Olfactory responses of Drosophila are encoded in the organization of projection neurons

  1. Kiri Choi  Is a corresponding author
  2. Won Kyu Kim  Is a corresponding author
  3. Changbong Hyeon  Is a corresponding author
  1. School of Computational Sciences, Korea Institute for Advanced Study, Republic of Korea

Abstract

The projection neurons (PNs), reconstructed from electron microscope (EM) images of the Drosophila olfactory system, offer a detailed view of neuronal anatomy, providing glimpses into information flow in the brain. About 150 uPNs constituting 58 glomeruli in the antennal lobe (AL) are bundled together in the axonal extension, routing the olfactory signal received at AL to mushroom body (MB) calyx and lateral horn (LH). Here we quantify the neuronal organization in terms of the inter-PN distances and examine its relationship with the odor types sensed by Drosophila. The homotypic uPNs that constitute glomeruli are tightly bundled and stereotyped in position throughout the neuropils, even though the glomerular PN organization in AL is no longer sustained in the higher brain center. Instead, odor-type dependent clusters consisting of multiple homotypes innervate the MB calyx and LH. Pheromone-encoding and hygro/thermo-sensing homotypes are spatially segregated in MB calyx, whereas two distinct clusters of food-related homotypes are found in LH in addition to the segregation of pheromone-encoding and hygro/thermo-sensing homotypes. We find that there are statistically significant associations between the spatial organization among a group of homotypic uPNs and certain stereotyped olfactory responses. Additionally, the signals from some of the tightly bundled homotypes converge to a specific group of lateral horn neurons (LHNs), which indicates that homotype (or odor type) specific integration of signals occurs at the synaptic interface between PNs and LHNs. Our findings suggest that before neural computation in the inner brain, some of the olfactory information are already encoded in the spatial organization of uPNs, illuminating that a certain degree of labeled-line strategy is at work in the Drosophila olfactory system.

Editor's evaluation

Choi et al. explore how olfactory information flows across the three major neuropils in the Drosophila brain – the antennal lobe (AL), mushroom Body (MB), and the lateral horn (LH). They use the two connectomes of adult Drosophila and 'inter-PN distances' to do this. Using this neuroanatomy based approach, they find support for a labeled-line strategy, which they subsequently test for with synaptic connectivity data for a subset of PNs. They find that while some labelled lines may exist, PNs generally participate in multi-channel integration at the MB and LH. This manuscript will be of interest to neuroscientists interested in olfactory processing and to those working on connectomic-level circuit analysis.

https://doi.org/10.7554/eLife.77748.sa0

Introduction

Anatomical details of neurons obtained based on a full connectome of the Drosophila hemisphere reconstructed from electron microscope (EM) image datasets (Bates et al., 2020; Scheffer et al., 2020) offer the wiring diagram of the brain, shedding light on the origin of brain function. Out of the immense amount of data, we study the second-order neurons, known as the projection neurons (PNs) of the olfactory system. It is the PNs that bridge the olfactory receptor neurons (ORNs) in the antenna and maxillary palp to higher olfactory centers where neural computation occurs for Drosophila to sense and perceive the environment (Hallem and Carlson, 2004a). The three neuropils, namely the antennal lobe (AL), mushroom body (MB) calyx, and lateral horn (LH), are the regions that abound with an ensemble of axonal branches of PNs and synapses (Figure 1). PNs can be classified as uniglomerular and multiglomerular PNs based on their structure and connectivity to other PNs. The uniglomerular PNs (uPNs) in AL constitute glomeruli that collect olfactory signals from ORNs of the same receptor type (Gao et al., 2000; Couto et al., 2005). uPNs innervating MB calyx and LH relay the signals further inside the brain through synaptic junctions with the Kenyon cells (KCs) and lateral horn neurons (LHNs), respectively. Multiglomerular PNs (mPNs), on the other hand, innervate multiple glomeruli, often contributing to the inhibitory regulation of signals relayed from ORNs to third-order olfactory neurons (Berck et al., 2016). PNs can functionally be categorized into either excitatory (cholinergic) or inhibitory (GABAergic), where a many GABAergic PNs tend to bypass MB calyx while innervating multiple glomeruli in AL (and hence are mPNs) (Schultzhaus et al., 2017; Shimizu and Stopfer, 2017).

A schematic of the Drosophila olfactory system.

uPNs comprising each glomerulus in AL collect input signals from ORNs of the same receptor type and relay the signals to MB calyx and LH. uPNs in MB calyx synapse onto KCs; and uPNs in LH synapse onto LHNs.

Since the seminal work by Cajal, 1911, who recognized neurons as the basic functional units of the nervous system, there have been a series of attempts at classifying neurons using different representations of neuronal morphologies and at associating the classified anatomies with their electrophysiological responses and functions (Uylings and van Pelt, 2002; Scorcioni et al., 2008; Jefferis et al., 2007; Seki et al., 2010; Gillette and Ascoli, 2015; Lu et al., 2015; Li et al., 2017; Kanari et al., 2018; Mihaljević et al., 2018; Gouwens et al., 2019; Laturnus et al., 2020). Systematic and principled analyses of neuronal anatomy would be a prerequisite for unveiling a notable link between the PN organization and olfactory representations. Several different metrics involving spatial projection patterns (Jefferis et al., 2007), electrophysiological properties (Seki et al., 2010; Gouwens et al., 2019), topological characteristics (e.g. morphometrics) (Uylings and van Pelt, 2002; Scorcioni et al., 2008; Lu et al., 2015; Mihaljević et al., 2018; Gouwens et al., 2019), intersection profiles (Gouwens et al., 2019), and NBLAST scores (Jeanne et al., 2018; Zheng et al., 2018; Bates et al., 2020; Scheffer et al., 2020) have been utilized in the past. More recently, machine learning approaches have been popularized as a tool for classification tasks (Vasques et al., 2016; Buccino et al., 2018; Mihaljević et al., 2018; Zhang et al., 2021).

Among a multitude of information that can be extracted from the neural anatomy associated with uPNs, the inter-PN organization draws our attention. To compare spatial characteristics of uPNs across each neuropil and classify them based on the odor coding information, we confine ourselves to uPNs innervating all three neuropils, most of which are cholinergic and follow the medial antennal lobe tract (mALT). Within this scope, we first calculate inter-PN distance matrices in each neuropil and study them in reference to the glomerular types (homotypes) to discuss how the inter-PN organization changes as the PNs extend from AL to MB calyx and from AL to LH.

In this study, we utilize two representative EM-based reconstruction datasets for the analysis (the latest FAFB Bates et al., 2020 and the hemibrain datasets Scheffer et al., 2020). The FAFB dataset specifically encompasses the Drosophila olfactory system, while the hemibrain dataset aims for a reconstruction of the entire right hemisphere of the Drosophila brain. The results based on the two datasets are largely consistent and interchangeable, which generalizes our findings.

We have conducted statistical analyses to unravel potential associations between the uPN organization and the behavioral responses of Drosophila to external stimuli encoded by glomerular homotypes, finding that certain odor types and behavioral responses are linked to a characteristic inter-neuronal organization. The map of synaptic connectivity between uPNs and the third-order neurons (KCs and LHNs in MB calyx and LH, respectively) complements the functional implication of the association between the inter-PN organization and olfactory processing. A ‘labeled-line design’ in olfaction is generally considered to exhibit a chain of neurons dedicated to encoding a single olfactory feature with no direct integration with other features as the signal is passed onto higher-order neurons. While we do not demonstrate the full architecture of labeled-line design in the Drosophila olfactory system as the signals from odor-sensing by ORNs are passed down to the inner brain for perception, our analysis shows that homotypic uPNs encoding particular odor types not only maintain their spatially localized and bundled structure throughout all three neuropils but also display synaptic connections that converge to a narrow subset of third-order neurons. The Drosophila olfactory system leverages the efficiency of the labeled-line design in sensory information processing (Min et al., 2013; Howard and Gottfried, 2014; Andersson et al., 2015; Galizia, 2014).

Results

Spatial organization of neurons inside neuropils

The inter-PN distance dαβ

First, we define a metric with which to quantify the spatial proximity between neurons. Specifically, the inter-PN distance dαβ represents the average taken over the minimum Euclidean distances between two uPNs α and β, such that dαβ is small when two uPNs are tightly bundled together (see Equation 1 and Figure 2—figure supplement 1A). Although metrics such as the NBLAST score (Costa et al., 2016) and others (Kohl et al., 2013) can be used to study the PN organization, these metrics take both the morphological similarity and the spatial proximity into account. The distance dαβ only measures the pairwise distance but not the dot product term (which measures the similarity of two neuronal morphologies), whereas the NBLAST score considers both. Therefore, while the distance dαβ is computationally comparable to the NBLAST score, it only measures the spatial proximity between two neurons. We notice that the features of the uPN organization captured by the NBLAST distance are not necessarily aligned with dαβ (see Figure 2—figure supplement 1B). The two distances are correlated but with significant dispersion, indicating that these two metrics are not the same. Since we are solely interested in the spatial proximity (or co-location) between two uPN innervations but not the morphological similarity between them (which the NBLAST score accounts for, a point also noted by Zheng et al., 2018), we deliberately chose the metric dαβ instead of the NBLAST score for our analyses.

The distances dαβ (Equation 1) between all the possible pairs (α and β) of 135 uPNs from the FAFB dataset are visualized in the form of a matrix (Figure 2). We perform hierarchical clustering on the distance matrix for uPNs in each neuropil (see the outcomes of dαβ-based clustering analysis in Figure 2—figure supplement 2 and Materials and methods for the details). Individual clusters from the hierarchical clustering of uPNs in MB calyx and LH from the FAFB dataset are visualized in Figure 3 and Figure 4 with the colors denoting the odor types encoded by the individual uPNs, which will be discussed in detail later.

Figure 2 with 3 supplements see all
The three matrices representing the pairwise distances dαβ in units of μm between individual uPN in AL, MB calyx, and LH.

The matrices are calculated based on uPNs available in the FAFB dataset. The diagonal blocks reflect the homotypic uPNs comprising the 57 glomerular homotypes defined in the FAFB dataset (Bates et al., 2020), labeled at the edges.

The dαβ-based clustering on uPNs based on the FAFB dataset in MB calyx resulting in 10 clusters.

The individual uPNs are color-coded based on the encoded odor types (Dark green: decaying fruit, lime: yeasty, green: fruity, gray: unknown/mixed, cyan: alcoholic fermentation, red: general bad/unclassified aversive, beige: plant matter, brown: animal matter, purple: pheromones, pink: hygro/thermo) (Mansourian and Stensmyr, 2015; Bates et al., 2020). The first and second columns illustrate the dorsal and the anterior view, respectively (D: dorsal, M: medial, P: posterior). The black line denotes the approximate boundary of MB calyx.

Figure 4 with 2 supplements see all
The dαβ-based clustering on uPNs based on the FAFB dataset in LH resulting in 11 clusters.

(inset) A cartoon illustrating the relative position between clusters CILH=C3LH and COLH=C4,5,6,9LH. The individual uPNs are color-coded based on the encoded odor types (Dark green: decaying fruit, lime: yeasty, green: fruity, gray: unknown/mixed, cyan: alcoholic fermentation, red: general bad/unclassified aversive, beige: plant matter, brown: animal matter, purple: pheromones, pink: hygro/thermo). The first and second columns illustrate the dorsal and the anterior view, respectively (D: dorsal, M: medial, P: posterior). The black line denotes the approximate boundary of LH.

Spatial proximity-based clustering results

In MB calyx, the hierarchical clustering divides the uPNs from the FAFB dataset into 10 clusters (Figure 3). Clusters C2MB and C10MB largely encompass the dorsal region and clusters C6MB and C7MB encompass the ventral region of the neuropil. The cluster C7MB shows a characteristic biforked pattern projecting to the lateral and medial regions. The cluster C3MB also exhibits the same structural pattern but is composed of a tight bundle of uPNs that are part of DL2d and DL2v (both of which encodes food-related odors). The cluster C8MB is located between the biforked innervation pattern of clusters C6MB and C7MB, and predominantly innervates the posterior region. Lastly, clusters C1MB, C4MB, and C5MB, innervate the anterior region of MB calyx, spatially separated from other uPNs.

In LH, 11 clusters are identified in the FAFB dataset (Figure 4). The cluster C3LH is the largest, which mainly innervates the dorsal posterior region of LH. Clusters C4LH, C5LH, C6LH, and C9LH display variable biforked projection patterns along the coronal plane, enveloping the boundary of the cluster C3LH. This creates a spatial pattern where a large blob of uPNs (CILH) are surrounded by a claw-like structure (COLH) (Figure 4, inset). Clusters C1LH, C2LH, and C7LH innervate the anterior-ventral region and display clear segregation from the other uPNs. Another group composed of clusters C10LH and C11LH innervates the posterior-ventral-medial region.

We use Pearson’s χ2-test (see Materials and methods for the details) to assess the likelihood of dependence between the dαβ-based clustering outputs for MB calyx, LH, and the glomerular labels (homotypes) statistically significant correlations are found in terms of both the p-value and the Cramér’s V (see Appendix 1—table 1 and Methods for a detailed explanation of the meaning behind the p-value and the Cramér’s V), the latter of which is analogous to the correlation coefficient for the χ2-test. The mutual information between the same set of nominal variables, which is calculated to verify our χ2-tests (see Materials and methods), offers a similar conclusion (see Appendix 1 and Appendix 1—table 3).

We also categorize the spatial organization of uPNs in reference to the glomerular labels. The homotypic uPNs constituting a tightly bundled glomerulus in AL manifest themselves as the block diagonal squares in the dαβ-matrix (Figure 2). This is apparent in the dendrogram constructed from the distance matrix for the uPNs at AL (Figure 2—figure supplement 3), where uPNs sharing the same glomerular label are grouped under a common branch, thereby demonstrating the spatial proximity between uPNs forming the same glomerulus. The dαβ-matrix indicates that such organizations are also preserved in MB calyx and LH. However, clear differences are found in the off-diagonal part of dαβ matrices (Figure 2).

The same hierarchical clustering analysis performed on the hemibrain dataset results in 14 clusters for uPN innervation in MB calyx and 13 clusters in LH. Despite the differences in the number of clusters, we find that spatial and structural characteristics of individual clusters observed from the FAFB dataset are well translated and comparable to those from the hemibrain dataset (see the clustering result in Figure 8—figure supplement 1). Furthermore, various statistical tests used in this study (e.g. Pearson’s χ2-test) on the hemibrain dataset lead to the same conclusion (see Appendix 1, Appendix 1—table 2, and Appendix 1—table 4).

The degrees of bundling, packing, and overlapping

To conduct a quantitative and concise analysis of dαβ matrices, we define the mean intra- and inter-homotypic uPN distances, d¯intra,X and d¯inter,X (see Methods for detailed formulation). The d¯intra,X is the average distance between uPNs in the same homotype and measures the degree of uPNs in the homotype X being bundled. Therefore, a smaller d¯intra,X signifies a tightly bundled structure of X-th homotypic uPNs (see Figure 5—figure supplement 1 for raw d¯intra,X values). Similarly, d¯inter,X, which measures the degree of packing (or segregation), is defined as the average distance between the neurons comprising the X-th homotype and neurons comprising other homotypes. Thus, a small value of d¯inter,X signifies tight packing of heterotypic uPNs around X-th homotype, while a large value indicates that the homotypic uPNs comprising the homotype X are well segregated from other homotypes (see Figure 5—figure supplement 1 for raw d¯inter,X values).

The degrees of bundling averaged over all homotypes (d¯intra=NX1XNXd¯intra,X4 μm) are comparable over all three neuropils (blue dots in Figure 5A and B). On the other hand, from d¯inter, which is defined as the mean inter-homotype distance averaged over all X s, we find that homotypic uPNs are well segregated from others in AL as expected, whereas spatial segregation among homotypes is only weakly present in MB calyx (orange dots in Figure 5A and B and the cartoon of Figure 5C). We also observe that the d¯intra and d¯inter are comparable for the two different datasets. A minor difference is observed in d¯intra, indicating a slightly tighter bundling structure for the hemibrain dataset.

Figure 5 with 2 supplements see all
Organization of homotypic uPNs in the three neuropils.

Plots of d¯intra (blue, degree of bundling), d¯inter (orange, degree of packing), and the ratio between the two distances λ (red, degree of overlapping) calculated based on (A) the FAFB dataset and (B) the hemibrain dataset. Error bars depict the standard deviation. (C) Diagram illustrating the overall organization of uPNs at each neuropil. Homotypic uPNs are tightly bundled and segregated in AL. Several groups of homotypic uPNs form distinct heterotypic spatial clusters at higher olfactory centers, extensively overlapping in MB calyx (see Figure 3).

Next, we take the ratio of mean intra- to inter-PN distances of X-th homotype as λX to quantify the degree of overlapping around X-th homotype (see Materials and methods). The term ‘overlapping’ is specifically chosen to describe the situation where different homotypes are occupying the same space. A large value of λX (particularly λX>0.4) suggests that the space occupied by the uPNs of the X-th homotype is shared with the uPNs belonging to other homotypes. The value λ(=NX-1XNXλX) averaged over all the homotypes (red in Figure 5A and B) suggests that the extent of overlapping between uPNs is maximal in MB calyx and minimal in AL (Figure 5C).

Figure 6A and B, Figure 7, and Figure 7—figure supplement 1 show individual values of λX for all homotypes in the three neuropils. We identify the following features: (i) In AL, λX0.4 for all homotypes except DL5 (a homotype encoding aversive odors), indicating that homotypic uPNs are tightly bundled and segregated from uPNs in other glomeruli. The same trend is observable in the hemibrain dataset (Figure 6B), but with λDL50.4.; (ii) In MB calyx, a large portion (65%) of λX’s exceed 0.4 and even the cases with λX>1 are found (VC5, DL5), implying that there is a substantial amount of overlap between different homotypes. In the hemibrain dataset, 76% of λX’s exceed 0.4.; (iii) Although not as significant as those in AL, many of uPNs projecting to LH are again bundled and segregated in comparison to those in MB calyx (see Figure 7B). (iv) The scatter plot of λX between MB calyx and LH (Figure 7C) indicates that there exists a moderate positive correlation (r=0.642,p<0.0001) between λX at MB calyx and LH. This implies that a higher degree of overlapping in MB calyx carries over to the uPN organization in LH. The result from the hemibrain dataset is similar (r=0.677, p<0.0001, see Figure 7—figure supplement 1).

Figure 6 with 1 supplement see all
Degree of overlapping between inter-homotypic uPNs, λX (X=VM4,VC5,,VP5).

The degree of overlapping (λX) for X-th homotype in AL, MB calyx, and LH (from lighter to darker colors) calculated from the uPNs based on (A) the FAFB dataset and (B) the hemibrain dataset. The homotype label is color-coded based on the odor types associated with the glomerulus obtained from the literature and is sorted based on the value of λX for each odor type at LH. Asterisks (*) mark homotypes composed of a single uPN while plus (+) mark homotypes composed of a single uPN under our selection criterion but are actually a multi-uPN homotype, whose intra-homotype uPN distance is not available. O (attractive) and X (aversive) indicate the putative valence information collected from the literature. The blue horizontal line denotes λX=0.4. (C) Two homotypes taken from the FAFB dataset, DL3 (purple) and DL5 (red), which are indicated by yellow triangles in (A), are highlighted along with other uPNs (gray).

Figure 7 with 1 supplement see all
Scatter plots depicting the relationships between λX s at two different neuropils calculated from the uPNs based on the FAFB dataset; (A) AL versus MB calyx, (B) AL versus LH, and (C) LH versus MB calyx.

The color code is the same as in Figure 6. The blue lines in (A) and (B) denote λX=0.4.

The entire neuron morphologies of uPNs from two homotypes with a small (X= DL3, which largely responds to pheromones) and a large (X= DL5) λX s in LH are visualized along with the other uPNs (gray) (Figure 6C). The homotype DL3, which seldom overlaps with others in AL (λDL30.07) and LH (λDL30.17), displays an increased overlapping in MB calyx (λDL30.31). Therefore, DL3 is tightly packed in AL and LH, whereas it is relatively dispersed in MB calyx. Meanwhile, the homotype DL5 displays a significant dispersion in all three neuropils, although the dispersion is the smallest in AL (λDL50.74) compared to that in MB calyx (λDL51.1) and LH (λDL51.5).

There are minor variations between the FAFB and the hemibrain datasets in terms of d¯intra,X, d¯inter,X, and λX, and they likely arise from the factors such as a minor mismatch in the glomerulus label annotations that sometimes affects the number of uPNs constituting a given homotype, and the difference in the number of uPNs between two datasets as a result of our selection criterion. Regardless, still present in both datasets are the spatial and organizational trends described above. Taken together, the organization of olfactory uPNs varies greatly in the three neuropils. The clear homotype-to-homotype segregation in AL no longer holds in MB calyx. Instead, the dαβ -based clustering suggests the presence of clusters made of multiple different homotypic uPNs (Figure 5C). For some homotypes, the well-segregated organizations in AL are recovered when they reach LH (compare Figure 7A and Figure 7B).

Relationship between neuronal organization and olfactory features

Now we explore how the structural features identified from our clustering outputs are associated with odor types and valences (behavioral responses). As briefly mentioned earlier, the color codes in Figure 3, Figure 4, Figure 6, and Figure 7 depict odor types encoded by corresponding homotypic uPNs, which follow the same categorical convention used by Mansourian and Stensmyr, 2015 and Bates et al., 2020. The O and X represent the putative valence, which indicates whether Drosophila is attracted to or repelled from the activation of specific homotypic uPNs. For example, DA2 responds to geosmin, a chemical generated from harmful bacteria and mold, which evokes a strong repulsion in Drosophila (Stensmyr et al., 2012). Similarly, VM3 is suggested to encode repulsive odors, while VM2 and VM7d encode attractive odors (Mansourian and Stensmyr, 2015; Bates et al., 2020). Overall, the following information is acquired from the literature (Hallem et al., 2004b; Galizia and Sachse, 2010; Mansourian and Stensmyr, 2015; Badel et al., 2016; Bates et al., 2020) and labeled accordingly:

  • DA1, DA3, DL3, DM1, DM4, VA1v, VA2, VA3, VC1, VC2, VM1, VM2, VM4, VM5d, VM5v, VM7d, and VM7v (17 homotypes) encode attractive (O) odor.

  • D, DA2, DA4l, DA4m, DC1, DC2, DC3, DC4, DL1, DL4, DL5, DM2, DM3, DM5, DM6, DP1m, V, VA5, VA6, VA7l, VA7m, VC3, VL2a, VL2p, and VM3 (25 homotypes) encode aversive (X) odor.

  • The remaining homotypes are characterized as either unknown, non-preferential, or conflicting valence information.

Collecting the glomerular types of tightly bundled homotypic uPNs with λX<0.4 in LH (Figure 6, Figure 7, and Figure 7—figure supplement 1), we explore the presence of any organizational trend.

  1. In LH, out of 37 homotypes composed of multiple uPNs (2n8) based on our selection criterion, 29 homotypes (DL2v, DL2d, VM1, VL1, DM6, VM7d, VA3, VM5v, DA3, VM2, DL1, VA7m, VC3, VM7v, VC4, V, DM2, VM3, DA2, D, DC2, VA5, VA1v, DA1, DC3, DL3, VA1d, VP1d, and VP1l) satisfy the condition of λX<0.4. In the hemibrain dataset, a couple of homotypes (VM1, DL5, DL3, and VP1l) are suggested to be single uPN homotypes based on our selection criterion.

