Oscillations and Sparsening of Odor Representations in the Mushroom Body Javier Perez-Orive, Ofer Mazor, Glenn C. Turner, Stijn Cassenaer, Rachel I. Wilson, and Gilles Laurent

1. Preparation and stimuli:

Results were obtained from locusts (Schistocerca americana) in an established, crowded colony. Young adults of either sex were immobilized, with one or two antennae intact for olfactory stimulation. The brain was exposed, desheathed and superfused with locust saline, as previously described (1). Odors were delivered by injection of a controlled volume of odorized air within a constant stream of dessicated air. Teflon tubing was used at and downstream from the mixing point to prevent odor lingering and cross-contamination. Odors were used at 10% vapor pressure (all PNs, 85% of KCs) or 100% vapor pressure (15% of KCs, all LHIs), further diluted in the dessicated air stream.

2. Tetrodes

Two types of tetrodes were used for extracellular recordings. Silicon probes were generously provided by the University of Michigan Center for Neural Communication Technology (www.engin.umich.edu/facility/cnct/). Wire tetrodes were constructed with insulated 0.0005" and 0.0004" wire (REDIOHM wire with PAC insulation, H.P. Reid). Four strands of wire were twisted together and heated to partially melt the insulation. The tip was cut with fine scissors and each channel tip was electroplated with gold solution to reduce the impedance to between 200 and 350k at 1kHz. The same custom-built 16-channel pre-amplifier and amplifier were used for both types of tetrodes. Two to four tetrodes were used simultaneously. The preamp has a unitary gain, and the amplifier gain was set to 10,000x. Data from each tetrode was acquired continuously from the four channels (15 kHz/channel, 12 bit), filtered (custom-built amplifiers, band-pass 0.3-6 kHz) and stored. Because of low baseline activity and low response probability in KCs (see results), fewer KCs than PNs were usually isolated in a typical recording session. Tetrodes were placed within the AL or MB soma clusters, peripheral to the neuropils at depths less than 200m. Cell identification was unambiguous because PNs are the only spiking neurons in the locust AL, (LNs do not produce sodium action potentials) (2), and because all the somata located above the MB calyx belong to KCs.

3. Extracellular data analysis

Tetrode recordings were analyzed as described in (3). Briefly, events were detected on all channels as voltage peaks above a pre-set threshold (usually 2.5 - 3.5 times each channel’s SD). For any detected event on any channel, the same 3ms window (each containing 45 samples) centered on that peak was extracted from each one of the four channels in a tetrode. Each event was then represented as a 180-D vector (4 x 45 samples). Noise properties for the recording were estimated from all the recording segments between detected events, by computing the auto- and cross- correlations of all four channels. A noise covariance matrix was computed and used for noise whitening. Events were then clustered using a modification of the expectation maximization algorithm. Because of noise whitening, clusters consisting of, and only of, all the spikes from a single source should form a Gaussian (SD=1) distribution in 180-D space. This property enabled us to perform several statistical tests to select only units that met rigorous quantitative criteria of isolation (Fig. S2).

