Somatostatin-Expressing Inhibitory Interneurons in Cortical Circuits — TEI-tagged view

YavorskapersName, WehrpersName (2016date). “Somatostatin-Expressing Inhibitory Interneurons in Cortical Circuits.” Frontiers in Neural Circuitstitle·j.

Object of study — SOMterm cells in cerebral cortexterm, with an emphasis on mice

Research question — whether the classification criteriaterm·var can vary independently or are correlated, which is what decides how many distinct subtypesterm there could be - and the assumptions are POORLY CONSTRAINED BY DATA BECAUSE MOST STUDIES HAVE MEASURED ONLY A SUBSET OF THESE CRITERIAterm

Contribution — an estimate of how many kinds of thing exist: THE TRUE NUMBER OF DISTINCT SUBTYPESterm OF SOMterm CELLS IS UNKNOWN, and it MOST LIKELY LIES SOMEWHERE BETWEEN 4 AND 100 SUBTYPESterm

Full text — tagged

The reference list is not part of the text we hold for this paper. Our copy was extracted from the publisher's page and the bibliography did not survive that extraction, so the article below ends before its references. Nothing has been removed by us.

Somatostatin-Expressing Inhibitory Interneuronsterm in Cortical Circuits Iryna YavorskapersName Iryna YavorskapersName 1 Institute of Neuroscience and Department of Psychology, University of Oregon, Eugene, OR, USA 1 , Michael WehrpersName Michael WehrpersName 1 Institute of Neuroscience and Department of Psychology, University of Oregon, Eugene, OR, USA 1, * 1 Institute of Neuroscience and Department of Psychology, University of Oregon, Eugene, OR, USA Edited by: Erika E. Fanselow, University of Pittsburgh, USA Reviewed by: Yoshiyuki Kubota, National Institute for Physiological Sciences, Japan; Luc J. Gentet, Centre de Recherche en Neuroscience de Lyon, France ✉ *Correspondence: Michael WehrpersName wehrpersName@uoregon.edu Received 2016 May 11; Accepted 2016 Sep 12; Collection date 2016.

Copyright © 2016 YavorskapersName and WehrpersName.

This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

PMCID: PMC5040712 PMID: 27746722 Abstract Cortical inhibitory neurons exhibit remarkable diversity in their morphologyterm, connectivity, and synaptic properties. Here, we reviewterm·mth the function of somatostatin-expressing (SOMterm) inhibitory interneuronsterm, focusing largely on sensory cortexterm. SOMterm neurons also comprise a number of subpopulations that can be distinguished by their morphologyterm, input and output connectivity, laminar location, firing propertiesterm, and expression of molecular markersterm. Several of these classesterm of SOMterm neurons show unique dynamics and characteristics, such as facilitating synapses, specific axonal projections, intralaminar input, and top-down modulation, which suggest possible computational roles. SOMterm cells can be differentially modulated by behavioral state depending on their classterm, sensory system, and behavioral paradigm. The functional effects of such modulation have been studied with optogeneticterm manipulation of SOMterm cells, which produces effects on learning and memory, task performance, and the integration of cortical activity. Different classesterm of SOMterm cells participate in distinct disinhibitory circuits with different inhibitory partners and in different cortical layers. Through these disinhibitory circuits, SOMterm cells help encode the behavioral relevance of sensory stimuli by regulating the activity of cortical neurons based on subcortical and intracortical modulatory input. Associative learning leads to long-term changes in the strength of connectivity of SOMterm cells with other neurons, often influencing the strength of inhibitory input they receive. Thus despite their heterogeneity and variability across cortical areas, current evidence shows that SOMterm neurons perform unique neural computations, forming not only distinct molecular but also functional subclasses of cortical inhibitory interneuronsterm.

Keywords: somatostatinterm, Martinottiterm, GINterm, x94term, x98term, VIP, parvalbumin, disinhibitory Introduction Inhibitory interneuronsterm represent about 20–30% of all cortical cells in mammals ranging from mice to humans (Hendry et al., 1987 ; Tamamaki et al., 2003 ; Markram et al., 2004 ; Sherwood et al., 2010 ). Interneuronsterm exhibit remarkable diversity in their morphologyterm, histochemistry, intrinsic membrane properties, and connectivity. This diversity strongly suggests that different types of interneuronsterm play distinct roles in cortical computation, although only the first glimmers of these functional roles have so far been brought to light. Although inhibitory interneuronsterm can be classified by many different characteristics, a widely used approach is to identify unique molecular markersterm such as neuropeptides or calcium binding proteins. This method has gained increasing popularity in recent years, because the promoters for such cell-type-specific genes provide access for targeting the expression of genetic tools to specific subsets of cells. Based on histochemical markersterm, we can divide cortical inhibitory cells into three non-overlapping categoriesterm in mice: those that express parvalbumin (PVterm), somatostatinterm (SOMterm), or vasointestinal peptide (VIP). These categoriesterm vary across species; in rats, for example, PVterm, SOMterm, and calretinin (CR) cells form non-overlapping categoriesterm (Gonchar and Burkhalter, 1997 ; Kawaguchi and Kubota, 1997 ), whereas mice show overlapping expression of SOMterm and CR (Freund and Buzsáki, 1996 ; Somogyi and Klausberger, 2005 ; Ascoli et al., 2008 ; Fishell and Rudy, 2011 ; Rudy et al., 2011 ). While these 3 major classesterm don't account for all inhibitory interneuronsterm (a small number of interneuronsterm express other less common markersterm), these 3 major classesterm do account for the vast majority (80–90%) of all inhibitory cells (Gonchar and Burkhalter, 1997 ; Rudy et al., 2011 ; Pfeffer et al., 2013 ).

In this reviewterm·mth, we focus on SOMterm cells in cerebral cortexterm, with an emphasis on mice. For an excellent recent reviewterm·mth of SOMterm interneuronsterm in cortical circuits, see Urban-Ciecko and Barth ( 2016 ). Because SOMterm cells differ markedly in many respects from PVterm and VIP cells, we will briefly reviewterm·mth some of the distinctive characteristics of those cell types. PVterm cells are by far the largest category of inhibitory cells, representing 30–50% of all inhibitory interneuronsterm (Tamamaki et al., 2003 ; Rudy et al., 2011 ). Although PVterm cells are not a homogenous population, they do appear to share several features. PVterm cells are found throughout cortical layers 2–6 and are typically fast spiking (FS) cells with narrow spike waveforms. However, not all PVterm cells are FS cells, and not all FS cells are PVterm cells (Cauli et al., 2000 ; Markram et al., 2004 ; Moore and WehrpersName, 2013 ). PVterm cells tend to target the somata and proximal dendrites of both excitatory cells and other PVterm cells (Kubota et al., 2016 ). They provide powerful inhibitionterm, but since they form depressing synapses, this inhibitionterm is relatively short-lived (Beierlein et al., 2003 ). Although it is still unclear whether PVterm cells perform similar functions in different sensory regions, current evidence suggests that they most likely provide gain controlterm in cortical networks by indiscriminately pooling locally available excitatory input and feeding this back to both PNs and other PVterm cells (Brumberg et al., 1996 ; Gabernet et al., 2005 ; Higley and Contreras, 2006 ; Cruikshank et al., 2010 ; Moore and WehrpersName, 2013 ). In sensory cortical areas with columnar organization, PVterm cells are thought to pool local input from similarly-tuned PNs, and are therefore well-tuned. For example, auditory cortexterm has columnar organization for sound frequency and PVterm cells there are well-tuned for frequency (Moore and WehrpersName, 2013 ). In contrast, mouse visual cortexterm does not have columnar organization for orientation, and PVterm cells pool input from heterogeneously-tuned PNs, and are therefore more broadly tuned for orientation than PNs (Niell and Stryker, 2008 ; Kerlin et al., 2010 ; Atallah et al., 2012 ).

While VIP cells comprise only 1–2% of all cortical cells, recent studies in a number of cortical areas have revealed that VIP cells provide weak inhibitionterm to PVterm networks and strong inhibitionterm to SOMterm networks, and thus indirectly regulate the activity of the local population of PNs (Lee et al., 2012 ; Hioki et al., 2013 ; Pfeffer et al., 2013 ; Pi et al., 2013 ; Karnani et al., 2016a ). In the context of understanding SOMterm networks, VIP cells are of particular interest because they target SOMterm cells strongly in layers 2/3 (and also weakly in layer 5), forming robust disinhibitory circuits (Lee et al., 2013 ; Pfeffer et al., 2013 ; Karnani et al., 2016a ; Figure 1 ). These disinhibitory circuits appear to be engaged under specific behavioral conditions including associative learning, reinforcement, locomotion, and attention (Uematsu et al., 2008 ; Letzkus et al., 2011 ; Pi et al., 2013 ; Fu et al., 2014 ; Kepecs and Fishell, 2014 ; Pala and Petersen, 2015 ). The axons of VIP cells extend vertically within a narrow column, thereby inhibiting mainly local SOMterm cells. VIP cells may therefore “open holes in the blanket of inhibitionterm” that is provided by SOMterm cells to local PNs (Karnani et al., 2016a ). VIP cells also belong to a subgroup of neurons that express the 5HT3a serotonin receptor, which also includes neurogliaform and late-spiking as well as a subset of cholecystokinin (CCK), CR, or neuropeptide Y (NPY) expressing neurons (Lee et al., 2010 ; Rudy et al., 2011 ).

Figure 1.

Summary diagram of the cortical circuits in which SOMterm cells participate . Numbers in yellow boxes refer to the references below. Some microcircuits combine results from different studies and include speculative connections. Connections not mentioned in the main text or not pertaining to SOMterm networks are omitted for simplicity. Thin lines indicate weaker connections. Figure references: 1. Adesnik et al. ( 2012 ); 2. Xu et al. ( 2013 ); 3. Kapfer et al. ( 2007 ); 4. Silberberg and Markram ( 2007 ); 5. Wang et al. ( 2004 ); 6. Cottam et al. ( 2013 ); 7. Pi et al. ( 2013 ); 8. Fu et al. ( 2014 ); 9. Karnani et al. ( 2016a ); 10. Kawaguchi et al. ( 1997 ); 11. Li et al. ( 2015 ); 12. Cruikshank et al. ( 2010 ); 13. Beierlein et al. ( 2003 ); 14. Fanselow et al. ( 2008 ); 15. Letzkus et al. ( 2011 ); 16. Tomioka et al. ( 2005 ); 17. Endo et al. ( 2016 ).

