Antennal lobe LNs and their neurotransmitter profiles
We used 10 Gal4 lines to genetically access antennal lobe LNs. The expression patterns of these lines span a range from hundreds of neurons near the antennal lobe to just a handful of LNs (). For simplicity, we refer to these Gal4 lines as Lines 1-10 (see Methods). In addition to LNs—which we define as neurons whose cell bodies are near the antennal lobe and whose processes are restricted to the antennal lobes—several lines also label ORNs, PNs and/or neurons near the antennal lobe that project to other areas of the brain (Supplementary Fig. 1
LN cell bodies are located in two clusters: a large but continuous cluster lateral and dorsolateral to the antennal lobe (most LNs labeled by Lines 1, 3-9), and a separate cluster ventral to the antennal lobe (most LNs labeled by Line10) (, Supplementary Fig. 1
online). To determine the number and the potential overlap of LNs labeled by these Gal4 lines, we counted the number of nuclei labeled by UAS-nuclear marker driven from individual Gal4s, as well as from their combinations in the same fly (Supplementary Table 1
online). These data suggest that some lines label largely non-overlapping LN populations, whereas other lines overlap partly or completely.
Line5 appears to include all cells that are labeled by Lines 6-9. All ~56 cells labeled by Line5 are LNs because 1) Line5 driven mCD8-GFP does not label any neurons that send processes out of the antennal lobe (), and 2) all 578 single cells from systematic MARCM analysis and biocytin fill (see below) from Line5 have their entire arborization within the antennal lobe. Based on the total number of LNs labeled by Lines 3, 4 and 5 in the same fly (Supplementary Table 1
online) and cell body positions from neuroblast clone analysis (Supplementary Fig. 1
online), we estimate that there are ~100 LNs labeled by Lines 1 and 3-9, which are mostly ipsilaterally projecting LNs in the lateral cluster (see below). Line10 labels mostly bilaterally projecting LNs (see below) in the ventral cluster. Therefore, the lower bound for Drosophila
LNs (labeled by our Gal4 lines) is ~100 ipsilaterally projecting and ~100 bilaterally projecting LNs for each antennal lobe.
Although LNs are traditionally considered to be inhibitory neurons that release GABA as their neurotransmitter, recent studies have suggested that some LNs can be excitatory14, 20
and cholinergic14, 17
. We examined the neurotransmitter type of different Gal4 lines by co-staining with antibodies against GABA and Choline acetyltransferase. Antibody specificity was validated by loss of staining in MARCM clones homozygous mutant for Glutamic acid decarboxylase 1
and Choline acetyltransferase
(Supplementary Fig. 2a-h
online). We found that the vast majority of LNs are GABAergic; however, there are a few cholinergic cells, and a larger minority that are neither GABAergic nor cholinergic (Supplementary Table 1
online; Supplementary Fig. 2i-l
online). Staining for additional neurotransmitter types identifies a few dopaminergic and a large number of (mostly ventral) glutamatergic LNs (Supplementary Fig. 3
online). These results indicate that although the large majority of LNs are GABAergic, some LNs use different neurotransmitters or their combinations, emphasizing the potential diversity in the functional roles of different LNs.
Morphological diversity of LNs
The MARCM method can be used to label single neurons born at specific developmental times21, 22
. We performed systematic MARCM labeling of LNs from these 10 Gal4 lines throughout development. shows representative images from 1,439 single cell LN clones. As is evident, LNs exhibit remarkable morphological diversity.
We identified 5 broad categories of ipsilaterally projecting LNs based on the extent of their arborization in the antennal lobe: the entire antennal lobe (pan-glomerular; ), all but a few glomeruli (), a continuous region of the antennal lobe (), discontinuous or patchy regions of the antennal lobe (), 1-3 glomeruli (few-glomerular; ). Within each category, however, there is considerable diversity. For example, although many LNs arborize nearly the entire antennal lobe, they differ in the density of arborization and the thickness of their processes (; Supplementary Fig. 4a-c
online). For LNs that skip a few glomeruli or arborize in a continuous region, different individuals skip different glomeruli or arborize in different regions of the antennal lobe (; Supplementary Fig. 5
online). Finally, bilaterally projecting LNs exhibit heterogeneous ipsilateral and contralateral arborization patterns ().