  2. Homotypes VA1v, DA1, DC3, DL3, and VA1d (colored purple in Figure 3, Figure 4, Figure 6, and Figure 7) encode pheromones involved with reproduction (Grabe et al., 2016; Bates et al., 2020; Dweck et al., 2015), and VM4, VM1, VM7d, DM1, DM4, VC2, VM5d, VA3, VM5v, DA3, and VM2 encode odors presumed to be associated with identifying attractive food sources (Couto et al., 2005; Semmelhack and Wang, 2009; Mohamed et al., 2019; Bates et al., 2020) (see Figure 6A). A previous work (Grosjean et al., 2011) has identified a group of glomeruli that co-process food stimuli and pheromones via olfactory receptor gene knock-in coupled with behavioral studies. The list of homotypes mentioned above is largely consistent with those glomeruli reported by Grosjean et al., 2011.

  3. Homotypes DM6, DM2, VM3, VL2p, DA2, and D are likely associated with aversive food odors. DA2 responds to bacterial growth/spoilage; VL2p, DM2, and VM3 to the alcoholic fermentation process; DM6 and D to flowers (Galizia and Sachse, 2010; Bates et al., 2020).

  4. Many homotypes responding to odors which can be described as kairomones, a type of odors emitted by other organisms (Kohl et al., 2015), are part of the 29 homotypes with λX<0.4. This includes the pheromone encoding groups (VA1v, DA1, DC3, DL3, and VA1d) and others such as DA2, VC3, and VA5, which respond to geosmin, 1-hexanol, and 2-methyl phenol, respectively (Hallem et al., 2004b; Galizia and Sachse, 2010).

Figure 8 recapitulates the cluster information from dαβ -based analysis along with homotypes, odor types (color-codes), and putative valence (attractive (O) and aversive (X) odors). Some points are worth making:

  1. Even though uPNs innervating MB calyx exhibit large λX s, the hierarchical clustering grouped homotypic uPNs together. This suggests the homotypic uPNs are still proximal in MB calyx, indicating the reduction in dinter is what is driving the increase in overlapping. This is already shown through d¯intra in Figure 5A, B and is supported by our statistical tests (see Appendix 1—table 1 and Appendix 1—table 3). The same is true for LH. The grouping of homotypic uPNs is also observable from the hemibrain dataset (Figure 8—figure supplement 1).

  2. In the FAFB dataset, 13 out of 57 homotypes are made of a single uPN (n=1, the asterisked glomeruli in Figure 6A and Figure 8), which tend to be characterized by comparatively dense branched structures (see Figure 6—figure supplement 1), suggestive of homotypic uPN number dependence for the neuron morphology. Among the 13 homotypes, 7 encode aversive stimuli (X), 4 encode attractive stimuli (O), and 2 have no known valence information (see Appendix 1—table 5). In the hemibrain dataset, 7 encode aversive stimuli (X), 5 encode attractive stimuli (O), and 1 has no known valence information (see Appendix 1—table 6). The relative prevalence of single-uPN homotypes encoding aversive stimuli is noteworthy.

  3. In LH, the cluster C1LH, located in the anterior-ventral region of the neuropil, is composed only of pheromone-encoding homotypic uPNs, DA1 and DC3. The cluster C2LH is also mostly composed of pheromone-encoding homotypic uPNs, DL3 and VA1d (Figure 4 and Figure 8), which is consistent with the results by Jefferis et al., 2007. In MB calyx, the majority of the uPNs encoding pheromones, except DL3, are grouped into the cluster C8MB (see Figure 3 and Figure 8). A similar trend is observed in the hemibrain dataset, although the arbitrary cluster labels differ (see clusters C4LH, C8LH, and C10MB in Figure 8—figure supplement 1).

  4. Hygro/thermo-sensing homotypes such as VP2 and VP4 are spatially segregated from other odor-encoding uPNs, which is observable through clusters composed predominantly of hygro/thermo-sensing homotypes (see Figure 8 and Figure 8—figure supplement 1). In MB calyx, these neurons rarely project to anterior region and are distributed along the base of the neuropil. This is in line with previous literature (Li et al., 2020). In LH, they are clustered in the posterior-ventral-medial region, hardly innervating the neuropil but covering the medial side of the neuropil (Figure 3 and Figure 4).

  5. Along with the clusters of uPNs visualized in Figure 3 and Figure 4, of particular note are the clusters formed by a combination of several homotypic uPNs. A large portion of uPNs innervating LH that encodes potentially aversive responses are grouped into clusters C4LH, C5LH, C6LH, and C9LH, which envelop the cluster C3LH where mostly food-related homotypes converge (Figure 4). In the hemibrain dataset, these correspond to C10LH and C11LH for the aversive responses and C6LH and C13LH for the food-related homotypes (Figure 8—figure supplement 1).

Figure 8 with 1 supplement see all
A diagram summarizing how the clusters of uPNs in MB calyx (10 clusters) and LH (11 clusters) from the FAFB dataset are associated with the odor types (Dark green: decaying fruit, olive: yeasty, green: fruity, cyan: alcoholic fermentation, red: general bad/unclassified aversive, beige: plant matter, brown: animal matter, purple: pheromones, gray: unknown, pink: hygro/thermo).

Asterisks (*) mark homotypes composed of a single uPN while plus (+) mark homotypes composed of a single uPN under our selection criterion but are actually a multi-uPN homotype, whose intra-homotype uPN distance is not available. O and X represent the putative valence information collected from the literature (O: attractive, X: aversive).

Given that the synaptic communications with KCs and LHNs are critical for neural computation in the inner brain, the specific type of uPN organization in each neuropil should be of great relevance. Indeed, it has been suggested that the spatial convergence, segregation, and overlapping of different homotypic uPNs within neuropil influence the information processing in higher olfactory centers (Grosjean et al., 2011).

According to previous studies (Jefferis et al., 2007; Liang et al., 2013; Kohl et al., 2013; Fişek and Wilson, 2014), uPN innervation in LH and LHNs are highly stereotyped in terms of connectivity and response. Homotypic uPNs are spatially organized in AL, and to a certain degree, in LH, based on the odor type and valence information (Min et al., 2013; Huoviala et al., 2020). The presence of tightly bundled anatomy of homotypic uPNs (λX<0.4) in both AL and LH (Figure 7B and Figure 7—figure supplement 1B) may imply that the Drosophila olfactory system dedicates a part of the second-order neural circuit on behalf of the ‘labeled-line’ design, which enables the organism to sense urgent chemical stimuli at the early stage of information processing without going through more sophisticated neural computation in the inner brain (Howard and Gottfried, 2014; Andersson et al., 2015; Min et al., 2013).

Labeled-line design of the higher order olfactory neurons

The concept of labeled-line design is widely considered at work at the ORN-PN interface (AL) as the signal generated from specific olfactory receptors converges to a single glomerulus (Vosshall et al., 2000; Couto et al., 2005; Fishilevich and Vosshall, 2005). A potential labeled-line strategy or separated olfactory processing of aversive odors encoded by DA2 has been extensively discussed (Stensmyr et al., 2012; Seki et al., 2017; Huoviala et al., 2020). It has been shown that pheromone-encoding homotypes in LH (Jefferis et al., 2007; Ruta et al., 2010; Kohl et al., 2013; Frechter et al., 2019; Bates et al., 2020; Das Chakraborty and Sachse, 2021) are at work in specific third-order olfactory neurons. So far, we have shown that the labeled-line design is present in the architecture of higher olfactory centers of second-order neurons, that is, MB calyx and LH, where homotypic uPNs are tightly bundled together despite the lack of glomerular structure. In this section, we will conduct a comprehensive analysis of the synaptic connectivity between PNs and third-order olfactory neurons (KCs and LHNs) using three demonstrations. We ask (i) whether the labeled-line strategy implied in the uPN organization is translated over to the third-order olfactory neurons, (ii) to what extent the signals encoded by different homotypic uPNs are integrated at synaptic interfaces with the third-order neurons, and (iii) whether the spatial properties of pre-synaptic neurons (PNs) play any role in signal integration by the third-order neurons.

Homotype-specific connections

For the analysis of the interface between homotypic uPNs and third-order neurons, we study the connectivity matrices CPNKC and CPNLHN (see Figure 9, Figure 9—figure supplement 1), which are extracted from the hemibrain dataset (Scheffer et al., 2020). The Cξ (ξ= PN-KC or PN-LHN) is a binary matrix (CX,iξ=0 or 1 dictating the connectivity) of synaptic connectivity between X-th homotypic uPNs and i-th third-order neuron (KC or LHN). It is observed that most of the KCs and LHNs integrate information from multiple homotypes, but that there are also a small number of KCs and LHNs that synapse only with a single homotype (Figure 10).

Figure 9 with 1 supplement see all
A schematic illustrating the connectivity between homotypes (X=A,B,,E) and third-order neurons (i=1,2,,7).

(A) The connectivity matrix C , where CX,i=1 when any uPNs in the X-th homotype and i-th third-order neuron synapses and CX,i=0 otherwise. (B) The number of X-th homotype-specific connections (NX,sp) and the total number of third-order neurons synapsed to any uPNs in the X-th homotype. (C) The common synapse matrix (S) whose element specifies the number of third-order neurons commonly connected between two homotypes. The homotype A is connected to three third-order neurons 1, 2, and 3 (NA,tot=3). Neuron 1 is not synapsing with any other homotype but A, and hence NA,sp=1; similarly, ND,sp=2 (the blue lines depict specific connections). The signals from the two homotypes B and C are shared by the third-order neurons 2, 3, and 5; therefore, SBC=3 in the common synapse matrix S.

Figure 10 with 1 supplement see all
Bar graphs depicting the number of KCs/LHNs that synapse with a specific homotype X (NX,sp, blue) and the percentage of KCs/LHNs that synapse with a specific homotype X (fX=NX,sp/NX,tot, red) at (A) PN-KC and (B) PN-LHN interfaces, with the synaptic weight of N=3 used as the threshold.

The ‘homotype-specific’ connections, defined as the number of third-order neurons that only synapses with a specific homotype but not with the others (see Figure 10 and Methods for more information) can be quantified in terms of the total number of third-order neurons in contact with X-th homotypic uPNs, and it can be obtained by counting the non-zero elements of the matrix C with fixed X. For the case of the PN-KC interface, this number can be obtained from NX,totPNKC=i=11754CX,iPNKC. Specifically, Figure 10A shows NX,spPN-KC and those normalized by NX,totPNKC (fX=NX,spPNKC/NX,totPNKC, see Materials and methods for the detailed algorithms behind the calculation), for all homotypes (X=VM4,VC5,,VP4). Compared to those in KCs, the ‘homotype-specific’ connections are much more prevalent in LHNs (Figure 10). Certain homotypic uPNs, in particular, the hygro/thermo-sensing homotypes are connected to the LHNs which are dedicated to process the signals from hygro/thermo-sensing homotypes (10% of PN-LHN connections made by homotypes).

To address the concern with potential false positives in the detected synapses, we reexamine our results based on the synaptic connectivity with a higher threshold (N=8). Figure 10—figure supplement 1 demonstrates that the homotype-specific connections tend to increase under a more stringent synapse selection criterion, especially in LH. This is most notable in homotypes DM1, DM4, DP1l, and VM6. The existence of these ‘homotype-specific’ third-order neurons suggests that a subset of olfactory processing may rely on the labeled-line strategy that extends beyond the layer of second-order neurons to the higher brain center.

Third-order neurons mediate signal integration

Figure 11A, B show the ‘common synapse matrices’ representing the number of commonly connected third-order neurons between two homotypes X and Y (SXYη with η= KC or LHN), which provide glimpses into the extent of signal integration mediated by KCs and LHNs (see Figure 9C and the caption for how these matrices are constructed from the connectivity matrix).

Figure 11 with 2 supplements see all
Common synapse matrices (A) SKC and (B) SLHN, each of which represents the extent of signal integration from homotypic uPNs to KCs and LHNs.

The color code represents SXY, which is the number of the third-order neurons (KCs or LHNs) synapsing with both homotypes X and Y. The black color is used when there is no third-order neuron-mediated signal integration (SXY=0) happening between two homotypes X and Y. See Figure 9C and its caption for how the common synapse matrices are calculated from the connectivity matrices provided in Figure 9—figure supplement 1.

  1. Overall, the number of synaptic connections between uPNs and KCs is greater than that between uPNs and LHNs (SXYKC>SXYLHN, see Figure 11).

  2. In MB calyx, the signals from food-related odors-encoding homotypes (e.g. Yeasty, Fruity, or Alcoholic Fermentation odor types) are shared by a large number of KCs, which constitute a few large clusters in SKC matrix, depicted in red (SXYKC35) and indicated by the blue arrows on the top in Figure 11A. Some KCs process signals almost exclusively from the hygro/thermo-sensing homotypes without sharing any signal from other homotypes (SXYKC=0 for the cases of X and Y homotype pairs without any signal integration, which are depicted in black in Figure 11). There are also homotypes with significantly less number of overall synaptic connections to KCs, dictated by the diagonal element of the matrix SKC (see Figure 11—figure supplement 2A). In comparison with SLHN, the SKC suggests a stronger but less organized signal integration between heterotypic uPNs by KCs and lends support to the previous literature pointing to the random synapsing of KCs with uPNs at MB calyx (Caron et al., 2013; Stevens, 2015; Eichler et al., 2017; Zheng et al., 2020).

  3. SLHN, on the other hand, demonstrates LHN-mediated signal integration localized to subsets of homotypes. When we collect LHNs connected to a particular homotype and check which other homotypes these LHNs are also synapsing (thereby analyzing the scope of signal integration happening at LH), we find a strong tendency of signals from pheromone and hygro/thermo-sensing uPNs to be integrated within the given odor/signal type (Figure 11). The fact that the pheromone-encoding and hygro/thermo-sensing homotypes share the synaptic connections to LHNs among themselves are demonstrated as the homotype-specific block patterns along the diagonal of the SLHN matrix (see purple and pink arrows on the side in Figure 11B). The SLHN matrix also shows that signals from various food-related odor encoding homotypes, such as DP1l, DP1m, VA2, and VL2p or DM1, DM4, and VA4 are also integrated (see blue arrows in Figure 11B). Many of these homotypes encode signals originating from esters, which is intriguing given the ester-encoding LHN cluster shown by Frechter et al., 2019. The results suggest that certain odor types are processed through common channels of LHNs that are largely dedicated to encoding a particular odor type.

A more stringent selection criterion for synaptic connectivity does not affect our conclusion on the signal integration by the third-order olfactory neurons (Figure 11—figure supplement 2). The only notable change is the general increase in the cases with no integration (SXY=0) in SLHN, especially for hygro/thermo-sensing homotypes. Thus, the extent of signal integration from homotypic uPNs to KCs and LHNs summarized in SKC and SLHN is robust.

Spatial proximity-based versus connectivity-based clustering

Next, we study the relationship between spatial proximity-based clustering and connectivity-based clustering results. Upon visual inspection, the connectivity-based clustering at MB calyx (Figure 12A on the right) appears less structured than the spatial proximity-based clustering (Figure 12A on the left). Specifically, many homotypic uPNs are grouped under a common branch in the tree structure obtained from the spatial proximity-based clustering, whereas such a feature is largely absent in the output of the connectivity-based clustering. Therefore, the spatially well-clustered uPNs at MB calyx (or stereotyped structure) do not necessarily translate into structured connectivity patterns (or stereotyped connectivity), which is consistent with the notion of randomized PN-KC connections (Caron et al., 2013; Stevens, 2015; Eichler et al., 2017; Zheng et al., 2020). In stark contrast to the outcomes for MB calyx, most homotypic uPNs are grouped in the connectivity-based clustering for LH (Figure 12B). This suggests that the spatially proximal uPNs synapse with a similar group of LHNs. The stereotyped organization and stereotyped connectivity of uPNs in LH have been suggested before (Jefferis et al., 2007; Liang et al., 2013; Kohl et al., 2013; Fişek and Wilson, 2014), and we demonstrate such stereotypies are, in reality, expressed throughout LH over all uPNs. In LH, spatial and organizational characteristics of uPNs are well translated to connectivity to LHNs.

Tanglegrams comparing the tree structures generated from the inter-PN distances-based (left) and the connectivity-based clustering (right) (A) between uPNs and KCs, and (B) between uPNs and LHNs.

The same uPNs in the two tree structures are connected with lines, which visualize where the uPNs clustered by one method end up in the clustering results of another. The labels for uPNs are representative of the homotype and are color-coded based on the encoded odor types (Dark green: decaying fruit, lime: yeasty, green: fruity, gray: unknown/mixed, cyan: alcoholic fermentation, red: general bad/unclassified aversive, beige: plant matter, brown: animal matter, purple: pheromones, pink: hygro/thermo).

A quantitative comparison of two trees based on statistical tests lends support to the notion that the spatial organization of uPNs can be indicative of connective properties, most evident in LH (see Appendix 2 for Baker’s Gamma index, entanglement, and cophenetic distance correlation).

Discussion

The inter-PN organization revealed in this study and its association with odor type/valence are reminiscent of the generally accepted notion that the form determines the function in biology. Previously observed stereotypes of neurons in the Drosophila olfactory system were largely based on the differentiation between pheromones and non-pheromones (Ruta et al., 2010; Kohl et al., 2013; Frechter et al., 2019; Das Chakraborty and Sachse, 2021), the whole-cell patch-clamp recording (Seki et al., 2017), and imaging studies suggestive of stimulus-dependent arrangement of neurons in LH (Marin et al., 2002; Wong et al., 2002; Jefferis et al., 2007). Our results are generally consistent with the previous studies, which suggest that a level of stereotypy in uPN organization in MB calyx and LH is universal throughout Drosophila, which can be captured through different metrics and methodologies. In line with Lin et al., 2007, our study finds that homotypes DL2v and DL2d constitute a bilateral cluster in MB calyx (C3MB), and that the dual organization of uPNs is present in MB calyx and LH, such that homotypes DC2, DL1, and VA5 are sorted into the same cluster in LH while sharing similar innervation pattern in MB calyx. Our clustering results in LH share similarities with the NBLAST score-based LH clusters (Bates et al., 2020). The uPNs that ended up in the same cluster or nearby clusters, such as homotypes DM1, DM3, DM4, VA4, and VM3 in the cluster C3LH, are also grouped in the NBLAST score-based clustering analysis (Bates et al., 2020). We find a significant correlation of dαβ with NBLAST score (see Figure 2—figure supplement 1) despite the fact that two metrics prioritize different aspects of neuronal anatomy.

Our inter-PN distances and clustering results suggest the spatial organization of uPNs differs greatly in each neuropil (Figure 5). Some of the tightly bundled organization of uPN homotypes are well preserved throughout the neuropils despite the lack of glomerulus in MB calyx and LH. The spatial segregation between different homotypes is, however, practically not present in MB calyx, leading to a high degree of overlapping. Therefore, the heterogeneity of homotypes at the PN-KC synaptic interface may physically assist the randomized sampling known to exist between uPNs and KCs (Caron et al., 2013; Stevens, 2015; Eichler et al., 2017; Zheng et al., 2020).

Our analysis suggests that LH is compartmentalized into four regions: (1) Posterior-dorsal region primarily occupied by food-related uPNs; (2) Anterior-ventral region occupied by pheromone-encoding uPNs; (3) Biforked bundle surrounding posterior-dorsal region largely occupied by food-related uPNs with an aversive response; (4) Posterior-ventral-medial region occupied by hygro/thermo-sensing uPNs. Previous attempts at identifying regions of odorant space in LH revealed compatible results. The three domains (LH-PM, LH-AM, and LH-AL) suggested by Strutz et al., 2014 seem to be a different combination of our clustering result (LH-PM and LH-AM correspond to the posterior-dorsal region and LH-AL corresponds to a combination of anterior-ventral region and the biforked bundle). Although not perfect, the study of the axo-axonic communities in LH yields results with comparable characteristics (Bates et al., 2020), understandably due to the necessity of inter-neuronal proximity to form synapses. For example, the community 12 by Bates et al., 2020 is predominantly composed of homotypes VP1l and DL5, which resembles our cluster C10LH. The community 6 contains a mixture of homotypes VA5, VC1, D, DA4l, DC2, DA3, and VA7m, which is reminiscent of our cluster C6LH.

Many homotypic uPNs that are spatially localized in LH can be associated with key survival features and a strong innate response (Seki et al., 2017). In this sense, the stereotyped localization of pheromone-encoding uPNs in C8MB, C1LH, and C2LH is of great interest. Our study not only lends support to the existing studies pointing to the labeled-line strategy in the Drosophila olfactory system but also suggests that there may exist an even more sophisticated level of spatial organization, which supersedes the pheromone versus non-pheromone segregation. Interestingly, while the spatial organization of uPNs in LH has a basis on the functionality of the odor type encoded, it does not seem to be directly translated to segregated chemical features seen in LHNs (Frechter et al., 2019). The apparent divergence observed at the PN-LHN interface, coupled with strongly stereotyped connectivity may contribute to a higher resolution of odor categorization.

The uPN organizations from FAFB and hemibrain datasets are consistent

Our analyses of both the FAFB and the hemibrain dataset (Scheffer et al., 2020) find that the results from both datasets are generally consistent. For example, d¯intra, d¯inter, and λ analyzed based on two different datasets are almost identical (see Figure 5A and B). d¯intra,X, d¯inter,X, and λX show slight differences due to a mismatch between the FAFB and the hemibrain dataset (on glomerulus labels and the number of uPNs based on our selection criterion) leading to a different number of uPNs per homotype (Figure 6A and B and Figure 5—figure supplement 1), but the correlation between λX s at MB calyx and LH are still observed (Figure 7C and Figure 7—figure supplement 1C). Most importantly, the clustering results are similar, where many spatial clusters in both datasets share the same set of homotypes. Additionally, odor type-dependent spatial properties are retained (Figure 8 and Figure 8—figure supplement 1), with all statistical tests supporting our hypothesis. In conclusion, the outcomes from our analyses of the two EM datasets lend support to the previous claims of stereotypy in the Drosophila brain and neuronal structures (Jenett et al., 2012; Jeanne et al., 2018; Schlegel et al., 2021).

Odor signal processing and labeled-lines

Our study suggests that while the primary connectivity motif of third-order olfactory neurons indeed integrates signals, there still exist several labeled lines. The synaptic connections at the PN-KC interface in MB calyx are largely integrative and randomized - with an exception of hygro/thermo-sensing homotypes that display stereotypy even in terms of the connectivity to the KCs. A similar observation has been made by Li et al., 2020, who employed NBLAST score to identify a structural segregation between odor-encoding and hygro/thermo-sensing homotypes. They found that specific KC types are preferentially targeted by hygro/thermo-sensing homotypes. Marin et al., 2020, who carried out connectome analysis specific to hygro/thermo-sensing homotypes, also identified that lateral accessory calyx (lACA), the anterior-dorsal part of the calyx, are primarily targeted by hygro/thermo-sensing homotypes (analogous to our clusters C1MB and C4MB in Figure 3), and found that a number of KCs are dedicated to encoding signals from these homotypes. The uPNs in LH are spatially segregated, which translates to connectivity in three different levels. First, certain LHNs are dedicated to encoding signals from a specific homotype. The number of these ‘homotype-specific’ LHNs varies across the homotype and can make up a significant portion of PN-LHN connections depending on the homotype (Figure 10). Second, synaptic connectivity maps between uPNs and LHNs indicate odor type-dependent integration occurs at LH (Figure 11B). Channels of LHNs predominantly encoding specific odor types are observed; one primarily integrates responses from certain food-related homotypes, one integrates pheromone-encoding homotypes, and another integrates hygro/thermo-sensing homotypes. Third, homotypic uPNs share similar connectivity to LHNs, unlike those in MB calyx. The signals relayed from the spatially well-organized (or tightly bundled) homotypes are localized into a specific group of LHNs, thereby forming a ‘homotype-specific’ connectivity motif (Figure 10, Figure 11, and Figure 12).

