Normalization for Sparse Encoding of Odors by a Wide-Field Interneuron

See allHide authors and affiliations

Science  06 May 2011:
Vol. 332, Issue 6030, pp. 721-725
DOI: 10.1126/science.1201835

This article has a correction. Please see:


Sparse coding presents practical advantages for sensory representations and memory storage. In the insect olfactory system, the representation of general odors is dense in the antennal lobes but sparse in the mushroom bodies, only one synapse downstream. In locusts, this transformation relies on the oscillatory structure of antennal lobe output, feed-forward inhibitory circuits, intrinsic properties of mushroom body neurons, and connectivity between antennal lobe and mushroom bodies. Here we show the existence of a normalizing negative-feedback loop within the mushroom body to maintain sparse output over a wide range of input conditions. This loop consists of an identifiable “giant” nonspiking inhibitory interneuron with ubiquitous connectivity and graded release properties.

Sparse coding, the properties and advantages of which have been known for decades (13) has recently found experimental support in a number of systems (48). In such representations, information is encoded by neurons that express rare, though not exclusive, responses. In some sparse encoding systems, such as the insect mushroom bodies (7, 9) or zebrafinch song control nuclei (6), the responses of individual neurons are also very brief (one or two action potentials over a background of 0), making these representations difficult to discover, but the spikes produced extremely informative. In locust, the principal neurons of the mushroom bodies, called Kenyon cells (KCs), respond to odors with high specificity (7, 10) and can express concentration- (11) and category-invariant properties. The baseline activity of KCs is close to 0 (7, 10, 12, 13), their responses to odors typically contain fewer than three action potentials, and the gain of their output synapses (the effectiveness of their rarely elicited spikes) is both high on average and modifiable by a Hebbian learning rule (14). Because each KC is, on average, connected to about half of its presynaptic population (the projection neurons or PNs) (15), small changes in the PN population’s output could affect the reliability of the KCs’ sparse output [supporting online material (SOM) text], inconsistent with experimental observations (11).

Earlier anatomical studies (16) identified a single “giant GABAergic neuron” (GGN) in each mushroom body—[that is, GGN contains and releases the neurotransmitter γ-aminobutyric acid (GABA)]—with extensive overlap with KC projections [Fig. 1A (i)]: the neurites of GGN in the peduncle and α lobe are fine and highly branched, consistent with dendrites—but varicose in the calyx, consistent with axonal projections (17) [Fig. 1A (ii)]. [This neuron appears similar to neuron APL recently described in Drosophila (18)]. Morphological data thus suggest that GGN is well suited to form a negative-feedback loop with KCs. Using numerical simulations, we verified that an all-to-all feedback system between KCs and GGN could solve the normalization problem described above (SOM text). We show experimentally that GGN in fact fulfills this role.

Fig. 1

Morphology of and responses to odors by GGN. (A) (i) GGN intracellular fill (5% biocytin) reveals extensive arbor in mushroom body calyx (c), α lobe (αL), and in the lateral horn (LH). (Left) Pasted intracellularly labeled KC image for comparison. p, pedunculus; β-L, β lobe; s, soma. (ii) Higher magnification of GGN axonal (top) and dendritic (bottom) fields. (B). GGN membrane voltage: (i) single traces, (ii) 14 superimposed trials (average in black) and mushroom body LFP, (iii) single trials, and (iv) spectral power in 10- to 30-Hz band (single trials are gray; average is black) recorded simultaneously in response to octanol concentration series (gray bar is odor). (C) (i) Cumulative integral of GGN voltage (mean and SD) in (B) (ii) against odor concentration. (ii) Scatter plot of LFP power versus GGN intracellular voltage integral over response duration. Dots represent single trials for each odor concentration (five experiments); circles represent averages for each concentration; light-to-dark gray or red shows increasing odor concentration. (D) GGN intracellular voltage (single trial, red) superimposed with simultaneous LFP (gray) and corresponding negative LFP envelope (black).

All experiments were conducted in vivo, in immobilized, nonanesthetized animals (19). GGN was impaled from one (sometimes two) neurite(s) in the calyx or peduncle with a sharp microelectrode after blind search (19). Our results are based on 80 such recordings in 55 animals. GGN, is a nonspiking neuron with a resting potential of –51 ± 5 mV. It responded to every odor tested (fig. S4) with graded potentials composed of superimposed excitatory and inhibitory postsynaptic potentials (E- and I-PSPS, respectively) [Fig. 1B (i)]. Overall, excitation dominated, and depolarization grew with stimulus concentration (tested over a million-fold) with a peak depolarization of 15 to 20 mV above rest [Fig. 1, B and C (i)]. The oscillatory power (15 to 30 Hz) of the mushroom body local field potential (LFP) increased with odor concentration [Fig. 1B (iii and iv)] (11). Simultaneously recorded LFP (power) and VGGN (∫Vdt) covaried over this concentration range (n = 364 pairs, linear fit, r = 0.93) [Fig. 1C (ii)]. In addition, the instantaneous variations of VGGN matched those of the LFP envelope (Fig. 1D). Hence, GGN output covaries with the global drive provided to the mushroom body.