Subcellular distribution of presynaptic terminals in LNs
Since all neurotransmitter release of LNs occurs within the antennal lobe, does each LN releases neurotransmitter throughout its arbor, or only in a restricted zone? To examine how presynaptic terminals are distributed within individual LNs, we performed a large subset of MARCM single cell labeling with an additional synaptotagmin-HA marker that labels presynaptic terminals23
. In general, synaptotagmin-HA labels puncta throughout an LN's processes (red staining in , Supplementary Fig. 4
online). These data suggest that as a general rule, LNs broadcast transmitter release across their arborization without notable glomerular selectivity. However, in exception to this rule, some few-glomerular and bilateral LNs have a more selective synaptotagmin-HA distribution (, Supplementary Fig. 4f
online). These LNs therefore appear to deliver information to only a subset of glomeruli they innervate.
Statistical analysis of glomerular innervation patterns
Given the diversity of morphologies, we next analyzed these data systematically with the aim of identifying organizational principles of the LN network. An important feature of an LN is its glomerular innervation pattern, as this determines which olfactory channels the LN may receive information from and send information to. We therefore started by analyzing the glomerular innervation patterns of individual LNs.
We scored 1,532 individual LNs (1,439 from MARCM single cell clones; 93 from dye fills during whole cell patch clamp recording) for their innervation of 54 glomeruli in the antennal lobe (). We employed hierarchical clustering to organize the 1,532 LNs according to their binary ipsilateral glomerular innervation patterns. This analysis revealed several distinct morphological classes (). For example, neurons within subcluster C tend to avoid certain glomeruli innervated by trichoid ORN classes that sense pheromones24
(). Neurons within subcluster D innervate the central antennal lobe and skip the dorsal and ventral glomeruli; we call these dumbbell cells after their shape (). Neurons within subcluster E exhibit patchy innervation ().
Statistical analysis of glomerular innervation patterns
Clustering all cells labeled by the same Gal4 (Supplementary Fig. 6
online) reveals that some Gal4 lines label a relatively uniform morphological population of cells (e.g. Lines 6-8), whereas others label a more heterogeneous group (e.g. Lines 1, 2, 3, and 5). This suggests that some Gal4 lines predominantly label LNs belonging to just a few functional classes, whereas others label a functionally heterogeneous population. We will examine this idea in more detail using electrophysiological recordings (see below).
Clustering all GABA+ vs. GABA- cells from our collection revealed that both subtypes include a diversity of morphological patterns (Supplementary Fig. 7
online). Because most LNs are GABAergic, we considered the hypothesis that LN innervation might be denser in glomeruli that receive more ORN input. We estimated the average lifetime activity of each ORN type based on a published dataset25
by averaging the responses of each ORN type across 110 odor stimuli. We found a significant positive correlation between LN innervation probability and mean odor-evoked ORN firing rate (). This would be consistent with the idea that the density of inhibitory innervation is adapted to the average level of afferent input to each glomerulus26
, assuming that these stimuli are representative of natural odors.
In addition to hierarchical clustering, we used principal component analysis (PCA) to identify emergent relationships in the glomerular innervation dataset without any arbitrary threshold of significance. The first and second principal components (PC1 and PC2; and Supplementary Fig. 8
online) for the entire set of innervation patterns accounted for 40% and 10% of the variance in the data, respectively, but 13 additional principal components were needed to account for an additional 25%. The high number of dimensions necessary to account for variations indicates that innervation patterns are not well described as the linear combination of any small number of basic glomerular relationships. PC1 is essentially a proxy for the number of glomeruli innervated (correlation coefficient=0.9993). Thus, a defining feature of an LN's innervation pattern is the number of glomeruli it connects to. LNs were bimodally distributed on PC2, with a distinctive subcluster containing mainly dumbbell cells (), confirming that the dumbbell cells constitute a distinctive morphological class.
