1. Neuroscience
Download icon

Feedback inhibition and its control in an insect olfactory circuit

  1. Subhasis Ray
  2. Zane N Aldworth
  3. Mark A Stopfer  Is a corresponding author
  1. Section on Sensory Coding and Neural Ensembles, NICHD, NIH, United States
Research Article
  • Cited 0
  • Views 1,035
  • Annotations
Cite this article as: eLife 2020;9:e53281 doi: 10.7554/eLife.53281

Abstract

Inhibitory neurons play critical roles in regulating and shaping olfactory responses in vertebrates and invertebrates. In insects, these roles are performed by relatively few neurons, which can be interrogated efficiently, revealing fundamental principles of olfactory coding. Here, with electrophysiological recordings from the locust and a large-scale biophysical model, we analyzed the properties and functions of GGN, a unique giant GABAergic neuron that plays a central role in structuring olfactory codes in the locust mushroom body. Our simulations suggest that depolarizing GGN at its input branch can globally inhibit KCs several hundred microns away. Our in vivorecordings show that GGN responds to odors with complex temporal patterns of depolarization and hyperpolarization that can vary with odors and across animals, leading our model to predict the existence of a yet-undiscovered olfactory pathway. Our analysis reveals basic new features of GGN and the olfactory network surrounding it.

Introduction

Olfactory information is transformed dramatically as it travels from the periphery to higher order brain centers. Multiple types of olfactory receptor neurons can respond to a given odor with vigorous bursts of action potentials, while neurons deeper in the brain, in the pyriform cortex (vertebrates) or mushroom body (insects), respond to the same odor with only a spike or two (Laurent and Naraghi, 1994b; Friedrich and Laurent, 2001; Perez-Orive et al., 2002; Cang and Isaacson, 2003; Bathellier et al., 2008; Poo and Isaacson, 2009). In these higher order neurons, information about odors is represented sparsely by the identities of the active neurons (population coding) and in the timing of the few spikes elicited in those neurons (temporal coding) (Perez-Orive et al., 2002; Poo and Isaacson, 2009; Stettler and Axel, 2009; Gupta and Stopfer, 2014). In both vertebrates and invertebrates multiple mechanisms interact to mediate these transformations, including important inhibitory contributions from GABAergic neurons (Luna and Pettit, 2010; Papadopoulou et al., 2011; Lin et al., 2014; Palmer and Harvey, 2014; Large et al., 2016; Large et al., 2018). Here, with intracellular recordings and a new large-scale biophysical model that includes tens of thousands of cells and spans multiple layers of olfactory processing, we focus on a singularly important inhibitory neuron to investigate the roles of input activity and feedback inhibition in creating a sparse spatio-temporal odor representation in a higher order brain center. Together, our recordings and models point to new functions, neural connectivity patterns, and mechanisms that underlie transformations in the format of olfactory information.

The locust olfactory system is well-studied because it is accessible, robust, and relatively simple. In the absence of environmental odors, olfactory receptor neurons (ORNs) in the antenna are spontaneously active. Odorants elicit increases or decreases in the spontaneous firing rates of ORNs, and simple patterns of spikes with sequences of excitation and inhibition, distributed heterogeneously across the ORN population. These responses vary with the odor and its concentration. Spontaneous activity in ORNs drives background spiking in PNs (Joseph et al., 2012). PNs respond to the variety of odor-evoked responses in ORNs with sometimes-elaborate firing patterns that are further shaped by inhibition from the antennal lobe’s local interneurons (LNs) (Raman et al., 2010). Spikes in PNs are also coaxed into rhythmic waves by fast reciprocal interactions between excitatory PNs and inhibitory LNs (MacLeod and Laurent, 1996; Bazhenov et al., 2001). Odor-elicited firing patterns distributed across the population of PNs are informative about the identity, concentration, and timing of the odor (Laurent et al., 1996; Stopfer et al., 2003; Brown et al., 2005). This information is carried by PNs to the mushroom body and the lateral horn.

The primary neurons within the mushroom body are Kenyon cells (KCs). Unlike the volubly spiking PNs, KCs are nearly silent at rest and respond very selectively to odors by generating very few spikes (Laurent and Naraghi, 1994b; Perez-Orive et al., 2002; Joseph et al., 2012). Thus, any given odor evokes responses in a small fraction of the KC population, and any KC responds to a small set of odors (Perez-Orive et al., 2002; Stopfer et al., 2003). This sparseness of activity in KCs is thought to arise mainly from two factors: specialized membrane conductances that amplify only strong, synchronized input; and a feedback circuit that tamps down their spiking with cyclic inhibition (Perez-Orive et al., 2002; Demmer and Kloppenburg, 2009; Papadopoulou et al., 2011; Lin et al., 2014). In the locust the main sources of this inhibition are the giant GABAergic neurons (GGNs), one on each side of the brain (Leitch and Laurent, 1996; Papadopoulou et al., 2011; Gupta and Stopfer, 2012). Thus, GGN plays a central role in shaping neural codes for odors.

GGN spans much of each brain hemisphere and branches very widely (Figure 1a). It is reported to receive excitatory input from all 50,000 KCs at synapses within the mushroom body’s α lobe and, in turn, provide inhibitory feedback to all KCs 400–500 μm away within the calyx. In addition, GGN receives inhibitory input from a spiking neuron aptly named ‘Inhibitor of GGN’ (IG) which itself receives inhibition from GGN (Figure 1a, right) (Papadopoulou et al., 2011). GGN is a non-spiking interneuron. Odor presentations, spiking in KCs, and intracellular current injections have all been shown to depolarize GGN, but none of these stimuli causes GGN to generate spikes; even strong intracellular current injections into GGN elicit only passive responses (Leitch and Laurent, 1996; Papadopoulou et al., 2011); and our own observations).

Figure 1 with 1 supplement see all
GGN is a very large neuron, one per hemisphere, spanning a large portion of the locust brain.

(a) Left hemisphere: Dye filled GGN (green) in a locust brain (magenta, visualized with nc82 antibody) with an overlaid KC tracing (orange, filled in a different brain). Dorsal: towards the viewer; ventral: into the page). Right hemisphere: Circuit diagram of the locust olfactory system. Arrows show known synaptic connections, + excitatory, - inhibitory. Scale bar: 200 μm. (b) Very thin, feather-like neurites from GGN wrap around and penetrate the peduncle of the mushroom body; these neurites are not included in our reconstruction. (c) GGN’s neurites in the calyx have many bouton-like protrusions (d) whereas GGN’s α lobe branches are relatively smooth. (b–d) scalebars: 100 μm. (e) 3D reconstruction of the same neuron shown in panel (a) viewed from (i) ventral and (ii) posterior side of the brain. Major branches shown in different colors. LCA: lateral calyx, MCA: medial calyx, LH: lateral horn, a: anterior, p: posterior, d: dorsal, v: ventral, l: lateral and m: medial.

GGN’s structure likely shapes its function. GGN is very large, and along its path from the α lobe to the calyx, its initially thick processes divide at myriad branch points into vanishingly thin fibers. Cable theory applied to neurons (Rall, 1964) predicts that a passive voltage signal within such a structure will attenuate dramatically as it encounters cytosolic resistance along the neurites, will attenuate further as it divides at the neuronal arbor’s branch points, and will leak out through ionic channels in the cell membrane. Thus, it seemed possible that GGN might lack the biophysical capacity to transmit signals long distances throughout its expansive arbors. It has been shown that GGN inhibits KCs (Papadopoulou et al., 2011). However, if signals originating in GGN’s α lobe attenuate to the extent that they cannot (as proposed) effectively and globally hyperpolarize KCs via synapses 400–500 μm away in the calyx, then GGN must be inhibiting KCs through a different mechanism; perhaps, for example, through local interactions entirely within the mushroom body calyx.

The extent of signal attenuation throughout GGN has not been measured directly. Our own efforts to achieve this by making simultaneous recordings from multiple branches of GGN, and by using optical techniques, yielded results that were difficult to interpret. Therefore, we developed a realistic computational model of GGN to characterize signal attenuation throughout its structure. Prior studies in invertebrates have shown that 2–5 mV depolarizations in nonspiking interneurons suffice to evoke changes in the membrane potentials of their post-synaptic neurons (Burrows and Siegler, 1978; Manor et al., 1997). Our model showed that, although electrical signals do undergo substantial attenuation as they travel through GGN, signals arising in its α lobe branches appear to remain strong enough to provide global inhibition to KCs in the calyx.

Because GGN plays an outsized and central role in the olfactory system, investigating its responses to odors can illuminate broader issues of olfactory function. Thus, to further understand the network determinants of GGN’s responses to odors, we recorded from it in vivo while delivering a variety of odors to the animal, and then developed a large-scale model including KCs, IG, and realistic olfactory inputs from PNs, to investigate the types of network activity needed to generate the membrane potential patterns we observed in GGN. We identified two novel features in the olfactory network. First, to generate the types of membrane potential patterns we observed in GGN, the synaptic connection strengths onto KCs must be heterogeneous. Second, our in vivo recordings of GGN revealed novel, complex response patterns not previously documented, including periods of hyperpolarization, that vary with the odorant. Although GGN receives reciprocal feedback from IG (Papadopoulou et al., 2011), the periods of hyperpolarization could not be explained by disinhibition of IG from GGN. Instead, our model predicts that this behavior could arise if, in addition to receiving input from GGN, IG also receives direct excitation from another, presently unknown odor-activated pathway. Additionally, our model replicated emergent features of the olfactory network not explicitly programmed into it, such as the appearance of a small portion of KCs spiking at relatively high rates.

Together, the results of our in vivo recordings and large-scale realistic computational models provide a more complete understanding of how different parts of the olfactory system interact. To generate odor-specific temporally patterned responses in GGN and in the mushroom body, temporally-patterned odor evoked excitation from PNs, feedback inhibition from GGN, and inhibition of GGN by odor-driven IG must all cooperate.

Results

GGN morphology

A valuable use of computational modeling is to answer questions about biological systems when those systems are too large, too complex, or too difficult to study with the tools of physiology alone. Earlier computational studies of the insect olfactory system used relatively simple models of neurons such as integrate-and-fire or map-based approaches that collapse entire neuronal structures into a single point (Perez-Orive et al., 2004; Papadopoulou et al., 2011; Arena et al., 2015; Kee et al., 2015; Peng and Chittka, 2017). However, GGN’s giant size, elaborate branching, and passive membrane properties raised questions about its function that could only be addressed by considering properties determined by its morphology. Thus, to understand how the size and shape of GGN affects electrical signal propagation, we constructed a detailed morphological model (available at neuromorpho.org).

To reconstruct the morphology of GGN, we first made intracellular recordings from it in vivo, filled it with dye, and obtained 3D confocal images of the dye-filled cell (Figure 1a, left; Figure 1—video 1). As previously shown, GGN has a reliable and unique location and morphology (Weiss, 1978; Leitch and Laurent, 1996; Papadopoulou et al., 2011; Gupta and Stopfer, 2012). On each side of the brain, its soma is located ventrally, just anterior to the optic nerve. A single neurite emerges from GGN’s soma, travels toward the posterior and dorsal side of the brain, and splits there into two branches, one innervating the α lobe and the other the mushroom body. Extending outward, these branching neurites expand in width, becoming much thicker than the primary neurite. The mushroom body branch further divides into two thick processes that innervate the medial and the lateral calyx. A thin neurite emerging from the lateral calyceal branch loops back to the lateral horn, close to the soma, and splits there into many branches. We further observed, for the first time, myriad thin fibers that emerge from the stems of the calyceal branches, splitting into very fine feather-like neurites that wrap densely around the peduncle, with some investing the peduncle core (Figure 1b). The neurites in the calyx and lateral horn are dotted with many irregular bouton-like structures (Figure 1c) whereas the branches in α lobe are relatively smooth (Figure 1d).

In two animals we traced and reconstructed the morphology of GGN from confocal image stacks (Figure 1e). Analyzing these traces, we found that the maximum path length of the neuronal trees (i.e., the maximum distance between any two points on the neuronal tree when traversed along the neurites) was on the order of 2 mm, and the maximum physical length (i.e., Euclidean distance between any two points on the neuron in three-dimensions) was on the order of 1 mm. Some neurites at their thickest were nearly 20 μm in diameter. The total traced branch length (i.e., the sum of the lengths of all the neurites) was about 65 mm (although many vanishingly thin branches were too fine to trace). Compared to 96,831 vertebrate and invertebrate neurons cataloged in the neuromorpho.org database, GGN fell into the 99.75th percentile for total branch length, and the 99.95th percentile for number of branch points. It is a really big neuron.

Signals traveling through GGN’s arbors are predicted to attenuate significantly

To investigate GGN’s electrical properties we constructed passive electrical model cells by transferring the morphology tracings of the two GGNs into the NEURON simulator (Carnevale and Hines, 2006). Simulations with both models produced qualitatively similar results; the model we describe here was derived from the second neuron we traced because it was imaged and mapped at higher resolution. To account for branches and changes in the diameters of processes that affect electrotonic distance, we segmented the model GGN into 5283 isopotential compartments. We set the membrane resistivity in the model to 33 kΩ-cm2 following published descriptions obtained from other non-spiking neurons in the locust (Laurent, 1991), the specific membrane capacitance to 1 μF/cm2, the approximate value for cell membranes (Curtis and Cole, 1938; Hodgkin and Huxley, 1952; Gentet et al., 2000), and the cytoplasmic resistivity to 100 Ω-cm, a typical order-of-magnitude value for neurons (Hodgkin and Rushton, 1946; Stuart and Spruston, 1998; Roth and Häusser, 2001).

Feedback signals are thought to travel passively from GGN’s α lobe branch to its calyceal branches. To test the extent of passive signal attenuation through GGN’s structure, we simulated a voltage clamp experiment in the GGN model (Figure 2a). In vivo, intracellular recordings (Papadopoulou et al., 2011; and our own) show GGN’s membrane potential rests at about −51 mV. Because strong odor stimuli depolarize GGN by about 10 mV in recordings made near the base of the α lobe branch (Papadopoulou et al., 2011) (also see Figure 4) we first stepped the clamp voltage in the α lobe branch to −40 mV, and, after holding it there for 450 ms, measured the resulting depolarizations throughout the model GGN (Figure 2a inset). Notably, the extent of signal attenuation was substantial and varied throughout the calyx, yielding depolarizations ranging from ~5–9 mV. The signal decreased with branch distance from the α lobe, leaving the least amount of signal at the medial portion of the lateral calyceal branch (Figure 2b,c, Figure 2—video 1).

Figure 2 with 4 supplements see all
Passive voltage spread through GGN.

(a) Simulation schematic: GGN’s α lobe branch was clamped at its stem to −40 mV (from a holding voltage of −51 mV). (b) Conditions described in panel a lead to different steady state depolarizations of the GGN model at different locations; color indicates voltage change from resting potential; axial resistivity RA = 100 Ωcm; membrane resistivity RM = 33 kΩcm2. (c) Violin plots showing distribution of depolarization in the α lobe and the calyceal dendrites of GGN for conditions described in panel a; width of gray area indicates density, white line median, and whiskers data range. (d) Schematic of simulation of multiple, random synaptic inputs to GGN: 500 synapses were connected to GGN’s α lobe branch. (e) Bottom: raster plot of incoming spike times at all 500 synapses. The spikes arrived at each synapse at random times at linearly decreasing rate, starting with 20/s down to 0 after 500 ms. Top: for conditions shown in panel d, membrane potentials at randomly sampled terminal neurites in different regions of GGN. (f) Violin plots of peak depolarizations in the α lobe dendrites and the calyceal dendrites in model described in panel d. (g) Violin plots of distribution of depolarizations in the calyx for different values of axial resistivity (RA, bottom) and membrane resistivity (RM, right) for the voltage clamp simulation described in panel a. Lines connect the median results for specified RMs (right).