We next tested the synaptic connections between KCs and GGN. Paired intracellular recordings were made from randomly chosen KC somata and a neurite of GGN. Superimposed VGGN sweeps (n = 139) triggered from the spikes of one KC are shown in Fig. 2A (i), together with their average (black). The spike-triggered averages for this and 10 other pairs are shown in Fig. 2A (ii). They all revealed waveforms typical of unitary EPSPs, with latencies consistent with monosynaptic connections after accounting for KC spike conduction delay (n = 1302 events). Unitary EPSPs were 1 ± 0.50 mV (n = 11 KCs), with some nearing 2 mV. Using extracellular stimulation of KC somata, we could progressively recruit increasing numbers of KCs and record increasingly large postsynaptic potentials in GGN, with a mean peak of 15 to 20 mV (Fig. 2, B and C, and fig. S5). These compound potentials had nonmonotonic falling phases, explained by an additional indirect inhibitory component (see below). We compared GGN responses evoked by odors—generated by periodic KC population input at the LFP frequency (~20 Hz) (red in Fig. 2D)—to ones evoked by direct extracellular electrical stimulation of KCs at the same frequency (blue in Fig. 2D). This comparison revealed a common depolarization and large unitary IPSPs, that counteracted depolarizing summation, especially in the odor-evoked response (red trace, Fig. 2D). The discrete nature of these IPSPs suggested that they might originate from a single inhibitory interneuron. We found this putative interneuron (named IG, for “inhibitor of GGN”); its action potentials led with a consistent latency the IPSPs in GGN, whether at baseline or during responses to odors [Fig. 2E (i to iii)]. IG itself received phasic inhibitory inputs that each corresponded to phasic depolarizations (compound EPSPs) of GGN [Fig. 2E (iii and iv)]. The amplitudes of the E- and I-PSPS in the two neurons were positively correlated [Fig. 2E (v)]. We conclude that GGN receives direct input from the KC population, that GGN is an inhibitory neuron [consistent with its GABA immunoreactivity (16)], that it releases neurotransmitter in a graded manner, and that GGN is itself reciprocally connected to a spiking inhibitory interneuron [Fig. 2E (vi)]. During an odor response, GGN receives both excitatory input (from KCs) and inhibitory input from IG, itself driven by KCs and possibly also antennal lobe projection neurons. Overall, the response of GGN to odors is depolarizing, but significantly less than the pure summation of KC-evoked EPSPs would suggest, at least in part, because of the action of IG on GGN. We now turn to the action of GGN on KCs.

Fig. 2

Synaptic inputs to GGN. (A) (i) Single (gray) and averaged (black) EPSPs caused by a single KC on GGN; spike-triggered sweeps and average (STA) from dual intracellular recording from KC soma and GGN neurite. KC spikes caused by direct current injection in KC soma. (ii) STA of 11 different KC-GGN pairs (gray), and their own average (black). Calibration as in (i). (B) (i) Compound GGN EPSP caused by extracellular stimulation of KCs (gray represents single sweeps; blue shows average). (ii) Same as (i), across stimulation intensities. Calibration as in (i). (C) Peak amplitude (i) and slope (ii) of compound EPSPs in (B) as a function of KC stimulation amplitude. (D) Comparison of GGN responses to odor (red) and 20-Hz KC stimulation train (blue). (E) Simultaneous intradendritic recordings of GGN (red) and of the source of its large, discrete IPSPs (neuron named IG, gray) (i) Spontaneous activity. (ii) Spike-triggered single (gray) and averaged (red) sweeps of GGN intracellular voltage, triggered on IG spikes. (iii) Response of simultaneously recorded GGN and IG to odor (gray bar), indicating antagonistic membrane potential fluctuations. Stippled lines indicate IG hyperpolarizing potentials coinciding with GGN EPSPs. (iv) Single sweeps of IG membrane potential (gray) triggered on GGN EPSPs, showing nonspiking, inhibitory synaptic transfer. (v) IG IPSP (absolute value) against GGN EPSP amplitude, showing positive correlation, which indicates graded release. (vi) Schematic of inferred KC-GGN-IG interconnectivity.