Next, we considered the LNs that arborize bilaterally. We compared ipsilateral innervation patterns (Supplementary Fig. 9a
online) with corresponding contralateral innervation patterns (Supplementary Fig. 9b
online). We found that many bilateral LNs have rather symmetric glomerular innervation patterns and overall, ipsi- and contralateral patterns were significantly correlated (Pearson's correlation coefficient, r
=0.46±0.05 (SEM), p<0.01, n
=38) (Supplementary Fig. 9c
online). In general, the exceptions were LNs with nearly pan-glomerular ipsilateral innervation.
Finally, based on clustering we identified a minimal number of LN classes with identifiable innervation patterns and further investigated birth timing and Gal4 expression for those cell types (Supplementary Fig. 5
online, Supplementary Table 2
online). We found that some (although not all) morphological types of LNs are labeled by only a few Gal4 lines and are born in limited time windows. Together, these analyses indicate that LNs are composed of genetically distinct groups of cells defined by a combination of morphology, birth timing and Gal4 expression.
Diversity and stereotypy of physiological properties
Do LN genetic categories (Gal4 lines) and morphological categories (glomerular innervation patterns) have any correlation with physiological properties? To address this, we selected five Gal4 lines that label relatively small subsets of LNs (Lines 5-9) and made whole-cell patch-clamp recordings from GFP-labeled somata in each line. We recorded the responses of each cell to a panel of diverse odors (see Methods), and filled each cell with biocytin via the patch pipette for visualization after recording. We successfully filled and scored the innervation patterns of 93 LNs.
All these LNs fired spontaneous action potentials, and their spiking was always modulated by odors. LN odor responses were remarkably diverse and typically varied more across cells than across odors within a cell (). Some LNs recorded in the same Gal4 line responded to odors in a similar manner (). For example, the LNs we recorded in Line7 were relatively stereotyped in their functional properties, as were Line8 LNs. In other cases, a Gal4 line could label LNs having very different response profiles (). Line5 and Line9 were particularly diverse in this respect.
Functional stereotypy and diversity among genetic classes
In all these LNs, we measured spontaneous firing rate, mean odor response, and maximum odor response. To quantify the time course of the odor response, we also measured the percentage of spikes fired during the first 100 msec of the odor response. All four of these properties were significantly dependent on Gal4 lines (). This implies that differences in Gal4 expression reflect physiological differences between LNs, in spite of the variability within each Gal4 line and the partial overlap in the cells that are labeled by these lines.
Physiological differences between morphological classes
Next, we asked whether LN physiological properties are correlated with their morphological class. Because the principal axis for morphological variation is the number of glomeruli each LN innervates (), we began by asking whether any response properties were correlated with this axis. Unexpectedly, LNs that innervate many glomeruli have lower odor-evoked firing rates than LNs that innervate fewer glomeruli. The number of glomeruli innervated by an LN is significantly (although weakly) anti-correlated with the average strength of its odor responses (Pearson's r=−0.2, p<0.05).
This finding motivated us to examine the physiology of pan-glomerular LNs in particular, because these cells innervate the largest number of glomeruli. This class comprised 28% of the cells we recorded from. On average, pan-glomerular LNs had significantly higher spontaneous firing rates than other LNs (). In the presence of an odor, spontaneous spiking in many pan-glomerular cells shut down completely, sometimes after a brief burst at odor onset, while others modestly increased their firing rate in the presence of odors (). Overall, odor-evoked changes in firing rate were significantly weaker in pan-glomerular cells that in other LNs ().
Functional differences between morphological classes
A second major class of LNs that we recorded from were LNs that selectively avoid glomerulus VA1d and frequently also avoid DL3 (). These glomeruli are innervated by trichoid ORNs thought to be selective for pheromones24
. This innervation pattern (with slight variations) accounts for ~15% of all LNs in our data set from Lines 1-7, notably those labeled by Line6 (; see below), and 10% of cells we filled in electrophysiological recordings. These cells differed from other LNs in having especially transient bursts of excitation at odor onset (). Overall, this morphological class fired a higher percentage of their spikes in the first 100 msec of the response, as compared to other LNs (). Thus, these LNs could create a transient pulse of GABAergic inhibition at odor onset. If so, then some pheromone glomeruli would evidently be selectively excluded from this transient pulse of inhibition.