How many distinct kinds of SOMterm cells are there?

SOMterm cells compose 30% of all inhibitory cells in the cortexterm (Xu and Callaway, 2009 ; Rudy et al., 2011 ), and these can be further subdivided into smaller distinct groups based on layer, physiology, morphologyterm, co-expression of other markersterm, and synaptic targets. These approaches typically produce partially overlapping categoriesterm, producing an inevitable tension between the tendency to be a “lumper” or a “splitter.” Anatomically, for example, the most distinctive type of SOMterm cell is the Martinotti cellterm·def (Martinottiterm, 1889 ; Karube et al., 2004 ; Wang et al., 2004 ; Ma et al., 2006 ). The most striking feature of Martinotti cellsterm·def is their characteristic axonal projection to layer 1, where they make extensive lateral arborizations (Wang et al., 2004 ; Ma et al., 2006 ; Gentet, 2012 ). All Martinotti cellsterm·def are SOM-positive, however not all SOM-positive cells are Martinotti cellsterm·def. Martinotti cellsterm·def are mostly located in supragranular layers 2 and 3, but can also be found sparsely in layers 4, 5, and 6. Their dendrites branch locally or down to deeper layers (Wang et al., 2004 ). Because Martinotti cellsterm·def make up the largest and best-studied category of SOMterm cells, it is tempting to to lump together all other SOMterm cells as “non-Martinotti,” a category that would include multiple anatomical classesterm such as subsets of basket cells, bitufted, horizontal, and multipolar cells as well as long-projecting GABAergic neurons (Rogers, 1992 ; Reyes et al., 1998 ; Ma et al., 2006 ; McGarry et al., 2010 ; Suzuki and Bekkers, 2010 ; Kubota et al., 2011 ).

A complementary categorization approach has been to take advantage of transgenic mouse lines such as the GINterm, X94term, and X98term lines (Ma et al., 2006 ). These three different lines of transgenic GAD67-eGFP mice were generated by pronuclear injection (i.e., not by knock-in), and fortuitously label subsets of SOMterm cells (most likely due to insertional effects depending on where GAD67-eGFP randomly inserted into the genome). These lines are an excellent tool for restricting GFP expression to SOMterm cells. But because they label only subsets of SOMterm cells, one must be careful not to infer relative proportions of SOMterm subtypesterm from studies using these lines. The GINterm line labels mostly Martinotti cellsterm·def, and most of these are found in L2/3, with sparse labeling in L5. GINterm cells account for 35% of SOMterm cells (Oliva et al., 2000 ; Ma et al., 2006 ). Targeted patching of these cells reveals that most of them have intrinsic firing propertiesterm characteristic of regular-spiking (RS) cells with generally depolarized membrane potentials, which distinguishes them from other types of SOMterm cells (Ma et al., 2006 ; McGarry et al., 2010 ; Kinnischtzke et al., 2012 ). L2/3 GINterm cells are also likely to be electrically coupled to each other (with 66% likelihood) and are strongly activated by cholinergic signaling (Fanselow et al., 2008 ).

The X98term line labels Martinotti cellsterm·def in L5 and upper L6, accounting for 20% of all SOMterm cells. X98term cells have distinctive intrinsic firing propertiesterm −40% of them are low-threshold spiking (LTSterm) cells. These cells are neither fast-spiking nor regular-spiking, but instead fire a characteristic rebound spike when depolarized from a relatively hyperpolarized holding potential, often in bursts. All LTSterm cells are inhibitory, but only about half of LTSterm cells are SOM-positive, and of those most are Martinotti cellsterm·def (Gibson et al., 1999 ; Beierlein et al., 2003 ; Goldberg et al., 2004 ). Morphologically, Martinotti cellsterm·def in L5 are mostly similar to Martinotti cellsterm·def in L2/3, except for a tendency to send their axons either to L4 or deeper layers in addition to L1 (Wang et al., 2004 ).

The X94term line labels only non-Martinotti cells in layers 4 and 5a. Thus X94term cells are a completely distinct population from GINterm and X98term cells, whereas the GINterm and X98term populations partially overlap with each other. X94term cells account for about half of L4 SOMterm cells and have a basket-like morphologyterm with mostly local axonal projections (unlike the striking L1 projection seen in Martinotti cellsterm·def). Unlike Martinotti cellsterm·def, which target PNs, X94term cells target PV-positive FS interneuronsterm in layer 4. X94term cells fire narrow action potentials at high rates, and therefore resemble FS cells, but unlike FS cells they have a distinctive “stuttering” firing pattern in which their spike trains are interrupted by seemingly random periods of silence. These firing propertiesterm have therefore been called “FS-like” or “quasi-FS” (Ma et al., 2006 ; Large et al., 2016 ). In piriform cortexterm, SOM-positive cells can also exhibit tonic fast spiking firing propertiesterm and PV-positive neurons can be stuttering fast spiking, which indicates that different cells types can have similar firing propertiesterm and thus this criterion alone can not be used to reliably identify a specific cell type (Large et al., 2016 ). Interestingly, the SOMterm cells in layer 4 that are not labeled in the X94term line are similar to X94term cells in all of these respects: they have only local axons that target L4 PV-positive cells, and have the stuttering FS-like firing type. Thus it appears that L4 SOMterm cells may be a single population, with no apparent functional correlates of their segregation into X94term and non-X94 categoriesterm (Ma et al., 2006 ; McGarry et al., 2010 ; Xu et al., 2013 ). However, some Martinotti cellsterm·def are found sparsely in L4, at least in rat (Wang et al., 2004 ).

Another genetic tool increasingly used to investigate SOMterm neurons is the SOM-Creterm line (Taniguchi et al., 2011 ; Lovett-Barron et al., 2012 ), which allows researchers to use optogeneticterm or other Cre reporters to manipulate or record SOMterm cell activity across cortical layers. The SOM-Creterm lines target all SOMterm cells, and have recently been used to study a number of specific brain regions (Cottam et al., 2013 ; Polack et al., 2013 ; Chen I.-W. et al., 2015 ; Neske et al., 2015 ; Sturgill and Isaacson, 2015 ). It is important to note that at least one SOM-Creterm line (Taniguchi et al., 2011 ; Jax Stock No. 013044) also erroneously marks a small subset (6–10%) of fast spiking PVterm neurons (Hu et al., 2013 ), perhaps due to transient SOMterm expression during embryonic development in a small subpopulation of FS PVterm cells (although this has not yet been tested). It is not yet known whether this is also true for the other SOM-Creterm line (Lovett-Barron et al., 2012 ). It is also important to note that about 5% of GFP-expressing neurons in the X94term and X98term lines, and 3% of cells in the GINterm line, are not SOM-positive (Ma et al., 2006 ). This could be attributed to low expression of SOMterm in some cells which might be below the detectability threshold for immunohistochemistryterm, highlighting the point that immunohistochemistryterm results should be interpreted with caution. Although the percentages of this off-target labeling are low, they should still be taken into consideration in experiments. More generally, it is important to note that immunohistochemistryterm is a relatively difficult technique, and standards for antibody validation have been adopted relatively recently. This likely contributes to sometimes contradictory results for peptide expression in different studies.

A small subset (6–9%) of SOMterm cells that express NPY, nitric oxide synthase (NOS), and the Substance P receptor (SPR) form a distinct morphological classterm with long distance axonal projections. Although few in number, they can project to multiple brain regions both horizontally and vertically, making them good candidates for synchronizing neural activity across multiple cortical and subcortical regions (Tomioka et al., 2005 ; Kubota et al., 2011 , 2016 ; Caputi et al., 2013 ; Kubota, 2014 ; Endo et al., 2016 ). These cells have high spine density early in development, which is greatly reduced during maturation (Kubota et al., 2011 ). Additionally, since NOS-positive neurons are highly active during sleep, while most SOMterm neurons are not, SOMterm/NOS/SPR neurons are likely to form a distinct subclass with different morphologyterm and activity patterns (Kilduff et al., 2011 ; for reviewterm·mth see Tricoire et al., 2012 ).

Adding to this diversity are distinct laminar distributions of many of these cell types, as detailed below. In addition, SOMterm cells co-express a variety of other molecular markersterm such as CB (calbindin; Kubota and Kawaguchi, 1994 ; Wang et al., 2004 ; Ma et al., 2006 ; Suzuki and Bekkers, 2010 ), NPY (Kubota and Kawaguchi, 1994 ; Ma et al., 2006 ), CR (Xu et al., 2006 ; Xu and Callaway, 2009 ), CCK (Gonchar and Burkhalter, 1997 ), NOS (Kubota and Kawaguchi, 1994 ; Gonchar and Burkhalter, 1997 ; Kawaguchi and Kubota, 1997 ; Xu et al., 2006 ; Kilduff et al., 2011 ; Perrenoud et al., 2012 ), and SPR (Tomioka et al., 2005 ; Kubota et al., 2011 , 2016 ; Caputi et al., 2013 ; Figure 2 ). Because most studies have used only a subset of these categorization methods (indeed, some methods are mutually exclusive, such as the transgenic lines), it is still not clear how many distinct combinatorial types of SOMterm cells are found in the cerebral cortexterm. Nevertheless, it may be informative to attempt to estimate upper and lower bounds on the number of distinct SOMterm subtypesterm. At a minimum there are 4 distinct types: (1) L2/3 Martinotti cellsterm·def, which are mostly regular-spiking, labeled by GINterm, and target PNs, (2) L5 Martinotti cellsterm·def, which include LTSterm, are labeled by X98term, and target PNs, (3) L4 SOMterm cells, which are FS-like and target L4 PVterm+ cells, and some of which are labeled by X94term, and (4) NOS/SPR long-projecting GABAergic cells that can make either cortico-cortical or corticofugal projections (Table 1 ). This estimate undoubtedly lumps together subtypesterm of cells that can be distinguished by at least some criteriaterm, but represents a lower bound on the number of functionally distinct SOMterm subtypesterm.

Figure 2.

Venn diagram of all molecular markersterm that are colocalized with SOMterm (CB, CR, NPY, NOS, CCK) . Numbers indicate the range of reported percentages of SOM-positive cells that coexpress a given markerterm across studies in different species (mouse or rat), cortical layer (2–6), and cortical region (frontal, visual, or somatosensory). For more details, see Table 2 . Area of each circle approximately represents the average range. Overlap of circles indicates known coexpression of depicted markersterm, however the area of the overlap does not indicate the extent of coexpression. Although there are no reports of SOMterm cells coexpressing more than two other molecular markersterm, this possibility has not been ruled out since it was rarely tested for.