Because GGN is a nonspiking neuron, we first performed double dendritic impalements (one for current injection, the other for voltage recording) and calibrated current injections to generate depolarizations commensurate with those evoked by odors (Fig. 3A and fig. S6). We could then assess the effect of depolarizing GGN on KC firing thresholds. A KC was impaled in the soma, and a short 70 to 300 pA pulse was injected to produce a few action potentials (n = 9 KCs, 85 trials) (Fig. 3B). This manipulation was subsequently combined with a depolarization of GGN, using increasing intensities [Fig. 3B (ii)]. In every pair, GGN depolarization beyond 5 mV reduced current-evoked firing of the recorded KC [Fig. 3, B and D (i)]. GGN thus exerts a direct, postsynaptic inhibitory effect on KCs. We then depolarized one KC (as in Fig. 3B) but depolarized GGN indirectly, by extracellular stimulation of other, unrelated KCs. A microelectrode was used to monitor simultaneously GGN membrane potential (Fig. 3C). As above, GGN depolarization counteracted current-induced spiking of the KC (indeed activating GGN synaptically was nearly twice as effective as via direct current injection) [Fig. 3, C and D (ii)]. Thus, GGN inhibits KCs postsynaptically in degrees correlated with membrane depolarization, itself a function of total KC population output.

Fig. 3

Action of GGN on KC responses to direct or synaptic depolarization. (A) (i) Schematic of experiment in (ii): GGN is impaled simultaneously with two intraneurite microelectrodes: One is used to inject direct current and the other to record resulting transmembrane voltage. (ii) Calibration of intracellular current pulse amplitude needed to depolarize GGN membrane (blue series) to values comparable to odor-evoked response (red; octanol, concentration 0.1). Current injected in GGN: 1.5, 5.5, 13.5, 15.5, 17.5, and 19.5 nA. (B) (i) Schematic of experiment in (ii): one intracellular electrode is used to depolarize GGN [to values determined in (A) (ii)]; another is used to record and depolarize a single test-KC above spike threshold. (ii) Pairing GGN and test-KC depolarizations reduces current-induced firing of KCs, which indicates graded postsynaptic inhibitory action of GGN onto KCs. Current injected in GGN from left to right: 3.5, 11.5, and 19.5 nA. (C) Similar experiment to that in (B), but GGN direct depolarization has been replaced with electrical stimulation of many KCs, the firing of which causes GGN depolarization by synaptic excitation. Current-evoked firing of test-KC is again reduced by KC-induced GGN depolarization, in a graded manner (left to right). Stimulation intensity (from left to right): 10, 20, and 30 μA. Downward deflections in KC traces (blue) are stimulation artifacts. (D) Quantification of the relation between KC firing rate reduction and GGN depolarization for experiments in (B) [D (i) five KCs] and (C) [D (ii) three KCs)]. See fig. S6 for x-axis calibration in (D) (i). For each experiment in (D) (ii), stimulation strength is expressed as percent of the observed odor-induced depolarization in the same location.

We next sought to manipulate GGN during odor presentation. During these experiments, we monitored LFPs in the mushroom body calyx. These LFPs result mainly from synaptic currents caused (directly and indirectly) by PN input onto KCs, and are strongly oscillating in the 20-Hz range during odor stimulation (12). Current-evoked depolarization of GGN during odor stimulation caused a strong and immediate reduction of the odor-evoked LFP oscillatory power [Fig. 4A (i), and fig. S7]. GGN hyperpolarization had a weaker but opposite effect [Fig. 4, A (ii) and D]. Replacing LFPs with intracellular KC recordings, we observed that odor-evoked KC membrane-potential oscillations were similarly affected by GGN polarization (Fig. 4B). One simple interpretation is that GABA released by GGN depolarization causes a conductance increase in KCs, shunting the odor-evoked synaptic currents (Fig. 4B) and the current loops responsible for the LFPs (Fig. 4A). It is possible, however, that GGN also affects KCs by presynaptic action on PN axons. Thus, although the results in Figs. 3 and 4 indicate a postsynaptic (shunting) action of GGN onto KCs, we cannot exclude the possibility that GGN also inhibits KCs presynaptically, by action ono PN axons.