In our study of the labeled-line strategy, we made several interesting observations, which are worth comparing with the concept of ‘fovea’ introduced by Zheng et al., 2020. A ‘fovea’ delineates deviations between experimentally observed connectivity matrices and connectivity under the assumption of random synapses in MB calyx, specifically for certain food-related uPNs (Zheng et al., 2020). A group of common KCs predominantly sampling ’food-related’ uPNs manifest themselves in the common synapse matrix SKC (see the group of homotypes comprising the clusters, highlighted by the blue arrows in Figure 11A). A subset of homotypic uPNs under the food-related ‘fovea’ reported by Zheng et al., 2020 are also spatially clustered (e.g. DM1, DM4, DP1m, DP1l, VA2, and VA4). While most of these homotypes are spatially proximal (the vast majority of the uPNs are located in clusters C6MB and C7MB), some homotypes under the food-related ‘fovea’ such as VA2 are sampled from spatially disparate clusters. Thus, it is likely that factors other than the spatial organization of uPNs in neuropils contribute to creating the ‘fovea’. Interestingly, the spatial proximity of pheromone-encoding homotypes in MB calyx may suggest the existence of pheromone-encoding ‘fovea,’ but most uPNs in these homotypes do not converge in connectivity-based clustering with an exception of VA1d. In fact, we suspect the spatial organization of pheromone-encoding homotypes in MB calyx, which is placed at the center of the neuropil, to facilitate the observed randomization of connections by increasing the accessibility of KCs to these homotypes. There is, however, a potential hygro/thermo ‘fovea,’ where homotypes such as VP1d and VP2 are spatially clustered together and the signals from these homotypes are relayed by the same set of KCs. Curiously, VL1 is part of this hygro/thermo ‘fovea’ (Figure 12A).

Multiglomerular PNs are spatially distinctive

Apart from uPNs primarily explored in this study, a host of local neurons (LNs) and multiglomerular PNs (mPNs) also constitute sophisticated neural circuits to regulate the signals received from ORNs (Sudhakaran et al., 2012; Bates et al., 2020), playing a significant role in the olfactory signal processing (Olsen et al., 2010; Jeanne and Wilson, 2015; Seki et al., 2017). A large portion of these mPNs is GABAergic and inhibitory (Berck et al., 2016; Tobin et al., 2017), whereas the role of interneurons can be both inhibitory and excitatory (Wilson et al., 2004; Turner et al., 2008). Electrophysiological measurements indicate that mPNs are narrowly tuned to a specific set of odor stimuli (Berck et al., 2016), which is significant given that PNs are generally thought to be more broadly tuned than presynaptic ORNs (Wilson et al., 2004). Several PNs do not follow the typical mALT, but mediolateral antennal lobe (mlALT) or lateral antennal lobe tracts (lALT) instead, thereby bypassing innervation through one of the higher olfactory centers (Schultzhaus et al., 2017; Zheng et al., 2018; Bates et al., 2020). As stated previously, we confined ourselves to uPNs innervating all three neuropils to compare the spatial organization across neuropils for each uPN. As a result, 28 uPNs present in the FAFB dataset are not explored in our study. In MB calyx, only two uPNs constituting VP3 were dropped as a result of our selection criterion, which ended up in an almost identical clustering output once hierarchical clustering was performed on the entire 137 uPNs that innervate MB calyx. Two missing uPNs were grouped into clusters C4MB and C6MB, along with other hygro/thermo-sensing homotypes. On the other hand, the addition of 27 uPNs constituting 15 homotypes innervating LH but not MB calyx created four new clusters when hierarchical clustering was performed (Figure 4—figure supplement 1). The additional uPNs changed the content of the individual clusters; that is, the tree-cutting algorithm broke down a few clusters that became larger due to the additional uPNs. Furthermore, when we calculate the d¯intra, d¯inter, and λ in LH for the 15 homotypes that included the 27 uPNs, we find that the d¯intra values increased when the 27 uPNs were included (see Figure 5—figure supplement 2). This suggests that the previously removed uPNs, most of which follow mlALT, are significantly different in terms of spatial and organizational characteristics and thus should be analyzed separately. Out of 27 additional uPNs in LH, 21 were in mlALT, 5 were in trans-lALT, and 1 was in mALT. Figure 4—figure supplement 2 illustrates how these 27 uPNs innervate LH which demonstrates the reason behind increased d¯intra values. These 27 uPNs are mostly GABAergic (21 are labeled as GABAergic, 1 as cholinergic, and 4 as unknown neurotransmitter type), covering 84% of all GABAergic uPNs available in the FAFB dataset. These uPNs innervate LH differently from other uPNs in the same homotype that follow mALT (see homotypes such as DA1, DC4, DL2d, DL2v, DP1l, VA1d, VA1v, VL2a, VL2p, and VP5 in Figure 4—figure supplement 2). Morphologically, inhibitory GABAergic neurons are often considered ‘smooth’ and aspiny (Douglas et al., 1989; Bopp et al., 2014; Gouwens et al., 2019), which are discernible from Figure 4—figure supplement 2.

The single-uPN homotypes may have different morphological properties

It is of great interest that many of the single-uPN homotypes, characterized by densely branched morphology, encode signals with aversive responses. Direct transmission of the associated signals across the three neuropils via a single PN might simplify the overall processing of the olfactory signals as well as reduce the energetic cost. Similarly, the morphological characteristics of uPN innervation at each neuropil are intriguing. Even though a structural difference exists between the single-uPN and multi-uPN homotypes, all uPN innervations within neuropil share a similar morphology regardless of the homotype (see Figure 6—figure supplement 1; Choi et al., 2022). A localized morphological diversity within a neuron may be a characteristic aspect of pseudo-unipolar neurons like uPN and suggests a fundamentally multi-scale characteristic of neuron morphology.

The Drosophila brain EM reconstruction project has evolved to its near completion since the EM image dataset was first released (Dorkenwald et al., 2022). The reconstruction of the majority of the Drosophila central brain as well as the corresponding connectome with detailed information of the individual synapses has become publicly available (Scheffer et al., 2020). Our analysis of the second-order neurons inside the Drosophila olfactory system may be translated to other parts of the nervous system in Drosophila as well as different organisms including the central nervous system (CNS) of humans. For the mammalian olfactory system, the details of analyses must be adapted, however, since the wiring scheme is much more complex than that of an insect (Maresh et al., 2008). For example, multiple glomeruli encoding the same olfactory signal exist in humans (Mombaerts et al., 1996). When analyzing the spatial properties, this can be accounted for by prioritizing the individual glomerulus over the homotypes. Then, homotypic PNs forming different glomeruli may be compared or averaged if one were to consider the homotype-dependent characteristics. According to the neurotransmitter map from a recent study (Dolan et al., 2019), sophisticated processes beyond neuronal anatomy are apparently at work in the olfactory signal processing. Thus, functional studies incorporating odor response profiles in PNs (Badel et al., 2016) and ORNs (Münch and Galizia, 2016; Bak et al., 2018) would supplement our findings. The extension of our study to the other regulatory interneurons and mPNs, morphological studies of second-order neurons, and spatial analysis of third-order neurons will be of great interest for a better understanding of the olfactory signal processing beyond the implication of the neural anatomy and connectivity studied here.

Materials and methods

Data preparation

Request a detailed protocol

We used the neuron morphology reconstruction of 346 Drosophila olfactory neurons from the FAFB dataset (Bates et al., 2020) traced from EM images. The neurons were extracted from the right hemisphere of the female Drosophila. Out of 346 neurons in the FAFB dataset, 164 neurons were uPNs. One uPN in the dataset (neuron ID = 1356477 forming VP3) did not have an associated reconstruction (.swc file) available and was therefore ignored. For this study, uPNs that innervate all three neuropils were chosen because our aim is (1) to compare spatial characteristics of the uPN innervation across each neuropil and (2) to classify each uPN based on the odor encoding information. Thus, out of 164 uPNs, a total of 135 uPNs constituting 57 homotypes were collected under this criteria, resulting in mostly cholinergic uPNs that follow mALT. Rest of the uPNs that did not innervate all three neuropils are collected for the supplementary analysis. In MB calyx, a total of 137 PNs are identified with two PNs constituting VP3 that do not innervate all three neuropils. On the other hand, in LH, a total of 162 PNs are identified, indicating that 27 PNs constituting 15 homotypes do not innervate all three neuropils. The morphological information of each neuron is stored as a set of 3D coordinates with the connectivity specified with the parent nodes. Complete reconstruction of neuron morphology was made by connecting data points based on their parent-child relationship.

The hemibrain dataset (Scheffer et al., 2020) was taken from the neuPrint database (Clements et al., 2020), where we collected from the right hemisphere of the female Drosophila a total of 120 uPNs forming 58 glomeruli based on the same criterion we used for the FAFB dataset (uPNs that innervate all three neuropils). Unlike the FAFB dataset, the neurons in the hemibrain dataset are labeled with regions of interest (ROI), which are used to query uPNs conforming to our selection criterion. The discrepancy in the number of uPNs between the two datasets most likely resulted from the difference between the neuropil boundary we used and the region defined by the hemibrain dataset. In fact, we find that the total number of uPNs in both datasets is comparable, with 164 uPNs in the FAFB dataset and 162 uPNs in the hemibrain dataset. The two datasets also had a minor mismatch in the glomerulus label annotations, sometimes affecting the number of uPNs constituting a given homotype. Among the 120 uPNs from the hemibrain dataset, five uPNs had ambiguity in terms of their glomerulus labels, which is presumably due to poorly formed glomerular structures. For these uPNs, we adopted the glomerulus labels of the FAFB dataset with the matching hemibrain neuron IDs.

Additionally, a recent community-led effort identified three glomeruli in both databases with conflicting glomerulus labels, which have been a source of confusion (Schlegel et al., 2021). After an extensive study, the community agreed to rename the glomeruli in both datasets labeled as VC3l, VC3m, and VC5 as VC3, VC5, and VM6, respectively (Schlegel et al., 2021). Thus, we have manually incorporated these labels into our analyses for both the FAFB and the hemibrain dataset.

Next, we systematically demarcated the regions of AL, LH, and MB calyx. The density of data points projected to each axis was used for the identification since the neuropils are featured with a much higher density of data points than the rigid backbone connecting them. The boundaries defining each neuropil were systematically chosen from local minima that separate neuropils from rigid backbones. Due to the unique structure of uPNs, sometimes the projection along a given axis cannot fully differentiate two neuropils. To resolve this issue, projections along each axis were sampled while rotating the data points along the reference axes at 5 increments to obtain multiple snapshots. The densities were analyzed to choose the optimal degrees of rotation along the reference axes that could best segment the neuropils. We used the smallest average and deviation value of density at the local minima as the criteria to choose the optimal rotation. The process has been repeated for each neuropil to identify a set of boundaries along multiple transformed axes with various degrees of rotations that optimally confine each neuropil. This information has been combined to create a set of conditions per neuropil for segmentation. The resulting neuropils were confirmed through visual inspection. We compared our neuropil segmentation boundaries with neuropil volume surface coordinates provided by Ito et al., 2014 via CATMAID (Saalfeld et al., 2009) and found the boundaries are comparable (data not shown). An overview of the segmentation process is available in Figure 13.

A diagram depicting the neuropil segmentation process.

The data points from skeletal reconstruction are projected to each axis to generate distributions from which local minima are obtained. The process is repeated while rotating the uPNs along each axis. A collection of histograms and corresponding local minima are surveyed to generate a set of optimal rotations and boundaries for individual neuropil. The resulting parameters are combined to form a collection of conditions to segment each neuropil.

The odor type and odor valence information were extracted from various literature (Hallem et al., 2004b; Galizia and Sachse, 2010; Stensmyr et al., 2012; Mansourian and Stensmyr, 2015; Badel et al., 2016; Bates et al., 2020) and we closely followed the categorical convention established by Mansourian and Stensmyr, 2015 and Bates et al., 2020. However, we note that the categorization of a uPN under a specific odor category may overshadow the complete spectrum of odorants a uPN might encode, especially if the uPN encodes ORs that are broadly tuned. Therefore, we focused on the well-separated pheromone/non-pheromone encoding types and valence information.

To test our labeled-line hypothesis, the connectivity information between uPNs and higher olfactory neurons such as KCs and LHNs was necessary. Since only the hemibrain dataset contains detailed connectivity information, all of our connectivity studies are done using uPNs, KCs, and LHNs queried from the hemibrain dataset. We chose KCs and LHNs that made at least three synaptic connections with any of the 120 uPNs from the hemibrain dataset. This resulted in 1754 KCs and 1295 LHNs, creating bipartite connectivity matrices at each neuropil.

Inter-PN distance

Request a detailed protocol

The ‘distance’ dαβ between two neurons, α and β, with different lengths (NαNβ) is quantified by calculating.

(1) dαβ2=1Nαi=1Nαmin[(riαrjβ)2],

where riα is an i-th coordinate forming the neuron α. Equation 1 is evaluated over all pairs of riα and rjβ with j=1,,Nβ that gives rise to the minimum value. This means that when NαNβ, for every i-th coordinate in the neuron α (riα), we find j-th coordinate in the neuron β (rjβ) that is the closest to riα. Then, the spatial proximity of a given pair of neurons is assessed by the dαβ that denotes the average of all the minimum Euclidean distances between the pair of coordinates.

The degree of bundling, packing, and overlapping

Request a detailed protocol

We define the mean intra- and inter-homotype distances as.

(2) d¯intra,X1Nα,βXNdαβ

and

(3) d¯inter,X1NαX,βXNdαβ,

where X denotes a homotype and N is the total number of uPN pairs to be averaged. The d¯intra,X is calculated over all the pairs of uPNs in the X-th homotype, quantifying the tightness of bundling of uPNs constituting the X-th homotype. On the other hand, d¯inter,X is calculated over the pairs of uPNs between α-th uPN belonging to the X-th homotype and β-th uPN in the Y-th homotype (YX), such that it measures the extent of packing of uPNs around the X-th homotype. The degree of overlapping for the X-th homotype, λX, is defined as the ratio of average intra- and inter-homotype distances,

(4) λX=d¯intra,Xd¯inter,X,

which represents how clearly the X-th homotype is segregated from other homotypes in a given space. A large value of λX (λX1) implies that the space spanned by the X-th homotype is not clearly discerned from other homotypes.

Spatial clustering of projection neurons

Request a detailed protocol

Hierarchical/agglomerative clustering was used to cluster the uPN innervation at each neuropil using the pairwise dαβ matrices. First, the linkage was decided based on the pairwise distance matrix built with the Farthest Point Algorithm (or ‘complete’ method), where uses the maximum distance between neurons to define the distance between two clusters. This criterion is used to build hierarchical relations (or nested clusters) in a bottom-up approach where each neuron is treated as a cluster at the beginning. The result is a fixed tree structure of individual neurons from which the finalized clusters are formed using an optimal tree-cutting algorithm. In the dendrogram from AL (Figure 2—figure supplement 3), homotypic uPNs are grouped together with high accuracy, suggesting our distance metric dαβ is adequate. We tested various tree-cutting criteria such as elbow method, gap statistics, maximum average silhouette coefficient, and dynamic hybrid cut tree method (Langfelder et al., 2008) to determine the optimal number of clusters. Among them, we selected the dynamic hybrid cut tree method, since it performed the best in giving the cluster number closest to the number of different odor types (which is 10) (Table 1). We deployed the dynamic hybrid cut tree method with the minimum cluster size of 4 neurons for the tree-cutting, following the neuron clustering procedure used by Gouwens et al., 2019.

Table 1
The optimal number of clusters of uPNs in the FAFB dataset determined by employing the dynamic hybrid cut tree method, elbow method, gap statistics, and maximum average silhouette coefficient.
Dynamic hybridElbowGapSilhouette
AL1914854
MB calyx101172
LH11972

Pearson’s x2-test of independence

Request a detailed protocol

The association between two categorical variables is assessed using Pearson’s χ2-test. For the test, a contingency table, which lists the categorical frequency of two variables, is created. For example, Oij of the i- and j-th element of the contingency table shown below is the frequency counting the putative valence i=1 (attractive), 2 (aversive), 3 (unknown), and the number of uPNs in one of the 10 clusters in MB calyx with j=1 (C1MB), 2 (C2MB), ... , 10 (C10MB).

C1MBC2MBC3MBC4MBC5MBC6MBC7MBC8MBC9MBC10MBTotal
Attractive04010541111844
Aversive1200412988347
Unknown478565123344
Total51386102214212214135

Then the χ2 value is evaluated based on the table using.

(5) χ2=i=1Rj=1C(OijEij)2Eij,

where R and C are the numbers of rows and columns, and Oij and Eij are the observed and expected frequencies of the event in the i-th row and j-th column, respectively. Eij is calculated from Oij as.

(6) Eij=Npipj,

where pi=jCOij/N and pj=iROij/N with N being the total count. Thus, Eij is the frequency expected by assuming that the two categorical data are statistically independent. Pearson’s χ2 test aims to check whether there is a significant difference between Oij and Eij.

In the χ2-test, the p-values are estimated using fk(x), the χ2-distribution with the degree of freedom k=(R-1)(C-1). If the test returns a χ2 value that gives rise to a p-value smaller than the defined significance level (α=0.01), the null hypothesis of independence between the two data sets should be rejected. As a result, the distribution of the categorical data is deemed significantly different from a randomly generated distribution, which concludes that the association between two sets of data is statistically significant.

For the above contingency table with k=18, which leads to χ266.1 (Equation 5), we get a p-value much smaller than the significance level (α=0.01), p=10χ2fk=18(x)dx2.016×107α=0.01.

When Pearson’s χ2 statistics are available, one can calculate Cramér’s V with bias correction, a measure of association between two categorical variables, as follows.

(7) V=ϕ2/Nmin(R1,C1),

where ϕ2=max(0,χ2/N-(R-1)(C-1)/(N-1)), R=R-(R-1)2/(N-1), and C=C-(C-1)2/(N-1). Similar to the Pearson correlation coefficient, the value V ranges between 0 and 1 where 0 indicates no correlation and 1 indicates a complete correlation between two categorical variables.

Mutual information

Request a detailed protocol

Mutual information (I) is used to verify the significance of association between nominal variables observed in Pearson’s χ2-test for independence. The I measures the information transfer or the similarity between two data. The concept can be extended to clustering outputs to check how two different clustering labels from the same data are similar to each other. Traditionally, the I between two jointly discrete variables A and B is given by.

(8) I(A;B)=i=1nAj=1nBP(Ai,Bj)log[P(Ai,Bj)P(Ai)P(Bj)],

where nA (or nB) is the number of clusters in A (or B). Numerically, the I between two clustering outputs A and B is calculated by evaluating P(Ai)=NAi/N, P(Bi)=NBi/N, and P(Ai,Bj)=NAiBj/N where N is the total count and NAiBj is the number of elements common in both clusters Ai and Bj.

The significance was assessed by comparing the observed I with the distribution of I s from randomly sampled variables. Specifically, the cluster label was randomly sampled 1000 times to generate a distribution of I under the assumption of independence. The value of observed I is considered significant if the approximated p-value is below 0.01 (p< 0.01).

Analysis of synaptic interfaces

Request a detailed protocol

We conducted three different analyses on the synaptic interfaces of uPNs with the third-order neurons (KCs or LHNs) from the hemibrain dataset.

(i) The ‘homotype-specific’ connections (NX,spξ with ξ= PN-KC or PN-LHN) are obtained by counting the number of third-order neurons that synapse with a homotype X but do not synapse with any other homotypes, the information of which is provided by the binarized connectivity matrix C. The total number of synaptic connections for a homotype X is simply the sum of the row of the connectivity matrix C (NX,totξ=Σi=1NξCXi).

(ii) To generate the S matrices, we counted the number of third-order neurons synapsing with a given homotype X that also synapses with other homotypes.

(iii) The tanglegram study required a hierarchical clustering of uPNs based on their connectivity to third-order neurons. The distances between uPNs in the connectivity matrix C represent the similarity of the connectivity patterns to third-order neurons between two uPNs. We utilized the metric of cosine distance, which is widely used for analyzing the connectivity matrix (Bates et al., 2019; Bates et al., 2020; Li et al., 2020; Eschbach et al., 2020; Schlegel et al., 2021). The cosine distance is defined as.

(9) dcos=1uv|u||v|,

where u and v are two vectors to be compared. After calculating the distances, we performed hierarchical clustering by Ward’s criterion, which minimizes the variance of merged clusters, to generate the tree structure. The results of hierarchical clustering using the spatial proximity (dαβ) and connectivity (dcos) are compared using a tanglegram (Figure 12) after untangling two trees using the ’step1side’ method (Galili, 2015).

Appendix 1

Monte Carlo approach to independence test

In this section, we describe an alternative method to the independence test inspired by the Monte Carlo significance test (Hope, 1968) to further support our Pearson’s χ2-test of independence. The procedure is as follows: (1) For a given contingency table, randomize the observation such that the marginal sum of each row remains the same as the observed contingency table. That is, for each row, randomize the vector with integers while the sum of the vector stays the same as the observed contingency table. This procedure randomly shuffles the distribution of the clusters while keeping the distribution of a particular categorical variable intact. (2) Calculate the χ2 value from the randomized contingency table. (3) Repeat steps 1 and 2 for 1000 times to generate a distribution of the χ2 values. (4) Obtain the mean and the standard deviation of χ2 values. The distribution of χ2 values is approximately normal. (5) If the χ2 value from the observed contingency table is more than 4σ different from the distribution, we consider the observed χ2 value statistically significant and reject the null hypothesis. Whenever we ran a Pearson’s χ2-test, we performed the above procedure alongside. The output of this procedure supported whichever conclusion we drew from Pearson’s χ2-test.

Identifying the agreement between two categorical data via mutual information

We verified our Pearson’s χ2-test of independence of two categorical variables by calculating the mutual information I (see Methods). In the FAFB dataset, the mutual information between glomerular labels and dαβ-based clustering output in MB calyx was equal to I(glo;CMB)=1.892, which is significantly (more than 4σ) different from the mean of randomly sampled I distribution, I(glo;CMB)rand=1.386±0.035. This result is consistent with our χ2-test, as the mutual information of the observed variables is significantly larger than the mutual information under the assumption of random sampling, suggestive of a statistically significant association between glomerular labels and MB calyx cluster labels. In LH, the mutual information between glomerular labels and dαβ-based clustering output was I(glo;CLH)=2.128 which deviated 4σ or more from the mean of the randomly sampled I distribution, I(glo;CLH)rand=1.466±0.035.