Table 1.

Four main subtypesterm of SOMterm cells (including 3 labeled by specific transgenic lines) with a minimal degree of overlap .

Line Morphologyterm Layer Firing propertiesterm Markerterm expression GINterm Martinottiterm L2/3, L5 RS 30% CB 27% NPY X94term Non-Martinotti L4, L5 FS-like None X98term Martinottiterm L5 LTSterm 96% CB 40% NPY None Non-Martinotti L2, L5+6 RS/BS 100% NPY+NOS+SPR Long-projecting LTSterm Subtypeterm differences are based on morphologyterm, cortical layer, firing propertiesterm, and co-expressed markersterm. Numbers in the Markerterm Expression column are the percentage of cells labeled in that subtypeterm that express a given markerterm .

What about the other extreme, an upper boundterm·stat estimate of the maximum possible number of SOMterm cell types? Estimating such an upper boundterm·stat requires several assumptions about whether classification criteriaterm·var can vary independently (producing all possible combinations) or whether there are correlations that reduce the number of combinations. These assumptions are poorly constrained by data because most studies have measured only a subset of these criteriaterm. In addition, we assume here that these criteriaterm are static, although it's likely that some criteriaterm could change over time. For example, cellular firing propertiesterm (such as burstingterm) can change under certain conditions in some cells (Bahrey et al., 2002 ), and molecular expression likely changes during development. Based on the ranges of classification criteriaterm·var (detailed in Table 2 ), we estimated an upper boundterm·stat of 100 on the number of possible subtypesterm of SOMterm cells. For example, in layer 2/3 there are two types of firing propertiesterm (RS and burstingterm) and 6 molecular markersterm (CB, CR, NPY, CCK, NOS, and SPR) in SOMterm cells. These belong mostly to one of two morphological classesterm, Martinotti cellsterm·def, and multipolar long-projecting neurons that express NOS/SPR. Cells are exclusively RS or burstingterm, but molecular markersterm can be expressed in various combinations (but are rarely tested for co-expression, providing little constraint on the number of possible combinations that could occur in SOMterm cells). Thus there are 2 morphological types, 2 6 potential combinations of molecular markersterm, and 2 firing-property classesterm, leading to 256 possible subtypesterm of SOMterm cells (2·2 6 ·2). Of these 256 possible combinations, burstingterm cells are known not to express NPY, and there is evidence against co-expression for 6 of the 15 possible binary combinations of markersterm (Table 3 ). This puts an upper boundterm·stat of 25 on the number of potential SOMterm subtypesterm in L2/3. Following similar logic for layers 2–6 (see Table 2 for details), we estimate the overall number of distinct subtypesterm of cortical SOMterm cells is likely no greater than 100. Additional data that further constrains the number of possible combinations would reduce this number. Thus while the true number of distinct subtypesterm of SOMterm cells is unknown, we estimate that it most likely lies somewhere between 4 and 100 subtypesterm.

Table 2.

Diversity of SOMterm cells across cortical layers 2–6 .

Layer Co-expression Morphologyterm Firing propertiesterm Targets Max Number 2/3 CB 6 (32–92% 7 ; 93% 9 ; 20% 11 ;86% 2 ; 85% 17 ) Martinottiterm 2 , 5 , 7 , 13 , 16 , 18 , 19 , 21 , 22 RS 5 , 7 L2/3 PNs 1 , 18 NPY 6 (15–27% 7 ; 18% 11 ) Non-Martinotti 20 , 21 , 22 , 23 Burstingterm 11 L2/3 INs 1 CR (57–96% 12 , 13 , 22 ) CCK (10% 2 , 11 , 22 ) NOS 8 , 20 , 24 (0.5% 2 ; 1–40% 12 ) SPR 23 , 20 (6.6%) 4 CB (31% 11 ) Non-Martinotti 7 , 18 , 19 LTSterm 14 , 15 L4 INs 18 NPY (5% 11 ) Martinottiterm 2 , 11 FS-like 7 L4 PNs 14 CR (10% 11 ; 18–44% 12 , 22 ) Burstingterm 11 CCK (8% 2 , 11 , 22 ) 5 CB 5 , 6 (49–96% 7 ;15% 11 ;92% 9 ;86% 2 ;92% 17 ) Martinottiterm 2 , 10 , 16 , 18 , 19 , 21 , 22 , 23 RS(<50%) 5 , 7 , 22 L5 PNs 10 NPY(1.4–41% 7 ; 40% 11 ) Non-Martinotti 7 , 21 LTSterm (<50%) 3 , 7 L2/3 PNs 4 CR(12–19% 12 , 22 ) FS-like 7 NOS 8 , 20 , 24 (0.5% 2 ; 1–40% 12 ) Burstingterm 5 , 11 , 22 SPR 23 (8.6%) 6 CB 6 (49–96% 7 ;92% 9 ;86% 2 ;92% 17 ) Non-Martinotti 9 , 21 , 23 RS 11 PNs 15 NPY 6 (1.4–41% 7 ; 40% 11 ) Martinottiterm 2 , 7 , 11 , 19 , 21 LTSterm 15 CR (15–17% 12 , 22 ) Burstingterm 11 CCK (40% 2 , 11 , 22 ) NOS 8 , 20 . 24 (0.5% 2 ; 1–40% 12 ) SPR 23 , 20 (8.4%) Total 100 No SOMterm cells are found in L1. The Co-expression column lists known molecular markersterm that co-localize with SOMterm in each layer, with the percentage of SOMterm cells that co-express each markerterm in parentheses. We included both burstingterm and LTSterm as distinct firing types, as reported in separate studies 3, 7, 11, 14, 15 , but considered these as a single category for the purposes of calculating an upper boundterm·stat on the number of distinct SOMterm cell types (Max Number column). We determined the upper boundterm·stat by counting the number of possible combinations of markersterm, morphologyterm, and firing type, given the constraints on markerterm co-expression detailed in Table 3 , known correspondences between morphologyterm and firing propertiesterm, and the observations that burstingterm SOMterm cells don't express NPY and that L4 non-Martinotti (X94term) cells don't express CB or NPY .

1. Cottam et al. ( 2013 ) mouse, visual; 2. Gonchar and Burkhalter ( 1997 ) rat, visual cortexterm; 3. Goldberg et al. ( 2004 ) mouse, visual and somatosensory cortexterm; 4. Kapfer et al. ( 2007 ) mouse, somatosensory cortexterm; 5. Kawaguchi and Kubota ( 1996 ) rat, frontal cortexterm; 6. Kawaguchi and Kubota ( 1997 ) rat, frontal cortexterm; 7. Ma et al. ( 2006 ) mouse (GINterm, X94term, X98term), somatosensory cortexterm; 8. Perrenoud et al. ( 2012 ) mouse, barrel cortexterm; 9. Rogers ( 1992 ) rat, visual cortexterm; 10. Silberberg and Markram ( 2007 ) rat, somatosensory cortexterm; 11. Wang et al. ( 2004 ) rat, somatosensory cortexterm; 12. Xu et al. ( 2006 ) mouse GINterm, frontal (high % NOS), somatosensory, and visual cortexterm; 13. Xu and Callaway ( 2009 ) mouse GINterm, somatosensory; 14. Beierlein et al. ( 2003 ) rat, barrel cortexterm; 15. Gibson et al. ( 1999 ) rat, somatosensory cortexterm; 16. Karube et al. ( 2004 ) rat, frontal cortexterm; 17. Kubota and Kawaguchi ( 1994 ) rat, frontal cortexterm; 18. Xu et al. ( 2013 ) mouse, somatosensory cortexterm; 19, McGarry et al. ( 2010 ) mouse, frontal, somatosensory, and visual cortexterm; 20. Endo et al. ( 2016 ) mouse, visual cortexterm; 21. Kawaguchi and Kubota ( 1998 ) rat, frontal cortexterm; 22. Uematsu et al. ( 2008 ) rat, frontal cortexterm; 23. Kubota et al. ( 2011 ) rat, frontal cortexterm; 24. Tomioka et al. ( 2005 ) mouse, motor, somatosensory, and visual cortexterm .

Table 3.

Co-expression of molecular markersterm in SOMterm cells .

NPY CR NOS CCK SPR CB + 7, 24 + 12, 13 + 2, 8, 24 + 11 + 23 NPY − 2 + 8, 24 − 11 + 23 − 2 − 2 − 23 NOS + 8 + 20, 23 CCK − 23 SPR Reported co-expression of binary combinations is indicated by +, reported absence of co-expression is indicated by −. Markersterm have been reported to co-express in at least some SOMterm cells for all binary combinations except NPY-CR, NPY-CCK, CR-NOS, CR-CCK, CR-SPR, and CCK-SPR. We pooled data from all layers and did not distinguish between NOS-1 and NOS-2. Numbers refer to references in Table 2 .

Another important means of categorizing SOMterm cells is by the layer in which their cell bodies reside. Here, we consider each of these layers in turn.

Layer 2/3 Multiple studies have shown that layer 2/3 SOMterm cells provide strong inhibitionterm to L2/3 PNs (Fino and Yuste, 2011 ; Adesnik et al., 2012 ; Cottam et al., 2013 ; Xu et al., 2013 ). However, there is disagreement over how much inhibitionterm L2/3 SOMterm cells provide to other inhibitory cells in L2/3. In somatosensory cortexterm, L2/3 SOMterm neurons provide only weak inhibitionterm to other inhibitory cells (Xu et al., 2013 ). A quite different pattern was observed in the visual cortexterm. There, SOMterm neurons strongly inhibit PVterm cells (twice as potently as PNs), and this inhibitionterm causes PVterm cells to be more tuned to orientation (Cottam et al., 2013 ; Pfeffer et al., 2013 ). The difference between these findings may indicate different circuits in somatosensory and visual cortexterm, but it is important to note that these studies also used different methods. Xu et al. performed slice experiments using current injections in single SOMterm neurons and optogeneticterm suppression of SOMterm neurons, whereas Cottam et al. looked at neural firing rate in vivo in behaving animals with and without presentation of visual stimuli while optogenetically activating many SOMterm neurons. Considering that brain states can strongly affect neural activity, it is quite possible that the reported differences in the inhibitory contribution of L2/3 SOMterm cells to neighboring neurons could be explained by different cortical states and the number of SOMterm cells recruited.

