Introduction

The brain is a complex network of neurons, and understanding their wiring strategies provides crucial insights into their roles in neuronal computation. Release sites in the presynaptic terminal of Drosophila neurons are decorated with electron-dense projections composed of the active-zone (AZ) scaffold protein Brp ^1–7. These Brp projections serve as a synapse marker enabling the whole-brain connectomes, comprehensive mapping of synaptic connections at the electron microscopic level ^8–12. While connectome-based approaches catalyzed the exploration of cell type distinctions, circuit motifs, and neurotransmitters, their application to comparisons across different conditions is still challenging due to the limited throughput of data acquisition and dense reconstruction. Since the synaptic structure and connectivity reflect substantial variability of development and experiences ^11,13–18, cross-individual comparisons are necessary. Fluorescence labeling offers high throughputs and contents, especially when comparing synaptic organizations across cell-types ¹⁹ or brain regions ²⁰. Nevertheless, characterizing endogenous synapses in specific cells using fluorescence-labeling remains challenging in the central nervous system (CNS) due to the dense synaptic connections from different neurons. This has driven growing interest in designing cell-type-specific fluorescence-tagging strategies for endogenous synaptic protein ²¹.

The circuit of Drosophila mushroom bodies (MBs) plays a central role in olfactory associative learning ²². The major intrinsic MB neuron, known as Kenyon cells (KCs) respond to odor stimulation and synapse onto the spatially segregated dendrites of MB output neurons (MBONs) that divide the MB into compartments. Distinct types of dopaminergic neurons (DANs) synapse onto specific compartments and thereby modulate KC-MBON synapses in the corresponding compartments ²³. Presynaptic calcium levels of these DANs undergo sub-compartmental GABAergic modulation and inform memory specificity, postulating the distinct synapse structures at individual release sites ²⁴ . Two types of giant interneurons, the anterior paired lateral neuron (APL) and the dorsal paired medial (DPM) neuron extensively ramify across the MB lobes and provide recurrent modulations to KC synapses ^25–32. Considering their comprehensive projections throughout the MB lobes, the modulation of local circuits requires intracellular synaptic tuning. These studies together underscore the importance to spatially distinguish individual synapses within a cell. However, the high AZ density in the MB lobes precluded conventional approaches of fluorescent microscopy from profiling synaptic structures of specific neurons⁹.

Here, we present a spatial analysis of individual AZs within the MB circuit at the single-cell resolution, achieved by cell-type-specific visualization of endogenous Brp using the CRISPR/Cas9-mediated split-GFP tagging system. We developed a high-throughput quantification pipeline to systematically profile Brp clusters of individual AZs in different MB-innervating neurons. Characterizing multiple structural parameters revealed a significant synaptic heterogeneity within single neurons and AZ distribution stereotypy across individuals. These cell-type-specific synapse profiles suggest that AZs are organized at multiple scales, ranging from neighboring synapses to across individuals.

Results

Establishing an experimental system for cell-type-specific profiling of active zone structures

As the molecular assembly of synapses is reported to be sensitive to the dosage of Brp, we visualized endogenous Brp instead of transgenic expression of tagged Brp ³³. To label endogenous Brp specifically in designated cell types, we employed the split-GFP tagging system ^34,35 . Using CRISPR/Cas9, we inserted the GFP₁₁ fragment (the eleventh β-strand of the super-folder GFP) just prior to the stop codon of brp gene. The GFP_1-10 fragment is expressed in cells of interest by the GAL4/UAS system. The self-assembly of the split-GFP fragments is therefore directed only in GAL4-expressing cells and visualizes endogenous Brp through reconstituted GFP fluorescence (Fig. 1A). As a proof of principle, we directed GFP reconstitution and expressed a plasma membrane marker in the single serotonergic DPM neuron using VT64246-GAL4 (Fig. 1B). Confocal microscopy revealed Brp::reconstituted GFP (Brp::rGFP) signals only to the MB lobes, aligning with DPM terminals ³². These signals were distributed as discrete puncta within plasma membrane varicosities resembling presynaptic boutons (Fig. 1B), suggesting that Brp::rGFP clusters labeled individual AZs in the DPM neuron.

Fig. 1.