The same method is applied to confirm that a statistically significant association exists between odor type and the clustering outputs, with I(odor;CMB)=0.819 and I(odor;CLH)=0.963, all of which differ by more than 4σ from the means of the randomly sampled I distributions, I(odor;CMB)rand=0.337±0.044 and I(odor;CLH)rand=0.372±0.043. For odor valence, we obtain I(val;CMB)=0.277 and I(val;CLH)=0.326, where both I(val;CMB) and I(val;CLH) differ significantly from the means of the randomly sampled I distributions, I(val;CMB)rand=0.073±0.026 and I(val;CLH)rand=0.081±0.026. Overall, the conclusion drawn from the association study based on mutual information is identical to Pearson’s χ2-test.

Appendix 1—table 1
Pearson’s χ2 tests of independence of variables in the FAFB dataset.

CZ indicates cluster labels from dαβ-based clustering in Z neuropil. Cramér’s V values are displayed on each cell and the corresponding p-values are shown in parentheses.

CLHGlomerular LabelsOdor TypeOdor Valence
CMB0.502 (1.149E-36)0.610 (1.255E-27)0.401 (3.303E-21)0.425 (2.016E-07)
CLH0.671 (2.266E-40)0.416 (1.980E-22)0.455 (2.586E-08)
Appendix 1—table 2
Pearson’s χ2 tests of independence of variables in the hemibrain dataset.

CZ indicates cluster labels from dαβ-based clustering in Z neuropil. Cramér’s V values are displayed on each cell and the corresponding p-values are shown in parentheses.

CLHGlomerular LabelsOdor TypeOdor Valence
CMB0.495 (6.635E-34)0.577 (3.461E-25)0.425 (9.400E-18)0.463 (6.283E-07)
CLH0.685 (1.523E-40)0.502 (6.072E-29)0.521 (2.932E-09)
Appendix 1—table 3
Mutual information (observed mutual information (top), randomly sampled mutual information (bottom) in each cell) from the association study using the FAFB dataset.

CZ is cluster labels from dαβ-based clustering at Z neuropil. The observed mutual information differs from the randomly sampled mutual information by more than 4σ.

CLHGlomerular LabelsOdor TypeOdor Valence
CMB1.076 

0.397±0.045
1.892 

1.386±0.035
0.819 

0.337±0.044
0.277 

0.073±0.026
CLH2.128 

1.466±0.035
0.963 

0.372±0.043
0.326 

0.081±0.026
Appendix 1—table 4
Mutual information (observed mutual information (top), randomly sampled mutual information (bottom) in each cell) from the association study using the hemibrain dataset.

CZ is cluster labels from dαβ-based clustering at Z neuropil. The observed mutual information differs from the randomly sampled mutual information by more than 4σ.

CLHGlomerular LabelsOdor TypeOdor Valence
CMB1.371

0.710±0.048
2.244 

1.783±0.033
1.036 

0.527±0.047
0.336

0.124±0.035
CLH2.344 

1.717±0.034
1.211 

0.493±0.048
0.434 

0.116±0.033
Appendix 1—table 5
Statistics of homotypes composed of a single uPN (or multiple uPNs) in the FAFB dataset and the corresponding putative valence.
AversiveAttractiveUnknownTotal
Single uPN Homotypes Count74213
Multiple uPN Homotypes Count18131344
Total25171557
Appendix 1—table 6
Statistics of homotypes composed of a single uPN (or multiple uPNs) in the hemibrain dataset and the corresponding putative valence.
AversiveAttractiveUnknownTotal
Single uPN Homotypes Count75113
Multiple uPN Homotypes Count18121545
Total25171658

Appendix 2

Testing the labeled-line hypothesis

We detail the analyses performed on the tanglegram and the respective outputs (Figure 12). First, we applied the dynamic hybrid cut tree method on the dendrogram generated from connectivity and conducted Pearson’s χ2 test. The results are shown in Table 1. The p-values for the connectivity-based clustering between uPNs and LHNs for glomerular labels, odor types, and odor valence were very small. For the connectivity between uPNs and KCs, we see a moderate to no association for the given categorical variables (Appendix 2—table 1).

The similarity between two tree structures from spatial proximity-based and connectivity-based clustering at a given synaptic interface is measured in several different ways to provide a comprehensive comparison. First, we quantified the similarity using Baker’s Gamma index (Baker, 1974), which is a measure of rank correlation (or ordinal relation) calculated from concordant and discordant pairs given by.

(10) GBaker=NconNdisNcon+Ndis,

where Ncon is the number of concordant pairs (the ordering of elements in two trees match) and Ndis is the number of discordant pairs (the ordering of elements in two trees do not match). Baker’s Gamma index ranges from -1 to 1 where 0 represents the ordering of two trees is completely dissimilar and 1 or -1 indicate the ordering of two trees match. We find GBakerMB=0.286 and GBakerLH=0.219 (which we double-checked using both the in-house code and ‘dendextend’ package in R). Baker’s Gamma index for LH is very similar to the one obtained by Bates et al., 2020 (GBakerLH=0.21), who conducted a similar study using the NBLAST score and connectivity. However, the fact that GBakerMB>GBakerLH when the tanglegram of MB calyx is seemingly more incoherent (Figure 12A) raises a question of whether Baker’s Gamma index alone is enough to describe the tanglegram.

Apart from the ordinal relations between two sets of leaves, we employed two additional metrics to compare the two trees: (1) entanglement, a measure spanning from 0 to 1 quantifying the number of lines crossing, and (2) cophenetic distance correlation, a measure spanning from 0 to 1 quantifying how similar the two branching structures are. The entanglement between two trees for MB calyx was 0.35 (higher entanglement), while the entanglement for LH was 0.26 (lower entanglement), which agrees with Figure 12. To calculate cophenetic distance correlation, we measured the pairwise cophenetic distances within each tree and calculated the Pearson correlation coefficient. The cophenetic distance between two leaves in the dendrogram is equal to the minimum distance (or height) to the branching point that contains both leaves. The Pearson correlation coefficient between cophenetic distances of the spatial proximity-based and connectivity-based tree structures was r=-0.032 (p>0.001) for MB calyx and r=0.236 (p0.001) for LH, reflecting the less disrupted tree structure in LH compared to MB calyx.

Appendix 2—table 1
Pearson’s χ2 tests of independence of variables on the connectivity-based clustering results.

Cramér’s V values are displayed on each cell and the corresponding p-values are shown in parentheses. Bold entries are used to specify statistically significant results.

Glomerular LabelsOdor TypeOdor Valence
CPNKC0.433 (2.472E-08)0.316 (9.978E-09)0.271 (0.012)
CPNLHN0.765 (1.410E-67)0.630 (1.519E-48)0.604 (4.055E-12)

Data availability

All data generated during this study and Python script are available in Drosophila Olfaction-main.zip included as the supporting file. They are also available at https://github.com/kirichoi/DrosophilaOlfaction, (copy archived at swh:1:rev:91dd60f4231a58590e2571e72b660c5dfee261b6).

The following previously published data sets were used

References

  1. Conference
    1. Buccino AP
    2. Ness TV
    3. Einevoll GT
    4. Cauwenberghs G
    5. Hafliger PD
    (2018) A deep learning approach for the classification of neuronal cell types
    Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering in Medicine and Biology Society. Annual International Conference. pp. 999–1002.
    https://doi.org/10.1109/EMBC.2018.8512498

Decision letter

  1. Sonia Sen
    Reviewing Editor; Tata Institute for Genetics and Society, India
  2. K VijayRaghavan
    Senior Editor; National Centre for Biological Sciences, Tata Institute of Fundamental Research, India
  3. Sonia Sen
    Reviewer; Tata Institute for Genetics and Society, India
  4. Alexander Shakeel Bates
    Reviewer; Harvard Medical School, United States
  5. Ann-Shyn Chiang
    Reviewer; National Tsing Hua University, Taiwan

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

Decision letter after peer review:

Thank you for sending your article entitled "Olfactory responses of Drosophila are encoded in the organization of projection neurons" for peer review at eLife. Your article is being evaluated by 3 peer reviewers, one of whom is a member of our Board of Reviewing Editors, and the evaluation is being overseen by K VijayRaghavan as the Senior Editor.

While appreciating the authors' work presented in this manuscript there were two major concerns that came up during the consultations:

1. Given that two EM datasets exist for the fly brain, determining the generalisability of what the authors describe should have been done.

2. Making claims about labelled line representation of odours or not using morphometric data alone is not justified when connectivity data exists for flies.

Based on this, and other comments, the essential revisions that were discussed were the following:

The writing: The text needs to be made more accessible to the reader, perhaps by providing more intuitive explanations, biological context, and emphasise their findings more than they currently do (which emphasises methods used more). We would also like to see their analyses in the context of what's been done before.

Reproducibility: Since the EM data in this analysis is based on a single fly, we recommend testing whether the authors' findings will be consistent across individuals. So, could the authors test their analysis on the hemibrain alongside the extant FAFB?

Number of uPNs used: The authors use 111 uPNs for their analysis whereas the number of uPNs is 164. (This might be why they obtain only 31 glomeruli). It's not clear why the rest were excluded. Could the authors conduct their analyses by including all cholinergic uPs (including those involved in hygro and thermo sensation)? (We assume the authors have omitted GABAergic neurons from their analyses because they do not innervate the MB? Including it would provide important biological insight, but if not, their reasoning needs to be stated.)

Hierarchical clustering: We are concerned that the inferences rely on a few clusterings that don't appear to be very robust. This is usually okay, but a lot of the paper is built around this. So, could the authors reassess their hierarchical clustering parameters and the normalisation for the λ value and state them explicitly for the reader?

Labelled line hypothesis: The claims made in this study about the labelled line hypothesis cannot be inferred from morphometrics alone. For this, the authors would need to compare these findings with connectivity. This could even be done at a small scale to understand whether their morphological groups also have more specific or segregated connectivity within the LH / AB.

(A possible pipeline for this could be: Use neuprint-python to pull connection data for each chosen PN and subset it to the select brain regions. They could then perform hierarchical clustering on that, and devise groupings that can be compared to the morphological grouping visually using a tanglegram. They might then compare the two using Baker's Γ Index or similar.)

In addition to this, please read the other (not essential) recommendations and if any of them are feasible to respond to, please do so.

Reviewer #1 (Recommendations for the authors):

The olfactory PNs have been studied using numerous approaches. In this study, the authors use a parameter called 'inter PN distance' to assess the relationship between all the uniglomerular PNs in the hemibrain connectome in the AL, MB, and LH. As odour information from the sensory neurons is brought into the AL via OSNs classes that innervate individual glomeruli, PNs that innervate the same glomeruli are considered homotypic. Using the inter PN distance and downstream statistical analyses, the authors infer that odour information is discretely represented in the AL and less so in the LH; however, in the MB, there is considerable overlap in odour representation.

The manuscript is well written and presented.

- PN information representation across the AL, MB, and LH has been explored before. While the authors do mention this, it would be useful if they discussed what each one of those approaches inferred about olfactory representation, and why they felt their approach was better. In the same vein, it would be nice if they finally concluded by comparing their findings with the others.

- The manuscript is statistically dense. It will help readers if the authors provided some intuitive explanations along the way – for example, for inter-PN distance, among others.

- Many of the legends are not very explanatory. Could the authors please elaborate on them? It would also help to have pointers in the figures to focus the attention of the reader. For example, the legend for S2 can be expanded. I can't tell which homotypic PNs are merging into a single cluster. There are similar issues in other figures as well – could the authors please add visual aids such as arrows in the figures?

Reviewer #2 (Recommendations for the authors):

Choi et al. competently quantify known aspects of excitatory, uniglomerular projection neuron (uPN) morphology, in order to demonstrate how neurons of this well-studied cell class differ in the three main brain regions in which they arborise. Classically, these three regions support three different functions: siloed olfactory channel processing (antennal lobe), ethological odour classification (lateral horn), and associative memory (mushroom body). They find that uPNs of the same cell type are morphologically similar in each region, but the nature of that similarity, and their similarity with other uPN cell types, differ as appropriate for the function of the neuron in that region. While not novel, their quantitive description contributes to the field. Their 'major' conclusion on labeled line processing is, however, insufficiently substantiated given that the olfactory system of the fly is post-connectomic and this question is best taken up with connectionist analysis. They also only use a single dataset from a single fly (though a second is available) and do not explore how robust their results are to other clustering approaches/decisions, making it unclear how general their description is.

The strength of this work is that it is able to use quantitative methods to demonstrate features of Drosophila olfactory projection neuron (PN) anatomy. To date, these features have been known and used in classification work on the PNs (e.g. Tanaka et al. 2012; Bates et al. 2020; Zheng et al. 2018; Bates et al. 2020; Jefferis et al. 2007). They have also been shown through (PN->target) connection analyses (Eichler et al. 2017; Schlegel et al. 2021). The findings are therefore largely not novel but are a good quantitative description and a nice synthesis in one place. This work would be cited in the field for the following point: within a cell type, PNs tile their input glomerulus with their dendrites, intermingle their axons with other PN cell types in the mushroom body, and overlap with each other, and to a lesser degree other PN cell types, in the lateral horn. The authors' other main conclusions would take more to support.

In this work, the authors first (A) choose the neurons they will examine, and then perform two main analyses, (B) pair-wise morphology similarity analysis within three major projection regions for these neurons and then (C) an analysis of 'homotypic' PN cell types and how similarity within and between the cell type varies between the three projection areas. They then (D) conclude that their results are indicative of 'form determining function' and 'labelled line' encoding at play in the Drosophila olfactory system, where 2nd and 3rd order neurons meet.

On A: The authors use data from the FAFB connectome. These are manually reconstructed neuronal skeletons, built using CATMAID. The dataset at present offers a sparse connectome, though a denser one using the same EM data is currently being built by the Flywire project (Dorkenwald et al. 2022), open to anyone for contributions. A dense connectome for the olfactory system already exists in a separate EM dataset, thanks to the hemibrain project (Scheffer et al. 2020). One of the limitations of the present work is that it assumes that the data it uses from FAFB (PNs on a single hemisphere of a single fly) are representative of all flies, and does not discuss this assumption (which is a fair enough assumption given how morphologically stereotyped the system appears to be (Jenett et al. 2012; Schlegel et al. 2021; Jeanne et al. 2018) but should be discussed; the degree of connection stereotypy is still unsettled, and there could be variation in how well neurons of a homotype may fasciculate across brains). The authors could better support their statements by re-running their analysis on the hemibrain data and comparing it to their CATMAID FAFB data. It would greatly strengthen their study to show that the patterns they observe hold across two EM data sets. (Light-level co-registered datasets also exist via the flycircuit project (Shih et al. 2015; Chiang et al. 2011; Costa et al. 2016), though registration offsets likely limit the effectiveness of the authors' present analyses – the authors could consider it).

The authors' analysis focuses on 111 PNs. This is close to the number of uniglomerular projection neurons (uPNs) in the right hemisphere. The actual count from Zheng et al. is 114. The updated count for the same dataset is 164 (Bates et al. 2020). The decision to not use a full complement of uPNs is not adequately justified. (The authors state that they dropped three neurons that are not PNs, including the APL – those neurons are not in the numbers I give above, and the Zheng et al. archiver must have given them the wrong label.) In addition, uPNs have two functionally important sub-divisions – there are excitatory uPNs (cholinergic) and inhibitory uPNs (GABAergic). This distinction, and how the two could differ in terms of which clusters they fall into, is significant because it might reveal how different odour channels can synergise with, or antagonise one another. I think the GABAergic uPNs are left out because they have no MB arbors, but the existence of these different PN groups and the choice to exclude some biological classes are not made clear in the main text.

Indeed, 6 neurons are said to be dropped because 'they did not project to any of the neuropils'. This does not seem like a great reason for dropping these cells completely from the paper. If the point the authors wish to make is that there are labeled lines among the olfactory PNs, surely the ones that stray off into other brain regions are the most evidential! While they do not say it directly, the authors are looking at cholinergic, olfactory uniglomerular PNs and excluding other PN classes – which is fine, but could be made clear. (There are a total of 347 PNs covering all classes, on the right hemisphere in FAFB).

On B: The authors use an all-by-all pair-wise morphology similarity metric to construct a score matrix and then perform hierarchical clustering. The metric is not adequately explained. In reviewing the methods, the metric appears, by the author's own note, to be similar to NBLAST (Costa et al. 2016). There are many different and valid reasons for using different metrics, but it is not clear to me why the authors did not use NBLAST, and how their own method differs. Perhaps results from both would be similar, and it does not really matter – or maybe they found that their own metric is more performant in this situation? The authors, I should be clear, do not claim to have invented a 'better' metric – but some explanation of their choice is warranted. The terms in the author's equation are not adequately explained, but it seems to be a distance score calculated between all points between two neurons in Euclidean space, that decays exponentially with distance, which is common practice (Schlegel et al. 2016; Strutz et al. 2014; Kohl et al. 2013; Masse et al. 2012). Some explanation to that effect should appear in the main text. (I opine that since there are now many studies examining PN morphology with similar metrics, it would be easier if the same algorithms were applied when there is not a good reason to do otherwise, so that results can be directly compared, NBLAST is currently the most popular choice).

In examining their scores, the authors do not give their clustering method, just stating 'a method of hierarchical clustering was used'. Drawing conclusions from hierarchical clustering is very sensitive to the cut height used. It would be nice if the authors showed the result of their clustering as, for example, a dendrogram with the cut height they used indicated by a horizontal line in the main sequence (this is currently S2). Their cut height was determined by a silhouette method for determining optimal clusters. Other common, valid methods include the elbow and gap statistic methods – perhaps in supplement, the authors could give or summarise how the result would have been using those, or in the methods, explain why they could not? How robust is their analysis to different clustering methods, different numbers of clusters, etc? The details of an unsupervised clustering analysis are often skimmed over in papers because we trust the authors were able to tweak things to bring their biological conclusions to the foreground, and that is fine. However, since this paper is all about the results of the authors' clustering analysis, the detail and how robust their analysis is, are very important. Their analysis yields 39 clusters in AL, 3 clusters in MB calyx, and 4 clusters in LH, and there is a nice visualisation in Figure S6 that could use some vertical lines to show the shown cluster numbers. However, those 4 LH clusters could easily be broken down further into smaller groups that look meaningful to a neuroanatomist. The authors use a Pearson's Chi-squared test to observe that their clusters are statistically different from a random sample of PNs, which is great. The authors could re-word this segment though, with simpler language, to let the reader qualitatively know what they are doing and why.

The authors do not consider their MB clusters in-depth or explain what one would expect from the field. The literature suggests that PN-KC connections in the field are either random or semi-random (Eichler et al. 2017; Zheng et al. 2020; Caron et al. 2013). The authors could discuss their results in this context, in particular, their conclusion that 'the distinction between the homotypes is only weakly present in MB calyx'. The structure of the PN-MB clusters might be compared to results from Zheng et al. 2020, who show that the calyx contains a 'fovea' of Kenyon cells oversampling particular PNs, specifically PNs that are from 'food-related' glomeruli. This PN group is probably captured somewhere in the authors' analysis. Interestingly, the authors show that pheromonal largely PNs cluster together in the MB, though they do not draw as much attention to this as the LH pheromonal grouping for some reason – could this be a pheromonal 'fovea'?.

In the LH, the authors focus on their findings of a pheromone-specific cluster within the LH. However, one result they also have is that the other clusters have a mixture of different odour class PNs, so the LH does not easily break down into siloed compartments like the AL. A few studies have tried to break down the LH into different regions that serve different sections of an ethological odour space, based on different analyses (Strutz et al. 2014; Bates et al. 2020; Frechter et al. 2019). If the authors could at least qualitatively compare their groups to the previous groupings, their results and contribution would make more sense in the context of the field.

On C: The authors are interested in looking at 'homotype' uPNs – neurons with dendrites in the same AL glomerulus. This is a novel and interesting approach. The intra-homotypic-PN-distance (a 'bundling' metric) and inter-homotypic-PN-distance scores (a 'packing' metric) they use are insufficiently explained in the text and are better understood looking at figure 4. Figure 4 is a good, concise summary. Because sister PNs are essentially copies of the same cell and occupy the same space – a phenomenon seen across the whole of the Drosophila nervous system – it is unsurprising that they arborise similarly in each of the three analysis regions – the authors show, however, that they are 'morphologically similar in different ways' between the AL, the MB and the LH, which is an interesting point. However, the text does not make it very clear to the reader what the authors have shown. On the intra-homotypic-PN-distance: sister neurons are similar in all three regions – though in the AL they seem to tile glomeruli rather than -- intermingle (Figure 4). This last point is not, I think, shown quantitatively, only schematically, – it would be best to formalise and plot somewhere. Considering the inter-homotypic-PN-distance: neurons interdigitate across homotypes in the MB, to a lesser degree in the LH, and barely in the AL, instead of tiling the whole AL.

The authors explore how their morphology findings might correlate with the odour response properties of PNs. They use odour categories from Bates et al. 2020, though in many cases a PN type was given multiple categories (see the supplemental figures), and it is not clear how the authors resolved this conflict. For example, the V glomerulus can also be viewed as 'aversive/bad' as it encodes CO2. Often a neuron cannot be given a singular 'odor scene' label, though they might be clustered using odour response data (Badel et al. 2016).

On D: In the conclusion, the authors do not discuss their own results in relation to the field that much. They only mention (i) one non-novel finding and (ii) one observation. The (i) finding is that certain pheromonal PNs segregate together in the LH (Ruta et al., 2010; Kohl et al., 2013, Frechter et al., 2019; Chakraborty and Sachse, 2021) – although their finding of the same in the MB is perhaps more novel? The authors say that "Our study not only lends support to the existing studies pointing to the labeled-line strategy in the Drosophila olfactory system but also suggests that an even more sophisticated level of spatial organization that depends on the homotypes and odor should be present"; it is unclear to me what the second half of the sentence means. The (ii) observation, implicit also in other work (Grabe et al. 2016) – that some PN cell types exist as singletons, often aversive ones. Certainly, in insects, there is some circuit-level difference between cell types that exist in multiples or as singletons – which are often larger, older neurons – but the present paper has no analysis on this point. Perhaps the authors could show how these singletons are morphologically special – e.g. they mention that they are more 'dense' but do not quantify this anywhere in the paper.

One of the authors' key conclusions and a point in their abstract, is that a labeled line strategy is at work in the Drosophila olfactory system at the point that second-order projections ramify in the lateral horn (LH). They make expansive statements such as: "Overall, our findings suggest that the Drosophila olfactory system leverages the efficiency of the labeled-line design in sensory information processing". A 'labeled line' is generally taken to be a chain of neurons that transfers a message about a single feature, onto higher-order neurons. This message may be modified and transformed along the way, but is generally not directly integrated with information about other features. The very peripheral part of the system in the antennal lobe (AL), ORNs -> PNs, is generally established to be a set of labeled lines (Couto et al., 2005; Fishilevich and Vosshall, 2005; Vosshall et al., 2000). Although significant cross-talk does exist, e.g. through AL local neuron computation, at this stage the PNs can still act as labeled lines (Olsen et al. 2010; Seki et al. 2017), which could support a somewhat labeled synfire chain up to the 3rd order (Jeanne and Wilson 2015).