4. Responses

Defining what constitutes a response quantitatively and equally accurately for PNs and KCs requires careful consideration. For example, a conventional mean firing rate measure applied to the entire “response period” is not appropriate, because PN responses are patterned; a typical PN response, such as one composed of subsequent excitatory and inhibitory epochs, often produces a mean rate no different from baseline, yet clearly constitutes an odor-specific response; reliability across trials thus needs to be taken into account. In addition, PNs and KCs have very different baseline firing statistics, implying that response criteria based on a change from baseline might not apply equally well to both populations. We thus analyzed the data using a variety of methods and display, in our paper, the results of one (Method A), applied identically to KCs and PNs. The analyses using other methods, summarized in Figs. S4 and S5, yielded nearly identical results. Our methods go as follows. First, for all methods, we used one of two response windows: short (0-1.4s) and long (0-3s after stimulus onset), with stimulus on for 1s in all cases. Method A used a 3s window. Second (Method A), a PN or KC was classified as responding if its firing behavior during the window met two independent criteria of response amplitude and reliability: a. Amplitude: the neuron’s firing rate (measured in successive 200-ms bins, averaged across all trials) had to exceed n SDs of the mean baseline rate in at least one bin within the response window. Baseline rate was measured for each cell-odor pair over a period of 3 to 5 s preceding stimulus onset and over all trials with that odor. We explored values of n from 2 to 4. If n was low (e.g., n = 2SDs) the rate of false responses detected in PNs prior to stimulation was unacceptably high (>35%). If n was high (n = 4SDs), the proportion of missed responses (as judged by visual inspection of PN rasters and PSTHs) during odor presentation was unacceptably high (>10%). Values of n of 3 or 3.5 gave low rates of both false positives (during baseline) and false negatives (during stimulation) in PNs. Values of n between (and including) 2 and 4 made no significant difference with KCs. We show the results with n = 3.5 (Method A, Figure 2); those obtained with other values of n are summarized in Fig. S4. b. Reliability: to ensure that responses detected were reliable even at low firing rates, we required that more than half of all trials with each odor contain at least one spike during the response window. We also analyzed the same data sets using different criteria for PNs and KCs, each adapted to each population’s baseline firing statistics. Despite this difference, the results (Figs. S4 and S5) are nearly identical to those shown in Fig. 2.

5. Sparseness

Data were analyzed using Matlab and Igor. The sparseness measures are taken from (4, 5). In brief, S_p = (1-{[^N_j=1 r_j/N]² / ^N_j=1 [r_j²/N]})/(1-1/N), where N is the number of units and r_j is the response of unit j. Lifetime sparseness, S_L, is calculated in the same way, except that index j now corresponds to each odor and N to the total number of odors tested with each cell. Analog response intensities for a given cell-odor pair were computed by first segmenting the recording into 200ms bins and computing the mean spike count in each bin, averaged over all trials with that odor. We then subtracted from all bin measures within the analysis window (1.4 or 3s), the mean baseline rate. All so-calculated values greater than 0 over the window (7 or 15 bins) were then added. S_p and S_L vary between 0 and 1 (1=sparsest).

6. Sharp pipette recordings and staining

Sharp electrode recordings of KCs (Figs 3, 5A) were made with borosilicate glass micropipettes (DC R: >300 M) filled with 0.2 or 0.5M K acetate or patch-electrode solution (see section 8 below). KC input resistance at the soma was usually around 1G. Intracellular recordings of LHIs (soma or dendritic impalement, Fig. 4) were made with borosilicate glass micropipettes filled with 0.5M K acetate (DC R:100-300 M). Intracellular staining of LHIs was carried out by iontophoretic injection of 2% neurobiotin in 0.5 M K acetate (0.5 s current pulses of –2.5 to –3.5 nA at 1 Hz for 30-60 min). Injected neurons were visualized in whole mounts using a diaminobenzidine-based chromogenic reaction (6). Local field potentials were always recorded in the mushroom body calyx, using saline-filled patch pipettes (DC R: 2-15M) or wire tetrodes. Electrical stimulation of PNs was carried out in the AL using 25m bipolar tungsten wires and a WPI stimulus isolator.