L2/3 SOMterm cells avoid inhibiting each other, and instead receive most of their inhibitionterm from VIP and PVterm cells (Pfeffer et al., 2013 ). Layer 2/3 SOMterm neurons have also been shown to participate in a form of lateral inhibitionterm in visual cortexterm, pooling excitatory input from adjacent PNs and thereby contributing to surround suppression (Adesnik et al., 2012 ). Indeed, stimulation of a single L2/3 PN in visual cortexterm can activate 30% of SOMterm neurons within a 100 micron radius (Kwan and Dan, 2013 ). This activation of SOMterm cells by PNs was more strongly distance-dependent than PN → PN activation (Kwan and Dan, 2013 ). Activity of the SOMterm network also increases supralinearly as the number of active L2/3 PNs increases. Compared to activating a single PN, activation of just two L2/3 PNs causes a 10-fold increase in the strength of recruited SOMterm inhibitionterm (Kapfer et al., 2007 ). Interestingly, in barrel cortexterm SOMterm cells are suppressed during stimulus presentation, and thus are anti-correlated with network activity, a feature which has not been observed in other sensory regions (Gentet et al., 2012 ). In contrast, strong activation of PNs is accompanied by a linear recruitment of PV-mediated inhibitionterm (Xue et al., 2014 ). Thus, it appears that the most important contribution of SOMterm inhibitionterm in L2/3 is likely to be driven by high-frequency activation of just a few surrounding PNs.

In visual and auditory cortexterm, L2/3 SOMterm cells respond much later than other neurons to input from L4, and have lower spontaneous firing rates than other inhibitory cells (Ma et al., 2010 ; Li et al., 2015 ). These late responses have been attributed to the fact that SOMterm cells in L2/3 do not receive input directly from L4, but rather pool from L2/3 PNs. In general, SOMterm cells respond with a delay, even if they are receiving input from neighboring L2/3 PNs (Kapfer et al., 2007 ; Kwan and Dan, 2013 ). This delay probably also arises in part from the integration of inputs from facilitating synapses.

L2/3 SOMterm neurons participate in both feedforward as well as feedback inhibitionterm, and primarily target the dendrites of L2/3 pyramidalterm neurons (Karnani et al., 2016a ). Paired in vitro recordings of SOMterm cells (in the GINterm line) revealed that both their subthreshold and suprathreshold activity is highly synchronous and they exhibit persistent firing more frequently than other cell types (Fanselow et al., 2008 ). During spontaneous activity in vivo , however, SOMterm cells do not correlate with oscillatory network activity, whereas PVterm cells do (Kwan and Dan, 2013 ). These results support the idea that PNs and PVterm cells receive similar input (dominated by thalamicterm activation), whereas SOMterm cells are modulated by distinct pathways (including top-down and subcortical input).

Layer 4 Most SOMterm cells in layer 4 are strikingly different from the typical Martinottiterm SOMterm cells in other layers. The X94term line sparsely labels L4 and L5a SOMterm neurons (Ma et al., 2006 ; Xu et al., 2013 ), and these neurons do not send their axons to layer 1, but instead target other inhibitory cells (i.e., PVterm cells) in layer 4. In fact, unitary IPSPs from SOMterm cells onto L4 PVterm cells are much larger than those in L4 excitatory cells (Xu et al., 2013 ). The morphologyterm of L4 SOMterm cells also differs from other SOMterm neurons. They are typically described as bitufted or multipolar cells that keep their axons and dendrites in the same layer (Gonchar and Burkhalter, 1997 ; Ma et al., 2006 ). These cells rarely co-express other markersterm and have been characterized as either quasi-FS (Ma et al., 2006 ) or burstingterm (Kubota and Kawaguchi, 1994 ; Wang et al., 2004 ). While thalamocortical axons in L4 provide strong and direct input to pyramidalterm and fast spiking cells, L4 SOMterm cells are only weakly excited by thalamicterm input (Cruikshank et al., 2010 ). Synaptically coupled SOM-FS pairs in L4 of barrel cortexterm can show synchronous spiking activity, even in the presence of glutamate antagonists (but not when GABA A receptors are blocked), indicating that inhibitory GABA A -mediated synaptic transmission is both necessary and sufficient to induce synchronous activity between SOMterm and FS cells (Hu et al., 2011 ). A small number of SOMterm cells with Martinottiterm morphologyterm are also found in L4, at least in juvenile rat (Wang et al., 2004 ). Due to differences in their targets, it appears that FS-like and RS SOMterm cells are members of different circuits in L4. In the frontal cortexterm, these cells also show different patterns of activity during a foraging task (Kvitsiani et al., 2013 ).

Layer 5 Layer 5 SOMterm cells represent 19% of all inhibitory cells in L5, some of which also co-express NPY and/or CB (Kawaguchi and Kubota, 1997 ; Ma et al., 2006 ) and which are labeled in the X98term mouse line. Layer 5 pyramidalterm neurons form disynaptic inhibitory circuits with one another via the L5 SOMterm network (Silberberg and Markram, 2007 ). Activation of a L5 PN typically produces inhibitionterm in neighboring PNs, and 40–90% of this inhibitionterm comes from a single L5 Martinotti cellterm·def (Silberberg and Markram, 2007 ). Layer 5 SOMterm cells also inhibit a subpopulation of L2/3 PNs, consistent with the translaminar projections that are the hallmark of Martinotti cellsterm·def (Kapfer et al., 2007 ; Xu and Callaway, 2009 ). 50% of SOMterm cells in L5 are low-threshold spiking cells. LTSterm SOMterm cells differ in their connectivity to one another compared with other SOMterm neurons. About 40% of LTSterm SOMterm cells make inhibitory connections with one another (Fino and Yuste, 2011 ). A small percentage of X94term line SOMterm neurons are also found in L5a (Ma et al., 2006 ). These neurons create a disinhibitory network by targeting PVterm cells and, at least in motor cortexterm, hyperactivity of SOMterm cells in L5 leads to excitotoxicity and death of excitatory neurons (Zhang et al., 2016 ).

Layer 6 SOMterm neurons in layer 6 consist mainly of Martinotti cellsterm·def that coexpress variable combinations of molecular markersterm such CB and NOS in rat (Kubota and Kawaguchi, 1994 ) as well as NPY and CCK in mouse (Wang et al., 2004 ). These cells send axons to layer 1, but about half of the SOMterm cells in L6 also make axonal arborizations in layers 5 and 6, suggesting less specific laminar targeting than layer 2/3 Martinotti cellsterm·def (Wang et al., 2004 ; Ma et al., 2006 ).

Several studies have now identified a small population of GABAergic projection neurons, i.e., GABAergic inhibitory neurons that are not interneuronsterm (McDonald and Burkhalter, 1993 ; Gonchar et al., 1995 ; Tomioka et al., 2005 ). These cells, while few in number (only 7–9% of SOMterm cells), project axons outside of the local area, can travel up to several mm, can cross areal boundaries, and in some cases project through the corpus callosum to the contralateral hemisphere (Gonchar et al., 1995 ). The vast majority of these cells express SOMterm, NPY, NOS, and substance P receptor (SPR), and are found in layer 6 (and to a lesser extent, in L2 and L5; Tomioka et al., 2005 ; Kubota et al., 2011 ; Caputi et al., 2013 ). Recent work has shown that these long inhibitory projections can regulate the output of medium spiny neurons in striatum, thereby modulating the activity of both direct (D1-type dopamine receptors) and indirect (D2-type dopamine receptors) reward pathways (Rock et al., 2016 ). NOS-positive projection neurons are a small subpopulation of neurons that is active during slow wave sleep, suggesting that they could play a major role in homeostatic sleep regulation by influencing global neuronal activity (Kilduff et al., 2011 ; for reviewterm·mth see Tamamaki and Tomioka, 2010 ).

Firing propertiesterm Cortical SOMterm cells differ in their firing propertiesterm and electrophysiology. In particular, several distinct categoriesterm have been reported, including regular spiking (RS) cells, LTSterm, burstingterm, and FS-like or stuttering cells (Kawaguchi and Kubota, 1996 ; Goldberg et al., 2004 ; Wang et al., 2004 ; Ma et al., 2006 ; Fanselow et al., 2008 ; Large et al., 2016 ). There is some disagreement about the prevalence of FS-like or stuttering SOMterm cells in different layers; one possible reason for this is that FS-like SOMterm cells might have been categorized as FS cells in some studies (Beierlein et al., 2003 ; Wang et al., 2004 ; Ma et al., 2006 ). SOMterm cells have also been classified as either accommodating or non-accommodating (Wang et al., 2004 ). Accommodating cells (the vast majority, at 90% of SOMterm cells) include both RS and burstingterm types, whereas non-accommodating cells (only 8% of SOMterm cells) have been described as analogous to FS cells based on their ability to fire at high rates without adaptation. Regular spiking and burstingterm firing types were originally described for cortical pyramidalterm neurons (Connors et al., 1982 ), and it is important to note that the firing propertiesterm of RS or burstingterm SOMterm cells are analogous but not identical to those classically observed in pyramidalterm neurons. For example, in vivo whole cell studies show that average action potential waveform of SOMterm neurons is somewhat narrower than pyramidalterm cells and wider than fast spiking cells recorded both in CRE-IRES-SOM and GINterm transgenic lines, thus putting them in a different category than RS neurons (Gentet et al., 2012 ; Polack et al., 2013 ). But in slice recordings from younger animals, the waveform of SOMterm neurons is comparable in width to excitatory neurons (McGarry et al., 2010 ). This discrepancy so far has not be investigated and could arise from either differences in recording methods or animal age.