The authors do not define the term 'labeled line' clearly, or whether the label in question should be glomerular identity, odour identity, or odour scene. They also do not clearly state (a) what neurons they mean to say are in the labeled line, nor (b) whether they think this is largely the case, or only in specific channels – their sweeping statements suggest they mean the former, but their data only weakly support the later. I assume, for (a) they are suggesting that some lateral horn neurons (LHNs) will be part of labeled lines, but not Kenyon cells of the mushroom body, because of the PN morphological properties that the authors quantify in the LH and MB. For (b), established work in the field suggests a murky division within the LH to support odour categorisation (Frechter et al. 2018; Jeanne et al. 2018) – a convergence scheme, and so, not a labeled line system – and the authors' own results show PNs of different, broad odor classes intermingling in 3/4 of their LH clusters. Their analysis, for (b), seems to show that only a small subset of PNs have the appropriate morphology to support labeled line connections in the LH. The main PNs that may form a labeled line, are the pheromonal PNs in the anterior-most region of the LH. This has been noted by the field, in terms of both morphology and connectivity (Ruta et al., 2010; Kohl et al., 2013, Frechter et al., 2019; Chakraborty and Sachse, 2021; Bates et al. 2020; Bates et al. 2020; Jefferis et al. 2007).

However, they do not demonstrate that they act as labeled lines. To make a statement about labeled lines is to make a statement about connectivity. It can be guessed by using morphology when connection data is absent. However, the connection analysis can now be done using data from the hemibrain connectome (Scheffer et al. 2020; Schlegel et al. 2021). Labeled line strategies may exist for some odour channels in Drosophila – in particular pheromonal channels such as DA1 (Kohl et al. 2013) or the aversive channels the authors note – but it is not possible to determine this using morphology alone. A preprint that examines DA2 – an aversive PN thought to be part of a labeled line (Stensmyr et al., 2012) – actually saw that in the connectome its targets experienced a lot of convergence from different PNs, only a few preserving a possible labeled line (Huoviala et al. 2020). Labeled lines are probably the exception, not the rule, and with DA2 a strong labeled line organisation seems to become a highly distributed representation at the lateral horn stage. If the authors could compare their work on morphology, to the reality in terms of connectivity, they would be better able to support their ideas and show that the features they quantify correlate with circuit structure.

The authors say that "our analysis for the second-order neurons inside the Drosophila olfactory system can be translated to the brain of different organisms including the central nervous system (CNS) of humans". However, I do not think this is strictly true. In Drosophila, neurons of the same cell type, the author's 'homotypes', are near isomorphic and occupy a similar space. Therefore, the authors can use, as they did, a metric defined in Euclidean space to compare the neurons, without spatially transforming the cells at all. In mammals, duplicate types can appear across, say, cortical columns and even in the olfactory system be very spatially segregated, since there is not just one glomerulus for each olfactory receptor, as in insects. Therefore, the spatial analysis that would need to be done in mammals would be different.

It is commendable that the authors have made their analysis in python available on GitHub (https://github.com/kirichoi/DrosophilaOlfaction), as well as preprinting their work (https://www.biorxiv.org/content/10.1101/2022.02.23.481655v1). The authors pre-process the skeleton data well, removing registration artefacts. The consensus lateral horn volume (Ito et al. 2014) is somewhat arbitrarily defined as the uPN terminus region, though the MB Calyx is more strongly defined by glial sheathing. The authors wisely do not use this volume but re-define their neuropils – perhaps more accurately regions of interest – by using density estimates built from the neurons' 3D points. It would have been nice to see, in a supplementary figure, how these ROIs differ from the Ito et al. 2014 standard neuropils – since when they use neuropil terms in the text, an informed reader will assume they used the Ito standards. I like the visual segmentation process description in Figure S5.

I think the paper does make many quantitative points on PN morphology well, and could either become a shorter manuscript that shows this information concisely, or a more involved one on 'labelled lines' and the validity of this idea for the olfactory system, where 2nd-order neurons meet 3rd order ones. In the first case, adding thermo sensory PNs, GABAergic uPNs, and mPNs to their analysis could make it more interesting, and produce some new and germane insights for their sub-regional analyses (see below). In the second case, the paper would need to include work that compares across different PN classes and EM datasets, including looking at connectivity in the hemibrain, to make its core claims on labeled lines.

In general, in the present work, the authors do not provide enough biological background information on the olfactory system and what is known about PNs already, the regions they innervate for a naive reader to understand their results, and why their results might be interesting. This could be easily fixed with a little re-writing. The use of mathematical notation within the main text is heavier than needs be and hampers the readers' understanding of what the authors are doing and intend to show. Plainer writing would help. The authors need to communicate their core findings, and their context in the field, more clearly.

Glomeruli identification through the literature can be tricky. There has been some confusion specifically about the identities of VM6 and VC5. The Zheng et al. 2018 paper had a few errors, later rectified by Bates and Schlegel et al. 2020 and then again by Schlegel and Bates et al. 2021. Given the confusion, no one can be faulted for mistakes here, but authors should make sure their labels are the same as in Schlegel and Bates et al. 2021 (that process was a multi-lab debate). I think they used the outdated Zheng et al. 2018 labels.

Suggestions

I have some suggestions for increasing interest in this work, that I humbly submit to the authors. The authors discuss 51 glomeruli of the antennal lobe. Of the olfactory glomeruli, there are 52, but there are actually 58 glomeruli in total in the antennal lobe (e.g. VP1-5), including the thermo-hygrosensory ones. Other work (Schlegel et al. 2021; Bates et al. 2020) has found that some of the most striking differences can be found between olfactory and thermo-sensory glomeruli, as well as some curious associations across the two modalities. Since intellectually this system is very similar – RNs contact PNs that then reach the LH and the zone just ventral to it – but not quite parallel – arbours intermingle with olfactory ones in the ventral LH, and antennal lobe local neurons cross compute between combinations across all 58 glomeruli – I strongly think including these neurons in the author's analysis would increase the interest in this work.

There are a total of 347 PNs in the FAFB dataset that project from the right-side antennal lobe. This is because there are 283 multi-glomerular PNs (mPNs). Bates et al. 2020 also make 58 uPNs from the left side of the brain available. All of the data is open and can be gotten by the authors here: https://fafb.catmaid.virtualflybrain.org/. Similar to the above, I think that including the mPNs as a comparison point, would increase interest in this work.

PNs could also have been clustered by their odour response profiles using data from PN response (Badel et al. 2016) or ORN response (Münch and Galizia 2016), rather than porting labels from Bates et al. 2020 / Mansuorian et al. 2015. The authors might consider whether this could add meaningfully to their analysis.

In addition, as noted in the public review, I think comparing against or using NBLAST could be informative. The authors could run NBLAST on their processed neuron skeletons and discover whether the results differ much from what they have in hand right now. NBLAST is now implemented in the navis python library: https://navis.readthedocs.io/en/latest/source/tutorials/nblast.html.

Lastly, a more out-there suggestion: PN morphology analysis using a skeleton representation has been common in the field. However, volumetric analyses based on neuronal meshes – now available for these neurons through the hemibrain and flywire projects – is almost non-existent. How do the volumes for neurons vary across types, clusters, etc? Using this information could add some simple points, to support the authors in their quantitative description of these neurons.

Reviewer #3 (Recommendations for the authors):

This work provides a quantitative evaluation of the spatial organization pattern of olfactory projection neurons (PN) in AL, CA, and LH based on the inter-PN distances using FAFB EM dataset. This NBLAST-based method does provide a simple way to cluster neurons with similar projection patterns and even predict underlying signal processing rules. However, the reliability of the clustering method needs to be improved since the method only got 39 clusters out of 51 well-segregated AL glomeruli serving as the ground truth. This result undermines the conclusions in the higher-brain centers which have a more complex organization. Moreover, the authors defined the neuropils by rotating the neurons along specific axes and then segmenting the dense innervation parts, which may hinder the accuracy of the boundary when the surface is convoluted and does not reveal inner subdomains. Nevertheless, serving as the first step to tackling complex connectomic data, the method is easy and potentially useful.

1. The current study analyzed only 111 uniglomerular PNs instead of the latest-released 164 uniglomerular PNs (Bates et al., 2020). To make the work more valuable, the authors should apply their analysis to the latest dataset. In addition, since the analysis was done from the EM data of a single fly, whether the preferential spatial distribution of PNs and the clustering are consistent in different individuals is unknown. It will be informative and more persuasive if a similar spatial distribution pattern can be observed in another fly EM dataset (Scheffer et al., 2020).

2. The segmentation of neuronal innervation in AL, Calyx, LH is achieved by rotating the neurons along specific axes and then identifying the dense innervation parts as the three neuropils (Figure S5). The methodology is convenient to define the neuropils but will hinder the accuracy of the boundary segmentation when the surface is convoluted (i.e. MB calyx) and does not reveal inner subdomains (i.e. AL glomeruli). As the original EM dataset has already offered neuropil surface point coordinates which can be downloaded from Catmaid website (https://catmaid-fafb.virtualflybrain.org/), attributing the innervation points directly to the defined neuropils should be more accurate.

3. Theoretically, the natural segregation of glomerular structures in AL would serve as the ground truth to test the reliability of the clustering method. Yet, the method only got 39 clusters out of 51 well-segregated glomeruli. This discrepancy undermines the effectiveness of the clustering strategy. The result of silhouette coefficient analysis also suggests that the coefficients are similar for clusters ranging from 30 to 50 in AL (the coefficients are very close almost reaching a plateau) (Figure S6). The author should justify why choosing the number of PN innervation clusters in AL as 39 and provide evidence that it is optimized to find the hidden pattern. Most importantly, if the method fails to reveal the remaining 11 AL glomeruli that can be visually distinguished, it is difficult to see how it can reveal the hidden pattern in MB calyx which is much less well defined.

4. The neuronal distance estimation algorithm may encounter a problem when projection neurons have very different innervation ranges and the total length of the innervation branches. For example, if there is a projection neuron, PNa, only innervates in the entry of Calyx where most PNs pass by, the distance between PNa and other PNs will be very small due to the algorithm only selects the shorter skeleton to evaluate the distance (see method and the symmetric matrices in Figure 3). Thus, the result may not faithfully reflect the fiber distance when two PNs have a drastic difference.

5. The authors estimate λ value which represents the ratio of the mean of intra-homotypic neuron distance and the mean of inter-homotypic neuron distance. (Figure 4, Figure 5, Figure 6, Figure S3). It is a creative way that gives us insight into how PNs overlap with each other in distinct neuropils. However, the volume and the aspect ratio of the three neuropils are different. AL occupies a much larger area compared to calyx and LH. Therefore, the current λ value may simply reflect the spatial distance presented in the neuropil but not the actual degree of overlapping (e.g. in two non-overlapping pairs, the one with long distance will have a low λ value while both pairs have no overlapping at all). To make the λ value reflect more to the degree of overlapping, they might need to normalize it based on the volume, compare the λ values between original data with the data after shuffling PNs into different clusters or take the innervation density of fibers in specific neuropils into account.

6. The authors annotate the reaction profile for a specific glomerulus using a dumb variable to do statistics. The result suggests that pheromone PNs will be clustered together in LH (into 2 clusters) and Calyx. Food-related PNs and aversive PNs also be clustered into different groups in LH but not in Calyx (Figure 7, Table S1). The author should elaborate and discuss more in terms of biology. For example, do food-related odor signals and aversive odor signals converge in MB calyx and diverge in LH? What biological properties may be implicated in the organization?

7. Although the authors have offered the python scripts for researchers to reproduce their results, a comprehensive spreadsheet that contains all the distance results along with functional annotation will greatly improve the accessibility of the analyzed results.

8. The authors should compare their cluster results with Lin et al. 2007 and Bates et al. 2020 and discuss the biological implication.

References:

Bates, A. S., Schlegel, P., Roberts, R. J., Drummond, N., Tamimi, I. F., Turnbull, R., … and Jefferis, G. S. (2020). Complete connectomic reconstruction of olfactory projection neurons in the fly brain. Current Biology, 30(16), 3183-3199.

Jefferis, G. S., Potter, C. J., Chan, A. M., Marin, E. C., Rohlfing, T., Maurer Jr, C. R., and Luo, L. (2007). Comprehensive maps of Drosophila higher olfactory centers: spatially segregated fruit and pheromone representation. Cell, 128(6), 1187-1203.

Lin, H. H., Lai, J. S. Y., Chin, A. L., Chen, Y. C., and Chiang, A. S. (2007). A map of olfactory representation in the Drosophila mushroom body. Cell, 128(6), 1205-1217.

Nishino, H., Iwasaki, M., Paoli, M., Kamimura, I., Yoritsune, A., and Mizunami, M. (2018). Spatial receptive fields for odor localization. Current biology, 28(4), 600-608.

Zheng, Z., Li, F., Fisher, C., Ali, I. J., Sharifi, N., Calle-Schuler, S., … and Bock, D. D. (2020). Structured sampling of olfactory input by the fly mushroom body. BioRxiv.

https://doi.org/10.7554/eLife.77748.sa1

Author response

While appreciating the authors' work presented in this manuscript there were two major concerns that came up during the consultations:

1. Given that two EM datasets exist for the fly brain, determining the generalisability of what the authors describe should have been done.

2. Making claims about labelled line representation of odours or not using morphometric data alone is not justified when connectivity data exists for flies.

Based on this, and other comments, the essential revisions that were discussed were the following:

The writing: The text needs to be made more accessible to the reader, perhaps by providing more intuitive explanations, biological context, and emphasise their findings more than they currently do (which emphasises methods used more). We would also like to see their analyses in the context of what's been done before.

In our revision, we have made a substantial update to the writing to make our manuscript more accessible to the readers. First, we now provide much more intuitive explanations for the concepts discussed in the manuscript. We have tried to move away from mathematical notations if possible and supplemented our descriptions with intuitive illustrations whenever we thought it was necessary. Therefore, most of the equations and detailed statistics are moved to either the Methods or the appendices, replaced by simple and intuitive descriptions. For example, one of the common comments across the reviewers was that we did not explain what our inter-PN distance meant. Now, the Result section starts with a simple explanation of what we meant by inter-PN distance, along with a demonstrative illustration (Figure 2—figure supplement 1A) and detailed formulations (in the Methods section). On the same note, our figures have been modified to make them more comprehensible, either by adding visual aids, increasing the size of figures and text, or providing a more explicit caption.

The general tone of the manuscript has been shifted as well, moving away from the methodology and statistics to explanations based on biological implications and comparisons to previous literature. Therefore, a substantial portion of the main text has been moved to either the Methods or the appendices. Our Discussion section has been significantly improved, discussing our results by comparing them with Bates et al. (2020), Lin et al. (2007), Strutz et al. (2014), Zheng et al. (2020), etc., covering almost all the references raised by the reviewers. Apart from providing a richer background in the context of what has been done before, we also emphasized some of our novel findings and their biological implications of them.

Reproducibility: Since the EM data in this analysis is based on a single fly, we recommend testing whether the authors' findings will be consistent across individuals. So, could the authors test their analysis on the hemibrain alongside the extant FAFB?

First, we would like to emphasize that our manuscript is now using the latest FAFB dataset by Bates et al., 2020, to answer other requests by the reviewers. Therefore, all our figures in the main text are replaced with the new results based on the updated FAFB dataset.

To test the generalizability of our result, we have performed a reproducibility study using the suggested hemibrain dataset (Scheffer et al., 2020). We have reproduced several key figures in the main text, which were based on the FAFB dataset, using the hemibrain dataset instead, and included them in Figure 13. Comparing the results based on the hemibrain dataset with our new results from the FAFB dataset, we found that the overall results from the two datasets are largely the same and the two datasets are interchangeable in supporting our claims. We also discussed the generality of the results from our analysis by commenting on the similarity and differences between FAFB and hemibrain datasets. Additional details are available in the Methods section and the Appendix 3. On our Github repository, we publicly shared the script we have used to query the neurons from the hemibrain dataset, and a copy of our analysis code tailored for the hemibrain dataset.

Number of uPNs used: The authors use 111 uPNs for their analysis whereas the number of uPNs is 164. (This might be why they obtain only 31 glomeruli). It's not clear why the rest were excluded. Could the authors conduct their analyses by including all cholinergic uPs (including those involved in hygro and thermo sensation)? (We assume the authors have omitted GABAergic neurons from their analyses because they do not innervate the MB? Including it would provide important biological insight, but if not, their reasoning needs to be stated.)

We would like to reiterate our reasoning behind why we chose a subset of uPNs in our original analysis. One of our primary goals in this manuscript was to analyze how the spatial characteristics of homotypic uPNs change as the uPNs innervate different neuropils. Because we wanted to compare the spatial characteristics of uPN innervation at each neuropil, we had to consider uPNs that innervate all three neuropils. Therefore, many GABAergic uPNs ended up not being included in our original analysis. We agree that the description and justification regarding our criterion might have been unclear in the original submission.

That said, the analysis of all uPNs would provide a more comprehensive view of the spatial organization of uPNs, making our study more interesting and satisfactory. Therefore, we have made the following changes to address this issue.

First, as stated before, we have used the latest FAFB dataset (Bates et al., 2020) for analysis, which contains a total of 164 uPNs composing 58 glomeruli. We have re-done all our calculations using the latest FAFB dataset and studied a set of 135 uPNs that innervate all three neuropils, including those involved in hygro and thermo sensation. Based on this result, we have updated the manuscript and the figures, pointing out several newly found properties.

Second, we performed a separate analysis on the ~30 uPNs that were originally left out based on our selection criterion (that they do not innervate all three neuropils). Most of these uPNs are GABAergic, following mlALT instead of mALT. These neurons are still crucial for the olfactory processing in Drosophila, but were largely incompatible with our analysis (e.g., we can’t calculate and compare λ across the neuropils). Therefore, we conducted a few compatible analyses that examine (1) the spatial difference between uPNs innervating all three neuropils and uPNs whose innervation is specific to a particular neuropil, and (2) neurotransmitter-based characteristics. The manuscript has been updated accordingly.

Hierarchical clustering: We are concerned that the inferences rely on a few clusterings that don't appear to be very robust. This is usually okay, but a lot of the paper is built around this. So, could the authors reassess their hierarchical clustering parameters and the normalisation for the λ value and state them explicitly for the reader?

We understand the concern regarding the robustness of the clustering result. The output of hierarchical clustering is indeed affected by the cut height and the specific tree-cutting method. However, we would like to point out that this is not a cause of serious concern, as the hierarchical relationships are generated purely from the distance metric we used and are fixed regardless of which tree-cutting method we choose. For example, when looking at the dendrogram for AL (Figure 2—figure supplement 3A), one might notice that the well-segregated glomeruli are expressed through homotypic uPNs under neighboring branches. Therefore, while the final tree-cutting step may affect the number of clusters, increasing the number of clusters will simply break down a large cluster into smaller ones. In fact, if our argument can still be explained by smaller clusters (or a large number of clusters), our argument may be considered better supported.

Nonetheless, we tested a total of four different tree-cutting methods/criteria: elbow method, gap statistics, maximum average silhouette coefficient, and the dynamic hybrid cut tree method (Langfelder et al., 2008) which has been successfully utilized for morphometric classification of mouse V1 neurons by Gouwens et al. (2019). We noticed that using the maximum average silhouette coefficient resulted in the most accurate cluster number in AL (N=54), but it showed a significant discrepancy in cluster number in MB calyx and LH when compared to the other three methods (see Table 1).

After an internal discussion, we have decided to switch our tree-cutting method to the dynamic hybrid cut tree method for the following reasons:

i) Choosing a tree-cutting method that returns a cluster number close to the number of different odor types (which is 10) best serves our goal.

ii) Since we are interested in the clusters of uPN innervations to MB calyx and LH (as the correct labels in AL are already given), it seemed reasonable to choose a method out of the three methods that produced comparable cluster numbers in MB calyx and LH.

iii) The dynamic hybrid cut tree method was proven to give satisfactory results and is methodologically systematic based on our previous experiences and previous literature.

iv) We wanted to test our hypothesis under a harsher condition.

Switching the tree-cutting method resulted in 10 clusters in MB calyx and 11 clusters in LH. We report that our argument still holds despite the higher cluster number. The trend of the uPNs under the same homotype falling into the same cluster became much more apparent (Figures 3, 4, and 8). Our various statistical tests also returned results in line with our findings. Additionally, we have updated our Method section to provide in-depth descriptions of how the hierarchical clustering was done, along with our reasoning behind the choices we made, to help the readers understand our clustering procedure without any ambiguity.

Concerning the issue with λ normalization, we would like to iterate that we devised the λ to be a scale-free ratio that quantifies the relative spatial innervation per homotype. This means that we intended not to scale our λ based on the volume. For example, let us consider a homotype composed of several uPNs with similar dintra values for both MB calyx and AL. Indeed, the volume of MB calyx is smaller than that of AL, so the dinter will be generally larger in MB calyx, leading to a higher λ value, which is what we intended. Our logic behind this is that in AL, despite all the additional space that the uPNs technically could have innervated, the uPNs ended up localized into a bundle with a small dintra. In MB calyx, the same uPNs had less space to innervate, to begin with, so the less ‘packing’ observed in MB calyx may not be as novel. We wanted our λ to differentiate the two, which is the reason why our λ is defined as such.

However, it seems like the term ‘overlapping’ is a bit elusive and can be defined in various ways, which we believe is why this issue has been brought up by the reviewers. We feel that our explanation and choice of words (overlapping) were not sufficiently clear. After a lengthy discussion, we have decided to keep the terminology but provide a much more comprehensive explanation of what we mean by the ‘degree of overlapping’ to make sure that the readers won’t get confused.

Labelled line hypothesis: The claims made in this study about the labelled line hypothesis cannot be inferred from morphometrics alone. For this, the authors would need to compare these findings with connectivity. This could even be done at a small scale to understand whether their morphological groups also have more specific or segregated connectivity within the LH / AB.

(A possible pipeline for this could be: Use neuprint-python to pull connection data for each chosen PN and subset it to the select brain regions. They could then perform hierarchical clustering on that, and devise groupings that can be compared to the morphological grouping visually using a tanglegram. They might then compare the two using Baker's Γ Index or similar.)

Following the suggestion, we decided to use the connectivity data to conduct an additional small-scale study to check how far the labeled-line principle holds. For a comprehensive connectivity dataset between PNs and higher olfactory neurons such as KCs and LHNs, we utilized the hemibrain dataset. Here, we tried to address the following questions: (1) Are there any KCs and LHNs that carry a specific type of information? If so, how prevalent are they? (2) Are there any insights we can gain from comparing the clustering outputs using spatial proximity (dαβ) and connectivity?

To answer these questions, we have conducted three additional analyses.

First, we collected a number of KCs and LHNs that synapse only with a single homotype (Figure 10). These ‘homotype-specific’ connections (Nx,spξ), defined as the number of third-order neurons that only synapses with a specific homotype but not with the others (see Figure 9 and Methods for more information), are much more prevalent in LHNs compared to KCs. Certain homotypes (e.g., DA1) have an especially high number of LHNs that only connect to the given homotype. The ‘homotype-specific’ neurons functionally carry a single type of information, thereby may be considered as an extension of the labeled-line strategy.

Second, we collected LHNs connected to a particular homotype and checked which other homotypes these LHNs are also synapsing (thereby analyzing the scope of signal integration happening at LH – see Figure 9 and Methods for more information). Based on this analysis, we found the homotype-specific labeled line predominate until the signal reaches the higher olfactory centers, some of which then transition into odor-specific channels where either a broad or a narrow integration occurs. In MB calyx, no such trend is observed, further supporting the previous literature on randomized connections. Even though we observed a strong per homotype bundling tendency at MB calyx, the high degree of overlapping (denoted by λ) seems to have a bigger impact on connectivity.

