimulus-locked IPSCs in pyramidal cells (Fig. 2e). Bath application of the GABA_A receptor antagonist SR95531 (gabazine, 10 μM) blocked these IPSCs (Fig. 2e), as did application of the sodium channel blocker tetrodotoxin (TTX, 0.5 μM, Fig. 2e). In contrast to the direct activation of synaptic release commonly seen in virally transduced or transfected neurons^34,35, which express ChR2 at sufficiently high levels in axons and synaptic terminals to make vesicle release resistant to TTX, only perisomatically triggered action potentials evoked transmission in our hands. Consistent with this interpretation, axonic stimulation of Purkinje cells in cerebellar slices failed to produce back-propagating action potentials or enhanced somatic depolarizations (Supplementary Fig. 3).

Illuminating the dendritic arbors of neocortical interneurons (Fig. 3a) also generated only insignificant photocurrents: Depolarization amplitudes and spiking probabilities decayed in a roughly radial-symmetric manner as a function of somatic distance, without stimulation hotspots at underlying dendrites (Fig. 3a,b). Our approach thus offers two hitherto unrealized advantages for mapping connectivity: First, it distinguishes synaptic inputs originating locally (that is, from neurons whose somata lie within the illumination cone) from axonal or dendritic projections into the illuminated volume. And second, the optical stimulus selectively addresses inhibitory neurons. Photolysis of caged glutamate, in contrast, activates both inhibitory^24,25 and excitatory^17,20,25 neurons and is therefore prone to mapping polysynaptic connections, because interneurons at some distance from the stimulation site may be recruited indirectly.

Open in a separate window

Figure 3

Optogenetic mapping of inhibitory connectivity

(a) Contour plots depict spiking probabilities (left) and depolarization amplitudes (right) of three interneurons as functions of stimulus location. Blue dots indicate stimulation points; arrowheaded scale bars of 100 μm point to the pial surface. Perisomatic illumination reliably elicits action potentials (left). Positioning the focus of the stimulating beam near dendritic branches does not cause higher spiking probabilities or larger depolarizations than illumination of dendrite-free neuropil equidistant from the soma (right). (b) Spiking probabilities of 9 interneurons as functions of the distance of the stimulation spot from the soma. Cells were recorded in the cell-attached (n=4 cells, green traces) or whole-cell configuration (n=5 cells, blue traces). (c) Sequential illumination of 10 different locations at 2.5 Hz (20 ms, 2 mW, grey tick marks). Illumination of sites marked by blue arrows gives rise to reproducible IPSCs in the recorded pyramidal cell. (d) Maps of inhibitory inputs to pyramidal neurons in V1 (top row) or S1 (bottom row), located in L2/3 (left column) or L5B (right column) at comparable depths (± 7 μm) from the surface of the slice. Two neurons were recorded simultaneously to test whether cells within the same local network could exhibit different connectivity patterns. Cell positions are indicated by triangles; a filled triangle denotes the postsynaptic target for each map. Color on a heat scale symbolizes the average amount of charge flowing during 100 ms following the onset of the IPSC, at a holding voltage of 0 mV.

However, it is formally possible also for an inhibitory input to propagate polysynaptically if the targets of inhibition generate rebound spikes. We tested this scenario and found no evidence in its favor (Supplementary Fig. 4). Pyramidal cells and fast-spiking as well as non-fast-spiking interneurons were hyperpolarized to increasingly negative holding potentials. Rebound spikes appeared in 4/19 pyramidal cells and 3/13 fast-spiking interneurons, but only when these neurons were released from holding potentials <−150 mV (Supplementary Fig. 4a,b). Non-fast-spiking interneurons emitted rebound spikes more frequently (11/24 cells) and readily (Supplementary Fig. 4c), but the spike latencies from the offset of shallow hyperpolarizations were so large (>100 ms) that any IPSCs caused by these spikes would have been excluded in our analysis because the temporal contingency between light pulse and postsynaptic event was broken. To generate short-latency (≤50 ms) rebound spikes, non-fast-spiking interneurons needed to be hyperpolarized below –117 mV (Supplementary Fig. 4c). GABAergic IPSCs could achieve this degree of hyperpolarization only at chloride reversal potentials of ≤−120 mV and, therefore, unphysiologically low³⁶ intracellular chloride concentrations. Our IPSC images thus depict the origins of monosynaptic inhibitory connections.

Optogenetic mapping of inhibitory input distributions

To analyze the laminar and columnar organization of local inhibitory connections, one or two excitatory neurons were voltage-clamped at 0 mV (to maximize chloride currents through GABA_A receptors and minimize spontaneous EPSC amplitudes), while a focused laser beam rastered a grid of locations spaced at 55-to-65-μm intervals. The grid spacing reflects the lateral resolution of photostimulation, limited by light scattering in the slice (Fig. 3a,b).