Since excitatory input typically arrives in GGN not at one location, but rather at many synapses from many KCs, we sought to test a more realistic form of simulated input to the α lobe arbor of GGN. Thus, we provided nonhomogeneous Poisson spike trains through 500 excitatory model synapses; each synapse had a maximum rate of 20 spikes/s that ramped down linearly to 0 over a 500 ms interval (Figure 2d,e). This stimulus set was calibrated to generate a peak depolarization in the thick branches of GGN in the same range we observed in vivo. This more realistic test also revealed significant attenuation of voltage in the neuron’s distant branches (Figure 2e,f) as well as delays in signal propagation (Figure 2—figure supplement 1).

Neither the membrane resistivity (RM) nor cytoplasmic (axial) resistivity (RA) has been measured definitively in GGN; yet, for a given morphology, these two parameters help determine signal attenuation. Thus, we explored a range of values for these two parameters with the voltage clamp simulation approach shown in Figure 2a. We based the RM value range on data obtained from many types of neurons provided by the neuroelectro.org database. For RA, neurophysiological data are sparse, so we explored broadly around the range of published values (Hodgkin and Rushton, 1946; Stuart and Spruston, 1998; Roth and Häusser, 2001). As expected, higher RA yielded greater signal attenuation, whereas higher RM yielded less signal attenuation (Figure 2g). This analysis suggests that signal transmission in GGN is robust: except for the most extreme values of this parameter range, signals from the α lobe remained strong enough throughout the calyx (>2–5 mV) to support synaptic transmission. The extent of depolarization throughout GGN’s calyceal arbor varied with location, as quantified in the extended lower lobe in the violin plots in Figure 2c,f and g.

Branches of GGN receiving weaker signals from the α lobe would be expected to provide less inhibition to their postsynaptic KCs. In a simplified model in which all KCs were strongly stimulated by identical input from PNs, the amount of KC spiking was indeed negatively correlated with local GGN voltage deflections (Figure 2—figure supplement 2). However, in a more realistic model of the mushroom body network that included variable excitatory input from PNs and variable strengths of inhibitory synapses between GGN and KCs, we found the negative correlation between depolarizations measured at presynaptic locations throughout GGN and postsynaptic KC activity was small and likely negligible (Figure 2—figure supplement 3). This suggests GGN’s inhibitory output has a surprisingly uniform influence across the population of KCs regardless of location.

Feedback inhibition is predicted to expand the dynamic range of KCs

We used our model to test other basic properties of GGN and the olfactory network surrounding it. In vivo, feedback inhibition from GGN sparsens the odor-elicited responses of KCs by increasing the KC spiking threshold and by restricting KC spiking to brief temporal windows defined by the oscillatory cycle established in the AL (Papadopoulou et al., 2011; Gupta and Stopfer, 2014). Large-scale feedforward inhibition has previously been shown to expand the dynamic range of cortical neurons (Pouille et al., 2009), and in Drosophila, feedback inhibition from APL, the analog of GGN, expands the dynamic range of KCs (Inada et al., 2017). Whether feedback inhibition from GGN has a similar effect on KCs is unknown. To test this, we extended our model to include, for simplicity, a single KC receiving feedback inhibition from GGN (Figure 3a). To simulate the KC in this test we used a single compartmental model with Hodgkin-Huxley type ion channels (Wüstenberg et al., 2004). Since a single KC would have negligible effect on GGN, we applied its spiking output to GGN’s α lobe branch via 50,000 synapses. To avoid unrealistic, strong synchronous input to GGN, we jittered the incoming spike times by applying random synaptic delays between 0 and 60 ms. Thus, after each spike generated by the model KC, GGN received 50,000 EPSPs spread over a 60 ms time window. We drove the KC model with a range of tonic current injections and compared its responses to those of an isolated KC model receiving the same input without feedback inhibition. As expected, baseline inhibition from spontaneous activity in GGN increased the KC’s threshold for spiking. Notably, though, the GGN-coupled KC continued to spike over a much larger range of current injection than the isolated KC, which quickly saturated to a level where it could no longer spike (Figure 3b,c). A similar result was obtained when we tested the KC by applying simulated GGN inhibition from a model of the olfactory network described later (Figure 3—figure supplement 1). These results suggested that feedback inhibition from GGN allows an individual KC to function effectively over a larger dynamic range of inputs. To quantify our results, we used a standard analysis from control systems literature in which dynamic range is characterized by the slope of the input-response curve, which quantifies the effectiveness of input for eliciting output. Expanding the dynamic range makes the slope of the input-response curve shallower, as we observed in our model (Figure 3c and d; see Materials and methods for the slope calculation). Thus, our model predicts that feedback inhibition from GGN expands the dynamic range of KCs.

Figure 3 with 1 supplement see all
Feedback inhibition from GGN extends the dynamic range of KCs.

(a) Model schematic: A single KC sends 50,000 excitatory synapses with random delays of 0–60 ms into GGN’s α lobe branch. The KC receives feedback inhibition via a graded synapse from GGN’s calyceal branch. A step current is injected into the KC from 0.5 to 3 s (up to 2 s is shown). (b) The model KC was tested with current pulses of increasing amplitude in three configurations: Isolated KC (left); KC receiving only spontaneous inhibition from GGN (middle), and KC receiving spontaneous and odor-elicited feedback inhibition (right). As the amplitude of the current pulse increases, the KC’s spiking first increases, and then, as the membrane potential nears saturation, decreases. (c) Comparison of the number of spikes evoked by current steps in KC in the three configurations shown in b). (d) Slopes of the rising parts of the KC-only (left, cyan) and KC with GGN feedback (right, orange) curves in c), normalized by the KC-only slope.

GGN responses can be complex, including hyperpolarization

Our recordings made in vivo from GGN frequently revealed depolarizations throughout an odor presentation, often with stronger depolarizing peaks upon the odor’s onset and offset (Figure 4, Animal 1, hexanol) (see also Papadopoulou et al., 2011). Notably, though, our recordings from GGN sometimes revealed more complex odor-elicited responses than previously reported, including combinations of depolarization and hyperpolarization (Figure 4, Animal 1, hexanal). Moreover, we found that the same GGN could respond differently when different odors were presented; for example, GGNs from Animals 2 and 3 shown in Figure 4 depolarized in response to one odor and hyperpolarized in response to another. Also, GGNs in different animals could respond differently to the same odor (Figure 4, hexanal). Almost a quarter of the odor-GGN pairs in our in vivo recordings showed reliable hyperpolarizations at some point in the odor response (40 out of 169). Earlier computational models (Papadopoulou et al., 2011; Kee et al., 2015) did not attempt to reproduce the variety of GGN responses we observed in vivo, including hyperpolarizations. To better understand the mechanisms underlying GGN’s odor-elicited responses (and by extension, novel features of olfactory circuitry), we placed our GGN model at the center of a more extensive mushroom body olfactory network.

In vivo, GGN’s responses to odors vary with odor and animal.

Examples from 3 animals and two odors (horizontal gray bars, 1 s) (black traces: average of 5 trials, gray: standard error of the mean, data low-pass filtered at 49 Hz). Response features include depolarization or hyperpolarization upon stimulus onset and/or offset, and depolarization or hyperpolarization evoked by different odors in the same GGN.

Sustained GGN responses require heterogeneous input to KCs

We sought to identify olfactory network characteristics required to generate realistic responses in GGN. Because GGN plays a central role in the olfactory system, this analysis could reveal gaps in our understanding of the network. We first tested whether simply simulating the average statistics of olfactory input to the mushroom body network would elicit responses from GGN like those observed in vivo. To test this, we extended our detailed GGN model with a full population of 50,000 simulated KCs. We used a relatively simple NEURON version of a single compartmental KC model (Wüstenberg et al., 2004). This approach was sufficient because KCs are very small, have few dendritic branches, and generate action potentials -- in contrast to the large, complex, and passive GGN, the morphologies of individual KCs are unlikely to differentially influence their odor coding properties. Each model KC was connected to GGN in the α lobe via an excitatory synapse, and, in turn each KC received inhibitory input from a random segment of GGN in the calyx via a graded synapse (Figure 5a). To provide excitatory input to the KCs, the firing patterns of 830 PNs were simulated as spike-trains, and, following observations made in vivo (Jortner et al., 2007), each KC was stimulated by 50% of these spike trains via excitatory synapses. The spike trains were designed to follow the statistics of PNs recorded in vivo (Mazor and Laurent, 2005; Jortner et al., 2007). Thus, 77% of the PN spike trains were assigned a spontaneous firing rate of 2.6 spikes/s; during odor stimulation, 20% of these PNs were switched to 20 spikes/s modulated by the 20 Hz oscillations generated in the antennal lobe and reflected in the local field potential (LFP), and 10% were inhibited (no spikes) (Figure 5b). In this model all the synapses between two given cell types had the same conductance. Simulations of this model produced a few highly synchronized bouts of spiking in the KC population (Figure 5c), and corresponding isolated peaks in GGN’s membrane potential (Figure 5d). These unrealistic responses were similar to those generated by earlier models (Papadopoulou et al., 2011; Kee et al., 2015).

Figure 5 with 2 supplements see all
Tuning the olfactory network by reference to GGN’s olfactory responses required heterogeneous synaptic strengths onto KCs and structured patterns of activity in PNs.

(a) Schematic of mushroom body network model. Each of the 50,000 KCs receives input from 50% of the 830 PNs, which are modeled as spike trains. All KCs excite GGN in its α lobe branch and receive inhibition from a random calyceal branch of GGN. (b–d) Model with simplified, homogenous firing patterns in PNs and uniform synaptic strengths generates unrealistic membrane potential in GGN. (b) Raster plot of model PN spike trains (67 shown); dots in each row mark spike times in a PN. (c) Raster plot of spike trains evoked in the KCs (397 shown) when all of them receive identically strong inhibitory connections from GGN. (d) Unrealistic membrane potential in GGN features a few peaks corresponding to highly synchronized bouts of activity in KCs. Light gray bar: 1 s odor stimulation. (e–g) Model in which subpopulations of PNs have different temporal patterns of spiking during and after odor stimulation generates unrealistic membrane potential in GGN. (e) Rasters show different firing patterns in different PNs. (f) Simulation of model with PN activity pattern in panel e) along with uniform synaptic strengths onto KCs generate synchronized spiking in KCs and (g) unrealistic membrane potential in GGN. Simulation of a model with both structured PN activity patterns in panel e) and heterogeneous synaptic strengths gives rise to (h) temporally diffuse spiking in KC population and (i) sustained depolarization of GGN similar to that observed in vivo (e.g. Figure 4 animal 1). Dark gray bar: 1 s odor stimulation; light gray bar: 0.5 s ‘off response’ period.

Since this simplified model network did not reproduce the long-lasting depolarization of GGN we had observed in vivo, we introduced more biological detail. Synaptic strengths follow a lognormal distribution (Song et al., 2005; Loewenstein et al., 2011; Buzsáki and Mizuseki, 2014). After we adjusted our network model to include this property, some KCs became weakly inhibited, allowing them to fire more volubly. We also adjusted the input to KCs provided by PNs. In our simplified model, input to the KCs included spikes from a fixed set of PNs that were all active throughout the duration of the odor stimulus. However, in vivo, spiking patterns including periods of excitation and inhibition in PNs evolve over the course of an odor presentation, and different PNs respond to the same odor in different ways (Laurent and Davidowitz, 1994a; Stopfer et al., 2003; Mazor and Laurent, 2005); thus, a given odor presentation activates sets of KCs that shift in composition over the course of the response.

To simulate these diverse responses, we divided the model’s PN population into five groups: four groups that responded to the stimulus and one that did not. In each of the responsive groups, odors activated changing subsets of PNs, updated with each LFP cycle (Figure 5e). Lacking the heterogeneity in synaptic strengths onto KCs described above, even this complex and changing input activity pattern elicited unrealistically synchronized bouts of activity in KCs, resulting in unrealistic isolated peaks in GGN’s simulated membrane potential (Figure 5f,g). Increasing the strength of input from PNs allowed KCs to spike throughout the stimulus duration and generate sustained depolarization of GGN, but, under this condition, the number of spiking KCs became much higher than observed in vivo (Perez-Orive et al., 2002). Increasing the diversity of synaptic delays from PNs to KCs did not alleviate overly-synchronized spiking in the KC population if heterogeneity in synaptic strengths and temporal patterns of PN activity were omitted from the model (Figure 5—figure supplement 1). However, a model combining heterogeneous connectivity with structured PN firing patterns gave rise to GGN voltage traces that included sustained depolarization and temporal dynamics more characteristic of responses we had observed in vivo (Figure 5h,i). Notably, even the steady PN activity shown in Figure 5b could produce sustained depolarization in GGN when the synaptic strengths were diversified (Figure 5—figure supplement 2). This analysis, centered on GGN’s responses, revealed important constraints for modeling the olfactory system.

Some KCs fire at high rates in a GGN-centered olfactory network model

The distribution of firing rates of the KC population in our simulations showed that, while most KCs produced 0–2 odor-elicited spikes, a few KCs spiked much more (Figure 6a and Figure 6—figure supplement 1). To compare the firing rate distribution of KCs in the model to that in vivo, we made patch clamp recordings from 147 KCs in 114 animals, obtaining results from 707 KC-odor pairs. On average, the spontaneous firing rates of these real KCs were very low (~0.09 Hz) and reached only somewhat higher rates during odor presentation and immediately after odor termination (~0.15 and~0.16 Hz, respectively, Figure 6b), as previously observed (Perez-Orive et al., 2002; Gupta and Stopfer, 2014). Some odor-elicited responses in KCs, though, consisted of many more spikes. Figure 6c shows a representative example of a hyperactive response in a KC, in which a 1 s odor pulse elicited an average of 7 spikes (Figure 6d and e). Figure 6f shows spike times obtained from 10 other KC-odor pairs. Overall, the distribution of spike counts in KCs we tested in vivo was clustered close to 0 but included a long rightward tail (Figure 6g), in agreement with our model (Figure 6a). KCs generating especially strong responses were not located in any particular part of the calyx, nor did they feature distinguishing morphologies (data not shown).

Figure 6 with 1 supplement see all
In vivo, some KC responses are hyperactive, as predicted by simulations.

(a) Distribution of KC firing rates in simulation producing realistic GGN depolarization. (b) KC firing rate averaged over 707 KC-odor pairs. Shaded region indicates standard error of the mean. Black horizontal bar: 1 s odor stimulation. (c) A hyperactive response KC filled with dye (mushroom body calyx and pedunculus outlined with dashed lines (scale bar: 50 μm) and (d) its recorded membrane potential in response to odor stimulus, and (e) its average firing rate elicited by this stimulus computed in a 100 ms window. (f) Sample KC spike trains. (g) Histogram of average firing rate of KCs across trials upon odor presentation measured in a 1.4 s window. The last bin in the rightward tail includes two different KCs.