Third, we performed a connectivity-based hierarchical clustering and compared the result against our spatial proximity-based clustering result. Tanglegrams are plotted to compare the dendrograms generated from the spatial proximity (dαβ) and from the connectivity (dcos) (see Figure 12). To check the relationship between the spatial proximity-based clustering and connectivity-based clustering results, we supplied Baker’s γ index (the ordinal relation), entanglement (transferability), and cophenetic distance correlation (the correlation between two tree structures). We report that the spatially well clustered uPNs at MB calyx do not precisely translate to structured connectivity patterns, consistent with the notion of randomized PN-KC connections, while in LH, spatial and organizational characteristics of uPNs are well-translated to connectivity to LHNs.

In addition to this, please read the other (not essential) recommendations and if any of them are feasible to respond to, please do so.

Reviewer #1 (Recommendations for the authors):

The olfactory PNs have been studied using numerous approaches. In this study, the authors use a parameter called 'inter PN distance' to assess the relationship between all the uniglomerular PNs in the hemibrain connectome in the AL, MB, and LH. As odour information from the sensory neurons is brought into the AL via OSNs classes that innervate individual glomeruli, PNs that innervate the same glomeruli are considered homotypic. Using the inter PN distance and downstream statistical analyses, the authors infer that odour information is discretely represented in the AL and less so in the LH; however, in the MB, there is considerable overlap in odour representation.

The manuscript is well written and presented.

- PN information representation across the AL, MB, and LH has been explored before. While the authors do mention this, it would be useful if they discussed what each one of those approaches inferred about olfactory representation, and why they felt their approach was better. In the same vein, it would be nice if they finally concluded by comparing their findings with the others.

We thank the reviewer for the suggestion. We substantially revised our paper, not only discussing how our results are in-line or different from the results by Bates et al. (2020), Lin et al. (2007), Strutz et al. (2014), etc., but also including new topics ranging from spatial innervation patterns to connectivity with third-order neurons. Also, our manuscript now better emphasizes some of the unique findings we made, thereby demonstrating the advantages of our approach.

- The manuscript is statistically dense. It will help readers if the authors provided some intuitive explanations along the way – for example, for inter-PN distance, among others.

Following the reviewer’s comment, we have updated our manuscript to make it more accessible to the readers. For example, we have added a simplified explanation of the distance metric dαβ to the main text before we discuss the spatial proximity-based clusters, along with a schematic depicting the step-by-step process of calculating dαβ (see Figure 2—figure supplement 1A). Most of the equations and detailed discussions of statistical test of independence are moved to either the Methods or the appendices, replaced by simple and intuitive descriptions. Whenever we feel our explanation might not be easy to follow, we added illustrations to supplement the text.

- Many of the legends are not very explanatory. Could the authors please elaborate on them? It would also help to have pointers in the figures to focus the attention of the reader. For example, the legend for S2 can be expanded. I can't tell which homotypic PNs are merging into a single cluster. There are similar issues in other figures as well – could the authors please add visual aids such as arrows in the figures?

Following the reviewer’s comment, we have updated our figures and captions to make them easier to read and understand. The size of the figures and texts is increased in general, and we added visual aids whenever we discuss a specific subset of a figure.

Reviewer #2 (Recommendations for the authors):

Choi et al. competently quantify known aspects of excitatory, uniglomerular projection neuron (uPN) morphology, in order to demonstrate how neurons of this well-studied cell class differ in the three main brain regions in which they arborise. Classically, these three regions support three different functions: siloed olfactory channel processing (antennal lobe), ethological odour classification (lateral horn), and associative memory (mushroom body). They find that uPNs of the same cell type are morphologically similar in each region, but the nature of that similarity, and their similarity with other uPN cell types, differ as appropriate for the function of the neuron in that region. While not novel, their quantitive description contributes to the field. Their 'major' conclusion on labeled line processing is, however, insufficiently substantiated given that the olfactory system of the fly is post-connectomic and this question is best taken up with connectionist analysis. They also only use a single dataset from a single fly (though a second is available) and do not explore how robust their results are to other clustering approaches/decisions, making it unclear how general their description is.

The strength of this work is that it is able to use quantitative methods to demonstrate features of Drosophila olfactory projection neuron (PN) anatomy. To date, these features have been known and used in classification work on the PNs (e.g. Tanaka et al. 2012; Bates et al. 2020; Zheng et al. 2018; Bates et al. 2020; Jefferis et al. 2007). They have also been shown through (PN->target) connection analyses (Eichler et al. 2017; Schlegel et al. 2021). The findings are therefore largely not novel but are a good quantitative description and a nice synthesis in one place. This work would be cited in the field for the following point: within a cell type, PNs tile their input glomerulus with their dendrites, intermingle their axons with other PN cell types in the mushroom body, and overlap with each other, and to a lesser degree other PN cell types, in the lateral horn. The authors' other main conclusions would take more to support.

In this work, the authors first (A) choose the neurons they will examine, and then perform two main analyses, (B) pair-wise morphology similarity analysis within three major projection regions for these neurons and then (C) an analysis of 'homotypic' PN cell types and how similarity within and between the cell type varies between the three projection areas. They then (D) conclude that their results are indicative of 'form determining function' and 'labelled line' encoding at play in the Drosophila olfactory system, where 2nd and 3rd order neurons meet.

On A: The authors use data from the FAFB connectome. These are manually reconstructed neuronal skeletons, built using CATMAID. The dataset at present offers a sparse connectome, though a denser one using the same EM data is currently being built by the Flywire project (Dorkenwald et al. 2022), open to anyone for contributions. A dense connectome for the olfactory system already exists in a separate EM dataset, thanks to the hemibrain project (Scheffer et al. 2020). One of the limitations of the present work is that it assumes that the data it uses from FAFB (PNs on a single hemisphere of a single fly) are representative of all flies, and does not discuss this assumption (which is a fair enough assumption given how morphologically stereotyped the system appears to be (Jenett et al. 2012; Schlegel et al. 2021; Jeanne et al. 2018) but should be discussed; the degree of connection stereotypy is still unsettled, and there could be variation in how well neurons of a homotype may fasciculate across brains). The authors could better support their statements by re-running their analysis on the hemibrain data and comparing it to their CATMAID FAFB data. It would greatly strengthen their study to show that the patterns they observe hold across two EM data sets. (Light-level co-registered datasets also exist via the flycircuit project (Shih et al. 2015; Chiang et al. 2011; Costa et al. 2016), though registration offsets likely limit the effectiveness of the authors' present analyses – the authors could consider it).

In the revised manuscript, we have used the latest FAFB dataset for the main analysis (Bates et al., 2020).

All our figures in the main text are replaced with the new results based on the updated FAFB dataset. Additionally, we have performed a reproducibility study using the suggested hemibrain dataset (Scheffer et al., 2020) and validated our results by repeating the calculation. When we compared the results based on the hemibrain dataset with our new results from the FAFB dataset, we found that the overall results from the two datasets are largely the same and the two datasets are interchangeable in supporting our claims. We have reproduced several key figures in the main text, which were based on the FAFB dataset, using the hemibrain dataset instead, and included them in Figure 13. We also discussed the generality of the results from our analysis by commenting on the similarity and differences between FAFB and hemibrain datasets. On our Github repository, we publicly shared the script we have used to query the neurons from the hemibrain dataset and a copy of our analysis code tailored for the hemibrain dataset.

The authors' analysis focuses on 111 PNs. This is close to the number of uniglomerular projection neurons (uPNs) in the right hemisphere. The actual count from Zheng et al. is 114. The updated count for the same dataset is 164 (Bates et al. 2020). The decision to not use a full complement of uPNs is not adequately justified. (The authors state that they dropped three neurons that are not PNs, including the APL – those neurons are not in the numbers I give above, and the Zheng et al. archiver must have given them the wrong label.) In addition, uPNs have two functionally important sub-divisions – there are excitatory uPNs (cholinergic) and inhibitory uPNs (GABAergic). This distinction, and how the two could differ in terms of which clusters they fall into, is significant because it might reveal how different odour channels can synergise with, or antagonise one another. I think the GABAergic uPNs are left out because they have no MB arbors, but the existence of these different PN groups and the choice to exclude some biological classes are not made clear in the main text.

Indeed, 6 neurons are said to be dropped because 'they did not project to any of the neuropils'. This does not seem like a great reason for dropping these cells completely from the paper. If the point the authors wish to make is that there are labeled lines among the olfactory PNs, surely the ones that stray off into other brain regions are the most evidential! While they do not say it directly, the authors are looking at cholinergic, olfactory uniglomerular PNs and excluding other PN classes – which is fine, but could be made clear. (There are a total of 347 PNs covering all classes, on the right hemisphere in FAFB).

We thank the reviewer for pointing out this issue. As the reviewer has noted, both cholinergic and GABAergic neurons play important role in olfactory signaling. Our intention was not to ignore GABAergic PNs under some arbitrary conditions. One of our primary goals in this manuscript was to analyze how the spatial characteristics of homotypic uPNs change as the uPNs innervate different neuropils. This made us only consider the uPNs that innervate all three neuropils, which is why many GABAergic uPNs ended up not being included in our original analysis. We agree with the reviewer that the description and justification regarding our criterion might have been unclear. We have updated our manuscript to better reflect the reasoning behind our choice.

That said, we noticed that there are several additional uPNs available from the latest FAFB dataset (Bates et al., 2020) that met our existing criterion, including those involved in hygro and thermo sensation. We considered that analyzing the additional uPNs will provide a more comprehensive view of the spatial organization of uPNs. We have re-done our calculations using the latest FAFB dataset and studied a set of 135 uPNs that innervate all three neuropils. We have updated the manuscript and the figures accordingly, pointing out several newly found properties.

Many GABAergic uPNs that do not innervate all three neuropils (there were 28 of them in the latest FAFB dataset) are still crucial for the olfactory processing in Drosophila. These neurons are largely incompatible with our analysis (e.g., we can’t calculate and compare λ across the neuropils), but we conducted a few compatible analyses that examine (1) the difference between uPNs innervating all three neuropils and uPNs whose innervation is specific to a particular neuropil, and (2) neurotransmitter-based characteristics.

There was a total of 137 uPNs innervating MB calyx, indicating our analysis only left two uPNs which constituted VP3. When the hierarchical clustering was performed on the entire 137 uPNs in MB calyx we ended up with an almost identical clustering output. Two missing uPNs were grouped into the clusters C4MB and C6MB, along with other hygro/thermo-sensing homotypes. On the other hand, there was a total of 162 uPNs innervating LH, analyzing the complete uPN innervation in LH more interesting. The addition of 27 uPNs constituting 15 homotypes innervating LH created four new clusters when the hierarchical clustering was performed (see Figure 4—figure supplement 1). The additional 27 uPNs changed the content of the individual clusters; that is, the tree-cutting algorithm broke down a few clusters that became larger due to the additional uPNs. Furthermore, when we calculated d̅intra, d̅inter, and λ for the 15 homotypes that included the 27 uPNs, we find that the LH d̅intra values increased after adding back 27 uPNs (see Figure 5—figure supplement 2). This suggests that the previously removed uPNs, most of which follow mlALT, are significantly different in terms of spatial and organizational characteristics and thus should be analyzed separately.

Out of 27 additional uPNs in LH, 21 followed mlALT, 5 followed trans-lALT, and 1 followed mALT. Figure 4—figure supplement 2 illustrates how these 27 uPNs spatially innervate LH and demonstrate why d̅intra values were increased, as they are indeed poorly bundled with the existing uPNs under the same homotype. These 27 uPNs are mostly GABAergic: they are composed of 21 GABAergic, 1 cholinergic, and 4 unknown neurotransmitter types, covering 84 % of GABAergic uPNs available in the FAFB dataset. These uPNs innervate LH drastically differently from other uPNs in the same homotype (see homotypes such as DA1, DC4, DL2d, DL2v, DP1l, VA1d, VA1v, VL2a, VL2p, and VP5 in Figure 4—figure supplement 2). Morphologically, inhibitory GABAergic neurons are often considered to be ‘smooth’ and aspiny (Douglas et al., 1989; Bopp et al., 2014; Gouwens et al., 2019), which are discernible from Figure 4—figure supplement 2.

On B: The authors use an all-by-all pair-wise morphology similarity metric to construct a score matrix and then perform hierarchical clustering. The metric is not adequately explained. In reviewing the methods, the metric appears, by the author's own note, to be similar to NBLAST (Costa et al. 2016). There are many different and valid reasons for using different metrics, but it is not clear to me why the authors did not use NBLAST, and how their own method differs. Perhaps results from both would be similar, and it does not really matter – or maybe they found that their own metric is more performant in this situation? The authors, I should be clear, do not claim to have invented a 'better' metric – but some explanation of their choice is warranted. The terms in the author's equation are not adequately explained, but it seems to be a distance score calculated between all points between two neurons in Euclidean space, that decays exponentially with distance, which is common practice (Schlegel et al. 2016; Strutz et al. 2014; Kohl et al. 2013; Masse et al. 2012). Some explanation to that effect should appear in the main text. (I opine that since there are now many studies examining PN morphology with similar metrics, it would be easier if the same algorithms were applied when there is not a good reason to do otherwise, so that results can be directly compared, NBLAST is currently the most popular choice).

First, we have added a simplified explanation of the distance metric dαβ to the main text before discussing the clusters, along with an illustration depicting the step-by-step process of calculating dαβ (see Figure 2—figure supplement 1A). Simply put, the distance metric dαβ only considers the pairwise distance but not the dot product term (which measures the similarity of two neuronal morphologies) used for the NBLAST score. Therefore, dαβ is computationally comparable to the NBLAST score, but it only measures the spatial proximity between two neurons. The reason why we have decided to use this metric is that we are predominantly interested in the spatial proximity (or co-location) between two uPN innervations but not the structural similarity between them, which the NBLAST score accounts for (a point which is also noted by Zheng et al., 2018). We agree that the NBLAST score is the community standard in terms of quantifying the similarity between two neurons, but we believe the distance metric dαβ is conceptually more physical and adequate for the primary aim of our study, which is to analyze the spatial and organizational characteristics of uPN innervation in each neuropil. For example, when we calculate the normalized NBLAST distance between uPN innervations and compare it against dαβ, the two distances are correlated but with significant dispersion, indicating that these two metrics are not exactly the same. There are some cases that inter-PN organization measured by dαβ can be distorted if NBAST score is used instead (see Figure 2—figure supplement 1B).

In examining their scores, the authors do not give their clustering method, just stating 'a method of hierarchical clustering was used'. Drawing conclusions from hierarchical clustering is very sensitive to the cut height used. It would be nice if the authors showed the result of their clustering as, for example, a dendrogram with the cut height they used indicated by a horizontal line in the main sequence (this is currently S2). Their cut height was determined by a silhouette method for determining optimal clusters. Other common, valid methods include the elbow and gap statistic methods – perhaps in supplement, the authors could give or summarise how the result would have been using those, or in the methods, explain why they could not? How robust is their analysis to different clustering methods, different numbers of clusters, etc? The details of an unsupervised clustering analysis are often skimmed over in papers because we trust the authors were able to tweak things to bring their biological conclusions to the foreground, and that is fine. However, since this paper is all about the results of the authors' clustering analysis, the detail and how robust their analysis is, are very important. Their analysis yields 39 clusters in AL, 3 clusters in MB calyx, and 4 clusters in LH, and there is a nice visualisation in Figure S6 that could use some vertical lines to show the shown cluster numbers. However, those 4 LH clusters could easily be broken down further into smaller groups that look meaningful to a neuroanatomist. The authors use a Pearson's Chi-squared test to observe that their clusters are statistically different from a random sample of PNs, which is great. The authors could re-word this segment though, with simpler language, to let the reader qualitatively know what they are doing and why.

We thank the reviewer for bringing up this topic. First, we admit that the hierarchical clustering was too briefly described. We have updated our Method section to provide in-depth descriptions of how the linkage was constructed and trees are cut. Additionally, we would like to point out that the different colored branches in our dendrograms (Figure 2—figure supplement 3) correspond to different clusters (in Figures 3 and 4).

We also understand the reviewer’s concern about the robustness of our clustering output. However, we do not think this is a cause of serious concern due to the nature of our clustering method. The output of hierarchical clustering is indeed affected by the cut height and the specific tree-cutting method. However, the hierarchical relationships are generated purely from the distance metric we defined, and the tree structure is fixed regardless of which method we choose to cut the leaves. Often, there is no ‘correct’ way to determine the number of clusters or cut height; different methods may offer slightly different suggestions. While we originally reported 39 clusters for AL, 3 clusters for MB calyx, and 4 clusters for LH, we do not consider these cluster numbers to be the ground truth with some biological implications, but rather a result of a particular clustering protocol and parameters we chose. This may be of considerable concern for clustering methods like K-means clustering, but for hierarchical clustering with a fixed distance matrix, the tree structure is retained regardless of the cut height. While the cut height changes the content of clusters, increasing the number of clusters will simply break down a large cluster into smaller ones. In this sense, if we can support our claim with smaller clusters (or a large number of clusters), our arguments might be better supported.

However, the reviewer brought up interesting questions that have made us reflect on our clustering procedure: (1) if we used a different method for tree-cutting, how variable the number of clusters would be? And (2) can our argument hold against smaller clusters (or a large number of clusters)? To answer the first question, we tested a total of four different tree-cutting methods/criteria: elbow method, gap statistics, maximum average silhouette coefficient, and the dynamic hybrid cut tree method (Langfelder et al., 2008) which has been successfully utilized for morphometric classification of mouse V1 neurons by Gouwens et al. (2019). We noticed that using the maximum average silhouette coefficient resulted in the most accurate cluster number in AL (N=54), but it showed a significant discrepancy in cluster number in MB calyx and LH when compared to the other three methods (see Table 1). After an internal discussion, we have decided to switch our tree-cutting method to the dynamic hybrid cut tree method for the following reasons: (i) choosing a tree-cutting method that returns a cluster number close to the number of different odor types (which is 10) best serves our goal, (ii) since we are interested in the clusters of uPN innervations to MB calyx and LH (as the correct labels in AL are already given), it seemed reasonable to choose a method out of the three methods that produced comparable cluster numbers in MB calyx and LH, (iii) the dynamic hybrid cut tree method was proven to give satisfactory results and is methodologically systematic based on our previous experiences and previous literature, and (iv) we wanted to test our hypothesis under a harsher condition to answer question 2.

Switching the tree-cutting method resulted in 10 clusters in MB calyx and 11 clusters in LH. We report that our argument still holds regardless of the higher cluster number. The trend of the uPNs under the same homotype falling into the same cluster became much more apparent (Figures 3, 4, and 8). Our various statistical tests also returned results in line with our findings.

As for the suggestion of a better explanation of Pearson’s Chi-squared test, we report that our manuscript is now significantly more accessible to the readers when it comes to communicating new concepts and introducing statistical quantities. Most of the equations and detailed statistics are moved to either the Methods or the appendices, replaced by simple and intuitive descriptions. Whenever we thought our explanation might be too difficult to follow, we added illustrations to supplement the text.

The authors do not consider their MB clusters in-depth or explain what one would expect from the field. The literature suggests that PN-KC connections in the field are either random or semi-random (Eichler et al. 2017; Zheng et al. 2020; Caron et al. 2013). The authors could discuss their results in this context, in particular, their conclusion that 'the distinction between the homotypes is only weakly present in MB calyx'. The structure of the PN-MB clusters might be compared to results from Zheng et al. 2020, who show that the calyx contains a 'fovea' of Kenyon cells oversampling particular PNs, specifically PNs that are from 'food-related' glomeruli. This PN group is probably captured somewhere in the authors' analysis. Interestingly, the authors show that pheromonal largely PNs cluster together in the MB, though they do not draw as much attention to this as the LH pheromonal grouping for some reason – could this be a pheromonal 'fovea'?.

We thank the reviewer for the suggestion. We are familiar with the literature brought up by the reviewer but actively decided not to pursue the subject in the draft due to the scope of our study. Originally, we did not analyze connectivity between uPNs and KCs, and without this information, we had to be very cautious when making any claims that may require PN-KC information to validate.

That said, we followed the reviewer’s suggestion and performed additional analyses to test the labeledline strategy using the PN-KC connectivity information. We now feel comfortable discussing the spatial properties of uPN innervation in MB calyx and the connectivity between uPNs and KCs. In our connectivity-based clustering, we observed a comparable ‘fovea’ of ‘food-related’ glomeruli seen by Zheng et al. (2020), based on the similarity in the connectivity pattern (see Figure 11). However, this ‘fovea’ does not seem to be completely driven by spatial proximity. We did observe a similarity between spatial proximity-based clustering and connectivity-based clustering for homotypes such as DM1, DM4, DP1m, DP1l, VA2, and VA4, all of which had a high deviation from the random bouton model in Zheng et al. (2020). While many of these homotypes are generally spatially proximal (the vast majority of the uPNs are located in clusters C6MB and C7MB), some homotypes under the food-related ‘fovea’ such as VA2 are sampled from spatially distinct clusters. Therefore, we believe the creation of ‘fovea,’ or the PN-KC connectivity in general, is only partly driven by the spatial properties of uPN in MB calyx, and it appears that other factors play a role. Interestingly, the fact that many pheromone-encoding homotypes are spatially proximal in MB calyx may suggest the existence of pheromone-encoding `fovea,' but most uPNs in these homotypes do not converge in connectivity-based clustering with an exception of VA1d. Therefore, we assume that a pheromonal ‘fovea’ does not exist in MB calyx. In fact, we suspect the spatial organization of pheromone-encoding homotypes in MB calyx, which is placed at the center of the neuropil, to facilitate the observed randomization of connections by increasing the accessibility of KCs to these homotypes. What we found is a potential hygro/thermo ‘fovea,’ where glomeruli such as VP1d and VP2 are both spatially and connectivity-wise clustered together (along with VL1, curiously).

In the LH, the authors focus on their findings of a pheromone-specific cluster within the LH. However, one result they also have is that the other clusters have a mixture of different odour class PNs, so the LH does not easily break down into siloed compartments like the AL. A few studies have tried to break down the LH into different regions that serve different sections of an ethological odour space, based on different analyses (Strutz et al. 2014; Bates et al. 2020; Frechter et al. 2019). If the authors could at least qualitatively compare their groups to the previous groupings, their results and contribution would make more sense in the context of the field.

We thank the reviewer for the valuable information. We have reviewed the references and updated the manuscript to reflect the literature. Bates et al. seem to focus on the axo-axonic PN connections which share certain similarities with our clustering result, presumably driven by the necessity of spatial proximity to form synapses. For example, community 12 by Bates et al. is largely composed of VP1l and DL5, which resembles the cluster C10LH. Community 6 contains a mixture of VA5, VC1, D, DA4l, DC2, DA3, and VA7m, which is reminiscent of the cluster C6LH. However, the results cannot be mapped one-toone, and we suspect many spatially proximate homotypes do not make axo-axonic connections.

The study by Strutz et al. is particularly interesting because their results are comparable to ours. The three domains (LH-PM, LH-AM, and LH-AL) suggested by Strutz et al. seem to be a different combination of our clustering result (LH-PM and LH-AM correspond to the dorsal-posterior region and LH-AL corresponds to a combination of ventral-anterior region and the biforked bundle). The results by Frechter et al. suggest the spatial compartments of uPN innervations in LH cannot be directly translated to chemical feature segregation. However, the spatial concentration of ester-encoding LHNs in the Drosophila brain identified by Frechter et al. is intriguing as many homotypes encoding signals originating from esters are integrated by a group of common LHNs (see Figure 11B).