It is still not clear whether LTSterm and burstingterm SOMterm cells form two distinct categoriesterm or instead are a single classterm that lie along a continuum. LTSterm cells are Martinotti cellsterm·def, and are found in cortical layers 4, 5, and 6 (Kawaguchi and Kubota, 1996 ; Beierlein et al., 2000 , 2003 ; Goldberg et al., 2004 ; Wang et al., 2004 ; Ma et al., 2006 ). They are so-named because they have very low thresholds for action potential initiation. Because of this low threshold, single-axon inputs to layer 5 LTSterm cells can generate spikes (Kozloski et al., 2001 ). In addition, LTSterm cells have a strong tendency to fire spikes or bursts on rebound from hyperpolarization. These rebound spikes/bursts are mediated by T-type calcium channels, similar to those found in thalamicterm relay neurons (Goldberg et al., 2004 ; Ma et al., 2006 ). Like relay neurons, L5 LTSterm cells can fire in either tonic or burstingterm mode, depending on their membrane potential. In contrast, L4 LTSterm cells fire tonically under control conditions, and only fire bursts in the presence of metabotropic glutamate agonists (Beierlein et al., 2000 , 2003 ; Goldberg et al., 2004 ). This suggests that LTSterm cells in layer 4 and 5 probably form two distinct classesterm.

Upon release from hyperpolarization, LTSterm cells can fire either a single spike or a burst of spikes, which in some studies has led to them being categorized as two distinct groups (Wang et al., 2004 ; Ma et al., 2006 ). However, it is possible that this difference is due only to variation in input resistance along a continuum (Ma et al., 2006 ), which would instead suggests that they form only a single group. Different studies have adopted different terminology (either burstingterm or LTSterm), and report somewhat different laminar distributions, which is further complicated by the fact that some studies are in rat while others are in mouse (Wang et al., 2004 ; Ma et al., 2006 ). Burstingterm SOMterm cells exhibit a prominent after-depolarization, which is almost certainly mediated by an I h current because these cells express HCN channel genes. Pharmacological blockade of the the I h current in GINterm and X94term cells eliminates rebound depolarization, indicating that I h likely contributes to burstingterm in those types of SOMterm cells. But blockade of the low-threshold T-type calcium channel (but not I h ) eliminates rebound burstingterm in only X98term cells, indicating that T-type channels are essential for this distinctive firing property of L5 LTSterm cells (Ma et al., 2006 ). Thus, based on the channels involved, LTSterm and burstingterm SOMterm cells described in different studies probably represent at least partially distinct populations. It seems likely that classification of SOMterm cells based on firing propertiesterm alone (such as the tendency to burst) could lump together distinct classesterm or erroneously split a single classterm, depending on the sample being studied.

LTSterm cells are notable because they have been shown to form gap-junction coupled networks in layer 4 of barrel cortexterm (Gibson et al., 1999 ; Beierlein et al., 2000 ; 2003 Amitai et al., 2002 ; Ma et al., 2011 ). Electrical coupling was also observed in L2/3 SOMterm neurons in the GINterm line, although these are usually not categorized as LTSterm cells (Fanselow et al., 2008 ). Electrical coupling of SOMterm cells likely has a substantial impact on the net effect of the SOMterm network, because SOMterm inhibitionterm increases supralinearly when more than one SOMterm neuron is activated (Kapfer et al., 2007 ). Adding to the importance of electrical coupling to SOMterm network activity is the fact that SOMterm cells don't receive thalamicterm input, and instead are mostly driven by intracortical input (Beierlein et al., 2003 ; Cruikshank et al., 2010 ). Although individual SOMterm neurons in L2/3 and L4 don't provide as strong or reliable inhibitionterm to PNs as do PVterm cells (Beierlein et al., 2003 ; Pfeffer et al., 2013 ), together as a unified network they may act as a powerful inhibitory force when activity in the local cortical excitatory population increases.

Synaptic physiology and input Unlike strongly depressing PN → FS and FS → PN synapses, L4 and L2/3 SOMterm cells typically receive strongly facilitating synaptic input from PNs and weakly facilitating synaptic input from FS and VIP cells (Thomson, 1997 ; Thomson and Deuchars, 1997 ; Markram et al., 1998 ; Reyes et al., 1998 ; Ma et al., 2012 ; Pi et al., 2013 ; Karnani et al., 2016b ). This suggests that SOMterm neurons are strongly but transiently inhibited at the onset of a new stimulus, but likely recover during prolonged activity. Facilitating input to SOMterm cells also means that they would be more sensitive to a sustained train of input from a single cell than to simultaneous but transient input from multiple PNs. Consistent with this, activation of SOMterm cells is quite different from other inhibitory cells depending on cortical network state. Synaptic input received by LTSterm SOMterm cells tends to be weaker and less reliable at low stimulation frequencies (<20 Hz) compared to input received by FS cells (Beierlein et al., 2003 ; Ma et al., 2012 ). At higher stimulation frequencies (>10–20 Hz), LTSterm cells are powerfully recruited at the same time that synapses onto other inhibitory cells become depressed. As a result, SOMterm cells are unlikely to be recruited during periods of low cortical activity, but become strongly activated during high cortical network activity. Congruent with these effects of network dynamics on SOMterm activation, SOMterm cells make weakly facilitating synapses onto both PNs and FS cells. Thus SOM-mediated inhibitionterm tends to be weak and unreliable at low frequencies, but will become robust at higher frequencies, making SOMterm cells an important player in shaping neural responses to prolonged stimuli (Gupta et al., 2000 ; Beierlein et al., 2003 ; Kapfer et al., 2007 ; Hayut et al., 2011 ; Ma et al., 2012 ; Pfeffer et al., 2013 ). Tonic firing of SOMterm neurons regulates spontaneous activity of principal cells via slow GABA B receptors (Urban-Ciecko et al., 2015 ), while also contributing fast GABA A -mediated synaptic input onto the distal dendrites of PNs (Wang et al., 2004 ; Silberberg and Markram, 2007 ). This indicates that SOMterm neurons contribute to spontaneous and evoked cortical responses.

Input to SOMterm cells is also unique since it does not appear to follow the canonical pattern of ascending thalamocortical information flow. Activation of thalamicterm fibers evokes a strong feedforward and depressing inhibitory current in L4 PNs, suggesting a recruitment of FS neurons by the thalamus. Intracortical stimulation, on the other hand, recruits disynaptic SOM-mediated inhibitionterm (Beierlein et al., 2003 ). Additionally, stimulation of thalamicterm projections evokes robust EPSPs in FS INs and RS PNs but only weak excitatory current in SOMterm neurons in L4 and L5/6 (Beierlein et al., 2003 ; Cruikshank et al., 2010 ). Whatever weak thalamicterm input makes it to SOMterm neurons is still mediated by depressing rather than facilitating synapses, making thalamicterm drive to SOMterm neurons even less impactful (Cruikshank et al., 2010 ). Within cortexterm, stimulation of L4 neurons evokes much stronger EPSPs in L2/3 PVterm cells and PNs than in SOMterm cells, whereas stimulation of surrounding L2/3 PNs leads to a strong recruitment of L2/3 SOMterm neurons (Adesnik et al., 2012 ). These results from visual cortexterm indicate that SOMterm cells in L2/3 are modulated more by within-layer input than translaminar input. SOMterm neurons in L2/3 of auditory cortexterm tend to have delayed EPSPs during presentation of sound, also suggesting that they are not directly connected to thalamorecepient neurons in L4, but rather pool input from neighboring L2/3 PNs (Li et al., 2015 ). Weak but facilitating input would take longer to integrate and produce a measurable EPSP, which could explain why excitatory currents in SOMterm cells are delayed. Although more systematic recordings from different layers and cortical regions are necessary to clarify the nature of the input that drives SOMterm neurons, it is clear that SOMterm neurons do not respond with the same dynamics as other neurons.

In visual cortexterm, different subtypesterm of interneuronsterm are frequently coactive with other neurons within their classterm. Thus SOMterm neurons are more likely to be recruited when other SOMterm neurons are firing. An interesting pattern emerges when co-inhibition between different inhibitory subclasses and their excitatory input are examined. Subclasses that exhibit strong co-inhibition (such as VIP-SOM) tend to receive non-overlapping excitatory input, whereas those with weak co-inhibition (such VIP-PV) have highly correlated membrane potentials (Karnani et al., 2016b ).

What does the somatostatinterm neuropeptide do?

Somatostatinterm is not just a cell-type specific markerterm, but also an inhibitory 14-amino-acid neuropeptide released by the subset of GABAergic neurons that express the somatostatinterm gene. Somatostatinterm activates 5 distinct G-protein coupled receptors (Hoyer et al., 1995 ). The cellular and synaptic effects of somatostatinterm are fairly well-understood, but less is known about the network, cognitive, and behavioral effects (for reviewterm·mth see Baraban and Tallent, 2004 ; Liguz-lecznar et al., 2016 ). Unlike GABA, which is released from conventional synaptic vesicles at axonal boutons, somatostatinterm is released from dense-core vesicles from both axons, and dendrites (Ludwig and Pittman, 2003 ). Neuropeptide release requires repetitive high-frequency firing (Kits and Mansvelder, 2000 ), suggesting that GABA and SOMterm are likely to be released under different conditions. Indeed, in the hippocampus, GABA and SOMterm are differentially released during the different oscillatory activities accompanying sleep and movement, suggesting the possibility of distinct functional roles (Katona et al., 2014 ). The functional interactions of SOMterm and GABA can be complex. In hippocampus, both SOMterm and GABA produce postsynaptic hyperpolarization, with SOMterm augmenting multiple K+ currents and reducing voltage-gated Ca++ currents (Pittman and Siggins, 1981 ; Moore et al., 1988 ; Ishibashi and Akaike, 1995 ; Viana and Hille, 1996 ; Schweitzer et al., 1998 ). SOMterm also acts via presynaptic receptors to inhibit glutamate release by excitatory neurons (Boehm and Betz, 1997 ; Tallent and Siggins, 1997 ; Dutar et al., 2002 ; Sun et al., 2002 ). Repeated release of SOMterm also reduces the density of dendritic spines and excitatory synapses, which depends on activation of SOMterm receptor subtypeterm 4 (Hou and Yu, 2013 ). All of these actions would be expected to work in concert with GABA release to reduce the firing probability of downstream neurons. However, there is some evidence that SOMterm also decreases GABA-mediated IPSPs and can lead to depolarization, which would counteract the inhibitory effects of GABA (Dodd and Kelly, 1978 ; Scharfman and Schwartzkroin, 1989 ; Greene and Mason, 1996 ; Leresche et al., 2000 ). The cellular and synaptic effects of SOMterm release in cortexterm have not been studied.