7. Immunocytochemistry

Anti-GABA immunostaining was carried out in whole locust brains using the following protocol [modified from (7)]. Partially desheathed locust brains were fixed for 1h in 5% formaldehyde, desheathed and washed for 20h in PBS. Brains were then dehydrated through an ethanol series, placed in propylene oxide for 20 min, rehydrated and then agitated for 5h in PBS containing 5% Triton X-100 and 0.5% bovine serum albumin (PBS 5% T 0.5% BSA). They were then washed for 30min in PBS 0.5% T 0.5% BSA, and transferred to fresh PBS 0.5% T 0.5% BSA containing anti-GABA at 1:100 dilution, or, for negative control, to PBS 0.5% T 0.5% BSA lacking primary antibody. After incubation at 47deg;C for 6 days, brains were washed for 2h in PBS at room temperature and transferred to PBS 0.5% T 0.5% BSA containing fluorescein isothiocyanate-conjugated goat anti-rabbit IgG at 1:20 dilution and incubated at 47deg;C for 4 days. They were then washed for 30 min in PBS, dehydrated through ethanol series, cleared in methyl salicylate and examined by confocal laser scanning microscopy. Fig. 4A is a projection along the z-axis of a stack of 30 optical slices each 2.7 m thick, constructed using the public domain ImageJ program (http://rsb.info.nih.gov/ij/). A movie containing the entire stack is shown in Fig. S6. Negative control brains showed diffuse background staining.

8. Patch-clamp recordings

Whole-cell patch-clamp recordings from KCs were obtained in a semi-reduced preparation. After the brain was exposed, it was removed from the head with antenna and eyes still attached, placed on a glass coverslip in a custom-built chamber, and immobilized using insect pins placed in the eyes. The brain was then desheathed as previously described (1). Recordings were obtained from KC somata under visual control using a microscope with IR-DIC imaging. Patch pipettes (5 to 6 M) were filled with a solution of (in mM): K gluconate 185, HEPES 10, EGTA 1, MgATP 4, Na₃GTP 0.5 (335 mOsm, pH 7.2). Glucose (10 mM) was substituted for an equimolar amount of sucrose in the external saline solution, and the saline was bubbled continuously with O₂. Hyperpolarizing current injections (10 pA) were used to continually measure intrinsic membrane properties, and the cell was accepted for recording as long as R_input>1G and R_access<40M. Data was acquired on an Axopatch 1D amplifier at 10kHz and low-pass filtered at 5kHz. Note: In whole-cell current-clamp mode, typical EPSP duration in controls at –60mV (Fig. 5B) was about twice that observed with sharp electrodes (Fig. 5A).

9. Picrotoxin injections

Patch pipettes were back-filled with a solution containing 1.67 mM picrotoxin and 0.3% Fast Green. After the pipette was introduced into the MB calyx (dendritic region of the MB), a pneumatic pico-pump (WPI) was used to apply a series of four to nine 100ms, 10psi pressure pulses. Each pulse injected ca. 1 pl of solution (as measured by previous injection into a drop of oil). Injected solution remained exclusively localized to calyx, as verified by dispersal of Fast Green.

Supplemental Figure 1. Olfactory circuits: Transverse section of the locust brain (left half, Bodian stain). Olfactory input originates from olfactory receptor neurons (ORNs) on the antenna. ORN axons terminate in the antennal lobe (AL), where projection neurons (PNs) act as relays, with projections to the mushroom body (MB) and the lateral horn (LH). OL: optic lobe. Calibration: 80m.

Medium version | Full size version

Supplemental Figure 2. Extracellular tetrode recordings and spike sorting. (A) Raw data traces with PN action potentials recorded in the AL (left), and KC action potentials recorded in the soma layer of the MB (right). Calibrations: 50V, 3s (top traces), 3ms (bottom traces). (B) Examples of two clusters: PN (left panel) and KC (right panel). In each panel the traces on the left show the superimposed events classified for that cluster (black) for each of the four tetrode channels, together with the average waveform (red). Calibration: 100V, 1ms. Two of the statistical tests used to evaluate the isolation of the cells in the model are shown in (B) and in (C): on the right side of each panel in (B) is the variance around the mean for each of the four channels, together with 95% confidence intervals which are based on the noise model. (C) Projection tests in which each pair of clusters in the model in 180-D space is projected onto the line connecting the cluster centers so as to evaluate their degree of isolation. All cluster centers are separated by at least 5 times the noise SD. All analyzed data were selected on these separation criteria.