On C: The authors are interested in looking at 'homotype' uPNs – neurons with dendrites in the same AL glomerulus. This is a novel and interesting approach. The intra-homotypic-PN-distance (a 'bundling' metric) and inter-homotypic-PN-distance scores (a 'packing' metric) they use are insufficiently explained in the text and are better understood looking at figure 4. Figure 4 is a good, concise summary. Because sister PNs are essentially copies of the same cell and occupy the same space – a phenomenon seen across the whole of the Drosophila nervous system – it is unsurprising that they arborise similarly in each of the three analysis regions – the authors show, however, that they are 'morphologically similar in different ways' between the AL, the MB and the LH, which is an interesting point. However, the text does not make it very clear to the reader what the authors have shown. On the intra-homotypic-PN-distance: sister neurons are similar in all three regions – though in the AL they seem to tile glomeruli rather than -- intermingle (Figure 4). This last point is not, I think, shown quantitatively, only schematically, – it would be best to formalise and plot somewhere. Considering the inter-homotypic-PN-distance: neurons interdigitate across homotypes in the MB, to a lesser degree in the LH, and barely in the AL, instead of tiling the whole AL.

As we understand, the reviewer is pointing out the structural (or morphological) similarity between uPNs under different homotypes within neuropil even though technically, the innervations are parts of the same neurons. We also noticed the difference in the morphology of arborization but decided not to pursue the subject in this manuscript, as our primary goal of this paper is to analyze the spatial and organizational characteristics of uPN innervation in each neuropil. We believe the suggested study may be a bit tangential to the scope of this paper.

We, however, have done a related analysis on the morphological features of PN innervations at each neuropil, and some of the results are currently in press. A preprint is available at

(https://www.biorxiv.org/content/10.1101/2022.04.07.487455v1), should the reviewer wish to see a further discussion on the morphological classification of PNs based on a quantitative measure we devised. We have supplemented our discussion with the reference to our preprint for the readers who might be interested in the subject. We would also like to emphasize that we are very much interested in this subject and are planning on a follow-up paper with the volumetric study of PN innervations included.

The authors explore how their morphology findings might correlate with the odour response properties of PNs. They use odour categories from Bates et al. 2020, though in many cases a PN type was given multiple categories (see the supplemental figures), and it is not clear how the authors resolved this conflict. For example, the V glomerulus can also be viewed as 'aversive/bad' as it encodes CO2. Often a neuron cannot be given a singular 'odor scene' label, though they might be clustered using odour response data (Badel et al. 2016).

As the reviewer has noted, labeling a homotype using a single odor category can be insufficient in describing the wide spectrum of odor molecules it might encode. This can be especially jarring for PNs receiving inputs from ORNs expressing receptors that are broadly tuned. We tried to follow the convention set by Bates et al. as much as possible although we did notice a few homotypes (e.g., VM7d) fall under multiple categories. We noticed that this issue is specifically prevalent in odor categories one might consider ‘food-related,’ somewhat unsurprisingly. Therefore, whenever we make comments on these odor categories, we tried to make qualitative assessments only over the general category of ‘foodrelated’ homotypes. Additionally, we decided to make this detail explicit to the readers by adding the following section to our manuscript (see page 18 in the revision): “The odor type and odor valence information were extracted from various literature (Hallem et al., 2004; Galizia and Sachse, 2010; Stensmyr et al., 2012; Mansourian and Stensmyr, 2015; Badel et al., 2016; Bates et al., 2020) and we closely followed the categorical convention established by Mansourian and Stensmyr (2015) and Bates et al. (2020). However, we note that the categorization of a uPN under a specific odor category may overshadow the complete spectrum of odorants a uPN might encode, especially if the uPN encodes ORs that are broadly tuned. Therefore, we focused on the well-separated pheromone/non-pheromone encoding types and valence information.”

On D: In the conclusion, the authors do not discuss their own results in relation to the field that much. They only mention (i) one non-novel finding and (ii) one observation. The (i) finding is that certain pheromonal PNs segregate together in the LH (Ruta et al., 2010; Kohl et al., 2013, Frechter et al., 2019; Chakraborty and Sachse, 2021) – although their finding of the same in the MB is perhaps more novel? The authors say that "Our study not only lends support to the existing studies pointing to the labeled-line strategy in the Drosophila olfactory system but also suggests that an even more sophisticated level of spatial organization that depends on the homotypes and odor should be present"; it is unclear to me what the second half of the sentence means. The (ii) observation, implicit also in other work (Grabe et al. 2016) – that some PN cell types exist as singletons, often aversive ones. Certainly, in insects, there is some circuit-level difference between cell types that exist in multiples or as singletons – which are often larger, older neurons – but the present paper has no analysis on this point. Perhaps the authors could show how these singletons are morphologically special – e.g. they mention that they are more 'dense' but do not quantify this anywhere in the paper.

When we stated that a more sophisticated level of spatial organization is present, we originally meant three things: (1) the spatial organization of uPNs differs greatly in each neuropil which can be quantitatively measured (Figure 5), (2) uPNs in a homotype are tightly bundled in higher olfactory centers (both MB calyx and LH) despite the lack of glomerular structures (Figure 8), and (3) the spatial organization that supersedes the pheromone vs non-pheromone exists and may depend on the odor types (Figures 3, 4, and 6). To address the ambiguity in the conclusion, we have overhauled and greatly enhanced our Discussion section to better discuss our findings in relation to previous literature.

As stated in the previous response, we have done a separate study on the structural aspect of the PN projections to each neuropil, in which the morphological characteristics of ‘singletons’ were touched upon. A more detailed study on this subject is planned.

One of the authors' key conclusions and a point in their abstract, is that a labeled line strategy is at work in the Drosophila olfactory system at the point that second-order projections ramify in the lateral horn (LH). They make expansive statements such as: "Overall, our findings suggest that the Drosophila olfactory system leverages the efficiency of the labeled-line design in sensory information processing". A 'labeled line' is generally taken to be a chain of neurons that transfers a message about a single feature, onto higher-order neurons. This message may be modified and transformed along the way, but is generally not directly integrated with information about other features. The very peripheral part of the system in the antennal lobe (AL), ORNs -> PNs, is generally established to be a set of labeled lines (Couto et al., 2005; Fishilevich and Vosshall, 2005; Vosshall et al., 2000). Although significant cross-talk does exist, e.g. through AL local neuron computation, at this stage the PNs can still act as labeled lines (Olsen et al. 2010; Seki et al. 2017), which could support a somewhat labeled synfire chain up to the 3rd order (Jeanne and Wilson 2015).

The authors do not define the term 'labeled line' clearly, or whether the label in question should be glomerular identity, odour identity, or odour scene. They also do not clearly state (a) what neurons they mean to say are in the labeled line, nor (b) whether they think this is largely the case, or only in specific channels – their sweeping statements suggest they mean the former, but their data only weakly support the later. I assume, for (a) they are suggesting that some lateral horn neurons (LHNs) will be part of labeled lines, but not Kenyon cells of the mushroom body, because of the PN morphological properties that the authors quantify in the LH and MB. For (b), established work in the field suggests a murky division within the LH to support odour categorisation (Frechter et al. 2018; Jeanne et al. 2018) – a convergence scheme, and so, not a labeled line system – and the authors' own results show PNs of different, broad odor classes intermingling in 3/4 of their LH clusters. Their analysis, for (b), seems to show that only a small subset of PNs have the appropriate morphology to support labeled line connections in the LH. The main PNs that may form a labeled line, are the pheromonal PNs in the anterior-most region of the LH. This has been noted by the field, in terms of both morphology and connectivity (Ruta et al., 2010; Kohl et al., 2013, Frechter et al., 2019; Chakraborty and Sachse, 2021; Bates et al. 2020; Bates et al. 2020; Jefferis et al. 2007).

However, they do not demonstrate that they act as labeled lines. To make a statement about labeled lines is to make a statement about connectivity. It can be guessed by using morphology when connection data is absent. However, the connection analysis can now be done using data from the hemibrain connectome (Scheffer et al. 2020; Schlegel et al. 2021). Labeled line strategies may exist for some odour channels in Drosophila – in particular pheromonal channels such as DA1 (Kohl et al. 2013) or the aversive channels the authors note – but it is not possible to determine this using morphology alone. A preprint that examines DA2 – an aversive PN thought to be part of a labeled line (Stensmyr et al., 2012) – actually saw that in the connectome its targets experienced a lot of convergence from different PNs, only a few preserving a possible labeled line (Huoviala et al. 2020). Labeled lines are probably the exception, not the rule, and with DA2 a strong labeled line organisation seems to become a highly distributed representation at the lateral horn stage. If the authors could compare their work on morphology, to the reality in terms of connectivity, they would be better able to support their ideas and show that the features they quantify correlate with circuit structure.

We thank the reviewer for the valuable opinion. First, we believe it is necessary to clarify what we meant by ‘labeled-line.’ It was not our intention to suggest the PN-KC or the PN-LHN connections follow the ‘labeled-line’ design in general. We used the term ‘labeled-line’ predominately to denote the homotypic bundling that largely extends to MB calyx and LH, instead of ending at the ORN-PN interface (AL). In AL, obvious glomerular structures and ORN convergence to specific glomerulus is already well-known. Our intra-, inter-PN distances, and the clustering results suggest the bundling of PN homotypes is generally well-preserved throughout the neuropils and spatially localized. We were suggesting the ‘labeled-line’ in terms of the glomerular labels (or homotypes) based on this result. Regardless of what the connections to KCs and LHNs are like, PN organization alone could be deemed to express the ‘labeled-line’ principle up until the synaptic interface.

However, the suggestion from the reviewer piqued our interest in this subject and we decided to use the connectivity data to conduct an additional small-scale study to check how far the labeled-line principle holds. For a comprehensive connectivity dataset between PNs and higher olfactory neurons such as KCs and LHNs, we utilized the hemibrain dataset. We queried KCs and LHNs with synaptic connections greater than or equal to 3 for the 120 PNs we tested for the reproducibility study, which resulted in 1754 KCs and 1295 LHNs. Here, we tried to address the following questions: (1) Are there any KCs and LHNs that carry a specific type of information? If so, how prevalent are they? (2) Are there any insights we can gain from comparing the clustering outputs using spatial proximity (dαβ) and connectivity?

From the connectivity data, we observed that most of the KCs and LHNs integrate information from multiple homotypes but there are also a small number of KCs and LHNs that synapse only with a single homotype (Figure 10). These ‘homotype-specific’ connections (Nx,spξ), defined as the number of thirdorder neurons that only synapses with a specific homotype but not with the others (see Figure 9 and Methods for more information), are much more prevalent in LHNs compared to KCs. Certain homotypes (e.g., DA1) have an especially high number of LHNs that only connect to the given homotype. We also noticed that hygro/thermo-sensing homotypes have a generally higher percentage of ‘homotype-specific’ LHNs. The ‘homotype-specific’ neurons functionally carry a single type of information, thereby may be considered as an extension of the labeled-line strategy.

When we collected LHNs connected to a particular homotype and checked which other homotypes these LHNs are also synapsing (thereby analyzing the scope of signal integration happening at LH – see Figure 9 and Methods for more information), we found a strong tendency of signals from pheromone and hygro/thermo-sensing uPNs to be integrated within the given odor/signal type (Figure 11). For example, signals encoded by pheromonal homotypes share many LHNs, integrating pheromone-related signals and forming odor type-specific ‘channels’ (see purple arrows in Figure 11B). Many food-related homotypes, such as DP1l and DL2v, also share common LHN channels. It seems like the homotype-specific labeled line predominate until the signal reaches the higher olfactory centers, some of which then transition into odor-specific channels where either a broad or a narrow integration occurs. This characteristic aligns with the spatial segregation we observed in our spatial proximity-based clustering study. In MB calyx, no such trend is observed, further supporting the previous literature on randomized connections. Even though we observed a strong per homotype bundling tendency at MB calyx, the high degree of overlapping (denoted by λ) seems to have a bigger impact on connectivity.

To address the second question, we performed a connectivity-based hierarchical clustering and compared the result against our spatial proximity-based clustering result. For the ‘distance’ between the connectivity patterns of two uPNs to third-order olfactory neurons, we used the cosine distance between two vectors, which has been previously used in the field to analyze the connectivity matrices (Bates et al., 2019, Bates et al., 2020, Eschbach et al., 2020). For the hierarchical clustering, we used Ward’s criterion, which minimizes the variance of merged clusters. Tanglegrams are plotted to compare the dendrograms generated from the spatial proximity (dαβ) and from the connectivity (dcos) (see Figure 12).

One thing we suspected of certain odor types (or scenes) is that the label extends further to encompass multiple homotypes with common functionality, based on the previous literature commenting on the segregation of pheromone and non-pheromone encoding PNs (Jefferis et al., 2007; Seki et al., 2017; Chakraborty and Sachse, 2021). Our connectivity analysis showed that the connection between PNs and LHNs is highly systematic (see Figure 12B). We quantitatively studied the output by (1) analyzing the relationship between neurons based on connectivity-based hierarchical clustering and (2) comparing the relationship between two hierarchical clustering outputs. We ran statistical tests on the glomerular labels, odor types, and odor valence against the connectivity-based clustering output. First, we found a statistically significant association between the glomerular labels/odor types/odor valence and the PNLHN connectivity (p ≪ 0.001). This is visually observable through the tanglegram, where PNs under the same homotypes are within small cosine distances and thus are neighboring in the dendrogram (Figure 12B).

PN-KC connectivity (see Figure 12A), on the other hand, has a much weaker association overall, with many homotypes that used to be spatially grouped into a cluster now distributed across multiple clusters. This is also in line with previous literature. We do see, however, a small subset of homotypes that are spatially clustered and carry similar connectivity patterns, such as D, V, and VM7d.

The two tree structures (one for spatial proximity and another for connectivity) are compared using several different metrics. First, we calculated Baker’s γ index, which is a measure of the rank correlation between two lists (corresponding to the leaves in a dendrogram) spanning from -1 to 1 (where 0 indicates the ordering of two trees are completely dissimilar and 1 or -1 indicate the ordering of two trees match). We found that GBakerMB = 0.286 and GBakerLH = 0.219. Baker’s γ index in LH is in line with the previous literature by Bates et al. (2020), who conducted a similar study using the NBLAST scores and reported GBakerLH = 0.21. What is surprising is GBakerMB, which had a higher value compared to that of LH. We assume that this is due to Baker’s γ index only considering the ordinal relation between leaves. In our opinion, to provide a comprehensive description of the tanglegram, the ordinal relation (quantified by Baker’s γ index) should be supplemented by additional metrics, such as transferability (quantified by entanglement) and the correlation between two tree structures (quantified by cophenetic distance correlation).

We, therefore, supplied two additional measures that provide a complete picture of the tanglegram. The first is the entanglement, a measure ranging from 0 to 1, which quantifies the number of lines crossing in a tanglegram. We report the entanglement at MB calyx to be 0.35 and at LH to be 0.26 (a lower entanglement score suggests a possibility of block-to-block mapping without disturbing the tree structure). Next, we computed the cophenetic distance correlation between two trees. The cophenetic distance between two leaves measures the minimum height of a tree that contains both leaves. We calculated the Pearson’s correlation coefficient using the pairwise cophenetic distance between all leaves.

We report r = −0.032 (p > 0.001) for MB calyx and r = 0.236 (p << 0.001) for LH.

The authors say that "our analysis for the second-order neurons inside the Drosophila olfactory system can be translated to the brain of different organisms including the central nervous system (CNS) of humans". However, I do not think this is strictly true. In Drosophila, neurons of the same cell type, the author's 'homotypes', are near isomorphic and occupy a similar space. Therefore, the authors can use, as they did, a metric defined in Euclidean space to compare the neurons, without spatially transforming the cells at all. In mammals, duplicate types can appear across, say, cortical columns and even in the olfactory system be very spatially segregated, since there is not just one glomerulus for each olfactory receptor, as in insects. Therefore, the spatial analysis that would need to be done in mammals would be different.

While we understand that duplicate but spatially segregated homotypes exist in mammals, we still believe a distance metric in Euclidean space can still work. Instead of per homotype categorization, the calculation can be done over each glomerulus. Different glomeruli under the same homotype can be compared, or if one wishes to consider features per homotype, glomeruli under the same types can be averaged. The detailed methodology must differ, but we believe a similar distance metric can be applied to the mammalian olfactory system in principle.

It is commendable that the authors have made their analysis in python available on GitHub (https://github.com/kirichoi/DrosophilaOlfaction), as well as preprinting their work (https://www.biorxiv.org/content/10.1101/2022.02.23.481655v1). The authors pre-process the skeleton data well, removing registration artefacts. The consensus lateral horn volume (Ito et al. 2014) is somewhat arbitrarily defined as the uPN terminus region, though the MB Calyx is more strongly defined by glial sheathing. The authors wisely do not use this volume but re-define their neuropils – perhaps more accurately regions of interest – by using density estimates built from the neurons' 3D points. It would have been nice to see, in a supplementary figure, how these ROIs differ from the Ito et al. 2014 standard neuropils – since when they use neuropil terms in the text, an informed reader will assume they used the Ito standards. I like the visual segmentation process description in Figure S5.

Unfortunately, we were unable to recover the surface point coordinates from the paper by Ito et al. (2014) as suggested. Instead, we present the comparison between our segmentation outputs to that provided in CATMAID. Our segmentations are largely within the boundary given by CATMAID. A slight difference is observed for MB calyx at the medial and posterior region of the neuropil. However, we do not think this is a serious issue since we are segmenting neuropils from a bundle composed solely of uPNs, where to goal is to simply remove the rigid PN backbone from the synaptically dense neuropils. Therefore, the accuracy of our boundaries is not as influential compared to segmenting neuropils out of a dense mixture of different types of neurons. The minor difference in the density we observed might be due to the lack of uPNs that does not innervate all three neuropils and mPNs (which were dropped when generating the boundaries) when we segmented the neuropils.

Author response image 1
Comparison of volumes defined in CATMAID (blue) and boundaries defined by our segmentation method (black) for MB calyx (top) and LH (bottom).

I think the paper does make many quantitative points on PN morphology well, and could either become a shorter manuscript that shows this information concisely, or a more involved one on 'labelled lines' and the validity of this idea for the olfactory system, where 2nd-order neurons meet 3rd order ones. In the first case, adding thermo sensory PNs, GABAergic uPNs, and mPNs to their analysis could make it more interesting, and produce some new and germane insights for their sub-regional analyses (see below). In the second case, the paper would need to include work that compares across different PN classes and EM datasets, including looking at connectivity in the hemibrain, to make its core claims on labeled lines.

In general, in the present work, the authors do not provide enough biological background information on the olfactory system and what is known about PNs already, the regions they innervate for a naive reader to understand their results, and why their results might be interesting. This could be easily fixed with a little re-writing. The use of mathematical notation within the main text is heavier than needs be and hampers the readers' understanding of what the authors are doing and intend to show. Plainer writing would help. The authors need to communicate their core findings, and their context in the field, more clearly.

We thank the reviewer for the suggestion. In the end, we decided to implement parts of both cases that were suggested by the reviewer. First, we have re-done our calculation using the latest FAFB dataset, including hygro/thermos-sensing uPNs. We have performed a comprehensive study on every uPNs that innervate all three neuropils and a separate analysis including uPNs that does not innervate all three neuropils. Many GABAergic uPNs that were originally left out are now incorporated into the manuscript. Additionally, we performed an extensive study on the labeled-line hypothesis using connectivity data with third-order olfactory neurons. Our connectivity-based studies examined the ‘homotype-specific’ KCs/LHNs, signal integration across different homotypes done by KCs/LHNs, and the comparison between spatial proximity-based and connectivity-based clustering.

We have updated our manuscript, along with its figures and captions, so that readers have less trouble digesting the content. We have considerably updated the manuscript by adding various biological backgrounds and previous literature the reviewers have mentioned.

Glomeruli identification through the literature can be tricky. There has been some confusion specifically about the identities of VM6 and VC5. The Zheng et al. 2018 paper had a few errors, later rectified by Bates and Schlegel et al. 2020 and then again by Schlegel and Bates et al. 2021. Given the confusion, no one can be faulted for mistakes here, but authors should make sure their labels are the same as in Schlegel and Bates et al. 2021 (that process was a multi-lab debate). I think they used the outdated Zheng et al. 2018 labels.

We thank the reviewer for the valuable information. Since we have re-done our calculation based on the latest FAFB dataset (Bates et al., 2020), we ended up predominantly using the labels by Bates et al. We have further updated our labels according to Schlegel et al. (2021), regarding glomeruli VC3m, VC3l, and VC5.

Suggestions

I have some suggestions for increasing interest in this work, that I humbly submit to the authors. The authors discuss 51 glomeruli of the antennal lobe. Of the olfactory glomeruli, there are 52, but there are actually 58 glomeruli in total in the antennal lobe (e.g. VP1-5), including the thermo-hygrosensory ones. Other work (Schlegel et al. 2021; Bates et al. 2020) has found that some of the most striking differences can be found between olfactory and thermo-sensory glomeruli, as well as some curious associations across the two modalities. Since intellectually this system is very similar – RNs contact PNs that then reach the LH and the zone just ventral to it – but not quite parallel – arbours intermingle with olfactory ones in the ventral LH, and antennal lobe local neurons cross compute between combinations across all 58 glomeruli – I strongly think including these neurons in the author's analysis would increase the interest in this work.

There are a total of 347 PNs in the FAFB dataset that project from the right-side antennal lobe. This is because there are 283 multi-glomerular PNs (mPNs). Bates et al. 2020 also make 58 uPNs from the left side of the brain available. All of the data is open and can be gotten by the authors here: https://fafb.catmaid.virtualflybrain.org/. Similar to the above, I think that including the mPNs as a comparison point, would increase interest in this work.

We thank the reviewer for the suggestion. As noted in previous responses, we have re-done our calculation using the latest FAFB dataset, from which we found 57 glomeruli (we couldn’t find the.swc file associated with neuron ID = 1356477 forming VP3 in the dataset and other uPNs in VP3 does not innervate all three neuropils) including those involved in hygro/thermos-sensing, and the hemibrain dataset, from which we recovered all 58 glomeruli. As pointed out by the reviewer, we observed hygro/thermos-sensing uPNs to innervate significantly different regions of MB calyx and LH. uPNs forming these glomeruli are spatially segregated from the odor/pheromone encoding uPNs. In MB calyx, we found these uPNs to rarely project ventrally, generally clustered to the base of the neuropil. In LH, we found these neurons to be clustered in the dorsal-ventral region of the neuropil, hardly innervating the neuropil but covering the medial side of LH. The manuscript has been updated to include this information. Additionally, as described in previous responses, we have updated the manuscript with the additional analyses of many GABAergic uPNs that were originally left out of our study.

PNs could also have been clustered by their odour response profiles using data from PN response (Badel et al. 2016) or ORN response (Münch and Galizia 2016), rather than porting labels from Bates et al. 2020 / Mansuorian et al. 2015. The authors might consider whether this could add meaningfully to their analysis.

We thank the reviewer for an interesting suggestion. The odor response/dose-response profiles are critical in functional and information-theoretic analyses of the olfactory system. A related study would be a valuable addition to our odor type-dependent analyses by providing a continuous physiological variable to study against the spatial characteristics of uPN innervations. Unfortunately, we have a limited time for the revision of this manuscript, and a satisfactory study involving the odor response profile may not be possible within this time frame.

In addition, as noted in the public review, I think comparing against or using NBLAST could be informative. The authors could run NBLAST on their processed neuron skeletons and discover whether the results differ much from what they have in hand right now. NBLAST is now implemented in the navis python library: https://navis.readthedocs.io/en/latest/source/tutorials/nblast.html.