Although relatively rare KC odor responses with many spikes have been observed before (Perez-Orive et al., 2002; Gupta and Stopfer, 2014), the origin and significance of these responses were unclear. We tested this with a series of simulations in which we systematically reduced the number of high spike-rate KCs by disconnecting them from the network. We found that when KCs producing more than five odor-elicited spikes were removed (Figure 7—figure supplement 1), a different set of KCs emerged to replace them by becoming more active. However, the replacement high spike-rate KCs were fewer in number than the original high-firing set (Figure 7a). As a result, the population spike time histogram and corresponding GGN voltage deflections became smaller (Figure 7b and c). As we reiterated this process, deleting newly-emerged high spike-rate KCs from the network with each round, the rightward tail of the KC population’s spike count distribution shortened, the number of low-spike KCs increased (Figure 7a), and, overall, the total number of spikes produced by the entire KC population decreased monotonically. Eventually, when no KCs in the reiterative simulation generated six or more odor-elicited spikes, we decreased the spike limit, first to three or more spikes, and then, to any spikes (Figure 7—figure supplement 1). Throughout this process, GGN’s odor-elicited response gradually decreased, shifting from extended depolarization to unrealistic, isolated peaks. Ultimately, no spiking KCs, or responses in GGN, remained (Figure 7c). Our analysis showed two things: first, the high spike-rate KCs in our model are not a static population, but rather change in identity, reflecting emergent features of the network’s distribution of synaptic conductances, PN input, and GGN feedback; and second, high-spike rate KCs augment the depolarization of GGN to realistic levels.

Figure 7 with 1 supplement see all
High spike-rate KCs contribute to sustaining GGN’s odor-elicited depolarization.

(a) Histogram of spike counts in KCs for one template network at the first and the last run of each spike count limit. (b) Population spike time histogram of KCs (100 ms bins) for the same simulations. (c) GGN membrane potential. See text for the procedure. Legend is same for a–c.

Odor-evoked hyperpolarization in GGN originates in IG spiking

Our intracellular recordings from GGN revealed extended periods of odor-elicited hyperpolarization (Figure 4), something not previously reported in recordings made in vivo nor explained by existing models of GGN within its olfactory network. We hypothesized that these periods of hyperpolarization might originate in the activity of IG, a neuron known to share reciprocal inhibition with GGN (Figure 1a, right) (Papadopoulou et al., 2011). Specifically, we hypothesized that an increase in IG activity might underlie the periods of hyperpolarization in GGN, with IG’s activity increase following release from inhibition from GGN. To test this, we first tried adding IG to our model according to the simple reciprocal connectivity plan shown in Figure 1a. We tested a range of synaptic strengths and time constants for GGN→IG, and IG→GGN synapses, and manipulated IG’s baseline firing rates, but none of these model variations could generate odor-elicited hyperpolarization in GGN (data not shown). Thus, our model suggested simple reciprocal interactions between these neurons cannot generate this response.

We sought to test this in vivo. We reasoned that if spiking in IG increased solely because inhibition from GGN decreased, then IG spiking should only increase after odor stimuli ended and activity in GGN returned to baseline. The location and most properties of IG are unknown, and we were not able to identify it in our recordings, making it difficult to directly test our idea. However, a previous report showed spikes in IG correlate one-to-one with IPSPs in GGN (Papadopoulou et al., 2011), suggesting we could infer IG’s activity and properties by examining GGN’s membrane potential. Our recordings made in vivo from GGN revealed IPSPs in the absence of odors, and that odor presentations caused the frequency of IPSPs to increase (Wilcoxon signed-rank test, N = 198 pairs, W = 2328.5, p<<0.001; responses from two animals are shown in Figure 8a, and responses from 1257 trials with several odors from 47 GGNs are shown as a peri-stimulus time histogram in Figure 8b and c.) Assuming these IPSPs originate as spontaneous and odor-elicited spikes in IG (Papadopoulou et al., 2011), these results show that IG’s responses to an odor pulse are delayed relative to the odor onset, and lengthy. Notably, as evident in Figure 8b, IG’s firing rate begins to increase before GGN’s membrane potential returns to baseline. This result rules out the possibility that IG’s odor response is driven solely by a reduction in inhibition from GGN, and suggests, rather, that IG receives odor-driven excitation from another pathway. It has been hypothesized that IG receives input from KCs (Papadopoulou et al., 2011). Using our model, we could indeed reproduce realistic hyperpolarization in GGN’s membrane potential by adding excitatory synapses to IG from either the KCs or the PNs (Figure 9a–e). Our simulations produced a variety of membrane potential profiles in GGN including hyperpolarizations; the profiles varied with the odor-driven activity patterns in PNs (Figure 9f–i).

IPSPs in GGN suggest IG spikes upon odor presentation and is responsible for hyperpolarizations in GGN.

(a) In vivo recordings of GGN’s membrane potential from two animals showing IPSPs (arrowheads) believed to originate as spikes in IG. Vertical scale bar: 10 mV. (b) Peristimulus time histogram (PSTH) shows IPSP peak times from 1257 odor-trials across 47 GGNs, presumably reflecting spikes in IG. Because IG begins to spike before GGN’s membrane potential returns to baseline, IG’s odor response cannot be driven by disinhibition from GGN. Black horizontal bar: odor presentation in panels (a) and (b), which share the time axis. (c) Same as panel (b) but showing full responses.

An additional, unidentified olfactory pathway to IG is needed to explain odor-elicited hyperpolarization in GGN.

(a) Schematic of model with IG receiving direct excitation from KCs and reciprocal inhibition from GGN. (b) Spike raster of PN activity producing GGN hyperpolarization. (c) Spike raster of KC population activity in the same simulation. (d) Simulated GGN membrane potential including odor-elicited hyperpolarization mimics responses observed in vivo (e.g. Figure 4, Animal 3, hexanal). (e) Corresponding simulated IG membrane potential. This simulation included a 200 ms synaptic delay from KCs to IG. (f–i) Varying the temporal pattern of PN population activity can produce different response pattern in the same network as e. (f) Raster plot of PN activity with a different temporal pattern from that in e. (g) Spike raster showing concomitant KC activity. (h) Simulated GGN membrane potential including odor evoked de- and hyperpolarization (similar to Figure 4, Animal one hexanal). (i) Corresponding simulated IG membrane potential. (j) Spontaneous activity in IG does not originate in the antennal lobe. Top two black traces show spontaneous IPSPs in GGN’s membrane voltage from two animals with intact olfactory systems. The bottom two red traces are from the left and the right GGN in another animal in which the antennal lobes had been silenced by cutting both antennal nerves. Vertical scale bar 5 mV, horizontal scale bar 1 s.

Our analysis of IG’s activity suggested that it spikes in the absence of odor stimulation (Figure 9j, top black traces). Thus, a remaining question concerned the source of excitation driving this spontaneous activity. KCs are nearly silent when odors are not present (Perez-Orive et al., 2002; Gupta and Stopfer, 2014), making it unlikely that they provide the sole source of excitation to IG. PNs, though, spike spontaneously because they receive direct, powerful input from spontaneously active ORNs (Joseph et al., 2012), suggesting a PN-driven pathway might be responsible for spontaneous activity in IG. To test this in vivo, we completely silenced PNs and KCs by bilaterally cutting the animal’s antennal and labral nerves (Joseph et al., 2012) and then recording intracellularly from GGN. We found that spontaneous IPSPs in GGN persisted in preparations with silenced PNs and KCs (Figure 9j, bottom red traces), demonstrating that IG’s spontaneous spiking either arises intrinsically or is driven by a source other than PNs or KCs.

Discussion

Inhibitory neurons play critical roles in regulating and shaping olfactory responses in vertebrates and invertebrates (Kay and Stopfer, 2006). In insects, these roles are performed by relatively few neurons that can be interrogated efficiently, revealing fundamental principles of olfactory coding. The unique giant GABAergic neuron GGN plays a central role in structuring olfactory codes in the locust mushroom body by regulating the excitability of KCs and parsing their responses into rhythmic bursts. We combined intracellular recordings from GGN and KCs, and developed a new morphologically detailed model of GGN as a focus of analysis to investigate GGN’s properties, inputs, and outputs. Further, we used a broader model of the olfactory system built around GGN to explore several basic properties of the olfactory network. Our computational model, based on our electrophysiological recordings, successfully reproduced the sparse activity of KCs and the membrane dynamics of GGN in the locust brain while providing new hypotheses about how the mushroom body circuit processes odor information.

It is clear that GGN inhibits KCs, but the pathway carrying this inhibitory signal through GGN is uncertain. It has been proposed that signals generated within GGN’s α lobe branch propagate to its distant calyceal branch, where they provide global inhibition to all KCs (Papadopoulou et al., 2011). But, the enormous size, extensive branching, and passive conduction of GGN brought us to hypothesize that signals arising in GGN’s α lobe branch would attenuate to such an extent as they travel to the calyx that they would be unable to effectively inhibit KCs. If this hypothesis were correct, GGN’s inhibition of KCs would presumably occur by a different mechanism, such as through local inhibition entirely within the mushroom body calyx. Non-spiking interneurons in insects are often large with complex splays of neurites in separate brain areas, suggesting their far-flung branches may be functionally isolated, serving separate local computations (Burrows, 1981). A variety of complex, local processing has been proposed to occur within the locust mushroom body calyx (Leitch and Laurent, 1996). Our simulation of GGN’s morphological and electrical properties shows that realistic, odor-elicited levels of depolarizations of ~10 mV in GGN’s α lobe branch do indeed attenuate greatly with distance to amplitudes as low as 5 mV in parts of the calyx. However, earlier studies of non-spiking neurons in invertebrates showed depolarizations of this amplitude could evoke enough neurotransmitter release to affect responses of postsynaptic neurons. For example, (Burrows and Siegler, 1978) showed that depolarizations of only about 2 mV in a non-spiking interneuron in the metathoracic ganglion of the locust can change the firing rate of its postsynaptic motor neuron. Similarly, (Manor et al., 1997) showed in a graded synapse in the lobster stomatogastric ganglion that voltage steps from −50 mV to −45 mV can reliably evoke postsynaptic effects. Thus, we conclude that GGN has the biophysical capacity to convey signals from KCs at GGN’s α lobe branches to the calyx, providing effective global inhibition to all KCs.

GGN’s arborizations in the calyx extend to different lengths, suggesting signals arising in the α lobe could attenuate more in some of its calyceal branches than in others. Our simulations indeed showed the amount of depolarization reaching GGN’s distant branches varied with their locations, but only by a few millivolts (Figure 2b), within the range of variation caused by other factors, such as synaptic strength. Consistent with this, imaging experiments in APL, the Drosophila analog of GGN (Lin et al., 2014) revealed different levels of Ca2+ activity in different spatial regions of its arbor when the animal was stimulated with low concentrations of odors (Inada et al., 2017). Inada et al. also found that some KCs are more effective than others at activating APL, and that APL, in turn, provides diverse levels of inhibition to different subsets of KCs in the same brain. By optogenetically activating a small subset of PNs, Inada et al. showed that APL could provide local inhibition to a subset of KCs, likely through dendrodendritic connections between KCs and APL within the calyx. Indeed, reconstructions of APL based on serial section electron microscope images show that it receives synaptic connections from KCs throughout its arbor, including within the calyx (Eichler et al., 2017; Zheng et al., 2018).

Could local and global inhibition of KCs operate in parallel? In vivo, strong odor stimuli appear to drive both GGN and APL to provide global inhibition to all KCs (Papadopoulou et al., 2011; Lin et al., 2014), and our model shows (Figure 2) that depolarizations of GGN evoked by strong input can spread robustly throughout the giant neuron’s arbor. We speculate that in the locust, weaker stimuli might evoke weaker depolarization in GGN’s α lobe branch, leading to attenuated signals in GGN’s distal branches too weak to inhibit KCs. If, GGN, like APL, forms reciprocal synapses with KCs within the calyx, it might be able to inhibit KCs there locally even in the absence of a global inhibitory signal. In this scenario, a KC driven to spike by a weak odor stimulus would effectively inhibit only its close neighbors via local interactions with GGN. Assuming randomly-connected input from PNs to KCs (Jortner et al., 2007; Caron et al., 2013; Eichler et al., 2017), we can speculate about possible impact of strong or weak odor stimuli on the circuit. Strong odor stimuli might drive this circuitry into a global, population-wide, winner-take-all scenario in which only KCs receiving the strongest inputs manage to spike. Weak stimuli, on the other hand, might recruit only local inhibition, allowing a distribution of winners to emerge locally, resulting in a center-surround-type of contrast enhancement.

Inhibition from GGN is known to sparsen the firing of KCs and to impose rhythmic time windows on their responses (Papadopoulou et al., 2011; Gupta and Stopfer, 2012). Notably, our model also predicts that feedback inhibition from GGN can expand the dynamic range of inputs able to activate KCs (Figure 3). This prediction is consistent with observations made by imaging calcium in Drosophila where blocking inhibitory inputs to KCs narrowed the dynamic range of their odor responses (Inada et al., 2017). Why might expanding the dynamic range of KCs benefit olfactory coding? In the brain, a wide range of synaptic strengths exists even within a given type of neuron, and, further, synaptic strength can change over time and with experience (reviewed in Barbour et al., 2007). As in other systems, it is possible that homeostatic mechanisms tune conductances in KCs to match their inputs, enabling them to respond appropriately to a broad and changing range of inputs (reviewed in Marder and Goaillard, 2006). Our results suggest an additional mechanism: feedback inhibition from GGN that helps KCs function robustly by expanding their sensitivities to a wider range of inputs, enabling them to generate consistent responses.

Although KCs generally respond very sparsely to any given odor, with very few cells firing just one or a few spikes (Laurent and Naraghi, 1994b; Perez-Orive et al., 2004; Mazor and Laurent, 2005), examples of KCs that fire substantially more spikes, spontaneously and in response to odors, have been reported (Perez-Orive et al., 2002; Gupta and Stopfer, 2014). Notably, our simulations that gave rise to realistic, sustained responses in GGN always included some high-spike rate KCs. When we removed these high-firing rate KCs from the network, new high-firing rate KCs arose to replace them (Figure 7). This analysis showed that high-firing rate responses in KCs are an emergent property of the olfactory network, determined by the balance of input KCs receive: distributed patterns of excitation from PNs; and rhythmic inhibitory feedback from GGN. Deleting all high spike-rate KCs from a network had a modest effect on GGN, slightly reducing the amplitude of its sustained, odor-elicited depolarization. Deleting KCs generating 3–6 spikes per stimulus reduced GGN’s depolarization to unrealistic levels, leaving only widely-separated peaks of activity (Figure 7).

Our intracellular recordings from GGN revealed more elaborate membrane potential dynamics than previously reported, including reliable and prolonged stimulus-dependent periods of hyperpolarization, that varied with odor and animal. In our model, simple reciprocal inhibition of GGN by the inhibitory neuron IG could not reproduce these features. But we found introducing odor-elicited excitatory input to the inhibitory neuron IG could give rise to hyperpolarizations of GGN’s membrane potential (Figure 10). The direct source of odor-elicited excitatory drive to IG is unknown, but could, in principle, be traced to KCs. Indeed, a version of our model in which all KCs synapse upon IG reliably reproduced fairly realistic odor-elicited hyperpolarizations in GGN. This result leads us to hypothesize that KCs, which require strong and coincident input from PNs to spike (Perez-Orive et al., 2002; Jortner et al., 2007) may serve as thresholding elements in an odor pathway linking the antennal lobe to IG, transferring odor-evoked activity, but not lower levels of spontaneous activity, from PNs. The identity of this proposed odor pathway is not known.