The effects of the somatostatinterm neuropeptide on network activity and cognition have been studied by intracerebral injections of agonists and antagonists, and also with SOMterm knockout mice. SOMterm appears to have an antiepileptic effect, reducing epileptiform activity and seizures in a number of different epilepsy models (Sun et al., 2002 ; Halabisky et al., 2010 ). This makes sense because of the inhibitory effects of SOMterm and the fact that it is released only under conditions of sustained high-frequency firing. Consistent with this, SOMterm knockout mice show increased severity of kainate-induced and sensory-triggered seizures (Buckmaster et al., 2002 ; Tomioka et al., 2014 ). Together these findings suggests that one function of SOMterm may be to act as a protective neuropeptide system to prevent runaway activity. SOMterm also appears to play a role in learning and memory. SOMterm receptor knockout mice show motor (Zeyda et al., 2001 ) and spatial learning deficits (Guillou et al., 1993 ; Dutar et al., 2002 ; Tuboly and Vécsei, 2013 ), whereas hippocampal or ventricular injections of SOMterm facilitate spatial learning in a variety of tasks (Vécsei et al., 1984 ; Vécsei and Widerlov, 1988 ; Lamirault et al., 2001 ). Cysteamine, a SOM-depleting agent, leads to deficits in memory retention in rats (Vécsei et al., 1984 ; Vécsei and Widerlov, 1988 ; Fitzgerald and Dokla, 1989 ; Nakagawasai et al., 2003 ). Cortical expression levels of SOMterm are reduced in aging in both rats and humans, and this is correlated with learning deficits in rats (Dournaud et al., 1995 ). SOMterm levels are also reduced in the brains of Alzheimer's disease patients examined postmortem, specifically in cortical layers 3 and 5 (Davies et al., 1980 ; Vécsei and Klivenyi, 1995 ). However, it's worth noting that many of these studies are relatively old and have not been replicated recently, despite broad interest in the topic. Based on the available evidence, it thus appears that SOMterm has a potential neuroprotective role in preventing epileptic activity, and also appears to be involved in both learning and memory retention.

Functional and computational roles of SOMterm cells Receptive field properties In general, the tuningterm of inhibitory neurons is very similar among different subtypesterm, and typically a bit more broad compared to excitatory cells. In L2/3, at least, SOMterm neurons appear to provide inhibitionterm to nearly all PNs in the local neighborhood, and likewise appear to pool input indiscriminately from the local population (Fino and Yuste, 2011 ). SOMterm cells are thus very likely to have the same tuningterm as the net tuningterm of the local population, such that their effect on the receptive fields of PNs may not be obvious just from measuring their tuningterm (Li et al., 2015 ; Karnani et al., 2016a ). In visual cortexterm, there is evidence that SOMterm cells are more orientation-tuned than PVterm neurons (Ma et al., 2010 ). Their frequency tuningterm in auditory cortexterm is also sharper than that of PVterm cells but does not differ from the tuningterm of excitatory cells (Li et al., 2015 ). These differences in tuningterm at the population level are quite small, however, and are much smaller in magnitude than the variability in tuningterm across individual cells. Together, these results are largely consistent with the idea that SOMterm cells, like PVterm cells, pool input from the local population. This “local pooling hypothesisterm·def” explains why inhibitory interneuronsterm tend to reflect the general tuningterm of the area and the neurons they pool from (Kerlin et al., 2010 ; Fino and Yuste, 2011 ; Moore and WehrpersName, 2013 ). This idea also explains the minor differences in tuningterm of IN subtypesterm, since there are some differences in their input and the size of the area they pool from. For example, PVterm neurons in L4 receive thalamicterm input as well as excitatory input from surrounding PNs within about 100 μm, and as a result are highly correlated with local network activity (Beierlein et al., 2003 ; Scholl et al., 2015 ). PVterm neurons provide equal inhibitionterm to PNs located both near and far from them (within a range of about 400 μm). VIP cells, in contrast, provide very local columnar inhibitionterm to SOMterm cells within about 120 μm (Zhang et al., 2014 ; Karnani et al., 2016a ). SOMterm neurons receive little or no thalamicterm input, and pool mostly from PNs in the same layer within about 550 μm (Fino and Yuste, 2011 ; Zhang et al., 2014 ). The peak of their inhibitory contribution is seen in PNs located approximately 200 μm away, making them ideal inhibitors of competing neural activity (Fino and Yuste, 2011 ; Zhang et al., 2014 ). SOMterm neurons often form disynaptic inhibitory circuits between PNs, thus contributing strongly to lateral inhibitionterm (Silberberg and Markram, 2007 ; Fino and Yuste, 2011 ; Adesnik et al., 2012 ; Zhang et al., 2014 ). For example, SOM-mediated lateral inhibitionterm contributes to surround suppression in L2/3 visual cortical neurons, conferring tuningterm for stimulus size (Adesnik et al., 2012 ). The fact that SOMterm cells form a disinhibitory network with VIP neurons, which are activated by locomotion, predicts that surround suppression should be modulated by locomotion (Ayaz et al., 2013 ; Fu et al., 2014 ). Indeed, locomotion alters spatial integration in V1 in mice, leading to a decrease in surround suppression (Ayaz et al., 2013 ). This suggests that effects of SOMterm cells on the receptive fields of neighboring PNs could depend in complex ways on behavioral state or task context, and may not be revealed by studies in anesthetized animals. For example, SOMterm neurons in visual cortexterm have been reported to be much less visually responsive under anesthesia (Adesnik et al., 2012 ). In addition, since SOMterm cells appear to integrate input from within their own layer, and have unique interlaminar connectivity, it is possible that specific contributions of SOMterm inhibitionterm to PN receptive field properties might only be seen under conditions in which cortical layers are differentially activated, as is seen for example during habituationterm in auditory cortexterm (Kato et al., 2016 ).

Divisiveterm and subtractiveterm inhibitionterm A number of studies have examined the functional role of SOMterm neurons and their effect on both PNs as well as interneuronsterm using optogeneticterm activation or suppression of SOMterm or PVterm cells (Atallah et al., 2012 ; Duguid et al., 2012 ; Lee et al., 2012 ; Wilson et al., 2012 ; Cottam et al., 2013 ; Seybold et al., 2015 ). A common theme is that SOMterm cells provide gain controlterm for cortical circuitry, but perhaps not surprisingly, a diversity of results have been reported. Gain controlterm in this context refers to the general enhancement or suppression of PN responses, independent of specific transformations of stimulus selectivity. For example, moderate suppression or activation of PVterm neurons in layer 2/3 causes a multiplicative scaling up or down of PN responses in visual cortexterm, without altering their orientation tuningterm (Atallah et al., 2012 ; Wilson et al., 2012 ). Such scaling is referred to as divisiveterm gain controlterm, which is often contrasted with subtractive gain controlterm·def. In subtractive gain controlterm·def, all activity (both spontaneous and evoked) is increased or decreased by a constant amount. In contrast to the findings of Atallah and of Wilson, stronger activation of visual PVterm neurons appears to produce a subtractiveterm instead of divisiveterm effect on PNs, which narrows their orientation tuningterm instead of leaving it unaffected (Lee et al., 2012 ). These conflicting results can best be understood in the context of the “iceberg effectterm,” which refers to the idea that firing rates cannot go below zero (i.e., the water level), which hides the subthreshold tuningterm curve. Stronger suppression of PN spiking can narrow tuningterm curves, even with purely divisiveterm inhibitionterm, because of this effect of spike threshold (El-Boustani and Sur, 2014 ; Lee et al., 2014 ; Xue et al., 2014 ). These studies illustrate that inferring the presence of divisiveterm or subtractiveterm inhibitionterm from optogeneticterm manipulations can be problematic (Kumar et al., 2013 ; Seybold et al., 2015 ).

Because PVterm cells provide fast and powerful proximal inhibitionterm, they appear to be ideally positioned to provide divisiveterm inhibitionterm, whereas the dendritic targeting by Martinotti cellsterm·def seems better suited to providing subtractiveterm inhibitionterm (Kubota et al., 2015 , 2016 ). Consistent with this, SOMterm cells provide subtractiveterm inhibitionterm to PNs in olfactory cortexterm, and moreover provide divisiveterm inhibitionterm to PVterm cells there (Sturgill and Isaacson, 2015 ). In visual cortexterm, conflicting results have been reported for activation of SOMterm cells. One study found that SOMterm activation sharpened PN orientation tuningterm, consistent with subtractiveterm inhibitionterm (Wilson et al., 2012 ), whereas a similar study reported that SOMterm activation reduced PN spiking without any effect on tuningterm, consistent with divisiveterm inhibitionterm (Lee et al., 2012 ). A key to reconciling these disparate results may lie in the relative timing, size, and durations of sensory and optogeneticterm stimulation. In particular, inhibitionterm may be more likely to be divisiveterm when it is co-active with strong PN activity, and more likely to be subtractiveterm when PNs and INs are not co-active. Importantly, one of these studies used brief activation (Wilson et al., 2012 ), whereas the other used prolonged activation (Lee et al., 2012 ). Brief SOMterm activation (which would result in less co-activation) produced a subtractiveterm effect, whereas prolonged SOMterm activation (with more co-activation) produced a divisiveterm effect (Lee and Dan, 2012 ; El-Boustani and Sur, 2014 ). Due to their facilitating input and the fact that they pool broadly from PNs, SOMterm cells are likely to be only weakly activated by brief or small visual stimuli, whereas prolonged and large visual stimuli are likely to strongly activate the SOMterm network. Indeed, SOMterm cells were found to provide late, subtractiveterm inhibitionterm to PNs for small visual stimuli, but switched to fast, divisiveterm inhibitionterm for large visual stimuli (El-Boustani and Sur, 2014 ). This illustrates that whether gain controlterm is divisiveterm or subtractiveterm is dynamic and stimulus-dependent, rather than a fixed property of a given cell type. In addition, SOMterm activation suppresses PVterm cells up to twice as powerfully as PNs, which will have its own effects on network activity and PN tuningterm (Cottam et al., 2013 ). A diverse mixture of subtractiveterm and divisiveterm effects during activation of SOMterm or PVterm neurons is seen in auditory cortexterm as well, even in simultaneously recorded neurons within the same column (Seybold et al., 2015 ). This degree of variability in the effects of SOMterm network perturbations makes sense given the broadly interconnected and recurrent nature of cortical networks. Indeed, modeling of these networks has shown that inhibitory gain controlterm can shift from being divisiveterm to subtractiveterm depending on the spike threshold of PNs and the strength of the optogeneticterm suppression. This suggests that these forms of gain controlterm may not be determined by the type of interneuronterm, but rather by intrinsic properties of a target neuron (Seybold et al., 2015 ). The picture that emerges from these studies is that whether gain controlterm is divisiveterm or subtractiveterm is a flexible and dynamic feature of inhibitory circuits.