Medium version | Full size version

Supplemental Figure 3. PN and KC baseline firing in the absence of odor stimulation (see text). Thirty-second rasters of 20 PNs (A) and 20 KCs (B) recorded with tetrodes. Note the exceedingly low baseline activity of KCs. (Empty rasters denote absence of action potentials during the randomly selected segment chosen for display. These rasters, however, of course originate from identified KCs, whose action potentials occurred at other times during the recording period.)

Medium version | Full size version

Supplemental Figure 4. Quantitative comparison of different methods of response detection. For each method, three statistics are computed for PNs and KCs. Response probability (P(R)) indicates the probability of a detected response, computed over all cell-odor pairs. The false positives value (False Pos.) is the percentage of responses detected when the method was applied to a window of baseline activity prior to odor onset (computed for all cell-odor pairs). The final statistic (Overlap), is a measure of similarity between a particular method and Method A (Section 4, SOM methods), defined as the percentage of cell-odor pairs for which the two methods either both detected or both did not detect a response. Methods B-D are identical to Method A, but use a different response amplitude threshold, ranging from 2 SDs to 4 SDs above baseline. Method E is the same as Method A, but uses only a 1.4s response window (0-1.4s after odor onset). Method F is based on Method A, but it uses a different reliability criterion that adapts to the cell’s baseline statistics. In this method, an odor response was deemed reliable if more than half of all trials contained at least one 200ms bin with a spike count higher than a threshold, specified as 1 SD above the mean baseline rate. Methods G and H are the methods of response detection for PNs and KCs, respectively, described in Fig. S5.

Medium version | Full size version

Supplemental Figure 5. Population responses and sparseness across PNs and KCs, calculated using different response criteria for PNs and KCs: a PN qualified as responding during the 3s following odor onset, if its firing rate increased to above 3.5 SDs of the pre-odor baseline rate (measured by a PSTH with 200ms non-overlapping bins). In contrast, a KC response occurred when over 50% of individual trials for a particular odor showed an increase from baseline activity anywhere in the 3s window. An increase in activity was defined as at least one 200ms bin with a spike count higher than 3 SDs above baseline (computed from the pre-odor period over all trials). (A) Left: Histograms displaying PN and KC response probability distributions. Response probabilities measured across all odors tested. Note opposite skews in KC and PN distributions. Right: Histograms displaying distributions of spike numbers in a response. Spike counts were computed only from cell-odor pairs with a significant excitatory response during the analysis window. (B) Excitatory responses (filled boxes) of individual PNs and KCs. Open squares denote inhibitory response (PNs only) or absence of a response; gray squares: not tested (see Fig. 2B legend for odors). Note similarity to Fig. 2, A and B.

Medium version | Full size version

To view this movie, download a QuickTime viewer.

• Movie 1   size = 2,192KB
  
  Z-series movie of the LHI neurite tract in an anti-GABA immuno-stained whole locust brain. The ca. 60 LHI cell bodies are seen to extend neurites through the Lateral Horn to the calyx of the MB, where they ramify into very fine axonal processes (video clip constructed from 61 optical slices each 2.7 m thick)

References

1. G. Laurent, M. Naraghi, J. Neurosci. 14, 2993 (1994).
2. G. Laurent, H. Davidowitz, Science 265, 1872 (1994).
3. C. Pouzat, O. Mazor, G. Laurent, "Model-based spike-sorting: an automated procedure for classification with confidence intervals," in preparation.
4. B. Willmore, D. J. Tolhurst, Network Comput. Neural Syst. 12, 255 (2001).
5. E. T. Rolls, M. J. Tovee, J. Neurophysiol. 73, 713 (1995); W. E. Vinje, J. L. Gallant, Science 287, 1273 (2000).
6. M. Wicklein, N. J. Strausfeld, J. Comp. Neurol. 424, 356 (2000).
7. D.V. Power, in Neurotransmitter Methods, R.C. Rayne, Ed. (Humana Press, Totowa, NJ, 1997).