Updated olfactory connectivity model where odor responses in the KC population are gated by feedback inhibition from GGN and inhibition of GGN via odor evoked spiking in IG, which itself receives input from an unknown olfactory pathway.

Why are hyperpolarizations of GGN elicited by only some odors in some animals, as we saw in vivo (Figure 4)? Our simulations suggest that odor-specific patterns of spiking across the PN population (Stopfer et al., 2003) may underlie this variability (Figure 9b–i). Another possibility, not tested here, is that different sets of KCs could drive GGN and IG differentially such that odor-driven KCs strongly activate IG, but only weakly activate GGN, causing GGN to hyperpolarize. Different classes of KCs have been reported to differentially activate APL in Drosophila (Inada et al., 2017), but this possibility remains unexplored in locust.

Our results indirectly suggest IG is active in the absence of odor stimuli (Figures 8 and 9). IG may generate this spontaneous activity on its own or may inherit it from other olfactory or non-olfactory neurons. KCs are nearly silent in the absence of odor stimuli, making them unlikely direct sources of this activity, but it is possible that even extremely sparse spontaneous activity distributed over the population of 50,000 KCs could suffice to drive spontaneous spiking in IG. This possibility would be difficult to test in vivo.

Several caveats apply to predictions of our model. Although we explored a wide range of values for GGN’s membrane and cytoplasmic conductances to test our model for robustness (Figure 2g), we cannot exclude the possibility that ion channels are distributed unevenly across GGN, potentially allowing parts of the neuron to behave in ways we could not observe in our recordings. We simplified our models of KCs as single compartments, and in some cases collapsed multiple synapses into a single equivalent synapse; however, these simplifications should not affect our predictions in a qualitative sense. Finally, our model excludes plasticity mechanisms, so our predictions apply best to odors that are novel or delivered at long intervals.

In sum, we used biophysically detailed simulations in combination with electrophysiology performed in vivo to explore the olfactory circuit of the locust mushroom body. Our intracellular and patch recordings revealed new features of GGN, a neuron that plays a central role in shaping olfactory responses, and of the KC population. These results extend our understanding of the olfactory system, highlighting ways different components interact, and provide new constraints and predictions for additional research.

Materials and methods

Key resources table
Reagent type (species) or resourceDesignationSource or referenceIdentifiersAdditional information
Software, algorithmNEURONhttps://neuron.yale.edu/neuron/RRID:SCR_005393Versions 7.4 and 7.5 with Python 2.7
Software, algorithmNeuroLucida 360https://www.mbfbioscience.com/neurolucida360RRID:SCR_016788
Chemical compound, drugNeurobiotinVector LaboratoriesSP-11201–5% in 2–3M potassium acetate
Chemical compound, drugAvidin-Alexa complexThermo Fisher ScientificS112260.01 mg/ml in 1X PBS

Dissection and electrophysiology

Request a detailed protocol

Newly eclosed adult locusts of both sexes picked randomly from our crowded colony were immobilized and the brain was exposed, desheathed and superfused with locust saline as described before (Brown et al., 2005). Our dissection included cutting the labral nerves and removing the suboesophageal ganglion (SOG), precluding stimulation via olfactory receptors in the palps. We usually obtained recordings from one GGN per animal and observed no systematic variation in results over the course of about a year. GGN is not visible from the brain’s surface. To record from GGN, a sharp glass micropipette filled with 2 or 3M potassium acetate with 5% neurobiotin was inserted into the peduncle region of the mushroom body; when impaled, GGN could be identified by its characteristic pattern of IPSPs in the voltage record (Figure 9j). At the end of the recording session, neurobiotin was injected into the cell iontophoretically using 0.2 nA current pulses at 3 Hz for 10 to 20 min, allowing the cell’s identity to be confirmed by subsequent imaging.

To test whether IPSPs in GGN originate from spontaneous activity in PNs, we first silenced the PNs by cutting both antennal nerves at the base (Joseph et al., 2012) and then searched for GGN, which could be identified by its ongoing unique pattern of IPSPs and confirmed by its morphology, revealed by subsequent filling with neurobiotin and imaging.

For patch clamp recordings from KCs, the initial dissection was performed as described above. Patch pipettes were pulled to between 7 and 12 MΩ, filled with locust internal solution (Laurent et al., 1993) as well as a neural tracer for subsequent histology (either 12 mM neurobiotin for later conjugation with an Avidin-Alexa complex, or 20 μM Alexa Fluor tracer with absorption wavelengths of 488, 568 or 633). Patch recordings were made in current clamp mode, and data were only analyzed if the observed membrane potential was within the previously reported range for KCs (−55 to −65 mV) and if either LFP or membrane potential oscillations were observed in response to odor stimulation. Firing rates were obtained by smoothing the PSTH with a 100 ms SD Gaussian window.

Stimulus delivery

Request a detailed protocol

For GGN recordings, odor pulses were delivered to the ipsilateral antenna as described in Gupta and Stopfer (2012). The odorants included 1-hexanol at the dilution of 1% v/v, 1-hexanol, hexanal, methyl benzoate, benzaldehyde, and cyclohexanone mixed at the dilution of 10% v/v in mineral oil, and 100% mineral oil.

For KC recordings, the following odors were delivered similarly: 1-hexanol, hexanal, cyclohexanol, 1-octananol, citral, geraniol, ethyl butyrate, 1-butanol, benzaldehyde, eugenol, 3-methyl-2-butenol, methyl jasmonate, decanal, methyl salicylate, linalool, limonene, pentyl acetate, (all 10% v/v in mineral oil, except methyl salicylate and limonene, at 40% each) and 100% mineral oil.

Histology and immunostaining

Request a detailed protocol

Brains were dissected from the head capsule and fixed in 4% paraformaldehyde overnight, then conjugated with Avidin-Alexa 568 or Avidin-Alexa 488. Some brains (Figure 1a) were first immunostained with mouse nc82 primary antibody (DSHB Cat# nc82, RRID:AB_2314866, deposited to the DSHB by Buchner, E.) and Alexa 568 conjugated anti-mouse IgG secondary antibody (Shimizu and Stopfer, 2017).

Imaging and neuronal tracing

Request a detailed protocol

Brains were dehydrated in an ethanol series and cleared with methyl salicylate or CUBIC (Susaki et al., 2014), mounted in methyl salicylate or mineral oil respectively, and imaged with a Zeiss LSM 710 confocal microscope. Two GGNs were traced in detail from 3D image stacks using NeuroLucida 360 software (MBF Bioscience, Williston, Vermont). The traces were converted to SWC format for further processing and cleanup using NLMorphologyConverter (www.neuronland.org). The two traces were very similar, and models based on both gave similar results. Here, we show results obtained from the GGN imaged at higher resolution.

Statistics

For each of 198 GGN-odor pairs, average IPSP rates were calculated over five trials in windows beginning 2 s before, and 2 s after, odor presentations. Wilcoxon-signed rank tests from the scipy package were used to compute the statistic and the two-sided p-value.

Computational model

Request a detailed protocol

GGN morphologies in SWC format were imported into NEURON and converted into passive models in NEURON’s hoc format. The resting membrane potential was set to −51 mV, as observed in vivo. The single compartmental KC model reported by Wüstenberg et al. (2004) was translated manually into a NEURON model. This Hodgkin-Huxley type model included fast and slow Na+ channels, and delayed rectifier, transient A type, and slow transient outward K+ channels. The passive reversal potential of the KCs was set to −70 mV. Custom Python scripts were written to set up network models and simulation experiments using NEURON’s Python interface. The simulations were run on NIH’s Biowulf supercomputer cluster (http://hpc.nih.gov) and simulation results were saved in HDF5 based NSDF format (Ray et al., 2016) and later analyzed with custom Python scripts.

PN activity model with a fixed responsive population

Request a detailed protocol

Modeled PN spike train rates were based on firing statistics reported in Figure 2c–e in Mazor and Laurent (2005) from which we infer 77% of PNs are spontaneously active. Odor presentations were set to evoke spiking in ~20% of PNs, each spiking at an average rate of 20 Hz (Mazor and Laurent, 2005; Jortner et al., 2007). A 20 Hz sinusoid with amplitude 0.4 times the average spiking rate was further superimposed on odor-elicited spiking to model oscillatory activity generated in the antennal lobe. Based on our own observations in vivo we set 10% of spontaneously active PNs to respond with inhibition to odor presentations. We used a non-homogeneous Poisson generator to create the spike trains based on these rates.

PN activity model with a shifting responsive population

Request a detailed protocol

To test the significance of varying temporal structure in PN firing patterns, we assumed ~30% of the PNs were unresponsive to a given odor, and divided the other 70% into four equally sized groups with the following odor-elicited sequences of excitation (E) and inhibition (I): EEI, EIE, IEI and IIE, with the last epoch occurring upon stimulus offset (Laurent and Davidowitz, 1994a). Within each group, excitation epochs featured shifting sets of active PNs, with new PNs recruited during each LFP cycle (Mazor and Laurent, 2005). Each group started its excitatory epoch with activation of 70% of its members and 10% were recruited in each of the next three LFP cycles. In this scheme at most ~30% of PNs were active at any given time during odor presentations.

Connectivity from GGN to KCs

Request a detailed protocol

Each KC received one graded inhibitory synaptic input from a randomly assigned point on GGN’s calyceal branches. The strength of each synapse was adjusted to keep the KC’s membrane potential close to −60 mV when bombarded by spontaneous activity from PNs, as observed in vivo (Joseph et al., 2012). The graded synapse was modeled as a NEURON mechanism based on published descriptions (Manor et al., 1997; Papadopoulou et al., 2011). In some simulations the individual synaptic conductances were selected from a lognormal distribution with the mean adjusted to produce realistic KC activity, that is to evoke spiking in about 10% of the KCs with most of them responding with fewer than five spikes to odor stimuli. The standard deviation of the lognormal distribution had the same value as the mean, based roughly on Song et al. (2005). In network simulations we considered KC activity to be realistic if ~10% or fewer KCs spiked in response to odor stimulus. Also, we considered KC activity to be unrealistically low if it failed to depolarize GGN.

Connectivity from PNs to KCs

Request a detailed protocol

Half of the PN population was randomly and independently selected as presynaptic partners for each KC. If two subsets of size m and n are randomly and independently selected from a set of size q, the expected size of their intersection is s = m * n/q. Thus, with 800 PNs and each KC receiving input from 400 PNs, the expected number of PNs shared by any two KCs is 400 * 400/800 = 200, that is they share about 50% of their presynaptic PNs, as shown in vivo (Jortner et al., 2007). We carried out a parameter search to obtain a preliminary estimate of a PN→KC synaptic conductance sufficient to evoke a few spikes in a single KC model when driven with the expected combined PN firing rate during odor stimuli. Later, in the full network model, we tested values around this conductance along with different GGN→KC conductances. For heterogeneous conductance models, the conductances of individual synapses were selected from a lognormal distribution with the mean adjusted to produce realistic KC activity, as described above.

To test whether diverse synaptic delays could alleviate synchronized spiking in KC population, we created a set of networks where PN→KC synapses were assigned a range of delays from a normal distribution with mean 7.5 ms and standard deviation 3 ms. We chose these values with the animal’s physiology in mind, by assuming a conservative speed of 0.2 m/s = 200 μm/ms for signal propagation through dendrites (Larkum et al., 1996). The length of PN axon is about 1500 μm, so the total time delay from PN to KCs should be under 1500/200 = 7.5 ms, consistent with the delay of 6 ms observed in vivo by Jortner et al. (2007). The maximum diameter of a large KC dendritic arbor is about 600 μm in the calyx and thus, the longest delay between two KCs receiving spikes from the same PN in their soma should be about 3 ms.

IG model and connectivity

Request a detailed protocol

To simulate IG, we used a single compartmental Izhikevich-type model of a regular spiking (RS) pyramidal cell from the Model DB repository (https://github.com/ModelDBRepository/39948) modified to include graded synaptic input from GGN. To model IG’s spontaneous firing, sufficient current was injected into it to generate ~7 spikes/s. A single inhibitory synapse with −80 mV reversal potential connected IG to one of GGN’s basal dendrite segments. The strength and time constant for this synapse were adjusted to produce IPSP amplitudes matching those observed in vivo. The same GGN segment was connected to IG via a graded synapse. In some simulations either the output of all PNs or the output of all KCs were connected to IG via excitatory synapses. The synaptic weights from KCs to IG were selected from a lognormal distribution.

Data analysis

Request a detailed protocol

Most analyses and 3D visualizations were carried out with custom Python scripts using published modules including numpy, scipy, networkx, matplotlib, h5py, pandas and scikitslearn.

Analyses of patch clamp recordings from KCs were carried out with custom MATLAB scripts.

References

  1. 1
  2. 2
  3. 3
  4. 4
  5. 5
  6. 6
    Local interneurones in insects
    1. M Burrows
    (1981)
    In: A Roberts, B. M Bush, editors. Neurons Without Impulses. Cambridge University Press. pp. 199–221.
  7. 7
  8. 8
  9. 9
  10. 10
  11. 11
  12. 12
  13. 13
  14. 14
  15. 15
  16. 16
  17. 17
  18. 18
  19. 19
  20. 20
    The electrical constants of a crustacean nerve fibre
    1. AL Hodgkin
    2. WAH Rushton
    (1946)
    Proceedings of the Royal Society of London. Series B - Biological Sciences 133:444–479.
    https://doi.org/10.1098/rspb.1946.0024
  21. 21
  22. 22
  23. 23
  24. 24
  25. 25
  26. 26
  27. 27
  28. 28
  29. 29
  30. 30
  31. 31
  32. 32
  33. 33
  34. 34
  35. 35
  36. 36
  37. 37
  38. 38
  39. 39
  40. 40
  41. 41
  42. 42
  43. 43
  44. 44
  45. 45
  46. 46
  47. 47
  48. 48
  49. 49
    Theoretical significance of dendritic trees for neuronal input-output relations
    1. W Rall
    (1964)
    In: R. F Reiss, editors. Neural Theory and Modeling. Stanford: Stanford University Press. pp. 73–97.
  50. 50
  51. 51
  52. 52
  53. 53
  54. 54
  55. 55
  56. 56
  57. 57
  58. 58
  59. 59
    Giant fiber associated with mushroom-body neuropiles in an insect brain a possible one-cell feedback loop
    1. MJ Weiss
    (1978)
    Society for Neuroscience.
  60. 60
  61. 61

Decision letter

  1. Kristin Scott
    Reviewing Editor; University of California, Berkeley, United States
  2. Ronald L Calabrese
    Senior Editor; Emory University, United States

In the interests of transparency, eLife publishes the most substantive revision requests and the accompanying author responses.

Acceptance summary:

This study uses a combination of electrophysiological recordings and modeling to examine how a wide-field inhibitory neuron contributes to the dynamics of odor responses. The work argues that heterogeneous excitation from a direct olfactory pathway as well as odor-driven inhibition both sculpt the response properties of the wide-field inhibitory neuron and contribute to odor coding. The combination of modeling and recordings provides new insights and testable hypotheses about how inhibition may expand the dynamic range of odor detection.

Decision letter after peer review:

[Editors’ note: the authors submitted for reconsideration following the decision after peer review. What follows is the decision letter after the first round of review.]

Thank you for submitting your work entitled "Feedback inhibition and its control in an insect olfactory circuit" for consideration by eLife. Your article has been reviewed by three peer reviewers, and the evaluation has been overseen by Kristin Scott as Reviewing Editor and a Senior Editor. The reviewers have opted to remain anonymous.