We have calculated the normalized NBLAST distance between uPN innervations in each neuropil and compared it against our dαβ (see Figure 2—figure supplement 1B). While the two distances are highly correlated, there are cases with the same NBLAST distance with different dαβ, as dαβ only measures the spatial proximity but not the morphological similarity between two neurons. As we stated before, the primary goal of this paper is to analyze the spatial and organizational characteristics of uPN innervation in each neuropil, and we believe the distance metric dαβ is conceptually more physical and adequate for the primary aim of our study.

Lastly, a more out-there suggestion: PN morphology analysis using a skeleton representation has been common in the field. However, volumetric analyses based on neuronal meshes – now available for these neurons through the hemibrain and flywire projects – is almost non-existent. How do the volumes for neurons vary across types, clusters, etc? Using this information could add some simple points, to support the authors in their quantitative description of these neurons.

As the reviewer pointed out, neuronal volumetric analyses of the brain are relatively rare, largely due to the lack of data to perform said analyses until now. We agree this is indeed a very interesting topic, but due to the limited time we have for this manuscript revision, we are afraid that a comprehensive study of the volumetric characteristics may not be possible at this time. However, we are quite enthusiastic about this subject, and we are eager to perform a follow-up study in the near future to answer this question.

Reviewer #3 (Recommendations for the authors):

This work provides a quantitative evaluation of the spatial organization pattern of olfactory projection neurons (PN) in AL, CA, and LH based on the inter-PN distances using FAFB EM dataset. This NBLAST-based method does provide a simple way to cluster neurons with similar projection patterns and even predict underlying signal processing rules. However, the reliability of the clustering method needs to be improved since the method only got 39 clusters out of 51 well-segregated AL glomeruli serving as the ground truth. This result undermines the conclusions in the higher-brain centers which have a more complex organization. Moreover, the authors defined the neuropils by rotating the neurons along specific axes and then segmenting the dense innervation parts, which may hinder the accuracy of the boundary when the surface is convoluted and does not reveal inner subdomains. Nevertheless, serving as the first step to tackling complex connectomic data, the method is easy and potentially useful.

1. The current study analyzed only 111 uniglomerular PNs instead of the latest-released 164 uniglomerular PNs (Bates et al., 2020). To make the work more valuable, the authors should apply their analysis to the latest dataset. In addition, since the analysis was done from the EM data of a single fly, whether the preferential spatial distribution of PNs and the clustering are consistent in different individuals is unknown. It will be informative and more persuasive if a similar spatial distribution pattern can be observed in another fly EM dataset (Scheffer et al., 2020).

Following the reviewer’s comment, we have re-done all our calculations using the PNs that meet our existing criterion in the latest FAFB dataset (Bates et al., 2020). We ended up analyzing 135 PNs that innervate all three neuropils in the latest FAFB dataset, including those involved in hygro and thermo sensation. Also, we have performed a separate analysis on the rest of the uPNs that did not innervate all three neuropils (28 uPNs), most of which are GABAergic. We have updated the manuscript and the figures, pointing out several newly found characteristics. Additionally, we have applied our calculation to the hemibrain dataset (Scheffer et al., 2020) to make sure that our results are reproducible and generalizable in different individuals. We generated several figures in the main text using the hemibrain dataset and compared them against the figures generated from the FAFB dataset. The results from both datasets are in support of our arguments (Figure 13).

2. The segmentation of neuronal innervation in AL, Calyx, LH is achieved by rotating the neurons along specific axes and then identifying the dense innervation parts as the three neuropils (Figure S5). The methodology is convenient to define the neuropils but will hinder the accuracy of the boundary segmentation when the surface is convoluted (i.e. MB calyx) and does not reveal inner subdomains (i.e. AL glomeruli). As the original EM dataset has already offered neuropil surface point coordinates which can be downloaded from Catmaid website (https://catmaid-fafb.virtualflybrain.org/), attributing the innervation points directly to the defined neuropils should be more accurate.

We appreciate the reviewer’s concern since systematically segmenting different parts of a brain is indeed a surprisingly challenging problem. However, we do not believe this issue will cause a serious problem for our study. Our segmentation process was done only on uPN bundles to separate neuropils from rigid PN backbones. This means that an accurate boundary is preferable but not strictly necessary since we do not have to worry about incorporating non-PNs into our study. We are also not too concerned about identifying the inner subdomains since our analyses do not necessitate to re-segment the AL but instead use the pre-determined labels offered by the original authors of the dataset. We have compared our boundaries with the volume surface provided in CATMAID and found that our methodology seems to be good enough for the purpose of detaching synaptically dense regions from the PN backbones.

3. Theoretically, the natural segregation of glomerular structures in AL would serve as the ground truth to test the reliability of the clustering method. Yet, the method only got 39 clusters out of 51 well-segregated glomeruli. This discrepancy undermines the effectiveness of the clustering strategy. The result of silhouette coefficient analysis also suggests that the coefficients are similar for clusters ranging from 30 to 50 in AL (the coefficients are very close almost reaching a plateau) (Figure S6). The author should justify why choosing the number of PN innervation clusters in AL as 39 and provide evidence that it is optimized to find the hidden pattern. Most importantly, if the method fails to reveal the remaining 11 AL glomeruli that can be visually distinguished, it is difficult to see how it can reveal the hidden pattern in MB calyx which is much less well defined.

As the reviewer pointed out, we have fewer clusters for AL than the known number of glomeruli available in the dataset. However, it should be noted that this is primarily an issue of a particular tree-cutting method we used, not the hierarchical clustering per se. When looking at the dendrogram for AL, one might notice that the well-segregated glomeruli are indeed expressed through uPNs forming the same glomerulus grouped under a common branch. We could arbitrarily choose a cut height that correctly recovers all known glomeruli available in the latest FAFB dataset, suggesting that our distance metric and the clustering algorithm are appropriate to show the hidden pattern in the spatial organization. If the tree structure correctly produces the expected grouping of uPNs in AL based on the glomerular labels, we believe the actual number of clusters formed through a specific tree-cutting method is less relevant – a smaller number of clusters is simply a manifestation of grouping several glomeruli together. In our application, clustering on AL is not necessary since we already know the correct labels for each uPN. On the other hand, we do not know the correct labeling for uPNs at MB calyx and LH, so we had to resort to some systematic approach to figure out the appropriate number of clusters. For this purpose, we decided to use a method that returns a cluster number close to the number of different odor types (which is 10), as we believe that would best serve our primary goal of this paper.

4. The neuronal distance estimation algorithm may encounter a problem when projection neurons have very different innervation ranges and the total length of the innervation branches. For example, if there is a projection neuron, PNa, only innervates in the entry of Calyx where most PNs pass by, the distance between PNa and other PNs will be very small due to the algorithm only selects the shorter skeleton to evaluate the distance (see method and the symmetric matrices in Figure 3). Thus, the result may not faithfully reflect the fiber distance when two PNs have a drastic difference.

We thank the reviewer for pointing out a valid concern. We believe the normalization term (the division by the number of coordinates in the neuron Nα) should alleviate this issue. This term will make sure that two long branches that are spatially well-bundled have a smaller value of dαβ than a short branch and a long branch. It is possible to have two long branches that are far apart having a similar dαβ to that of a short branch and a long branch, but we believe both cases should be treated as poorly bundled.

5. The authors estimate λ value which represents the ratio of the mean of intra-homotypic neuron distance and the mean of inter-homotypic neuron distance. (Figure 4, Figure 5, Figure 6, Figure S3). It is a creative way that gives us insight into how PNs overlap with each other in distinct neuropils. However, the volume and the aspect ratio of the three neuropils are different. AL occupies a much larger area compared to calyx and LH. Therefore, the current λ value may simply reflect the spatial distance presented in the neuropil but not the actual degree of overlapping (e.g. in two non-overlapping pairs, the one with long distance will have a low λ value while both pairs have no overlapping at all). To make the λ value reflect more to the degree of overlapping, they might need to normalize it based on the volume, compare the λ values between original data with the data after shuffling PNs into different clusters or take the innervation density of fibers in specific neuropils into account.

We thank the reviewer for their valuable opinion. First, we would like to iterate that we devised the λ to be a scale-free ratio that quantifies the relative spatial innervation per homotype. This means that we intended to not scale our λ based on the volume. Let us consider a homotype composed of several uPNs and has similar dintra values for both MB calyx and AL. Indeed, the volume of MB calyx is smaller than that of AL, so the dinter will be generally larger in MB calyx, leading to a higher λ value. This is what we intended and our logic behind this is that in AL, despite all the additional space that the uPNs technically could have innervated, the uPNs ended up localized into a bundle with a small dintra. In MB calyx, the same uPNs had less space to innervate, to begin with, so the less ‘packing’ observed in MB calyx may not be as novel. We wanted our λ to differentiate the two, which is the reason why our λ is defined as such.

We believe both definitions provide valuable information on spatial organization and therefore important. However, it seems like the term ‘overlapping’ is a bit elusive and can be defined in various ways, which we believe is why the reviewer might have been confused by our definition. We feel that our explanation and choice of words (overlapping) were not sufficiently clear. After a lengthy discussion, we have decided to keep the terminology but provide a much more comprehensive explanation of what we mean by the ‘degree of overlapping’ to make sure that the readers won’t get confused.

6. The authors annotate the reaction profile for a specific glomerulus using a dumb variable to do statistics. The result suggests that pheromone PNs will be clustered together in LH (into 2 clusters) and Calyx. Food-related PNs and aversive PNs also be clustered into different groups in LH but not in Calyx (Figure 7, Table S1). The author should elaborate and discuss more in terms of biology. For example, do food-related odor signals and aversive odor signals converge in MB calyx and diverge in LH? What biological properties may be implicated in the organization?

We believe there is a much more subtle uPN organization going on at each neuropil. First, our inter-PN distances and clustering results suggest the bundling of uPN homotypes is generally well-preserved throughout the neuropils despite the visual lack of glomerulus at MB calyx and LH. However, in MB calyx, the spatial segregation between different homotypes is practically non-existent, leading to a high degree of overlapping. In LH, the spatial segregation between homotypes becomes much more pronounced, leading to a reduction in the degree of overlapping. Therefore, in terms of connectivity, the surface presented by MB calyx is much more diverse, perhaps to assist the randomized sampling known to exist at the PN-KC interface. In LH, we find four stereotyped conformations (both spatially and morphologically) overall: (1) dorsal posterior region largely occupied by food-encoding uPNs, (2) ventralanterior region largely occupied by pheromone-encoding uPNs, (3) biforked bundle surrounding dorsalposterior region largely occupied by food-encoding uPNs with an aversive response, and (4) dorsalanterior-medial region largely occupied by hygro/thermo-sensing uPNs. The spatial localization seems to be related to the sensory signal encoded by the uPNs, as suggested in previous literature.

To gain insight into the biological implications of the spatial properties we observed, analysis based on the connectivity to high-order olfactory neurons is necessary. Our connectivity-based analyses observed a substantial number of LHNs synapsing only with particular homotypic uPNs, which seems to be correlated with the spatial segregation of uPN homotypes in LH. Additionally, LHNs tend to be much more specific about signal integration at LH compared to KCs. That is, LHNs tend to be selective about which homotypic uPNs to synapse with and the connectivity is correlated with the odor types. The thirdorder olfactory neurons that take inputs only from a specific homotype are indeed exceptions in the PNKC or the PN-LHN interface, as most of the KCs and LHNs tend to integrate signals from diverse homotypic PNs. But these special cases may be considered as an extension of the ‘labeled-line’ design seen in the second-order neurons (PNs). These ‘labeled-line’ designs have significant biological implications. The olfactory information encoded by neurons part of the ‘labeled-line’ design may be processed under a different principle from other odors. We have updated the manuscript with the above information.

7. Although the authors have offered the python scripts for researchers to reproduce their results, a comprehensive spreadsheet that contains all the distance results along with functional annotation will greatly improve the accessibility of the analyzed results.

We have shared the pre-computed pairwise distance matrices for PNs in every neuropil with the respective neuron IDs annotated for both the FAFB and the hemibrain dataset in.csv files. Furthermore, we have created spreadsheets containing some of the functional information (e.g., glomerulus labels, clustering labels, etc.), all of which are available in our repository.

8. The authors should compare their cluster results with Lin et al. 2007 and Bates et al. 2020 and discuss the biological implication.

Following the reviewer’s suggestion, we have added a paragraph in the Discussion section comparing our cluster results with that of Lin et al. (2007) and Bates et al. (2020) Our results are generally consistent with the previous studies that have clustered olfactory neurons using various forms of data. For example, in our study, homotypes DL2v and DL2d constitute a bilateral cluster in MB calyx (C3MB), and the dual organization of uPNs is present in MB calyx and LH, such that homotypes DC2, DL1, and VA5 are sorted into the same cluster in LH while sharing similar innervation pattern in MB calyx, all of which are in line with Lin et al. (2007)

The uPNs under the same cluster or nearby clusters in our study are clustered together in the NBLAST score-based clustering analysis by Bates et al. (2020). The uPNs that ended up in the same cluster or nearby clusters, such as homotypes DM1, DM3, DM4, VA4, and VM3 in the cluster C3LH, are also grouped in the NBLAST score-based clustering analysis by Bates et al. (2020). This suggests that a level of stereotypy of uPN organization in MB calyx and LH that is universal, which can be captured through different metrics and methodologies.

References:

1. Badel, L., Ohta, K., Tsuchimoto, Y., and Kazama, H. (2016). Decoding of context-dependent olfactory behavior in Drosophila. Neuron, 91(1), 155-167.

2. Bates, A. S., Janssens, J., Jefferis, G. S., and Aerts, S. (2019). Neuronal cell types in the fly: singlecell anatomy meets single-cell genomics. Current opinion in neurobiology, 56, 125-134.

3. Bates, A. S., Schlegel, P., Roberts, R. J., Drummond, N., Tamimi, I. F., Turnbull, R., ... and Jefferis, G. S. (2020). Complete connectomic reconstruction of olfactory projection neurons in the fly brain. Current Biology, 30(16), 3183-3199.

4. Bopp, R., Maçarico da Costa, N., Kampa, B. M., Martin, K. A., and Roth, M. M. (2014). Pyramidal cells make specific connections onto smooth (GABAergic) neurons in mouse visual cortex. PLoS biology, 12(8), e1001932.

5. Das Chakraborty, S., and Sachse, S. (2021). Olfactory processing in the lateral horn of Drosophila. Cell and Tissue Research, 383(1), 113-123.

6. Choi, K., Kim, W. K., and Hyeon, C. (2022). Polymer physics-based classification of neurons. bioRxiv.

7. Douglas, R. J., Martin, K. A., and Whitteridge, D. (1989). A canonical microcircuit for neocortex. Neural computation, 1(4), 480-488.

8. Eschbach, C., Fushiki, A., Winding, M., Schneider-Mizell, C. M., Shao, M., Arruda, R., ... and Zlatic, M. (2020). Recurrent architecture for adaptive regulation of learning in the insect brain. Nature Neuroscience, 23(4), 544-555.

9. Galizia, C. G., and Sachse, S. (2010). Odor coding in insects. The neurobiology of olfaction, 35, 70.

10. Gouwens, N. W., Sorensen, S. A., Berg, J., Lee, C., Jarsky, T., Ting, J., ... and Koch, C. (2019). Classification of electrophysiological and morphological neuron types in the mouse visual cortex. Nature neuroscience, 22(7), 1182-1195.

11. Hallem, E. A., Ho, M. G., and Carlson, J. R. (2004). The molecular basis of odor coding in the Drosophila antenna. Cell, 117(7), 965-979.

12. Ito, K., Shinomiya, K., Ito, M., Armstrong, J. D., Boyan, G., Hartenstein, V., ... and Insect Brain Name Working Group. (2014). A systematic nomenclature for the insect brain. Neuron, 81(4), 755-765.

13. Jefferis, G. S., Potter, C. J., Chan, A. M., Marin, E. C., Rohlfing, T., Maurer Jr, C. R., and Luo, L. (2007). Comprehensive maps of Drosophila higher olfactory centers: spatially segregated fruit and pheromone representation. Cell, 128(6), 1187-1203.

14. Langfelder, P., Zhang, B., and Horvath, S. (2008). Defining clusters from a hierarchical cluster tree: the Dynamic Tree Cut package for R. Bioinformatics, 24(5), 719-720.

15. Lin, H. H., Lai, J. S. Y., Chin, A. L., Chen, Y. C., and Chiang, A. S. (2007). A map of olfactory representation in the Drosophila mushroom body. Cell, 128(6), 1205-1217.

16. Mansourian, S., and Stensmyr, M. C. (2015). The chemical ecology of the fly. Current opinion in neurobiology, 34, 95-102.

17. Scheffer, L. K., Xu, C. S., Januszewski, M., Lu, Z., Takemura, S. Y., Hayworth, K. J., ... and Plaza, S. M. (2020). A connectome and analysis of the adult Drosophila central brain. Elife, 9.

18. Seki, Y., Dweck, H. K., Rybak, J., Wicher, D., Sachse, S., and Hansson, B. S. (2017). Olfactory coding from the periphery to higher brain centers in the Drosophila brain. BMC biology, 15(1), 1-20.

19. Stensmyr, M. C., Dweck, H. K., Farhan, A., Ibba, I., Strutz, A., Mukunda, L., ... and Hansson, B. S. (2012). A conserved dedicated olfactory circuit for detecting harmful microbes in Drosophila. Cell, 151(6), 1345-1357.

20. Zheng, Z., Li, F., Fisher, C., Ali, I. J., Sharifi, N., Calle-Schuler, S., ... and Bock, D. D. (2022). Structured sampling of olfactory input by the fly mushroom body. Current Biology.

21. Zheng, Z., Lauritzen, J. S., Perlman, E., Robinson, C. G., Nichols, M., Milkie, D., ... and Bock, D. D. (2018). A complete electron microscopy volume of the brain of adult Drosophila melanogaster. Cell, 174(3), 730-743.

https://doi.org/10.7554/eLife.77748.sa2

Article and author information

Author details

  1. Kiri Choi

    School of Computational Sciences, Korea Institute for Advanced Study, Seoul, Republic of Korea
    Contribution
    Conceptualization, Data curation, Software, Formal analysis, Funding acquisition, Validation, Investigation, Visualization, Methodology, Writing - original draft, Writing – review and editing
    For correspondence
    ckiri0315@kias.re.kr
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-0156-8410
  2. Won Kyu Kim

    School of Computational Sciences, Korea Institute for Advanced Study, Seoul, Republic of Korea
    Contribution
    Conceptualization, Funding acquisition, Validation, Investigation, Visualization, Methodology, Writing – review and editing
    For correspondence
    wonkyukim@kias.re.kr
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-6286-0925
  3. Changbong Hyeon

    School of Computational Sciences, Korea Institute for Advanced Study, Seoul, Republic of Korea
    Contribution
    Conceptualization, Resources, Supervision, Funding acquisition, Validation, Investigation, Visualization, Methodology, Writing - original draft, Project administration, Writing – review and editing
    For correspondence
    hyeoncb@kias.re.kr
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-4844-7237

Funding

KIAS individual grant (CG077001)

  • Kiri Choi

KIAS individual grant (CG076002)

  • Won Kyu Kim

KIAS individual grant (CG035003)

  • Changbong Hyeon

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

Acknowledgements

We thank Dr. Ji Hyun Bak for helpful discussions. This study was supported by KIAS Individual Grants CG077001 (KC), CG076002 (WKK), and CG035003 (CH). We thank the Center for Advanced Computation in KIAS for providing the computing resources.

Senior Editor

  1. K VijayRaghavan, National Centre for Biological Sciences, Tata Institute of Fundamental Research, India

Reviewing Editor

  1. Sonia Sen, Tata Institute for Genetics and Society, India

Reviewers

  1. Sonia Sen, Tata Institute for Genetics and Society, India
  2. Alexander Shakeel Bates, Harvard Medical School, United States
  3. Ann-Shyn Chiang, National Tsing Hua University, Taiwan

Version history

  1. Received: February 9, 2022
  2. Preprint posted: February 25, 2022 (view preprint)
  3. Accepted: September 18, 2022
  4. Accepted Manuscript published: September 29, 2022 (version 1)
  5. Version of Record published: October 28, 2022 (version 2)

Copyright

© 2022, Choi et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 990
    Page views
  • 288
    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)

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

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

  1. Kiri Choi
  2. Won Kyu Kim
  3. Changbong Hyeon
(2022)
Olfactory responses of Drosophila are encoded in the organization of projection neurons
eLife 11:e77748.
https://doi.org/10.7554/eLife.77748

Further reading

    1. Computational and Systems Biology
    2. Neuroscience
    George Zacharopoulos, Francesco Sella ... Roi Cohen Kadosh
    Research Article

    Previous research has highlighted the role of glutamate and gamma-aminobutyric acid (GABA) in perceptual, cognitive, and motor tasks. However, the exact involvement of these neurochemical mechanisms in the chain of information processing, and across human development, is unclear. In a cross-sectional longitudinal design, we used a computational approach to dissociate cognitive, decision, and visuomotor processing in 293 individuals spanning early childhood to adulthood. We found that glutamate and GABA within the intraparietal sulcus (IPS) explained unique variance in visuomotor processing, with higher glutamate predicting poorer visuomotor processing in younger participants but better visuomotor processing in mature participants, while GABA showed the opposite pattern. These findings, which were neurochemically, neuroanatomically and functionally specific, were replicated ~21 mo later and were generalized in two further different behavioral tasks. Using resting functional MRI, we revealed that the relationship between IPS neurochemicals and visuomotor processing is mediated by functional connectivity in the visuomotor network. We then extended our findings to high-level cognitive behavior by predicting fluid intelligence performance. We present evidence that fluid intelligence performance is explained by IPS GABA and glutamate and is mediated by visuomotor processing. However, this evidence was obtained using an uncorrected alpha and needs to be replicated in future studies. These results provide an integrative biological and psychological mechanistic explanation that links cognitive processes and neurotransmitters across human development and establishes their potential involvement in intelligent behavior.

    1. Computational and Systems Biology
    2. Neuroscience
    Huu Hoang, Shinichiro Tsutsumi ... Keisuke Toyama
    Research Article Updated

    Cerebellar climbing fibers convey diverse signals, but how they are organized in the compartmental structure of the cerebellar cortex during learning remains largely unclear. We analyzed a large amount of coordinate-localized two-photon imaging data from cerebellar Crus II in mice undergoing ‘Go/No-go’ reinforcement learning. Tensor component analysis revealed that a majority of climbing fiber inputs to Purkinje cells were reduced to only four functional components, corresponding to accurate timing control of motor initiation related to a Go cue, cognitive error-based learning, reward processing, and inhibition of erroneous behaviors after a No-go cue. Changes in neural activities during learning of the first two components were correlated with corresponding changes in timing control and error learning across animals, indirectly suggesting causal relationships. Spatial distribution of these components coincided well with boundaries of Aldolase-C/zebrin II expression in Purkinje cells, whereas several components are mixed in single neurons. Synchronization within individual components was bidirectionally regulated according to specific task contexts and learning stages. These findings suggest that, in close collaborations with other brain regions including the inferior olive nucleus, the cerebellum, based on anatomical compartments, reduces dimensions of the learning space by dynamically organizing multiple functional components, a feature that may inspire new-generation AI designs.