Stimulus-locked IPSCs could be evoked by focal illumination of certain locations but not others, revealing the somatic positions of inhibitory interneurons presynaptic to the recorded excitatory cell(s) (Fig. 3c,d). The distributions of input sources were reproducible for the same postsynaptic neuron during repeated sweeps of the stimulation grid (Fig. 3c) but differed for two simultaneously recorded cells located nearby in the same slice (Fig. 3d).

A cautionary remark, however, is appropriate. Holding the excitatory targets of inhibition at 0 mV could lead to depolarization-induced suppression of inhibition (DSI), which might mask inhibitory inputs^37,38. DSI is mediated by retrograde endocannabinoid signals³⁸ that communicate the depolarization of postsynaptic neurons to inhibitory presynaptic terminals, which downregulate GABA release upon activation of cannabinoid receptor-1 (CB-1). To test for a possible effect of DSI, we compared the inhibitory input maps of the same six pyramidal cells when these cells were voltage-clamped at −70 and 0 mV (Supplementary Fig. 5). No differences in the number and distribution of inhibitory inputs were found (P=0.821). In a second set of control experiments, we compared the inhibitory input maps of 17 cells in the presence and absence of the CB-1 antagonist AM251 at 2 μM (Supplementary Fig. 5). Again, no differences in input number or map structure were detected (P=0.379).

Optical raster stimulation allowed us to identify sources of inhibitory inputs to excitatory cells in M1, S1, and V1 and assemble input maps for 6–22 excitatory neurons per layer and cortical area (Supplementary Table 2), yielding a data set of 3,823 inhibitory-to-excitatory connections. The identity and gross integrity of all recorded neurons were confirmed post hoc by visualizing spines on the neurobiotin-filled arbors (Supplementary Fig. 6).

Columnar center-surround structure of inhibitory circuits

Sources of inhibition were confined within horizontal domains spanning at most 550 μm, or 3 somatosensory barrels (179±28 μm per barrel; n=17, Fig. 4), suggesting that inhibitory microcircuits interconnect adjacent columns (Supplementary Table 3). Because the columnar structure is visible in S1, this interpretation could be tested directly by recording simultaneously from pairs of L2/3 pyramidal neurons in the same (n=5 pairs) or adjacent (n=6 pairs) barrel-related columns. While the input maps of L2/3 neurons in the same column overlapped to a large degree (Fig. 4a), those of cells in adjacent columns were shifted by about one barrel diameter (Fig. 4b). Cross-correlation analysis (see Methods) confirmed that the input maps of two neurons in the same barrel-related column were displaced by slightly less than the average distance between the two cells (0.32±0.28 vs. 0.39±0.11 barrel widths, means±s.d.), whereas the maps of cells in different columns were shifted by an average of 0.94±0.09 barrel widths (P=0.002). As would be expected if map displacements occurred in discrete, column-sized steps, they were better described by a step function with a level change at the barrel septum than a linear function of intercell distance (Supplementary Fig. 7). Inhibitory connections thus appear to follow a pericolumnar “center-surround” arrangement, which provides an anatomical substrate by which activity in one column can suppress that of its immediate neighbors^1,39.

Open in a separate window

Figure 4

Horizontal (columnar) organization of inhibitory connections

(a, b) Overlay maps of inhibitory inputs to pairs of simultaneously recorded pyramidal neurons in layer 2/3 of S1. The maps depict the locations of inhibitory inputs but not their strength and have been scaled to the size of a standard somatosensory barrel (yellow outlines). Cell positions are marked by triangles. Data from the left cell in each pair are coded in blue, data from the right cell in red. (a) Pairs of pyramidal neurons in the same barrel-related column. (b) Pairs of pyramidal neurons in adjacent barrel-related columns. (c) Horizontal profiles of input distributions show the interpolated number of inhibitory inputs as a function of horizontal distance from the center of the input distribution for each layer, ignoring the laminar (vertical) location of these inputs. Horizontal distances were scaled to the size of a standard somatosensory barrel (yellow outlines) and center-aligned; the number of inhibitory inputs to an excitatory neuron in a given layer at a given distance from the map center was then averaged. The intensity of blue color symbolizes the density of input sources. See Methods and Supplementary Table 3 for statistical detail.

The horizontal reach of inhibitory connections varied from layer to layer, giving rise to hourglass-shaped input profiles with notable constrictions in L4 of sensory cortices (Fig. 4c, Supplementary Table 3). This observation is consistent with the idea that lateral inhibition constrains the flow of information to columnar units in supra- and infragranular layers³⁹, whereas columns in L4 are largely defined by the parcellation of thalamocortical projections.