To systematically profile Brp::rGFP clusters, we developed an image processing pipeline (Fig. 1A). In brief, image deconvolution was applied to reduce out-of-focus light and enhance signal sharpness. Individual Brp::rGFP clusters were then segmented using the 3D suite plugin in Fiji ³⁶, generating the 3D region-of-interests (3D ROIs) that encircle each cluster. Brp::rGFP signal intensity, volume and location of individual clusters were measured using the 3D ROIs. To optimize the segmentation, we systematically varied the parameters and calculated the F-score to measure detection accuracy against manually defined ground truth (see Materials and Methods). The F-score is near 1.0 in segmenting Brp::rGFP clusters within DPM and APL neurons and achieved > 0.90 in KCs with certain combinations of segmentation parameters (Fig. 1B).

To confirm that Brp puncta represent AZs, we quantified the numbers of Brp::rGFP clusters in DPM and APL neurons specifically and compared them to the connectome data. We found that cluster counts closely matched the number of AZs annotated by the hemibrain connectome (Fig. 1C) ⁹. This result validated our method that accurately isolated Brp::rGFP clusters corresponding to single AZs.

AZ profiles within KCs are compartmentally heterogeneous

Brp accumulation at individual AZs is known to be heterogeneous even within a single motor neuron ^5,6,37,38. To study such AZ heterogeneity in CNS neurons, we quantified the signal intensity of individual Brp::rGFP clusters in MB lobes. Using KC subtype-specific drivers to direct split-GFP reconstitution in γ KC (MB009B-GAL4), α’/β’ KCs (MB370B-GAL4) and α/β KCs (MB008B-GAL4). we found that within-cluster Brp::rGFP intensities are highly diverse in all KC subtypes (Fig. 2A).

Fig. 2.

To visualize spatial patterns of Brp::rGFP intensity in each KC subtype, we reconstructed 3D distribution of Brp::rGFP clusters and color-coded them by signal intensities (Fig. 2B). We found that clusters showed noticeable intensity differences between compartments in all subtypes. Even adjacent compartments such as γ1 and γ2, β’2 and β’1, α2 and α3, showed drastic differences (Fig. 2B). Importantly, these compartmental patterns were stable across individuals (Fig. 2C) ³⁹. This result suggests that AZs in different compartments have distinct structures. Since individual KCs of each subtype synapse onto all compartments ^8–10,40, this Brp compartmental heterogeneity is likely formed intracellularly.

Intracellular heterogeneity of Brp concentration at individual AZs and cell-type dependent diversity

Super-resolution microscopy has shown that the Brp molecular density at individual AZs is dynamically adjusted in motor neurons ^41,42. We hypothesized that the amount of Brp does not necessarily correlate with AZ cluster size. To verify this, we calculated the Brp concentration by quantifying both the volume (number of voxels within a 3D ROI) and the Brp::rGFP intensity of each cluster. We found that clusters with similar sizes can exhibit vastly different signal intensities in the DPM neuron, suggesting the heterogeneity of Brp concentration at individual AZs (Fig. 3A). Correlation analysis between the Brp::rGFP intensity and cluster volume revealed a substantial variability among cell types (Fig. 3B-E). There were even significant differences among KC subtypes. These results suggest cell-type dependent variability of the Brp concentration of individual clusters.

Fig. 3.

Stereotypy of intracellular distribution of AZs

To study the spatial distribution of Brp clusters within single neurons, we focused on DANs. MB504B-GAL4 labels four types of single PPL1 DANs projecting to distinct compartments ⁴³. Using this driver, we analyzed Brp::rGFP distribution in each compartment and found it uneven in terminals of both PPL1-α2α’2 and α3 (Fig. 4A-B). To better characterize cluster distribution, we calculated the “AZ density” defined as the number of Brp::rGFP clusters surrounding a particular cluster within a specified radius (see Fig. 1A and Materials and Methods for detail). By color-coding Brp::rGFP clusters according to their AZ density, we found that AZs in PPL1-α3 were more localized to the core of the α3 compartment. This pattern was consistent across individuals (Fig. 4C-D), suggesting a stereotyped AZ distribution within PPL1-α3 neurons.

Fig. 4.

We extended the analysis to two other single pairs of neurons innervating the MBs, the DPM and APL neurons. Visualizing AZ density revealed a stereotyped and compartment specific distribution in the DPM neuron (Fig. 5A). Specifically, the AZ density was constantly high in the α’/β’ lobes across all individuals, supporting the known functional importance of DPM branches in the α’/β’ lobes ³⁰. In contrast, this compartmental pattern was less pronounced in the APL neuron. Notably, the ratio of AZ density between α3 and α’3 is strikingly different among individuals (Fig. 5B). Taken together, these data suggest that the individual variability in AZ distribution depends on the cell type.

