Odor representation in the antennal lobe and mushroom body
To characterize odor representations in the antennal lobe, we made intracellular recordings from projection neurons and analyzed their responses to odor pulses presented to the antenna. In all cases, we confirmed the cell type by dye injection and subsequent histological analysis (Supplementary Fig. 1 online
). Consistent with earlier studies in locusts2,4,5
, we found that, over the course of an odor pulse, different projection neurons responded with slowly changing temporal patterns of spikes and periods of inhibition (). These distributed, time-varying firing patterns were reliable over repeated trials and varied greatly with the odor. A standard test for information content4
showed that these odor-elicited patterns were sufficiently reliable and distinctive to allow for classification far exceeding chance (Supplementary Fig. 2 online
); thus, these firing patterns could carry information about the odors. We were particularly interested in characterizing responses to relatively lengthy pulses of odor, which match the conditions in which moths naturally learn about food sources and which have often been used to test perception, learning and memory in insects, including moths18,20
. In projection neurons, responses to 4-s odor pulses generally consisted of lengthy trains of spikes, with 51% of odor-evoked spikes occurring in the first 0.6 s after odor arrived at the antenna (; timing determined by reference to an electroantennogram, data not shown). We also found that odors evoked the oscillatory synchronization of projection neurons, which, in turn, regulated the fine timing of spiking in the Kenyon cells (I. Ito et al.
, Soc. Neurosci. Abstr.
Figure 1 Projection neurons respond reliably to odors, and different odors evoke different temporally structured patterns of activity. (a) Examples of intracellular recordings of projection neurons (PN) responding to 4-s odor pulses (stimulus duration indicated (more ...)
To systematically examine the neural representation of odors by populations of Kenyon cells in the moth, we made intracellular recordings from Kenyon cells and extracellular recordings from the mushroom body with tetrodes (see Methods). Using 4-s pulses of each of a panel of 21 odors, we tested a set of 117 Kenyon cells (recorded extracellularly, 2,457 Kenyon cell–odor combinations, 10 trials per odor, each trial was 12 s long with an intertrial interval of 20 s, ; a smaller set of intracellular recordings from Kenyon cells revealed the same response properties, ).We detected extremely little spontaneous activity in Kenyon cells in the pre-stimulation period (2 s) of each trial; in 24,570 trials (49,140 s), we observed only 203 spikes. This spontaneous firing rate (mean ± s.d., 0.0041 ± 0.0122 Hz; range, 0–0.1696 Hz; n = 117) was ~2,000-fold lower than the spontaneous firing rate that we observed in the projection neuron population (measured from intracellular recordings; mean ± s.d., 8.046 ± 5.899 Hz; range, 0–26Hz; n = 15). Despite the strong and constant convergent and excitatory drive from spontaneously active projection neurons, Kenyon cells remained inactive.
Figure 2 Odor-elicited spiking in Kenyon cells is brief and sparse. (a) Examples of Kenyon cells (KC) responding to a panel of 21 odors. KC50a, KC41a and KC21a responded very sparsely, with either spikes at odor onset or offset. KC3a responded to a broader set (more ...)
We found that Kenyon cells responded mainly to the onset of lengthy odor pulse: 72% of spikes evoked by an odor occurred in the first 0.6 s of a 4-s odor presentation (we refer to these early spikes as the ‘on response’) (). Additional spikes sometimes occurred just after an odor’s offset (21% of spikes were off responses) and very few spikes occurred between these on and off responses (7% in the 3.4-s ‘middle response’ period). During the on responses, the mean firing rate, averaged over odors and trials, significantly increased (P < 0.0001, Wilcoxon signed rank test, n = 117 Kenyon cells, 0.6-s response bracket) about 21.5-fold from the basal firing level (activity during 2 s before odor stimulation). The mean firing rate during the off responses increased 3.5-fold (P < 0.0001, 3.4-s bracket) and increased by 1.3-fold during the middle responses (P < 0.005, 1.5-s bracket).
Most Kenyon cells responded to only a few of the 21 odors that we tested, although a subset of Kenyon cells responded to a broader range (). In some experiments, we presented pulses of clean air as control stimuli. These presentations evoked no reliable responses (see Methods) in any of the 42 Kenyon cells that we tested this way. To characterize odor responses across the Kenyon cell population, we computed population sparseness (SP
) and lifetime sparseness (SL
(see Methods). These measures, which take into account all of the odor-evoked spikes in all of the tested Kenyon cells, range from 0 to 1, where 1 is sparsest. Mean population sparseness SP
(full) was 0.79 (), indicating that a given odor elicited responses in very few cells. Similarly, mean lifetime sparseness was 0.72 (), indicating that a given cell responded to a narrow range of odors, although a subset of Kenyon cells was more broadly tuned, as in the locust6
. Most odor responses consisted of a single spike per trial and the maximum number of spikes in one responsive trial was 5 (). These results indicate that odor representations in the moth mushroom body are extremely sparse: they consist of very few spikes in very few neurons.