Our decision has been reached after consultation between the reviewers. Based on these discussions and the individual reviews below, we regret to inform you that your work will not be considered further for publication in eLife at this time. The reviewers were positive about the significance of the study but had several concerns, detailed in individual reviews below. While many of the concerns may be addressed with changes to the text, one experiment regarding testing the model in Figure 3 seems likely to require additional time to complete beyond a two month time window allowed for revision. However, the reviewers do feel the work could be made acceptable after the reviewers' concerns are addressed. Should you choose to submit a new version of this work, we will endeavor to have it reviewed by the same BRE member and reviewers.

Reviewer #1:

Overall I like this combination of modelling and experiments – it covers an important topic and reaches interesting conclusions. I have some concerns that the authors should address before publication.

Ray et al. construct a morphological model of the locust's giant GABAergic neuron (GGN) based on its reconstructed anatomy, and use it to model the spread of depolarization within GGN. The model predicts that although GGN is very large and doesn't spike, signals should still spread readily from one part to the other, though they get attenuated. Although we don't know if this prediction is correct because the authors don't actually measure the spread of depolarization experimentally (presumably because this is technically impossible), the model itself is a valuable advance. In particular, the model generated an interesting prediction – that GGN can only show the experimentally measured sustained odor-evoked depolarization if some KCs fire at high rates due to more intense/prolonged excitatory input from PNs – and the authors test the prediction experimentally, finding that a small previously overlooked fraction of KCs do in fact fire at high rates. Finally, the authors analyze the timing of IPSPs in GGN to infer that the "inhibitor of GGN" must be directly excited by odors (via KCs or PNs). While this KC->IG or PN->IG excitation was hypothesized in the original description of GGN (Papadopoulou et al), the timing of IPSPs in GGN provides novel evidence for this connectivity.

1) I am less convinced by Figure 3. This is certainly an interesting prediction of the model. But unlike Figures 4-6, which culminate in experimental verification of the prediction, here there is no experimental evidence that KCs dynamic range really is improved by inhibition. It's a prediction based on a model without experimental validation. While the model itself is useful, I feel that Figure 3 might be "a bridge too far". Could they experimentally test this model ? They could repeat Figure 3C of Papadopoulou et al., 2011 with a twist: use field stimulation to activate many KCs, while injecting varying amounts of current into one KC. The field stimulation should expand the dynamic range of the single KC that they inject current into, i.e. its spike rate as a function of current injected should have a shallower slope. Unfortunately, this would not test one of the weird results of the model, where the feedback inhibition prevents spiking at 14-16 pA even though the KC is not spiking and therefore can't trigger feedback inhibition, because this experiment isn't really feedback inhibition as defined in Figure 3, because the KC being measured doesn't affect its own activity at all (it's only being affected by the mass of other KCs). But, this can still be a useful result, that inhibition onto a KC expands KCs' dynamic range, and they could change their model to match the experiment. (The difficulty of testing their model directly as currently presented is itself a reason to change the model.) Plus, an increased dynamic range would suggest that GGN acts as "input gain control" rather than "response gain control" (Figure 3, Olsen et al., 2010 Neuron).

2) I also don't understand what happens in Figure 3B in the KC with GGN feedback at 14-16 pA. Presumably it doesn't spike (compared to the KC without GGN) because it's getting inhibition. But how can GGN be activated if the KC isn't spiking? Doesn't the KC need to spike at least once? If the model GGN releases GABA spontaneously, they should show that spontaneous inhibition isn't enough for increasing the dynamic range and you need odor-evoked activity in GGN.

3) Figure 5-6: Can the authors quantitatively compare the experimentally measured distribution of KC spike rates (6e) with the distribution in the model (5h)? How well do they match? And if they force the KC spike rates in the model to match the experimentally measured spike rates, does it reproduce sustained GGN odor responses? This is important to show that their experimental data match the computational prediction.

4) How can the authors be sure that their two outlier KCs in Figure 6E weren't weird experimental flukes? (Where they even 2 cells? maybe 2 recordings from the same cell?)

5) Could the synchronization of KC spikes in Figure 5C be alleviated by introducing variability in timing of KC inputs? For example, in Figure 7, KC->IG synapses have a 200 ms delay. Could KC inhibitory inputs also be delayed in some KCs more than others (more so than caused by the delay of passive spread of activity within GGN), eg based on the shape of KC dendrites? At least, ruling out this possibility is important for the implicit logic "they only way to get sustained GGN responses is if some KCs spike at high rates".

6) Figure 5F,G, 7G – I'd like to see the KC population spiking in these models (as in 5c). Presumably in Figure 5G the KCs won't be quite as synchronized as in 5c; and in 7g the KC spikes should precede the period of IG activity by 200 ms.

7) Figure 2—figure supplement 2, final paragraph of subsection “Signal attenuation in GGN” – does each KC get inhibition from exactly one GGN segment? That seems unrealistic. Also can the authors put error bars on # of spikes per KC?

8) Final paragraph of subsection “Signal attenuation in GGN”: “data not shown” – this seems like an interesting conclusion – can the authors make a figure of this?

9) “Earlier computational models(Kee et al., 2015; Papadopoulou et al., 2011) did not reproduce sustained responses in GGN” – this seems unfair to Kee et al., whose Figure 2D looks quite similar to Animal 1 in this manuscript's Figure 4, or the simulation in Figure 7G.

10) Materials and methods: “the synaptic conductances were selected from a lognormal distribution with the mean adjusted to produce realistic KC activity” Can the authors elaborate on this? Was the KC spiking threshold set to a particular value? What is “realistic KC activity”? How did they set the variance of the lognormal distribution?

Reviewer #2:

This manuscript describes a model of the locust olfactory system that is accompanied by electrophysiological recordings from the giant GABAergic neuron and the Kenyon cells. The model is constructed to test the role of the GGN in shaping KC output. The authors construct a model and then go through a series of steps to validate the model. Model validation was accomplished through observational recordings of the GGN or the KCs during olfactory stimulation.

As written, the manuscript reads like an unfolding story, where the authors create the model, test it, and then find it does not reproduce what is observed in the olfactory system. At each step, they alter a parameter of the model (e.g. synaptic strength of the KCs) to identify the contribution of this parameter to the output of the modelled neurons. This is informative, but I also found this style difficult to read. This style also makes the manuscript feel like it is not hypothesis driven. There are portions of the manuscript where the paragraphs ramble; the manuscript would be strengthened if the authors restructure these paragraphs. For example: they present a paragraph about how they had to adjust the model to include heterogeneity in synaptic strength and PN activation (subsection “GGN responses suggest some KCs fire at high rates”). But the reader doesn't find this conclusion until the end of the paragraph, and few details are reported as data in the paper. The paragraph should start out with the conclusion or a hypothesis and then present supporting evidence. The way it is written, the rationale feels ad hoc.

One of the main conclusions they claim from the model is that they discovered that the spike rates of some of the KCs was relatively high. (Introduction). As far as I could tell within the manuscript, the authors only tested a model in which the KC spike rates were all active or a model in which 1/5 of the neurons were inactive to show that when some of the KCs are silent and the synapse strength is heterologous, some KCs fire more than others. We know that the pattern of KC responses is odor-specific. It would have been good to test the model with different patterns of KC input (and presumably different proportions of inactive KCs) – as if stimulating with different odors – to show whether this outcome depended on the proportion of inactive KCs. A test of different synaptic strengths across the KCs would also have made this portion of the manuscript more convincing.

I thought the portion of this manuscript was the data/modelling of the IG neuron and its impact on the GGN was interesting, but I'm concerned about their conclusion that they have evidence for another olfactory pathway in the brain. The authors are forced to assume that the data of Papadapoulou et al. is correct and that spikes in GGN directly correlate with spikes in IG because they could not record from the IG neuron. However, when they cut the antennal nerve, they find that the GGN continues to spike – as normal – ruling out input from PNs or KCs to the IG. They conclude that the spiking in the IG is intrinsic or that it is generated from another source. What other olfactory source could it be, if it wasn't from the antenna….? Confusingly, later in the manuscript, they conclude that their model predicted that spiking in the IG could be driven by the KCs (Discussion paragraph seven). The contradictions in rationale need to be addressed and clarified.

I thought the discussion could be strengthened if the authors took more space in the discussion to compare their data/model to the recent paper on Drosophila from Inada et al., 2017.

Reviewer #3:

This study explores the function of the Giant GABAergic Neuron (GGN) in the locust using a combination of neural recordings and computational models. The authors conclude that the GGN contributes to an increase in the dynamic range of Kenyon Cells (KCs; the principle cells of the MB) and suggest the existence of a direct olfactory input pathway onto IG, an inhibitory neuron that directly synapses onto GGN. They report some novel functional properties of GGN (e.g. suppressive responses to odor), provide additional support for some known properties of GGN, and formulate some concrete hypotheses from their model that can be tested experimentally. Overall, this manuscript offers some additional detail about the function of this neuron within its network and some new testable hypotheses. My main concerns are twofold. (1) some of their findings are presented as novel when in fact they have already been shown before. These results need to be presented more clearly within the context of the existing literature. (2) some of their conclusions are not well-supported by the results that they provide. I detail these issues below.

1) Signal attenuation in GGN: The author's statement "Together, these features made it unclear whether this giant neuron has the biophysical capacity to perform its suggested function of carrying effective signals passively from the α lobe to distant points in the calyx" appears to not consider previous studies of this neuron that clearly show this capacity experimentally (Papadopoulou et al., 2011, Figure 3C), whereby activation of KCs is sufficient to propagate through the GGN to suppress KC spiking. Their anatomical characterization and compartmental model of GGN is a useful resource, but their conceptual findings should be more thoroughly contextualized in the existing knowledge of this neuron. Similarly, the authors should be more clear about the alternative hypotheses of GGN signal propagation their model is intended to disambiguate.

2) Subsection "Feedback inhibition expands the dynamic range of KCs". Given that this conclusion is entirely based on their model, this assertion should be changed to reflect that it is a strong and testable hypothesis, rather than an explicit test of that hypothesis. Moreover, the authors should better justify their model that uses a single KC with 50,000 synapses, rather than a more realistic model with 50,000 KCs. This seems like it would drive especially strong and synchronized (i.e. non-physiological) activation of GGN. Alternatively, since they subsequently enhance their model to include a larger complement of KCs, it would be relatively straightforward to consider that model here as well.

I am also a bit confused by the model of KC-GGN feedback to illustrate the capacity for expanding the dynamic range of KCs. The threshold for evoking spikes is increased when the KC is reciprocally connected to the GGN. The authors call this "feedback inhibition." But if KCs only influence the GGN via spikes (rather than graded release), then the GGN is not being activated by the KC with injected current ranging from 13-16pA (Figure 3B right). The only explanation I can see is that spontaneous activity from GGN is inhibiting the KC. However, if this is the case, then it's not really "feedback" inhibition, but just inhibition. If the authors want to claim that feedback inhibition is responsible for expanding the dynamic range of KCs, then they should determine what aspect of GGN activity is actually originating from the KC and not spontaneous activity. Because they have a detailed computational model, this analysis should be straightforward. Without it, their conclusion that feedback inhibition expands the dynamic range of KCs is not well supported by their model.

3) The finding that some KCs fire at high rates (subsection “GGN responses suggest some KCs fire at high rates”) also does not fully consider previous results. For instance, Perez-Orive et al., 2002, show that some KC-odor pairs (e.g. their Figure 1B and 2A) can elicit relatively high rates, as high as 16spikes/sec. Thus their predictions and data are entirely consistent with previously reported data, and their text should be updated to reflect this point, so that readers are aware of the relevance of existing literature. Perhaps more interestingly, I think that the authors have an excellent opportunity to test the function of these high-rate responses in the KC-GGN circuit. With their existing model, they could easily omit those high-rate KCs and observe the effects on the GGN. Are these particular responses critical for GGN operation? Do they serve any particular function? This would be relatively easy to explore and could unveil an unappreciated importance of these neurons.

Subsection “Odor evoked spiking in IG can explain GGN hyperpolarization”: Additional intuition for why the simple model of reciprocal connectivity between IG and GGN needs to be provided. The authors state that "a broad range of IG properties" was tested, but since this analysis is not shown, the reader cannot evaluate the capabilities of IG in driving odor-evoked hyperpolarization in GGN.

Final paragraph of subsection “Odor evoked spiking in IG can explain GGN hyperpolarization”: They authors state they "completely" silenced PNs and KCs by cutting the antennal nerves. Since locusts also have maxillary palps which contain ORNs that innervate the antennal lobe, I do not understand how cutting the antennal nerve silences all PNs and KCs. In principle, excitation from the palp could still provide enough excitation in PNs and KCs to drive spontaneous IPSPs in GGN. In order to make this claim, the authors need to additionally cut the maxillary nerve.

[Editors’ note: the decision after re-review follows.]

Thank you for resubmitting your work entitled "Feedback inhibition and its control in an insect olfactory circuit" for further consideration by eLife. Your revised article has been evaluated by Ronald Calabrese (Senior Editor), Kristin Scott (Reviewing Editor) and the three original reviewers.

The manuscript has been improved but there are some remaining issues that need to be addressed before acceptance, as outlined below:

1) Signal attenuation in GGN. The authors' justification in their response for why it is uncertain how signal may be propagating between KCs and GGN (based on prior data from Papadopoulou et al., for instance) makes sense to me. However, the text in the manuscript still is a bit sparse. I would suggest incorporating some of this explanation into the manuscript so that readers will understand the nuances of interpreting the results of the authors' model within the context of existing experimental data.

2) Experimental test of dynamic range expansion (Figure 3E-G). It is very difficult to assess the experimental data underlying the newly added results to test their theoretical finding that GGN expands the dynamic range of KCs. This is largely due to the fact that the experimental data are presented in such a highly processed form (Figure 3G) that it is impossible to assess directly. Ideally, more than two current levels would be injected into the KC (something more akin to the simulation in 3C). At the very least, the authors should show an example of the KC voltage traces (already measured) that went into Figure 3G in the main Figure 3 (not supplement) and make the raw data accessible as a Supplementary file. Given the details of the predictions in 3B, it is essential to show whether the texture of the experimental data reflects that of their model, to enable the reader to see similarities and any differences. Simply showing some of the voltage traces that they already have would help tremendously.

3) How did the authors determine the "high" and "low" amounts of current injected into GGN – were those amounts directly proportional to the "high" and "low" amounts of current injected into the KC?

4) Figure 3E, the color scheme could be clearer – perhaps the "A" pulses should be teal and the "B" pulses orange? And in 3G it would be clearer not to re-use colors that were used to indicate different models (eg purple and orange)

5) In Figure 3—figure supplement 1 I suggest using the same y-axis scale for 14 pA injection as for the other rows.

6) Discussion: "Given stronger odor stimuli (or in the absence of local KC-GGN reciprocal connections in the calyx), signals originating in the α lobe would produce inhibitory output throughout GGN's arbor, dominate local inhibition, and allow KCs receiving inhibition from the same region of GGN to fire together, reducing the contrast among their responses."

I had a hard time understanding the last part of this sentence. Why would global inhibition from GGN make KCs fire together – is it because KCs could only fire in time windows when GGN activity is low? If this only happens for KCs receiving inhibition from the same region of GGN, how is this different from local inhibition? And by "reducing the contrast among their responses" do they mean that global GGN inhibition is actually bad (for separating KC odor responses)? Perhaps the authors could clarify what they mean here.