Surprisingly, functional properties of SOMterm neurons in barrel cortexterm appear strikingly different from those in auditory and visual cortexterm. Unlike SOMterm cells in V1, which have unremarkable responses to visual stimuli, SOMterm cells in L2/3 of S1 are tonically active in the absence of whisker stimulation but become hyperpolarized and cease firing in response to either active or passive whisker stimulation (Gentet et al., 2012 ). Optogeneticterm activation of SOMterm cells during stimulus presentation might therefore produce unnatural results that would be markedly different from what is seen in the non-perturbed circuit. Optogeneticterm suppression of SOMterm cells in S1, however, is easier to interpret, and leads to increased burst firing in nearby PNs (Gentet et al., 2012 ). Tonically active SOMterm cells likely provide tonic inhibitionterm to cortical neurons, especially to apical dendrites (the distinctive target of Martinotti cellsterm·def). Tonic inhibitionterm remains poorly understood in cerebral cortexterm, but has been shown to improve the signal-to-noise ratio in cerebellum, allowing reliable transmission of sensory information (Duguid et al., 2012 ). Reducing tonic inhibitionterm in cerebellum mainly results in increased spontaneous activity, with little effect on evoked responses, consistent with results seen in barrel cortexterm (Gentet et al., 2012 ). In hippocampus, silencing SOMterm (but not PVterm) neurons increased the probability of burst spiking in PNs (Royer et al., 2012 ), similar to the effect seen in barrel cortexterm. Burst spiking in cortical regions has been hypothesized to carry more information than single spikes (Livingstone et al., 1996 ; Lisman, 1997 ) and is associated with improved stimulus detection in the visual system (Guido et al., 1995 ; Mukherjee and Kaplan, 1995 ). This suggests a mechanism by which the hyperpolarization of SOMterm cells by whisker stimulation might enhance sensory information processing.

In cortexterm, excitation is typically balanced by inhibitionterm that is proportionally scaled depending on the strength of excitatory input (Anderson et al., 2000 ; WehrpersName and Zador, 2003 ; Wilent and Contreras, 2005 ; Okun and Lampl, 2008 ). In order for inhibitionterm to scale proportionately with excitation, both sources must receive at least some overlapping sensory input. Since PVterm neurons, but not SOMterm cells, receive thalamicterm input, they appear to be best positioned to provide balanced inhibitionterm via feedforward circuitry, as has been shown in the visual cortexterm (Xue et al., 2014 ). SOMterm cells, on the other hand, seem more likely to provide modulatory inhibitory input that is pooled from the activity of surrounding PN population, hence providing feedback inhibitionterm (Murayama et al., 2009 ). Both sources likely contribute to balanced inhibitionterm in PNs.

Gain controlterm by locomotion Behavioral states can profoundly change how sensory neurons respond to a stimulus (Niell and Stryker, 2010 ; Kato et al., 2016 ). A powerful new model for studying these effects has been to study the effects of locomotion on sensory processing, typically by recording stimulus-evoked responses in mice that are free to run on a ball or wheel. The effects of locomotion are strikingly different across sensory systems. In the visual cortexterm, for example, running enhances neural responses without changing their orientation tuningterm (Niell and Stryker, 2010 ). The opposite effects are seen in the auditory cortexterm, where projections from secondary motor cortexterm suppress sensory responses during locomotion (Schneider et al., 2014 ). Similarly, both the neural circuitry and the neuromodulatory systems underlying locomotion effects also appear to differ across sensory regions. In visual cortexterm, running depolarizes both PNs and inhibitory cells. The resulting increase in both excitation and inhibitionterm in PNs reduces membrane potential variance, and leads to more stimulus-evoked spikes without any increase in spontaneous activity (Polack et al., 2013 ). Whereas cholinergic input affects membrane potential fluctuations during quiescent periods, the effect of locomotion on membrane potential variance is mostly dependent on noradrenergic input. Interestingly, SOMterm neurons do not show decreased membrane potential variability during running, suggesting a differential influence of norepinephrine on SOMterm neurons and PNs (Polack et al., 2013 ). Different classesterm of inhibitory neurons show marked differences in how they are modulated by locomotion in the visual cortexterm. VIP neurons are depolarized throughout the entire running period, while PVterm cells only respond transiently at the beginning. SOMterm neurons are typically suppressed during running, and fire mostly at the end of the running period (Fu et al., 2014 ). These results suggest that the effect of locomotion is mediated by a disinhibitory circuit, in which VIP cells inhibit SOMterm cells and thereby increase the activity of neighboring PNs. VIP cells are known to be activated by basal forebrain stimulation, via nicotinic acetylcholine receptors (nAChRs). The basal forebrain projects extensively to V1, and nAChR antagonists strongly reduce the locomotion-induced depolarization of VIP cells. These results suggest that cholinergic projections are a key element of the circuitry underlying the locomotion effect in V1, but because nAChR blockade did not completely abolish this effect, there must be additional pathways involved (Fu et al., 2014 ). Multiple interacting modulatory pathways could explain the apparent contradiction between the results of Polack and of Fu about the relative importance of norepinephrine and acetylcholine for the locomotion effect in V1. Acetylcholine and norepinephrine modulation can interact in complex ways; for example, it is possible that acetylcholine predominantly affects the gain of evoked responses, while norepinephrine produces shifts in baseline activity, as seen in barrel cortexterm (Constantinople and Bruno, 2011 ). Interestingly, ACh also affects SOMterm neurons, but acts through muscarinic receptors (Kawaguchi et al., 1997 ; Fanselow et al., 2008 ; Xu et al., 2013 ), suggesting that the influence of ACh on activity of cortical neurons may be complex and depend on the type of activated receptor and neuronal subtypeterm.

One exciting but still speculative possibility is that this disinhibitory circuit operates in much the same fashion to increase gain during selective attention or similar top-down enhancement. For example, VIP cells have been proposed to mediate attentional enhancement by opening local holes in the blanket of inhibitionterm (Fomby et al., 2014 ; Zhang et al., 2014 ; Karnani et al., 2016a ). One difference between this idea and the disinhibitory effects of locomotion is that all VIP cells in V1 are activated by running, which is consistent with the observation that the effects of locomotion are distributed broadly across all of visual cortexterm. This contrasts with the idea of very local disinhibition achieved by activating one or a few VIP cells. It is possible that the same underlying circuitry could operate in these two distinct modes in different functional contexts.

A similar (but not identical) disinhibitory circuit modulates activity in barrel cortexterm. Unlike V1, somatosensory cortexterm receives strong M1 input, particularly to VIP cells that express 5HT3aR serotonin receptors (Lee et al., 2013 ; Fu et al., 2014 ). While PNs and other types of INs also receive weak input from M1, VIP neurons in all layers are strongly recruited by M1 activation. Whisking reliably activates VIP neurons that in turn suppress SOMterm activity (Lee et al., 2013 ). This disinhibitory circuit therefore explains how SOMterm cells cease their tonic firing and become hyperpolarized during whisking (Gentet et al., 2012 ). This suggests that the same disinhibitory circuit motif that underlies the locomotion effect in V1 also modulates S1 sensory responses during active whisking, although the source of VIP activation is different in the two systems. Considering that 5HT3aR-expressing VIP cells are strongly depolarized by serotonin and acetylcholine as well, an intriguing possibility is that other behavioral states and contexts could additionally influence sensory responses during whisking (Moreau et al., 2010 ; Rudy et al., 2011 ; Lee et al., 2013 ).

Although auditory neurons also exhibit depolarization and decreased membrane potential variability during running, effects of locomotion on auditory cortexterm are distinct, since locomotion mostly suppresses sound-evoked responses instead of enhancing them (Schneider et al., 2014 ). These changes also tend to precede movement, indicating that modulation comes from a motor planning region rather than from muscle feedback. Interestingly, in auditory cortexterm, M2 projections inhibit PN responses via the PVterm network, bypassing the VIP → SOMterm disinhibitory circuit (Schneider et al., 2014 ). While running desynchronizes auditory cortexterm and depolarizes PNs, optimal performance on an auditory task is associated with an intermediate state of arousal and hyperpolarized membrane potentials in PNs, in an attentive but quiescent behavioral state (McGinley et al., 2015 ). Thus the state of arousal falls along a continuum, and different points of this spectrum are likely mediated by different modulatory systems, not all of which involve SOMterm inhibitory networks. Taken together, these studies demonstrate that each sensory system integrates information about movements through different local and global circuits.

Salience and behavioral relevance A number of recent studies have looked at the responses of SOMterm neurons during more complex forms of contextual stimulus presentation. In auditory cortexterm, SOMterm inhibitionterm contributes to stimulus-specific adaptationterm and habituationterm. In both of these phenomena, SOMterm neurons appear to be sensitive to the statistics of stimulus presentation, and suppress the responses to frequently presented tones. However, the time scales and contextual structure of the two paradigms suggests that they engage distinct processes. Stimulus-specific adaptationterm describes how auditory neurons respond in an “oddball” paradigm, in which a frequent stimulus is interleaved with a rare stimulus. Responses to the frequent stimulus are suppressed, but responses to the rare stimulus are not. Stimulus-specific adaptationterm is seen across all cortical layers, and in all cell types, including SOMterm neurons. The phenomenon can be seen in anesthetized animals, and develops within a few presentations of brief tone stimuli (Szymanski et al., 2009 ; Chen I.-W. et al., 2015 ; Natan et al., 2015 ). Suppression of SOMterm neurons reduces stimulus-specific adaptationterm in other cortical neurons, increasing PN responses to the frequent tone. Thus SOMterm cells contribute to stimulus-specific adaptationterm, even while experiencing stimulus-specific adaptationterm themselves (Natan et al., 2015 ). SOMterm neurons also contribute to a form of habituationterm to tones that develops over several days. In this paradigm, daily exposure to repeatedly presented long tones (9 s duration) gradually reduces tone-evoked responses in L2/3 PNs. This habituationterm can be partially relieved if mice are engaged in a sound detection task (Kato et al., 2016 ). Tone-evoked responses in L2/3 SOMterm cells are increased as both PN and PVterm cell responses are decreased, suggested that SOMterm cells mediate the habituationterm in other cortical neurons. This increase in SOMterm responsiveness contrasts with the decreased SOMterm responsiveness during stimulus-specific adaptationterm. However, what causes the increase in SOMterm inhibitionterm is still unclear. One possibility is that frequently-repeated stimuli might increase SOMterm activation via facilitating synapses. This mechanism could explain short-term stimulus-specific adaptationterm in cortical neurons, but cannot explain habituationterm effects that persist across days. This suggests that there could be long-term cellular or synaptic changes that lead to a “memory” of a frequent tone. Another possibility is that SOMterm cells might receive specific input that is not adapted during habituationterm paradigms. Interestingly, however, thalamorecipient L4 neurons did not show the habituationterm seen in L2/3 PNs, which makes them unlikely candidates for enhanced SOMterm activation (Kato et al., 2016 ). This is consistent with the fact that L2/3 SOMterm cells get little or no input from L4 (Li et al., 2015 ), and also suggests that this form of habituationterm is not inherited from subcortical auditory structures. Yet another possibility is that enhancement of SOMterm cell responses during habituationterm may be generated by top-down input. For example, SOMterm cells in visual cortexterm receive weak but measurable input from cingulate cortexterm (Zhang et al., 2014 ). Modulation of VIP neurons, as occurs in different behavioral states (Pi et al., 2013 ; Fu et al., 2014 ; Karnani et al., 2016a ), is also a plausible candidate for the habituationterm signal.