Line6 LNs: coarse stereotypy and fine variability
Does the diversity we observe in LN morphology and physiology arise from variability of the same LNs across different individual animals, or from many diverse LN types each with stereotypic patterns in all individuals? The first alternative is implied by our finding of 847 distinct innervation patterns in 1,489 ipsilateral-projecting LNs (), far exceeding the number of estimated total ipsilateral-projecting LNs (~100) within an individual antennal lobe. LN arborization patterns therefore cannot be completely stereotyped across animals.
We sought to address this question by dense sampling of a Gal4 line (Line6) that labels a small population (7 LNs per antennal lobe, Supplementary Table 1
online). Of 131 labeled single cells for Line6, we observed 76 distinct glomerular innervation patterns (). Since there are many more innervation patterns than the number of cells per animal, individual Line6 labeled LNs cannot be identical across animals.
Variability and stereotypy of Line6 LNs
Electrophysiological recordings showed that Line6 LNs also have diverse functional properties ( and ). Some cells () showed a transient burst at the onset of almost every odor, followed by inhibition. Other cells () exhibited sustained excitation, off-excitation, and more odor-specific tuning. Thus, neither the morphology nor the physiology of Line6 LNs is highly stereotyped.
Despite the clear evidence that individual LNs labeled by Line6 are not rigidly stereotyped, the properties of these LNs are far from randomly distributed. For example, the odor response properties of Line6 cells are significantly less varied across cells than the odor response properties of all LNs. Mean and maximum odor-evoked firing rates were less variable within the Line6 population (n=16) than across the LN population as a whole (n=92, p<5×10-5, F-tests). In addition, spontaneous firing rates were also less variable within the Line6 population than across all LNs (p<5×10-6, F-test).
The glomerular innervation patterns of Line6 LNs are also far from random. Each Line6 LN innervates 51.1 glomeruli (n=131, SEM=0.22) on average. If the innervation pattern were completely random, then every glomerulus would have a 94.4% probability of being innervated. Assuming a binomial probability distribution, the number of samples innervating a glomerulus should vary by the standard error, ±2%. The actual distribution clearly deviates from this prediction (). Most glomeruli are almost always innervated, whereas a subset of glomeruli is often missed. Among this latter group of glomeruli, VA1d is most often omitted. Indeed, most of the biocytin-filled LNs that selectively avoided VA1d () were Line6 LNs (7 of 9).
Since our binary scores for innervation patterns are only a coarse measure of a LN's morphology, we additionally measured the innervation density of Line6 LNs for selected glomeruli that are either always innervated or often missed. We found that DM1 (always innervated) has a significantly higher innervation density compared to VA1d or DA1 (occasionally innervated), whereas randomly selected pan-glomerular LNs exhibit similar innervation density of DM1, VA1d and DA1. Innervation density of Line6 neurons in DA1 and VA1d glomeruli is significantly reduced compared with that of pan-glomerular LNs (). These data strengthen the distinction of these two groups of glomeruli with regard to Line6 LN innervation. One group is always innervated, while the other group has highly reduced innervation if not completely avoided.
In summary, analysis of Line6 indicated a principle likely applicable to most other morphologically and genetically-defined classes of LNs. LNs exhibit marked variability within a class, and some of this variability must represent variations in the same cells across different brains. Yet the properties of these LNs are clearly not drawn randomly from the entire distribution of LN properties. This picture is consistent with the idea that the coarse properties of these cells are genetically pre-programmed, whereas their finer-scale properties may also reflect factors such as developmental plasticity and sensory experience.
Patchy LNs: a potential mechanism for variability
The variability of glomerular innervation patterns is exemplified in a small subset of LNs from Lines 1, 3, and 5 that exhibit a distinctive pattern of patchy innervation (, ). Strikingly, comparison of the glomerular innervation patterns of 161 patchy cells () indicates that none of these single cells innervate identical sets of glomeruli. Given that these patchy cells represent only a small fraction of cells labeled by LN Gal4s and a small fraction of the ~100 ipsilaterally projecting LNs, it is clear that the innervation patterns of patchy LNs are highly variable across different animals.