[Editors' note: further revisions were suggested prior to acceptance, as described below.]

Thank you for resubmitting your work entitled "Feedback inhibition and its control in an insect olfactory circuit" for further consideration by eLife. Your revised article has been evaluated by Ronald Calabrese (Senior Editor), Kristin Scott (Reviewing Editor) and two of the original reviewers.

The manuscript has been improved but there are some remaining issues that need to be addressed before acceptance. As experimental evidence for dynamic range expansion is no longer included, the manuscript must be revised to discuss the model's predictions and limitations, as outlined below:

1) The authors have not demonstrated that "feedback inhibition expands the dynamic range of KCs" in the locust MB. Their conclusion about the dynamic range implies that they have experimental evidence, even though they are no longer including it. They need to make clear that their model predicts the dynamic range result, rather that stating definitively that feedback increases dynamic range.

2) As the authors now conclude with the model, I am concerned that they have provided neither a discussion of the potential caveats to interpretation of the model's predictions, nor any analysis of the robustness of their model's predictions to variations in parameters. An expanded discussion of the model's predictions and limitations would be valuable.

3) Abstract: "Our in vivo recordings and simulations of our model suggest that depolarization of GGN at its input branch can globally inhibit KCs several hundred microns away." This finding comes entirely from the model, not recordings.

4) Discussion paragraph five: they should change "revealed" to "predicts".

https://doi.org/10.7554/eLife.53281.sa1

Author response

[Editors’ note: the authors submitted for reconsideration following the decision after peer review. What follows is the decision letter after the first round of review.]

Reviewer #1:

Overall I like this combination of modelling and experiments – it covers an important topic and reaches interesting conclusions. I have some concerns that the authors should address before publication.

Ray et al. construct a morphological model of the locust's giant GABAergic neuron (GGN) based on its reconstructed anatomy, and use it to model the spread of depolarization within GGN. The model predicts that although GGN is very large and doesn't spike, signals should still spread readily from one part to the other, though they get attenuated. Although we don't know if this prediction is correct because the authors don't actually measure the spread of depolarization experimentally (presumably because this is technically impossible), the model itself is a valuable advance. In particular, the model generated an interesting prediction – that GGN can only show the experimentally measured sustained odor-evoked depolarization if some KCs fire at high rates due to more intense/prolonged excitatory input from PNs – and the authors test the prediction experimentally, finding that a small previously overlooked fraction of KCs do in fact fire at high rates. Finally, the authors analyze the timing of IPSPs in GGN to infer that the "inhibitor of GGN" must be directly excited by odors (via KCs or PNs). While this KC->IG or PN->IG excitation was hypothesized in the original description of GGN (Papadopoulou et al), the timing of IPSPs in GGN provides novel evidence for this connectivity.

We thank the reviewer for finding our model “a valuable advance,” and for noting that it generated “an interesting prediction.” The reviewer is correct that testing the spread of depolarization within GGN in vivo proved technically impossible – several approaches we tried gave unclear results. Thus, we decided to use a computational model. We have now performed additional computational experiments, detailed below. Our new simulations extend our earlier conclusions. As we explain below, we also demonstrate high firing rates in KCs are an emergent property of the olfactory network, and contribute modestly to sustain odor-elicited activity in GGN.

1) I am less convinced by Figure 3. This is certainly an interesting prediction of the model. But unlike Figures 4-6, which culminate in experimental verification of the prediction, here there is no experimental evidence that KCs dynamic range really is improved by inhibition. It's a prediction based on a model without experimental validation. While the model itself is useful, I feel that Figure 3 might be "a bridge too far". Could they experimentally test this model ? They could repeat Figure 3C of Papadopoulou et al., 2011 with a twist: use field stimulation to activate many KCs, while injecting varying amounts of current into one KC. The field stimulation should expand the dynamic range of the single KC that they inject current into, i.e. its spike rate as a function of current injected should have a shallower slope. Unfortunately, this would not test one of the weird results of the model, where the feedback inhibition prevents spiking at 14-16 pA even though the KC is not spiking and therefore can't trigger feedback inhibition, because this experiment isn't really feedback inhibition as defined in Figure 3, because the KC being measured doesn't affect its own activity at all (it's only being affected by the mass of other KCs). But, this can still be a useful result, that inhibition onto a KC expands KCs' dynamic range, and they could change their model to match the experiment. (The difficulty of testing their model directly as currently presented is itself a reason to change the model.)

Following the reviewer’s suggestion, we have now performed additional experiments to test our prediction in vivo. As we describe below, the results confirm the prediction made from our model: increased inhibition originating in GGN onto a KC increases the dynamic range of the KC.

We tested the dynamic range of KCs by making intracellular recordings from them in vivo while injecting low or high amplitude current pulses. Simultaneously, we varied GGN-dependent inhibition onto the KCs by depolarizing GGN with intracellular current injections, also of low or high amplitude. Then, we measured the slope of KC firing rate in response to intracellular current stimuli in the KC with and without increased inhibition from GGN. (Details of this procedure are given in the text Materials and methods.)

We counted the number of spikes in each tested KC to obtain its average firing rate (f) in response to low and high current injections over multiple trials, with and without GGN depolarization. Then we computed the slope with respect to current amplitude by dividing the change in firing rate by the difference between the magnitudes of the two current levels (i1 and i2), i.e., [f(1,A) – f(2,A)] / (i1 – i2) for KC-only and [f(1,B) – f(2,B)]/(i1 – i2) for KC+GGN stimulation. Thus, a decrease in the slope upon GGN depolarization indicates an increase in the dynamic range of the KC.

In 5 of 5 experiments (p<.033, binomial test) we observed a decrease in the slope with GGN depolarization, consistent with the prediction from our model (shown in Figures 3C and D). We now provide this new information in Figure 3 E, F and G, and in the updated corresponding Results and Materials and methods sections.

As we now note in the text, our results are consistent with a general principle of Control Theory that negative feedback increases the linear range of a system [see, e.g. Astrom and Murray, “Feedback Systems”, v3.0i, Chapter 2, eqns 2.2 and 2.5. URL: http://www.cds.caltech.edu/~murray/books/AM08/pdf/fbs-principles_30Sep18.pdf]. Although our model is relatively complex with non-linear conductances generating spiking activity in KCs and graded negative feedback from GGN, this general principle seems to extend to this circuit. As can be seen in Figure 3C in our manuscript (and also in Figure 2.2 in Astrom and Murray) a wider dynamic range is accompanied by shallower slope in a plot of response vs input. In our case, the input is the current injected into a KC and the response is its spike rate.

Plus, an increased dynamic range would suggest that GGN acts as "input gain control" rather than "response gain control" (Figure 3, Olsen et al., 2010 Neuron).

We do not feel comfortable making a strong statement of input vs response gain control. The idea of “input gain control” and “response gain control” originated in visual neuroscience (McAdams and Maunsell, 1999; Reynolds and Desimone, 1999; Ayaz and Chance, 2009). The mathematical idea behind this differentiation seems to be to contrast input-output curves of the form

(1) y=a*x/(1+x) … (response gain control)

to those of the form

(2) y=x/(1+a+x) … (input gain control)

where x is input, and y is output or response. Curve (1) reaches the asymptotic value `a` for large `x`, thus the saturation level firing rate will be lower with inhibition (a < 1). Curve (2) reaches the asymptotic value of 1, same as without the factor `a`. In that sense our simulations suggest that GGN indeed modulates input gain, as the KC firing rates reach the same levels at extreme current injection. However, in our in vivo recordings we limited the current injection below saturation level for fear of damaging the cell. It is easy to show mathematically that the slope of both curves (1) and (2) are less than the curve with unity gain: x/(1+x) for 0 < a < 1. Thus, a change in slope alone is insufficient for distinguishing between the two.

We discuss the absence of spiking in 14-16 pA in the next comment.

2) I also don't understand what happens in Figure 3B in the KC with GGN feedback at 14-16 pA. Presumably it doesn't spike (compared to the KC without GGN) because it's getting inhibition. But how can GGN be activated if the KC isn't spiking? Doesn't the KC need to spike at least once? If the model GGN releases GABA spontaneously, they should show that spontaneous inhibition isn't enough for increasing the dynamic range and you need odor-evoked activity in GGN.

The GGN model indeed includes some inhibition in the absence of odorant. This is meant to reflect conditions in vivo. As shown in Figure 9J, GGN’s membrane potential activity changes very little after the antenna is cut, a manipulation known to silence PNs and KCs (Joseph et al., 2012). This observation suggests GGN itself is spontaneously active, or that it receives input from other neurons that are themselves spontaneously active or activated by stimuli other than odors. To clarify our results we have now updated Figures 3B and c with model results showing the effect of spontaneous inhibition from GGN on a single KC when the KC-to-GGN connection is removed. Spontaneous inhibition increases the spiking threshold of the KC, and slightly increases its dynamic range (purple traces). By contrast, the complete feedback circuit increases the dynamic range of the KC much more (orange traces). To make it clear that GGN provides inhibition to KCs even at rest we updated the text as follows:

“As expected, baseline inhibition from spontaneous activity in GGN increased the KC’s threshold for spiking.”

For increased clarity we have also replotted Figure 3C.

3) Figure 5-6: Can the authors quantitatively compare the experimentally measured distribution of KC spike rates (6e) with the distribution in the model (5h)? How well do they match? And if they force the KC spike rates in the model to match the experimentally measured spike rates, does it reproduce sustained GGN odor responses? This is important to show that their experimental data match the computational prediction.

We thank the reviewer for this interesting question that led us to test additional simulations.

Our initial simulations generated distributions of KC spike rates qualitatively similar to distributions we obtained in vivoin that both had long rightward tails. However, standard statistical tests (Mann-Whitney U and Kolmogorov-Smirnov) showed that the in silico and in vivo distributions were not identical. Further modeling in which we simulated many instances of the mushroom body network (using different random number generator seeds such that individual synaptic connections and strengths were randomized but, overall, these features kept the same mean and standard deviation across the networks) yielded KC firing rate distributions that were qualitatively similar to each other, but, as above, statistical tests showed that these distributions were not identical. This result shows that the exact distribution of KC firing rates is a complex emergent property of each instantiation of the network. The result also shows it would be difficult or impossible to force the KC spike rates in our model to exactly match the experimentally measured spike rates. To make this state of affairs clear to readers we have revised the text and have added a new supplementary figure, Figure 6—figure supplement 1, to illustrate more examples of KC spike count distributions and their associated GGN membrane voltages.

4) How can the authors be sure that their two outlier KCs in Figure 6E weren't weird experimental flukes? (Where they even 2 cells? maybe 2 recordings from the same cell?)

The two outliers in Figure 6E are indeed two different Kenyon cells; they were separately visualized and targeted with patch electrodes, filled with dye and reconstructed. We now emphasize this in the legend. We don’t think these results are weird experimental flukes because, as we should have made clear in our manuscript, others have obtained similar results: see Figure 2A in Perez-Orive et al., 2002 and Figures 1, B1 and B2 in Gupta and Stopfer, 2014. We now cite these previous observations in our manuscript: “Although relatively rare KC odor responses with many spikes have been observed before (Gupta and Stopfer, 2014; Perez-Orive et al., 2002), the origin and significance of these responses were unclear.”

5) Could the synchronization of KC spikes in Figure 5C be alleviated by introducing variability in timing of KC inputs? For example, in Figure 7, KC->IG synapses have a 200 ms delay. Could KC inhibitory inputs also be delayed in some KCs more than others (more so than caused by the delay of passive spread of activity within GGN), eg based on the shape of KC dendrites? At least, ruling out this possibility is important for the implicit logic "they only way to get sustained GGN responses is if some KCs spike at high rates".

We thank the reviewer for this interesting suggestion which led us to perform additional modeling experiments. As suggested, we investigated whether variability in synaptic delays to KCs could alleviate their overly-synchronized spiking. Instead of introducing additional delays in the graded synapses from GGN to KCs (which would be difficult to implement), we introduced a set of variable synaptic delays in the PN-to-KC connections and simulated the model with PN input patterns similar to those shown in Figure 5B. However, we found the variations we tested in this arrangement produced overly-synchronized spiking in KCs (see Figure 5—figure supplement 1) and did not generate sustained GGN responses. We want to clarify that having some KCs spike at high rates alone is not enough to generate realistic responses in GGN. For example, we also tried introducing some KCs firing at high rates by setting high PN→KC conductances in simulations that did not include temporal patterning of PN activity. These simulations, however, also yielded KCs whose firing was overly-synchronized, resulting in unrealistic oscillatory peaks in GGN’s membrane potential.

To clarify this we have updated the text in the Results as follows:

“Increasing the diversity of synaptic delays from PNs to KCs did not alleviate overly-synchronized spiking in the KC population if heterogeneity in synaptic strengths and temporal patterns of PN activity were omitted from the model (Figure 5—figure supplement 1).”

We updated the Materials and methods section as follows:

“To test whether diverse synaptic delays could alleviate synchronized spiking in KC population, we simulated a set of networks where PN -> KC synapses were assigned a range of delays from a normal distribution with mean 7.5 ms and standard deviation 3 ms. We chose these values with the animal’s physiology in mind, by assuming a conservative speed of 0.2 m/s = 200 μm/ms for signal propagation through dendrites (Larkum, Rioult, and Luscher, 1996). The length of PN axon is about 1500 µm, so the total time delay from PN to KCs should be under 1500 / 200 = 7.5 ms (also, (Jortner et al., 2007) found a delay of about 6ms). The maximal diameter of a large KC dendritic arbor is about 600µm in the calyx and thus, the longest delay between two KCs receiving spikes from the same PN in their soma should be about 3 ms.”

6) Figure 5F,G, 7G – I'd like to see the KC population spiking in these models (as in 5C). Presumably in Figure 5G the KCs won't be quite as synchronized as in 5C; and in 7G the KC spikes should precede the period of IG activity by 200 ms.

The reviewer is correct in both points: KCs are less synchronized, and KC spikes precede the period of IG activity by 200 ms. We have updated Figure 5 to show KC spike rasters and have moved the KC histogram to Figure 6. (We’ve also corrected a small error in Figure 5E-G: we originally gave off-response duration as 200 ms, now corrected to 500 ms).

Figure 7 is now Figure 8. We redesigned this figure by adding KC spike rasters and splitting off other content to a new Figure 9. Figure 8 now shows GGN’s membrane potential in vivo with IPSPs and the inferred PSTH of IG spikes. New Figure 9 shows the simulated GGN and IG membrane potentials along with PN and KC spike rasters, and the spontaneous IPSPs recorded in vivo from GGNs of animals with intact or cut antennal nerves.

7) Figure 2—figure supplement 2, final paragraph of subsection “Signal attenuation in GGN” – does each KC get inhibition from exactly one GGN segment? That seems unrealistic.

In our model, indeed, each KC receives inhibition from one GGN segment. The number and locations of GGN’s synapses onto individual KCs are unknown, but in the locust, KCs have multiple dendritic branches, so it is likely that each KC receives GGN input from multiple regions of the calyx. In our model we merged all possible GGN-KC synapses into a single synapse per KC as a manageable simplification.