Inhibitionterm of SOMterm cells via VIP neurons, as seen during locomotion, also seems to play a central role in modulating SOMterm activity during task performance and behavioral relevance (Pi et al., 2013 ; Zhang et al., 2014 ; Karnani et al., 2016a ). Although locomotion changes activity broadly across visual cortexterm, disinhibitory effects of VIP neurons might also be local to a region of specific tuningterm (Karnani et al., 2016a ). A highly localized disinhibitory network could provide a mechanism for selectively enhancing visual processing in a small part of the visual field, without affecting inhibitionterm in other regions. A similar circuit motif can also enhance the signal-to-noise ratio in cortical neurons during task performance. Indeed, V1 receives strong localized input from the cingulate cortexterm that enhances VIP activity (Zhang et al., 2014 ). A top-down control signal from an executive region would be an ideal candidate to selectively modulate visual responses. In auditory cortexterm, neurons that are tuned to target frequency display enhanced selectivity during performance of a tone-in-noise detection task, while neurons that are tuned to other frequencies suppress their responses (Atiani et al., 2009 ). All cortical neurons showed a dramatic gain reduction during the task, which, in combination with sharpened receptive fields for target frequency, leads to a dramatic reduction in noise and better task performance (Atiani et al., 2009 ; Otazu et al., 2009 ; Sadagopan and Wang, 2010 ). Although it is unclear whether the VIP → SOMterm network is responsible for this short-term receptive field plasticity, reinforcement signals are known to activate VIP-positive neurons in auditory cortexterm, which in turn results in strong suppression of SOMterm neurons and of a small fraction of PVterm cells (Pi et al., 2013 ).

The modulatory signals that underlie differential recruitment of various cell types during task performance are still unclear. In visual cortexterm, VIP neurons receive modulatory input from cingulate cortexterm, basal ganglia, and to a weaker extent M1 (Lee et al., 2013 ; Fu et al., 2014 ; Zhang et al., 2014 ). Slice experiments show that ACh increases input resistance in regular and burst spiking VIP and SOMterm neurons, resulting in increased firing, whereas fast and late spiking cells remain unaffected (Kawaguchi et al., 1997 ). Moreover, norepinephrine increases spike probability in regular spiking and burstingterm SOMterm neurons in rat frontal cortexterm (Kawaguchi and Shindou, 1998 ). Additionally, VIP cells belong to a subgroup of neurons that express 5HT3a receptors, thus making them a primary target of serotonergic projections (Rudy et al., 2011 ). Adding to this complexity, these modulatory inputs also have different effects on cortical activity. ACh and serotonin desynchronize cortical network activity, whereas norepinephrine synchronizes cortical activity (for reviewterm·mth see Lee and Dan, 2012 ). Cortical neurons can also be desynchronized by tonic glutamatergic input from the thalamus (Hirata and Castro-Alamancos, 2010 ). Interestingly, ACh can synchronize coupled LTSterm cells, although it is important to remember that LTSterm cells are only a subset of SOMterm neurons (Beierlein et al., 2000 ). Awake behaving mice typically show desynchronized cortical activity, whereas quiet wakefulness, sleep, and anesthesia are associated with synchronized activity featuring up and down states. Evidence from barrel cortexterm suggest that non-fast-spiking inhibitory cells tend to correlate with the membrane potential of excitatory cells during quiet wakefulness states, but show dramatic increases in depolarization and firing during active whisking, thus suppressing cortical responses (Gentet et al., 2010 ). By receiving either direct or disynaptic input from other brain regions, SOM-mediated inhibitionterm could play a role in selective attention of sensory processing during task engagement and dictate changes in cortical network activity during specific behavioral states. Norepinephrine and acetylcholine could have opposing effects because the former activates SOMterm cells, whereas the latter inhibits them through the VIP → SOMterm circuit. In summary, an intriguing but still speculative possibility is that different neuromodulators might act via SOMterm networks to promote synchronized or desynchronized network states, in much the same way as they modulate gain during locomotion.

Learning SOMterm neurons also appear to play an important role in memory formation. Classical fear conditioning of a whisker stimulus increases the number of inhibitory synapses in L4 of the corresponding barrel in S1 (Siucinska, 2006 ; Jasinska et al., 2010 ). Recently, upregulation of GABA was shown to be accompanied by an increase in the number of SOM-expressing neurons in L4 of barrel cortexterm after conditioning (Gierdalski et al., 2001 ; Cybulska-Klosowicz et al., 2013 ). This presumably results from an increase in SOMterm expression by SOMterm cells, bringing them above immunostaining detection threshold, rather than an increase in the number of SOMterm cells per se . Thus SOMterm cells express more GABA and somatostatinterm after associative learning, suggesting that increased SOM-mediated inhibitionterm may play a role in circuit plasticity during learning. In cortical slices from naive animals, SOMterm neurons exhibit lower levels of GABA expression compared to other classesterm of inhibitory cells (Gonchar and Burkhalter, 1997 ). Upregulation of GABA after associative learning may allow the SOMterm inhibitory network to shape cortical responses that represent a newly behaviorally-relevant stimulus after learning. A quite different pattern of changes has been demonstrated in motor cortexterm, where motor learning induces reorganization of dendritic spines on the apical tufts of L2/3 PNs. This reorganization coincides with a decrease in axonal boutons of SOMterm cells in layer 1 shortly after the beginning of training (Chen S. X. et al., 2015 ). Indeed, SOMterm activation during motor learning destabilizes spines on PN apical tufts, whereas SOMterm suppression hyperstabilized those spines. These two results—fewer SOMterm synapses after motor learning, but more synapses after associative learning—appear at first glance to be contradictory. However, the decrease after motor learning was seen for L1 synapses onto L2/3 apical dendrites, whereas the increase after associative learning was seen in L4. The former are almost certainly synapses from Martinotti cellsterm·def, which target PNs, whereas the relevant L4 SOMterm cells could be non-Martinotti cells that target PVterm interneuronsterm in L4. The net effect of these changes could therefore be in the same direction—a disinhibition of PNs. Much remains to be understood about exactly how these SOMterm inhibitory networks are involved in learning. Yet the differential activation of these networks during learning paradigms illustrates that SOMterm cells do not merely relay sensory information, but rather modify cortical sensory responses based on an animal's previous experience.

Challenges in studying interneuronterm populations Although new optogeneticterm tools have been indispensable in understanding the role of specific cell types in intact cortical circuits in vivo , they do have important limitations that must be considered when interpreting the results. Variability in illumination intensity and duration, transgenic vs. viral expression, and the details of sensory stimuli or task parameters can interact in complex ways that affect how neurons respond to optogeneticterm manipulation even in the same cortical region. In the case of Arch-mediated suppression, for instance, depending on the cell type, and region of inactivation, it can be extremely difficult to completely silence neural responses. For example, spontaneous or low-amplitude evoked activity can show 65–80% suppression, whereas strong evoked burstingterm activity remains unchanged even with high-power Arch activation (Cardin, 2012 ). Expression of ChR2 can also produce variable responses of a given cell to the same light pulse, ranging from robust spiking responses to less reliable prolonged responses (Cardin, 2012 ). Misinterpretation of ChR2 manipulation may also come from atypical firing patterns evoked in cells that show late or suppressed responses under control conditions. In S1 SOMterm cells, for example, which are normally hyperpolarized by sensory stimulation, optogeneticterm activation would produce highly unnatural activity (Gentet et al., 2012 ). Recent advances in genetic calcium imaging techniques (such as GCaMP) have provided a new approach to the study of inhibitory neuronal populations such as SOMterm networks ((Jackson et al., 2016 ; Karnani et al., 2016a ),b). Two caveats to this approach are that SOMterm cells are known to have high basal firing rates, and TdTomato labeling in high concentrations can contaminate the fluorescence signal. Both of these caveats could lead this technique to underestimate neuronal firing rates, and thus care should be taken when inferring neural activity from fluorescence changes. Lastly, high levels of interconnectivity in cortical circuits make it challenging to study the effects of only a single interneuronterm subtypeterm. In particular, it is virtually impossible to perturb the activity of the SOMterm network without changing firing patterns in other inhibitory neurons, which may lead to conflicting and misleading results (Gibson et al., 1999 ; Cottam et al., 2013 ). Nevertheless, the past few years have seen tremendous advances in knowledge about interneuronterm function in general, and SOMterm cell function in particular. It is important to remember that SOMterm cells are not a single population, but rather consist of several distinct classesterm of inhibitory cells; new methods for targeting these specific subpopulations will only accelerate discoveries in the near future. It is also exciting that much of the progress in recent years has come not from studying SOMterm cells and sensory cortexterm in isolation, but rather from taking account of behavioral state, task context, learning, and sensorimotor integration—in other words, how SOMterm cells participate in large-scale interactions between sensory cortexterm and the rest of the brain.

Author contributions IY and MW wrote the paper.

Funding This work was funded by NIH R01 DC011379.

Conflict of interest statement The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.