Area-specific laminar organization of inhibitory circuits

In contrast to a horizontal structure that appeared similar across cortical areas (Fig. 4c, Supplementary Table 3), the vertical organization of inhibitory connections showed clear area-specific variations. With the exception of neurons in L6, excitatory cells in different areas displayed distinct laminar source distributions of inhibitory inputs (Fig. 5, Supplementary Tables 4–7). While the dominant source of inhibition was usually intralaminar—that is, originating from interneurons residing in the same layer as the soma of the excitatory neuron^19,24—striking exceptions to this rule were found, particularly in V1. Here, pyramidal cells in L5B were strongly inhibited by interneurons in L6 (Fig. 5d, Supplementary Tables 4–6) and cells in L2/3 and L4 by interneurons in L5, especially L5B (Fig. 5a,b; Supplementary Tables 4–6). Quantitatively, the total inhibitory charge flow (Fig. 5) as well as the normalized numbers of inhibitory inputs (Supplementary Table 5) from the dominant translaminar source (L6 for L5B, and L5B for L2/3 and L4) were at least twice as large in V1 as in M1 and S1 (see also Fig. 6c).

Open in a separate window

Figure 5

Vertical (laminar) organization of inhibitory connections

(a–e) Average strength of inhibitory input from the indicated source layers (rows) to excitatory neurons located in L2/3 (a), L4 (b), L5A (c), L5B (d) and L6 (e). An example of a neurobiotin-filled excitatory neuron, recovered after recording, is shown to the left of each panel. The figure summarizes data from 30 neurons in M1, 54 neurons in S1, and 53 neurons in V1. The strength of a connection is expressed as the average percentage of inhibitory charge flow arising from identified inputs in a layer. L5 represents the sum of L5A and L5B. Values are represented numerically (s.d. in parentheses) and by the intensity of grey shading. Colored boxes indicate significant differences (P < 0.05), either between two cortical areas (red) or between one area and the other two (blue), as determined by one-way ANOVA and Tukey-HSD post-hoc tests (Supplementary Tables 2 and 4).

Open in a separate window

Figure 6

Area-specific differences in the laminar organization of inhibitory connections

(a) Discriminant analysis of the laminar source distributions of inhibitory inputs to excitatory neurons in different target layers of M1, S1, and V1 (Supplementary Table 8). Neurons are represented as points in the coordinate system spanned by the discriminant functions Y1 and Y2. Borders between colored areas indicate decision boundaries for assigning neurons to M1 (blue), S1 (yellow), and V1 (red). Data points whose fill color matches the background color are classified correctly. Black squares indicate the centroid positions for all cells in each cortical area. (b) Abundance of interneuron subtypes expressing parvalbumin (P), somatostatin (S), and VIP (V) in in the respective cortical layers and areas; pie chart diameters represent overall interneuron densities (Supplementary Table 9). (c) Percentages and absolute numbers of inputs, charge flow, and failure rates (means+s.d.) of the four translaminar motifs exhibiting area-specific differences. Colored bars symbolize data for M1 (blue), S1 (yellow), and V1 (red). Asterisks denote statistically significant differences between cortical areas (P < 0.05), as determined by ANOVA and Tukey-HSD post-hoc tests (Supplementary Tables 5, 6, 7 and 10).

Evidence of putative feedforward inhibition between layers, or of descending inhibition from superficial to deeper laminae, was scant, with two notable exceptions. In S1 and V1, an inhibitory L4-to-L2/3 connection¹⁸ ran parallel to the flow of excitatory signals⁸ from spiny stellate neurons of L4 (Fig. 5a). And in S1, a 5-fold larger amount of L4-derived inhibition than in V1 entered L5A, via marginally faster synaptic connections (Fig. 5c; 20-80 % rise time=5.8±1.1 ms in S1 vs. 7.6±1.5 ms in V1, P=0.06).

Area-specific differences in vertical inhibitory connectivity (Figs. ​(Figs.55 and ​and6a)6a) were so characteristic that discriminant analysis could assign >70 % of excitatory neurons of all layers correctly to their cortical areas of origin, based on the laminar distribution of their inhibitory inputs alone (Fig. 6a, Supplementary Table 8). This was true irrespective of whether inhibition was quantified by measuring inhibitory charge flow or the number of synaptic inputs from a layer (Supplementary Table 8). The classification was most accurate in M1 (96 %) and V1 (75 %) but less accurate in S1 (63 %; averages across layers 2/3 to 5B).