Fig. 4

Nature of GGN action on KCs and consequences on mushroom body output. (A) Action of GGN on odor-evoked LFP oscillations. (i) LFP traces (top) and spectral power in 10- to 30-Hz band (bottom) in control and GGN-depolarized conditions (20 interleaved trials; individual trials and averages are shown in lighter and darker colors, respectively). Note massive reduction in LFP amplitude and power. (ii) Same as (i) but with intracellular hyperpolarization of GGN. Note enhancement of LFP, indicating graded release of GABA at rest. (Artifacts caused by GGN current pulses are grayed out.) (B) Same as in (A), but with single KC intracellular membrane potential replacing LFP. Note the effect of GGN on KC membrane potential oscillations, indicative of postsynaptic shunt. [(B) (i): 18 trials, (B) (ii): 8 trials; (19).] (C) Intracellular recording of βLN responses to odors in control conditions [red in (i), (iii), and (v)] and during GGN current injections [depolarizing, blue in (ii); and hyperpolarizing, green in (iv)]. Asterisks: spikes have been clipped. (vi) Schematic of the experiment. (vii) Quantification of effect of GGN polarization on βLN instantaneous firing rate in response to odor (19). (D) Summary of all experiments in (A) to (C), as well as control experiments evaluating the effect of positive (light gray) or negative (dark gray) current injection from 50 to 100 μm outside of GGN on LFP, after withdrawal of GGN microelectrode. (E) Relation and fit between LFP or βLN outputs and depolarizing current injected in GGN. Note the steeper action on βLN.

Thus far, we have assessed the effects of GGN only on individual KCs. Because thousands of KCs converge on a small number of extrinsic neurons in the output lobes of the mushroom body (14, 20), we can use β-lobe neurons (βLNs) as assays of GGN action onto the KC population. We impaled βLNs in a dendrite (n = 10 βLNs) to monitor odor-evoked activity. Manipulating GGN membrane potential during the odor pulse changed the recorded βLN’s responses to the odor (Fig. 4, C to E): A large GGN depolarization could silence the βLN [Fig. 4C (ii)]; conversely, hyperpolarizing GGN (moderately) increased βLN firing rate [Fig. 4C (iv and vii)]. The action of GGN on the βLN was not direct, for GGN had no effect on βLN firing evoked by current injection (fig. S8). Increasing depolarization of GGN during an odor caused a progressive reduction of βLN firing and LFP power (Fig. 4E), consistent with GGN’s effect on current-induced firing of KCs (Fig. 3, B to D). Hence, GGN affects βLNs indirectly by its actions on the KC population output.

Using simultaneous intradendritic, intrasomatic, and extracellular recordings in vivo and in nonanesthetized animals, we assessed directly and specifically the functional connectivity and actions of a single, identifiable wide-field interneuron (GGN) in a structure implicated in learning and memory in insects. This single neuron forms the negative arm of a feedback loop by KCs onto themselves and thus regulates KC excitability adaptively, a function required to maintain the sparseness of odor representations by KCs (SOM text, figs. S1 to S3). The effects of this neuron are such that it can, on its own, shut down entirely the output of the mushroom body. Conversely, its hyperpolarization can increase mushroom body output. Because GGN is also under the influence of at least one other inhibitory neuron (IG), however, the gain of the KC negative-feedback loop can, in principle, itself be modulated. This attribute is highly desirable in a circuit involved in memory for it could allow the lowering or raising of KC firing threshold and thereby increase the probability of—and degrees of refinement in—object recognition during recall.

GGN lacks action potentials, a property common in insect interneurons (21). Although the implementation we described may be specific to invertebrate brains (see fig. S9 for intracellular recordings from the Drosophila analog of GGN, for example), the underlying principles may be widespread among circuits with equivalent requirements for sparse representations. GGN acts as an integrator, similar in function to that of a population of spiking neurons: the membrane potential of GGN can be thought of as equivalent to the post stimulus time histogram of a population of spiking interneurons, smoothed with an EPSP-like kernel. Just as functionally equivalent feed-forward inhibitory loops have been found in insect mushroom bodies (7) and in mammalian piriform cortex (22), we may find in mammalian olfactory cortex a global negative-feedback loop comparable to the one we describe here.

Supporting Online Material

Materials and Methods

SOM Text

Figs. S1 to S9


References and Notes

  1. Materials and methods are available as supporting material on Science Online.
  2. Acknowledgments: We thank G. Turner for help with early locust experiments; V. Jayaraman, G. Turner, and G. Jefferis for help with the Drosophila recordings (fig. S9); and the Caltech Imaging Center for use of a confocal microscope. This work was funded by the National Institute for Deafness and Communications Disorders (G.L.), the Lawrence Hanson Fund (G.L.), the Max Planck Society (G.L.), the Office of Naval Research (grants N00014-07-1-0741 and N00014-10-1-0735 to G.L. and S.C.), a grant from Evolved Machines, Inc. (G.L. and M.P.), a Research Council of UK Academic Fellowship, and a grant from the Biotechnology and Biological Sciences Research Council (UK grant number BB/F005113/1) (T.N.).

Stay Connected to Science

Navigate This Article