Spatiotemporal odor representations in Kenyon cells
When driven by a lengthy odor stimulus, the great majority of spikes that form the odor representation in the mushroom body occur at the onset and to a lesser extent the offset of an odor pulse (). Are the Kenyon cells that fire at the odor onset the same ones that fire at odor offset? To analyze how spiking patterns in the mushroom body change over time, we divided the odor response time into three 1.5-s periods that together captured about 95% of all spikes (). We chose to focus on responses of Kenyon cell–odor combinations that were relatively strong and reliable, which consisted of at least three responsive trials out of ten (see Methods for rate and reliability criteria). Our set of Kenyon cell–odor combinations elicited 145 reliable on and 39 reliable off responses; of these, an odor elicited reliable spiking in the same Kenyon cell both during onset and offset in only six cases. We observed only 13 reliable Kenyon cell–odor combinations during the middle time period, with two overlaps with the on response and two overlaps with the off response (). Together, these findings indicate that odor responses in the mushroom body are spatially distributed and vary over the course of the stimulus. Thus, the moth olfactory system appears to use a time-varying, distributed spatiotemporal code to represent odors both in the antennal lobe and in the mushroom body.
To examine the effect of odor-pulse duration on Kenyon cells, we analyzed all of the spikes that we observed in another set of experiments () and found that the probability of off response spiking increased with the length of the odor pulse. We almost never observed off responses following odor pulses of less than 750 ms (examples of Kenyon cells selected for their prominent off responses are shown in ). Odor pulses of at least 4 s produced the most off responses (). On and off responses elicited in Kenyon cells by long odor pulses (4 and 18 s) were distributed almost exclusively around two narrow time ranges, 0–600 ms after the odor arrived at the antenna (determined by reference to electroantennogram recordings, data not shown) and 0–800 ms after the odor was removed by vacuum (). On response spiking was maximal at around 65 ms after odor arrival.
Figure 3 Kenyon cells responded only to the onset of brief odor pulses and to the onset and offset of long pulses. (a) Kenyon cell responses varied with odor pulse duration. Pulses at least 4 s long were most likely to induce odor-specific off responses. Briefer (more ...)
STDP alone cannot mediate odor learning in Kenyon cells
Hebbian STDP mechanisms require the temporal convergence of activated neural pathways. Do the spikes that we observed in Kenyon cells constitute the odor representation that coincides with reinforcement that supports learning? To test this, we examined the relative timing of odor-elicited spiking in Kenyon cells and sucrose reinforcement in the context of a learning procedure. We trained several groups of moths and compared the amount of learning elicited by procedures in which we varied the temporal intervals between the odor and the reward ().
Figure 4 Greater temporal overlap between odor-elicited spiking in Kenyon cells and reinforcement delivery did not lead to more learning. (a) Diagrams illustrate PER conditioning procedures used to vary temporal overlap between spiking in Kenyon cells and sucrose (more ...)
Effective appetitive conditioning in honeybees19,21,28
generally occurs when the unconditioned stimulus, a sucrose reward, is presented a few seconds after the onset of a lengthy conditioned stimulus, an odor pulse. Using a computer-controlled delivery system identical to (and frequently calibrated with) the olfactometer used for our electrophysiology experiments (see Methods), we precisely regulated the timing of both the conditioned and unconditioned stimuli in all procedures ().
The control ‘unconditioned stimulus alone’ procedure group received five trials of 3-s unconditioned stimulus presentations alone (n = 33; ). This repeated delivery of sucrose alone may have caused some sensitization, as the spontaneous PER probability slightly increased from the baseline of 0 to 6.1% (not significant, P = 0.5, McNemar’s exact test; ).
For all associative conditioning procedures, the unconditioned stimulus duration was 3 s and the conditioned stimulus was paired with the unconditioned stimulus five times with 5-min intertrial intervals. Short-term memory was assessed 5 min after training by delivering only the conditioned stimulus. Our ‘on/off response’ procedure (), one that is commonly used for training honeybees and moths, consisted of a 4-s conditioned stimulus18,21
and a 2-s inter-stimulus interval (ISI) from the onset of conditioned stimulus to the onset of unconditioned stimulus18,21,28
. Moths in the on/off response group (2-s ISI, n
= 64) attained a 34.4% PER probability (). This amount of appetitive learning is typical for moths18,20,22
, which, having fattened as caterpillars, do not need to eat as much as adults. Another group of moths trained with the on/off response procedure (2-s ISI, n
= 23) and then tested with a different, non-trained odor did not respond to the different odor (Supplementary Fig. 3 online
). This result indicates that learning was specific; moths learned to associate the odor, rather than unintended cues, with the reward. The amount of learning elicited by the on/off response procedure (2-s ISI) was significantly greater than that of the control, unconditioned stimulus alone procedure group (Fisher’s exact test, P
Notably, in this effective and commonly used learning procedure, sucrose reinforcement was delivered ~1.2 s after the end of the on response in the Kenyon cells, as we knew from our physiology experiments. Thus, successful conditioning occurred in the absence of any overlap between odor-elicited on response spikes in Kenyon cells and the sucrose reward.