This simplification does not compromise our result that the effect of voltage differences distributed across GGN’s arbor on KC activity is small. With a single synapse per KC, two KCs can receive quite different levels of inhibition if the presynaptic terminals are on GGN neurites carrying different voltages. On the other hand, if a KC receives inhibition at multiple dendritic branches contacting different parts of GGN’s arbor that carry different voltages, then the total effect observed at the spike initiation zone should be a superposition of the effects due to each synapse, reducing the variation between different KCs. As we point out in the next line in the manuscript, this is a small effect either way.

Also, after adding a figure showing the distribution of signal transmission delays across GGN.

Also can the authors put error bars on # of spikes per KC?

Agreed and done.

8) Final paragraph of subsection “Signal attenuation in GGN”: “data not shown” – this seems like an interesting conclusion – can the authors make a figure of this?

Yes, we have now added Figure 2—figure supplement 3 showing Spearman’s rank correlation coefficient for the different parameters influencing KC spiking in three different sets of simulations. (Earlier Supplementary Figure S4 is now Figure 5—figure supplement 2)

9) “Earlier computational models (Kee et al., 2015; Papadopoulou et al., 2011) did not reproduce sustained responses in GGN” – this seems unfair to Kee et al., whose Figure 2D looks quite similar to Animal 1 in this manuscript's Figure 4, or the simulation in Figure 7G.

We agree and have now rephrased this to more specifically describe our observations:

“Earlier computational models (Papadopoulou et al., 2011; Kee et al., 2015) did not attempt to reproduce the variety of GGN responses we observed in vivo, including hyperpolarizations.”

10) Materials and methods: “the synaptic conductances were selected from a lognormal distribution with the mean adjusted to produce realistic KC activity” Can the authors elaborate on this? Was the KC spiking threshold set to a particular value? What is “realistic KC activity”? How did they set the variance of the lognormal distribution?

These are good questions. We have now updated our Materials and methods section with more details. Briefly, in our model KC the spike generation mechanism is implemented as a set of Hodgkin-Huxley (HH) type ionic conductances. In contrast to simpler models such as Integrate and Fire, HH type models do not have an explicit threshold voltage; rather, the threshold emerges from channel dynamics (Izhikevich, 2007, Dynamical Systems in Neuroscience, p 238 ff). So, rather than setting an explicit threshold value we carried out a parameter search to obtain a preliminary estimate of PN→KC synaptic conductances sufficient to evoke a few spikes in a single KC model. Later, in the full network model, we tested values around this conductance along with a range of GGN→KC conductances. In the network simulations we considered KC activity to be realistic if around 10% or fewer neurons in the KC population spiked in response to odor stimulus, as has been documented in vivo. Also, we considered KC activity too low to evoke depolarization in GGN to be unrealistic. In the lognormal distributions for synaptic strengths, the standard deviation was set to (1.0 x mean) to follow roughly Song et al., 2005.

Reviewer #2:

This manuscript describes a model of the locust olfactory system that is accompanied by electrophysiological recordings from the giant GABAergic neuron and the Kenyon cells. The model is constructed to test the role of the GGN in shaping KC output. The authors construct a model and then go through a series of steps to validate the model. Model validation was accomplished through observational recordings of the GGN or the KCs during olfactory stimulation.

As written, the manuscript reads like an unfolding story, where the authors create the model, test it, and then find it does not reproduce what is observed in the olfactory system. At each step, they alter a parameter of the model (e.g. synaptic strength of the KCs) to identify the contribution of this parameter to the output of the modelled neurons. This is informative, but I also found this style difficult to read. This style also makes the manuscript feel like it is not hypothesis driven. There are portions of the manuscript where the paragraphs ramble; the manuscript would be strengthened if the authors restructure these paragraphs. For example: they present a paragraph about how they had to adjust the model to include heterogeneity in synaptic strength and PN activation (subsection “GGN responses suggest some KCs fire at high rates”). But the reader doesn't find this conclusion until the end of the paragraph, and few details are reported as data in the paper. The paragraph should start out with the conclusion or a hypothesis and then present supporting evidence. The way it is written, the rationale feels ad hoc.

We appreciate these suggestions and have revised our text throughout for clarity. Our goals in organizing our manuscript have been to be informative and to make our reasoning as clear as possible. To address the specific example provided by the reviewer, we have changed the title of the first subsection in Results section to “Signals travelling through GGN’s arbor attenuate significantly,” and have updated this section to state our initial hypothesis at the beginning. The rest of the subsections are now titled with their respective conclusions.

One of the main conclusions they claim from the model is that they discovered that the spike rates of some of the KCs was relatively high. (Introduction). As far as I could tell within the manuscript, the authors only tested a model in which the KC spike rates were all active or a model in which 1/5 of the neurons were inactive to show that when some of the KCs are silent and the synapse strength is heterologous, some KCs fire more than others. We know that the pattern of KC responses is odor-specific. It would have been good to test the model with different patterns of KC input (and presumably different proportions of inactive KCs) – as if stimulating with different odors – to show whether this outcome depended on the proportion of inactive KCs. A test of different synaptic strengths across the KCs would also have made this portion of the manuscript more convincing.

Here we think the reviewer may have mistaken results from PNs for results from KCs. We have now revised our text to make this clear.

As we show in Figures 5 (B) and (E), we activated either 1/5th or 2/3rd of PNs, not KCs, to simulate the odor-elicited input ultimately reaching KCs. Although the compressed nature of the figure may give the impression that KC activity in Figures 5 (C), (F) and (H) was dense, our model was constrained such that only about 5000, or 10% of the 50000 KCs, were spiking.

We did indeed simulate different odor responses by changing the temporal patterns of PN activity in the model (see Figure 9), but we realize this was not clear enough from the text and the figures. We have now added spike rasters of the PN population for two different temporal patterns in Figure 9 (B) and (F).

I thought the portion of this manuscript was the data/modelling of the IG neuron and its impact on the GGN was interesting, but I'm concerned about their conclusion that they have evidence for another olfactory pathway in the brain. The authors are forced to assume that the data of Papadapoulou et al. is correct and that spikes in GGN directly correlate with spikes in IG because they could not record from the IG neuron. However, when they cut the antennal nerve, they find that the GGN continues to spike – as normal – ruling out input from PNs or KCs to the IG. They conclude that the spiking in the IG is intrinsic or that it is generated from another source. What other olfactory source could it be, if it wasn't from the antenna….? Confusingly, later in the manuscript, they conclude that their model predicted that spiking in the IG could be driven by the KCs (Discussion paragraph seven). The contradictions in rationale need to be addressed and clarified.

Again, we have revised our text to clarify a point of confusion. The possibility of spontaneous activity in IG does not contradict the hypothesis that KCs excite IG. A cell can be spontaneously active (see Hausser et al., 2004. J. Neurosci.), and this activity can increase following stimulation; for example, ORNs are spontaneously active, but odor presentations increase (or decrease) this activity (Joseph et al., 2012). The drive for spontaneous activity in IG may be intrinsic or it may arise in sensory modalities other than olfaction, such as vision. Here is our line of thought: Because KCs require strong and coincident input from PNs to spike, spontaneous activity in the antennal lobe evokes little activity in KCs (Laurent and Naraghi, 1994; Perez-Orive et al., 2002; Joseph et al., 2012) and thus does not affect spontaneous activity in IG. Odor stimulation, however, increases activity in the antennal lobe, raising the KCs above spiking threshold and generating activity across the KC population, which, in turn, drives an increase in IG spiking.

Another possibility, as pointed out by reviewer # 1 in comment # 13, is that even extremely rare spontaneous activity in KCs, when summed over a population of 50,000 cells, might drive spontaneous spiking in IG. However, as we explain above, we believe KCs are completely silent after antennal nerves are cut (Joseph et al., 2012), but we are unable to test this directly.

We have now updated our text to make this clear: “This result leads us to hypothesize that KCs, which require strong and coincident input from PNs to spike (Perez-Orive et al., 2002; Jortner et al., 2007) may serve as thresholding elements in the odor pathway from the antennal lobe to IG, transferring odor-evoked activity, but not lower levels of spontaneous activity, from PNs.”

And “KCs are nearly silent at rest, making them unlikely direct sources of this activity, but it is possible that even extremely sparse spontaneous activity distributed over the population of 50,000 KCs could suffice to drive spontaneous spiking in IG. This possibility would be difficult to test in vivo.”

I thought the discussion could be strengthened if the authors took more space in the discussion to compare their data/model to the recent paper on Drosophila from Inada et al., 2017.

We agree and have updated this part with a more detailed discussion of Inada et al., 2017 and how our results relate to their findings.

Reviewer #3:

[…] My main concerns are twofold. (1) some of their findings are presented as novel when in fact they have already been shown before. These results need to be presented more clearly within the context of the existing literature. (2) some of their conclusions are not well-supported by the results that they provide. I detail these issues below.

1) Signal attenuation in GGN: The author's statement "Together, these features made it unclear whether this giant neuron has the biophysical capacity to perform its suggested function of carrying effective signals passively from the α lobe to distant points in the calyx" appears to not consider previous studies of this neuron that clearly show this capacity experimentally (Papadopoulou et al. 2011, Figure 3C), whereby activation of KCs is sufficient to propagate through the GGN to suppress KC spiking. Their anatomical characterization and compartmental model of GGN is a useful resource, but their conceptual findings should be more thoroughly contextualized in the existing knowledge of this neuron. Similarly, the authors should be more clear about the alternative hypotheses of GGN signal propagation their model is intended to disambiguate.

Our work considers and fully incorporates earlier findings that GGN effectively inhibits KCs through a feedback mechanism. Papadopoulou et al., 2011 suggested, based on the morphologies of GGN’s neurites, that the neuron receives input from KCs in its α lobe branches and sends its output via its calyceal branches. However, nothing in Papadopoulou et al. excludes the possibility that feedback from GGN to KCs occurs through local interactions within the calyx. As we explain in our text, given the non-spiking nature of GGN, the distance between the α lobe and the calyx, and other factors, we hypothesized that depolarizing signals originating in the α lobe will greatly attenuate by the time they reach the calyx, possibly to the point of being ineffective. If this idea is correct, then the inhibition of KCs by GGN demonstrated by Papadopoulou et al. would have to occur via a different mechanism, such as through reciprocal connections between KCs and GGN within the calyx. Several other lines of evidence are consistent with our idea. Papadopoulou et al. noted that activating GGN synaptically was more effective than activating it through direct current injection in counteracting current-induced spiking of KC. Assuming Papadopoulou et al. impaled GGN in one of its relatively thick branches in the peduncle or the α lobe (these are by far the most successful target regions), their observation is consistent with the idea that effective local feedback may occur within calyx. Further, a recent study found evidence for pre- and post-synaptic terminals within calyceal branches of the APL neuron, the Drosophilahomologue of GGN (Zheng et al., 2018).

Thus, the goal of our simulations was to investigate whether the extent of signal attenuation in GGN would prioritize the idea of local rather than global feedback, a possibility that would raise interesting questions about the spatial structure or lack thereof in KC↔GGN connections. Our model suggested that, although signals originating in GGN’s α lobe branches do attenuate significantly, enough signal remains in the calyx to hyperpolarize KCs there. This finding made further exploration of alternative hypotheses of GGN signal propagation moot.

2) Subsection "Feedback inhibition expands the dynamic range of KCs". Given that this conclusion is entirely based on their model, this assertion should be changed to reflect that it is a strong and testable hypothesis, rather than an explicit test of that hypothesis.

Yes, we agree. We have now tested this prediction in vivo, and the results support our hypothesis. The results are described above in a response to reviewer 1 and in the revised manuscript

Moreover, the authors should better justify their model that uses a single KC with 50,000 synapses, rather than a more realistic model with 50,000 KCs. This seems like it would drive especially strong and synchronized (i.e. non-physiological) activation of GGN. Alternatively, since they subsequently enhance their model to include a larger complement of KCs, it would be relatively straightforward to consider that model here as well.

The goal of our simulations with a single KC was to isolate the specific effect of negative feedback from the complexity generated by the full network. As we describe in the text, we avoided synchronous input by jittering the synaptic delays. To clarify this we have now revised the manuscript as follows:

“Moreover, to avoid unrealistic, strong synchronous input to GGN we jittered the incoming spike times by applying random synaptic delays between 0 and 60ms. Thus, after each spike generated by the model KC, GGN received 50,000 EPSPs distributed over a 60ms time window.”

In addition, we took the reviewer’s suggestion and carried out new simulations. In the full-scale network, the spiking of a single KC would have negligible effect on GGN depolarization. Thus, testing a single KC receiving inhibition with the same time course as in an extended network simulation should closely approximate its response in the full-scale model. Therefore, we simulated in our model a dynamic clamp experiment: we played the GGN membrane potential from a simulation of the extended network to a single KC via a graded inhibitory synapse. We injected a range of currents to this KC in the same way as shown in Figure 3. The results were consistent with our above findings, as shown in a new supplementary figure, Figure 3—figure supplement 1.

I am also a bit confused by the model of KC-GGN feedback to illustrate the capacity for expanding the dynamic range of KCs. The threshold for evoking spikes is increased when the KC is reciprocally connected to the GGN. The authors call this "feedback inhibition." But if KCs only influence the GGN via spikes (rather than graded release), then the GGN is not being activated by the KC with injected current ranging from 13-16pA (Figure 3B right). The only explanation I can see is that spontaneous activity from GGN is inhibiting the KC. However, if this is the case, then it's not really "feedback" inhibition, but just inhibition. If the authors want to claim that feedback inhibition is responsible for expanding the dynamic range of KCs, then they should determine what aspect of GGN activity is actually originating from the KC and not spontaneous activity. Because they have a detailed computational model, this analysis should be straightforward. Without it, their conclusion that feedback inhibition expands the dynamic range of KCs is not well supported by their model.

Reviewer 1 raised this point as well (see comment # 2). As we describe in detail above, reflecting the situation in vivo, the model does indeed include spontaneous inhibition, but this inhibition is greatly enhanced by feedback from KC activity. To clarify this point, we have now run additional simulations, revised our text, and added new figures (Figure 3B, middle panel and updated Figure 3C) showing responses of KCs when they receive only spontaneous inhibition from GGN, with no feedback inhibition. These results provide strong support for our conclusion that feedback inhibition expands the dynamic range of KCs.

3) The finding that some KCs fire at high rates (subsection “GGN responses suggest some KCs fire at high rates”) also does not fully consider previous results. For instance, Perez-Orive et al., 2002 Science, show that some KC-odor pairs (e.g. their Figure 1B and 2A) can elicit relatively high rates, as high as 16spikes/sec. Thus their predictions and data are entirely consistent with previously reported data, and their text should be updated to reflect this point, so that readers are aware of the relevance of existing literature. Perhaps more interestingly, I think that the authors have an excellent opportunity to test the function of these high-rate responses in the KC-GGN circuit. With their existing model, they could easily omit those high-rate KCs and observe the effects on the GGN. Are these particular responses critical for GGN operation? Do they serve any particular function? This would be relatively easy to explore and could unveil an unappreciated importance of these neurons.

We thank the reviewer (and reviewer 1 as well) for highlighting these previous results. We have now revised our text to discuss them.

We have also run new simulations, as suggested by the reviewer, to explore questions about high-firing-rate KCs. We found that these KCs inevitably emerge from activity in the feedback network, and that the identities of high-firing-rate KCs change from odor to odor in any given network. In the simplest model specification, we only set the mean and standard deviation of the distribution from which the synaptic strengths were sampled. Thus, we could reduce the mean strength of PN→KC synapses and/or increase the same for GGN→KC synapses to restrict the entire KC population to fire only small numbers of spikes. In this condition we found the KCs began to spike in extreme, unrealistic synchrony, and failed to produce sustained depolarization of GGN.