The patchy nature of these cells suggested that these innervation patterns might be established through cell-cell interactions among LNs. To examine the relationships between innervation patterns of different LNs in the same animal, we designed a genetic method to visualize a pair of sister patchy cells with two different colors in the same animal (). Although the efficiency of this method was extremely low, yielding only 5 pairs of patchy cells after thousands of dissected brains, it is evident that processes of sister patchy cells avoid each other to carve out their innervation domains, sometimes by splitting a single glomerulus (outlined in , Supplementary Fig. 10
online). By contrast, differentially labeled sister cells of other LN types inter-mix at much finer scales (). We confirmed the qualitative difference by quantifying the degree of overlap between the signals of the two cells after the signals had been systematically dilated to varying degrees (see Methods). Non-patchy cells exhibited a great deal of overlap with just a small amount of dilation, while patchy cells exhibited similar degrees of overlap only after they had been dilated much more (). These observations suggest that sister patchy LNs may tile the antennal lobe the way ganglion cells tile the vertebrate retina27
and sensory neurons tile the Drosophila
. We also estimated that each patchy LN occupies 13.1±1.6% (mean±SD, n
=8) of the antennal lobe volume, implying that eight patchy LNs could tile an entire antennal lobe. These findings suggest that LN-LN interactions play a role in establishing patchy LN morphologies.
Development and maintenance of LN innervation patterns
Finally, we tested the contributions of ORNs to the glomerular arborization pattern of LNs. During the construction of the antennal lobe circuit, PN dendrites initiate pattern formation by targeting dendrites to positions similar to where they are found in the adult antennal lobe before the arrival of pioneering axons. ORN axon invasion into the antennal lobe and subsequent ORN axon-PN dendrite recognition results in the formation of distinguishable glomeruli29, 30
. The development of LNs has not previously been described.
We first tested whether LN arborization during pupal development requires ORN axons. We made use of the fact that Hedgehog signaling is required for antennal disc proliferation, and clonal loss of Smoothened, a component essential for Hedgehog signaling, results in occasional loss of maxillary palps while leaving the antennae intact30
. Under this condition, ORN axons from the antennae would innervate glomeruli that are normal targets of antennal ORN axons, but glomeruli that are normal targets of maxillary palp ORNs are devoid of ORN input30
. By introducing Line5 Gal4 driven mCD8-GFP into this genetic background and screening for flies with bilateral loss of maxillary palps (~1 in 250; ), we found that LN processes are not present in the glomeruli that are normal targets of maxillary palp ORNs. LN processes still innervate the nearby glomeruli that are targets of antennal ORNs (, bottom panels, compared to , top panels; ). By contrast, PN dendritic processes are mostly still present in glomeruli devoid of ORN axons (). This experiment suggests that glomerular innervation of LNs requires the presence of ORN axons.
Development but not maintenance of LN arborization depends on ORNs
To test whether the ORN axons are required for the maintenance of LN processes, we bilaterally removed the maxillary palps in adults, well after antennal lobe wiring is completed. We examined the processes of Line6 LNs (which are a subset of Line5 LNs) at least 10 days after removal of the maxillary palps to allow complete degeneration of ORN axons. We found that LN processes still innervate maxillary palp glomeruli (). The volume of one of the two glomeruli quantified is significantly reduced (, top panel), likely because ORN axons contribute to the glomerular volume31
. However, the total length of LN processes and the total number of presynaptic terminals as marked by synaptotagmin-HA puncta are unaffected by the adult removal of ORNs (, middle and bottom panels). We also compared variability of glomerular innervation patterns of Line6 LNs between wild-type flies and those with ORNs removed. We found no significant differences in the number of glomeruli innervated or unique innervation patterns (Supplementary Fig. 11
online), suggesting that variability of Line6 LNs is not dependent on the presence of ORN axons in adult. Taken together, these experiments indicate that ORN axons are essential for LN innervation during development, but are not required for the maintenance of their glomerular innervation in adulthood.