To further explore the timing relationship of on-response spikes in Kenyon cells and sucrose reinforcement, we then used an on/off response procedure (3.75 s ISI) in which conditioned stimulus and unconditioned stimulus were spaced further apart in time (4 s conditioned stimulus duration, 3.75 s ISI). We found that this group (n = 58) learned as well as that receiving the on/off response procedure with 2 s ISI (34.5%, ), a level of learning significantly greater than that shown by the control, “unconditioned stimulus alone” procedure group (Fisher’s exact test, P = 0.0021).
Our matching electrophysiology experiments found that brief odor pulses elicited only on response spikes in Kenyon cells (). To test the importance of overlapping on response spikes in Kenyon cells with sucrose reward, we conditioned a group of moths with brief (0.5 s) odor pulses, which were followed 0.25 s later by reinforcement (on response procedure, 0.25-s ISI, n = 61). Shifting the timing of the reward presentation closer to the on response spikes in Kenyon cells actually resulted in decreased learning (18.0% PER probability, which was not significantly different from that elicited by the unconditioned stimulus alone procedure, P = 0.1299; ).
Among these three groups, only the on response procedure (0.25-s ISI) elicited exclusively on response spikes in Kenyon cells, and was ineffective for learning. This raised the possibility that off response spiking in Kenyon cells (and possibly some middle response spiking) in the other two groups (on/off response procedures with 2-s and 3.75-s ISIs) may have contributed substantially to successful conditioning. To test this, we trained moths with a brief odor pulse in a trace procedure (on response procedure with 3.75-s ISI, 0.5-s conditioned stimulus duration, n = 23). Notably, conditioning with this procedure yielded learning (43.5%, significantly different from unconditioned stimulus alone procedure, Fisher’s exact test, P = 0.0018) that was similar to that elicited by other conditioning procedures including off response spikes in Kenyon cells (on/off response procedures with 2-s and 3.75-s ISIs; ). This suggests that the off response spikes contributed little or nothing to conditioning efficacy. We counted the number of spikes evoked in Kenyon cells during the time of unconditioned stimulus presentation in these procedures (). The on response procedure (3.75-s ISI) group, which elicited the highest learning rate, corresponded to the fewest spikes in Kenyon cells during the time of reinforcement (). To examine the limits of the interval between on response spikes in Kenyon cells and reinforcement to effectively support conditioning, we tried spacing the conditioned and unconditioned stimuli further and further apart. When we set the conditioned stimulus duration to 0.5 s to induce almost exclusively on response spikes () and gradually increased the interval between the conditioned and unconditioned stimuli, we found that the PER probability peaked at the 3.75-s ISI (43.5%) and gradually decreased (at the 9.75-s ISI, 31.8%, n = 22, not significantly different from the unconditioned stimulus alone procedure, Fisher’s exact test, P = 0.022, not significant after Bonferroni correction) and reached the control level at around the 20-s ISI (9.1%, n = 22, not significantly different from the unconditioned stimulus alone procedure, Fisher’s exact test, P = 1; ).
Notably, effective conditioning was possible even when sucrose reinforcement was delivered many seconds after the spiking responses in Kenyon cells had returned to baseline levels. These results indicate that appetitive olfactory conditioning in the moth Kenyon cells cannot be mediated by a Hebbian STDP process that requires the near-overlap of spikes elicited by the odor stimulus and spikes elicited by the reinforcement; spiking in Kenyon cells cannot be the representation that coincides with appetitive reinforcement during associative conditioning.
Finally, we asked whether off response spikes alone in Kenyon cells could support associative learning. Drawing on the results of our on response procedures, we used an ISI that was long enough to separate the onset and offset spiking in Kenyon cells by an interval that exceeded that which can support trace conditioning (); we used an extra-long conditioned stimulus (18 s, which induced small off responses similar to those elicited by 4-s odor pulses; ) and delivered the unconditioned stimulus at a 20-s ISI (2 s after the beginning of the off response). This allowed us to selectively reinforce the off response spikes, but not the on response spikes (off response procedure group, 6.7%, n
= 30; ). This procedure did not lead to PER conditioning that was significantly different from the control level (Fisher’s exact test, P
= 1). Therefore, we concluded that off response spiking alone cannot support learning. This absence of learning may result because off responses were generally small, consisting of far fewer spikes than the on responses (). As middle response spikes were much less frequent than off response spikes, we conclude that only the on response spikes contributed substantially to learning. Responses occurring after the on response could be important for other tasks that require temporal integration. The apparent importance of odor onset for associative conditioning suggests that, at least for our simple learning task, moths were prepared to make rapid behavioral choices. Consistent with this analysis, we found that moths tended to respond rapidly with proboscis extension on the onset of an odor pulse regardless of its duration or time of reinforcement during training (Supplementary Fig. 4 online