We further followed the suggestion of the reviewer and simulated 10 network model instances including the one used in Figure 5I, but after specifically removing the excitatory synapses to all the KCs that had produced more than 5 spikes in the original simulation, thus effectively deleting them from the network. Notably, removing these high-firing-rate KCs did not produce substantial differences in GGN’s response because another set of KCs arose within the network to replace them and became high-spiking. However, the number of high spiking KCs in the modified model was smaller than the original. Therefore, we ran a reiterative series of simulations, removing the KCs that produced more than 5 spikes in the previous model until no remaining KCs fulfilled the removal criterion. As high-spike rate KCs were removed from the network, the simulated GGN response changed from realistic levels of sustained depolarization to widely-separated peaks of activity, suggesting these KCs augment GGN’s depolarizing response. We added these results as new Figure 7. To be clear, we did not explicitly specify these firing rates into our model KCs – rather, as appears to be the case in vivo, they emerged on their own. This convergence of observations made in vivo and, in our model, gives us confidence that our model is reasonably realistic. We have revised our text to describe these new analyses, and have updated the Discussion section to clarify the emergence and roles of high-spike rate KCs.

Subsection “Odor evoked spiking in IG can explain GGN hyperpolarization”: Additional intuition for why the simple model of reciprocal connectivity between IG and GGN needs to be provided. The authors state that "a broad range of IG properties" was tested, but since this analysis is not shown, the reader cannot evaluate the capabilities of IG in driving odor-evoked hyperpolarization in GGN.

We agree and have now updated this section with the following more intuitive explanation of our reasoning:

“We tested a range of synaptic strengths and time constants for GGN→IG, and IG→GGN synapses, and manipulated baseline firing rates of IG, but none of these efforts could generate odor-elicited hyperpolarization in GGN (data not shown). This hinted at a discrepancy in the causal connection between GGN and IG activity. Intuitively, if IG spiking increased solely because inhibition from GGN decreased, then IG spiking could only increase after GGN returned to baseline. However, as evident in Figure 8, sometimes GGN, recorded in vivo, hyperpolarized at the odor’s onset, ruling out the possibility that disinhibition from GGN is the sole driver of spiking in IG.”

Final paragraph of subsection “Odor evoked spiking in IG can explain GGN hyperpolarization”: They authors state they "completely" silenced PNs and KCs by cutting the antennal nerves. Since locusts also have maxillary palps which contain ORNs that innervate the antennal lobe, I do not understand how cutting the antennal nerve silences all PNs and KCs. In principle, excitation from the palp could still provide enough excitation in PNs and KCs to drive spontaneous IPSPs in GGN. In order to make this claim, the authors need to additionally cut the maxillary nerve.

The reviewer is correct, but facts we inadvertently omitted convince us that palp input did not contribute to our results. The locust preparation we used here included cutting the labral nerves and removing the suboesophageal ganglion. In previous work (Joseph et al., 2012) we confirmed that cutting the antenna in this preparation does in fact completely silence PNs and KCs. We have now clarified this in the Materials and methods:

“Our dissection included cutting the labral nerves and removing the suboesophageal ganglion (SOG), precluding stimulation via olfactory receptors in the palps.”

We also made this explicit in the text, now reads:

“… we completely silenced PNs and KCs by bilaterally cutting the animal’s antennal and labral nerves (Joseph et al., 2012) …”

[Editors’ note: the decision after re-review follows.]

The manuscript has been improved but there are some remaining issues that need to be addressed before acceptance, as outlined below:

1) Signal attenuation in GGN. The authors' justification in their response for why it is uncertain how signal may be propagating between KCs and GGN (based on prior data from Papadopoulou et al., for instance) makes sense to me. However, the text in the manuscript still is a bit sparse. I would suggest incorporating some of this explanation into the manuscript so that readers will understand the nuances of interpreting the results of the authors' model within the context of existing experimental data.

We agree and have now brought more of this explanation into the Introduction:

“If signals originating in GGN’s α lobe attenuate to the extent that they cannot effectively hyperpolarize KCs, then GGN must be exerting its influence on KCs through a different mechanism. One such possibility is that GGN could inhibit KCs through local interactions entirely within the mushroom body calyx. Consistent with this idea, synaptically stimulating GGN has been found to be more effective than direct depolarization by intracellular current injection in counteracting current induced spiking of a KC (Papadopoulouet al., 2011). Also, APL, the Drosophilahomologue of GGN, has been found to have both pre- and post-synaptic terminals within its calyceal branches(Zhenget al., 2018). These findings lend support to the hypothesis that GGN may inhibit KCs via local reciprocal connections within the calyx.”

2) Experimental test of dynamic range expansion (Figure 3E-G). It is very difficult to assess the experimental data underlying the newly added results to test their theoretical finding that GGN expands the dynamic range of KCs. This is largely due to the fact that the experimental data are presented in such a highly processed form (Figure 3G) that it is impossible to assess directly. Ideally, more than two current levels would be injected into the KC (something more akin to the simulation in 3c). At the very least, the authors should show an example of the KC voltage traces (already measured) that went into Figure 3G in the main Figure 3 (not supplement) and make the raw data accessible as a Supplementary file. Given the details of the predictions in 3b, it is essential to show whether the texture of the experimental data reflects that of their model, to enable the reader to see similarities and any differences. Simply showing some of the voltage traces that they already have would help tremendously.

Since the last submission of our manuscript, hoping to increase our “N” and obtain a compelling illustration, we performed additional replicates of the in vivo experiment to test KC dynamic range. Unfortunately, the new results have convinced us that we lack the experimental control needed to appropriately test this prediction. In brief, the experiments described by the reviewer require prolonged paired recordings and a long series of intracellular current injections in tiny Kenyon cells and, simultaneously, in the unique neuron GGN. We have learned that, in the context of this experiment, small variations in the membrane potentials of either neuron lead to uninterpretable results. Because we cannot perform the necessary tests, we now conclude this part of the manuscript with a statement of our model’s prediction. We believe this is appropriate for several reasons: it is not unusual for authors of a computational model to conclude their work with predictions; as we point out (here and in the manuscript) our model’s prediction is consistent with basic principles of control theory; and comparable experimental results have been reported in other model systems where more stable preparations were possible (Olsen et al., 2010 Neuron; Inada et al., 2017). We note that this prediction of our model represents only a small part of our manuscript.

3) How did the authors determine the "high" and "low" amounts of current injected into GGN – were those amounts directly proportional to the "high" and "low" amounts of current injected into the KC?

We have now removed this experimental test from our manuscript. Please see our response to comment #2, above.

4) Figure 3E, the color scheme could be clearer – perhaps the "A" pulses should be teal and the "B" pulses orange? And in 3g it would be clearer not to re-use colors that were used to indicate different models (eg purple and orange)

Please see response to comment #2 above. We have now removed this experimental test and corresponding Figures 3E, F and G from our manuscript.

5) In Figure 3—figure supplement 2 I suggest using the same y-axis scale for 14 pA injection as for the other rows.

We agree and have updated the figure as suggested.

6) Discussion: "Given stronger odor stimuli (or in the absence of local KC-GGN reciprocal connections in the calyx), signals originating in the α lobe would produce inhibitory output throughout GGN's arbor, dominate local inhibition, and allow KCs receiving inhibition from the same region of GGN to fire together, reducing the contrast among their responses."

I had a hard time understanding the last part of this sentence. Why would global inhibition from GGN make KCs fire together – is it because KCs could only fire in time windows when GGN activity is low? If this only happens for KCs receiving inhibition from the same region of GGN, how is this different from local inhibition? And by "reducing the contrast among their responses" do they mean that global GGN inhibition is actually bad (for separating KC odor responses)? Perhaps the authors could clarify what they mean here.

We thank the reviewers for helping us clarify our meaning. Our intention was to note that local inhibition could enhance the influence of weak input, not to argue that global inhibition would somehow be bad. We have now rephrased this as follows:

“Assuming randomly-connected input from PNs to KCs(Jortneret al., 2007; Caronet al., 2013; Eichleret al., 2017), we can speculate about possible impact of strong or weak odor stimuli on the circuit. Strong odor stimuli might drive this circuitry into a global, population-wide, winner-take-all scenario in which only KCs receiving the strongest inputs manage to spike. Weak stimuli, on the other hand, might recruit only local inhibition, allowing a distribution of winners to emerge locally, resulting in a center-surround-type of contrast enhancement.”

[Editors' note: further revisions were suggested prior to acceptance, as described below.]

The manuscript has been improved but there are some remaining issues that need to be addressed before acceptance. As experimental evidence for dynamic range expansion is no longer included, the manuscript must be revised to discuss the model's predictions and limitations, as outlined below:

1) The authors have not demonstrated that "feedback inhibition expands the dynamic range of KCs" in the locust MB. Their conclusion about the dynamic range implies that they have experimental evidence, even though they are no longer including it. They need to make clear that their model predicts the dynamic range result, rather that stating definitively that feedback increases dynamic range.

We agree. We want to be perfectly clear that we are offering a prediction about dynamic range expansion and have now revised our manuscript to ensure this. Specifically, we changed the title of this Results section to: “Feedback inhibition is predicted to expand the dynamic range of KCs;” and, as noted below, in the Discussion, we changed “revealed” to “predicts.”

2) As the authors now conclude with the model, I am concerned that they have provided neither a discussion of the potential caveats to interpretation of the model's predictions, nor any analysis of the robustness of their model's predictions to variations in parameters. An expanded discussion of the model's predictions and limitations would be valuable.

We agree and have now expanded the Discussion with the following paragraph about caveats to our model and its predictions:

“Several caveats apply to predictions of our model. Although we explored a wide range of values for GGN’s membrane and cytoplasmic conductances to test our model for robustness (Figure 2G), we cannot exclude the possibility that ion channels are distributed unevenly across GGN, potentially allowing parts of the neuron to behave in ways we could not observe in our recordings. We simplified our models of KCs as single compartments, and in some cases collapsed multiple synapses into a single equivalent synapse; however, these simplifications should not affect our predictions in a qualitative sense. Finally, our model excludes plasticity mechanisms, so our predictions apply best to odors that are novel or delivered at long intervals.”

3) Abstract: "Our in vivo recordings and simulations of our model suggest that depolarization of GGN at its input branch can globally inhibit KCs several hundred microns away." This finding comes entirely from the model, not recordings.

We agree and have updated the Abstract to correct this statement:

“Our simulations suggest that depolarization of GGN at its input branch can globally inhibit KCs several hundred microns away. Our in vivo recordings show that GGN responds to odors with complex temporal patterns of depolarization and hyperpolarization that can vary with odors and across animals, leading our model to predict the existence of a yet-undiscovered olfactory pathway.”

4) Discussion paragraph five: they should change "revealed" to "predicts".

We agree and have updated this sentence accordingly:

“Notably, our model also predicts that feedback inhibition from GGN can expand the dynamic range of inputs able to activate KCs (Figure 3).”

https://doi.org/10.7554/eLife.53281.sa2

Article and author information

Author details

  1. Subhasis Ray

    Section on Sensory Coding and Neural Ensembles, NICHD, NIH, Bethesda, United States
    Contribution
    Conceptualization, Data curation, Software, Formal analysis, Investigation, Visualization, Methodology
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-2566-7146
  2. Zane N Aldworth

    Section on Sensory Coding and Neural Ensembles, NICHD, NIH, Bethesda, United States
    Contribution
    Data curation, Formal analysis, Investigation, Visualization, Methodology
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-0647-8465
  3. Mark A Stopfer

    Section on Sensory Coding and Neural Ensembles, NICHD, NIH, Bethesda, United States
    Contribution
    Conceptualization, Resources, Supervision, Funding acquisition
    For correspondence
    stopferm@mail.nih.gov
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-9200-1884

Funding

Eunice Kennedy Shriver National Institute of Child Health and Human Development (Intramural grant)

  • Subhasis Ray
  • Zane N Aldworth
  • Mark A Stopfer

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

We thank Tianming Li and Ian McBain for help with tracing GGN morphology from confocal image stacks, Nitin Gupta (IIT Kanpur, India) for providing a confocal image stack of GGN for a pilot model and helpful advice for making recordings from GGN. We thank the members of the Stopfer lab for feedback and suggestions and Diantao Sun and Kui Sun for excellent animal care. We also thank Vincent Schram and Lynn Holzclaw of the NICHD Microscopy and Imaging Core for help with confocal imaging, George Dold and Bruce Pritchard of NIMH Section on Instrumentation for help with electrophysiology data acquisition setup, and Theodor Usdin, NIMH/NIH, for suggestions on tissue clearing and for providing initial reagents for the CUBIC protocol. We thank the Developmental Studies Hybridoma Bank for the nc82 antibody. This work utilized the computational resources of the NIH HPC Biowulf cluster (http://hpc.nih.gov). This work was funded by an intramural grant from NIH-NICHD to MS.

Senior Editor

  1. Ronald L Calabrese, Emory University, United States

Reviewing Editor

  1. Kristin Scott, University of California, Berkeley, United States

Publication history

  1. Received: November 3, 2019
  2. Accepted: March 9, 2020
  3. Accepted Manuscript published: March 12, 2020 (version 1)
  4. Version of Record published: April 9, 2020 (version 2)

Copyright

This is an open-access article, free of all copyright, and may be freely reproduced, distributed, transmitted, modified, built upon, or otherwise used by anyone for any lawful purpose. The work is made available under the Creative Commons CC0 public domain dedication.

Metrics

  • 1,035
    Page views
  • 162
    Downloads
  • 0
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Download citations (links to download the citations from this article in formats compatible with various reference manager tools)

Open citations (links to open the citations from this article in various online reference manager services)

Further reading

    1. Neuroscience
    Víctor J López-Madrona et al.
    Research Article Updated

    Hippocampal firing is organized in theta sequences controlled by internal memory processes and by external sensory cues, but how these computations are coordinated is not fully understood. Although theta activity is commonly studied as a unique coherent oscillation, it is the result of complex interactions between different rhythm generators. Here, by separating hippocampal theta activity in three different current generators, we found epochs with variable theta frequency and phase coupling, suggesting flexible interactions between theta generators. We found that epochs of highly synchronized theta rhythmicity preferentially occurred during behavioral tasks requiring coordination between internal memory representations and incoming sensory information. In addition, we found that gamma oscillations were associated with specific theta generators and the strength of theta-gamma coupling predicted the synchronization between theta generators. We propose a mechanism for segregating or integrating hippocampal computations based on the flexible coordination of different theta frameworks to accommodate the cognitive needs.

    1. Neuroscience
    Kyle Jasmin et al.
    Research Article

    Individuals with congenital amusia have a lifelong history of unreliable pitch processing. Accordingly, they downweight pitch cues during speech perception and instead rely on other dimensions such as duration. We investigated the neural basis for this strategy. During fMRI, individuals with amusia (N=15) and controls (N=15) read sentences where a comma indicated a grammatical phrase boundary. They then heard two sentences spoken that differed only in pitch and/or duration cues, and selected the best match for the written sentence. Prominent reductions in functional connectivity were detected in the amusia group, between left prefrontal language-related regions and right hemisphere pitch-related regions, which reflected the between-group differences in cue weights in the same groups of listeners. Connectivity differences between these regions were not present during a control task. Our results indicate that the reliability of perceptual dimensions is linked with functional connectivity between frontal and perceptual regions, and suggest a compensatory mechanism.