Multiple factors could underlie these regional differences in vertical inhibitory connectivity, singly or in combination. The source layers of translaminar inhibition (L4, L5B, and L6) could be populated with interneurons, or particular subtypes of interneurons, at different densities. Alternatively, the translaminar connection probabilities of interneurons and excitatory cells could vary between areas. Finally, regional variations could be due to differences in the properties of these connections, such as their release probabilities, synaptic conductances, and the distribution of synapses along the somatodendritic axis of pyramidal cells. To distinguish between these possibilities, we began by comparing the abundances of the three principal interneuron subtypes defined cytochemically³³ in L4, L5B, and L6 of V1, S1, and M1. No consistent covariation between the density of inhibitory neurons or the subtypes expressing parvalbumin, somatostatin, or VIP and the strength of inhibition originating in a layer was detected (Fig. 6b, Supplementary Table 9). We next compared the absolute and normalized numbers of identified translaminar connections (Fig. 6c, Supplementary Tables 5 and 6) as well as three of their functional properties (Fig. 6c, Supplementary Tables 7, 10, and 11): the frequency of transmission failures (an index of release probability, given the uniform light sensitivity of interneurons across areas, as documented in Supplementary Fig. 2); the inhibitory charge flow per IPSC (which is proportional to synaptic conductance); and the 20-80% rise time of IPSCs (which can reflect the electrotonic distance of a synapse from the soma, and is therefore a proxy of anatomical distance). Regional differences in inhibitory connectivity were to a large extent caused by variable connection probabilities; in one instance (the L4-to-L5A motif), there was an additional increase in inhibitory charge flow per IPSC (Fig. 6c). The relative prominence of translaminar inhibition from L6 to L5B in V1, in contrast, was not due to an absolute increase in the number or strength of L6-to-L5B connections but rather a reduction of inputs from other laminar sources (Supplementary Table 5). Differences in transmission failures (Fig. 6c) or the somatodendritic locations of inhibitory synapses, as inferred from IPSC rise times measured in individual trials (Supplementary Table 11), played no discernible roles.

Are inhibitory neocortical microcircuits canonical?

We return to this question, which served as our point of departure, with an ambiguous answer. If the term “canonical” is interpreted in the strictest possible sense, to imply that precisely the same laminar and tangential organization of inhibitory connections is repeated across the entire neocortical mantle, our discovery of characteristic area-specific variations in inhibitory connectivity clearly refutes the notion of canonical wiring. Still, we find that the same distinctly recognizable—perhaps even “canonical”—motifs of inhibitory-to-excitatory connectivity recur in most of the cortical areas we have examined: It is the frequency with which these structural elements are present, not their configuration, that varies between regions.

The most common circuit motif is lateral, intralaminar inhibition of excitatory neurons by interneurons located in either the same column or an immediate neighbor. The abundance of this motif in V1, S1, and M1 supports the traditional view that inhibition is largely local, intralaminar, and uniform across areas^{14,17-19,21-26}. However, there is also evidence that interneuron axons ramify extensively beyond laminar borders^16,40 and form synaptic contacts across layers^18,21,25. For example, although the ascending axons of somatostatin-positive, non-fastspiking Martinotti cells of L5 are widely thought to connect preferentially with the apical dendrites of L5 pyramidal cells^22,26,41, there are now at least three electrophysiologically documented cases of L5 Martinotti cells targeting L2/3 pyramidal neurons²¹.

The high throughput of our optogenetic mapping technique has facilitated the detection of these rarer connectivity patterns and allowed us to supplant anecdotal evidence from pairwise recordings with quantitative estimates of motif frequencies in a large data set encompassing three cortical areas. Nine translaminar motifs were found, four of which varied between areas (Supplementary Fig. 8). These motifs include inhibitory-to-excitatory connections from L6 to L5B, from L5 (in particular, L5B) to L2/3, from L5B to L4, and from L4 to L5A. The same structural elements are generally present in different cortical regions, but in each region only a subset of excitatory neurons takes part in them (Fig. 7): Like excitatory synapses^18,20,42, inhibitory terminals choose postsynaptic targets with high selectivity²⁵.

The most striking examples of translaminar inhibition were seen in V1, where some pyramidal cells in L5B were strongly inhibited by neurons in L6 and cells in L2/3 and L4 by interneurons in L5, especially L5B (Fig. 5). The likely conductors of these inhibitory signals are Martinotti cells, which are thought to provide normalizing dendritic inhibition through axonal arbors extending above their home layers^21,22,40,41. Because Martinotti cells are driven intracortically^21,22,26,43 rather than by thalamic afferents⁴⁴, these translaminar inhibitory connections may form part of inhibitory feedback loops from the canonical target layers^6-9 of excitatory output: One such loop connects L5 back to L2/3; another loop links L6 back to L5B (Fig. 5, Supplementary Fig. 8).

We thank P. Chambon, F. Costantini, S. Dymecki, G. Nagel, and P. Soriano for plasmids and T. Ellender, M. Kohl, K. Lamsa, W. Nissen, T. Nottoli, R. Roorda, Y. Tan, and L. Upton for assistance and/or discussions. This work was supported by the Medical Research Council (G.M.), NIH (G.M.), the Dana Foundation (G.M.), the Office of Naval Research (G.M.), the Boehringer Ingelheim Fonds (D.K.), and the Christopher Welch Scholarship Fund (D.K.).