Fig. 5.

Regulation of spatial configuration of AZs

Characteristics of Brp clusters and AZ distribution in KCs revealed spatial organization at the level of compartments. Recent studies further showed intra-compartmental variability of synaptic plasticity upon associative learning, prompting us to analyze sub-compartmental organization of AZs ^44,45. We thus examined potential local patterns among AZs in proximity by calculating the correlation between nearest neighbor clusters in terms of their Brp::rGFP intensities. This analysis revealed strong correlations in KCs and APL but not in DPM (Fig. 6A).

Fig. 6.

We further analyzed the sub-compartmental AZ patterns by examining the spatial distribution of Brp::rGFP intensity. To reduce the high frequency variability of Brp::rGFP intensities (Fig. 2B), we calculated the “local intensity” for each cluster by applying mean filtering (the average Brp::rGFP intensity of all surrounding clusters within a specified radius r, see Materials and Methods for detail). By visualizing the local intensity in γ KCs, we identified AZ “hot spots”, sub-compartment-sized small groups of AZs with high local intensities, on top of the compartmental differences (Fig. 6B). Since Brp molecules are clustered more in AZ-dense boutons of a single motor neuron ⁵, we hypothesized that these hot spots correspond to regions with higher AZ densities. Indeed, we found that hot spots appear in high AZ densities regions in γ KCs (Fig. 6B). Consistently, correlation between the local intensity and the AZ density was high in all KC subtypes. In contrast, for DPM and APL neurons, the local intensity of Brp::rGFP was not associated with AZ density (Fig. 6C-D). Taken together, these results suggest previously unidentified sub-compartmental synaptic configuration organized across individual KCs.

Associative learning re-organizes sub-compartmental synaptic configuration

Sub-compartmental synaptic configuration may undergo experience-dependent changes, such as through associative learning ^17,18. To examine such structural plasticity, we visualized Brp::rGFP specifically in KCs using R13F02-GAL4 (Fig. 7A) and presented flies with odor (4-methylcyclohexanol) with a concomitant (paired) or shifted (unpaired) electric shock (Fig. 7B). The difference between the paired and unpaired groups therefore represents the effect of associative learning. We prepared the brains at different time points (ranging from 3 min to 1 day) after conditioning and measured the correlation between the AZ density and local intensity of individual Brp::rGFP clusters. Strikingly, we found that associative learning induced a transient modification to the KC synaptic configurations in a compartment-specific manner (Fig. 7C). This learning-induced structural plasticity was specific at 90 min after conditioning and disappeared within one day (Fig. 7D), consistent with previous studies reporting short-lived AZ remodeling in the MBs by conditioning ^17,18. These results showcase the advantage of split-GFP system in reporting the dynamics of local synapse subsets with high throughputs and contents. Altogether, we suggest that associative learning induces transient memory traces of local structural plasticity in AZ configuration.

Fig. 7.

Discussion

By leveraging the CRISPR/Cas9 genome editing and split-GFP technique, we systematically profiled presynaptic structures of MB neurons in a cell-type-specific manner. This approach enabled high-throughput analysis of multiple brain samples and detailed characterizations, revealing previously inaccessible structural features of AZs in CNS neurons.

Single-cell level analysis of AZ profiles revealed cell-type-specific stereotypy in spatial regulation (Fig. 4-6). We found that AZ density in the DPM neurons is consistently high in the α’/β’ lobes across individuals (Fig. 5). This AZ stereotypy of DPM explains the functional significance of the branches in the α’/β’ lobes. Overexpression of a Down syndrome cell adhesion molecule (DScam) variant in the DPM neurons was shown to disrupt their innervation except in the α’/β’ lobe, while memory formation remained unaffected ³⁰. Similarly, our analysis identified the intracellular stereotypy in AZ distribution of the PPL1-α3 DAN within the α3 compartment (Fig. 4). This sub-compartmental organization of the single DAN and layered projections of α/β KCs may underlie the differential dopaminergic modulation of KC subtypes in associative learning ⁴⁶. While such stereotypy is less pronounced in APL, we found strong correlations between Brp::rGFP intensities of neighboring AZs in contrast to DPM (Fig. 6D). This suggests that AZs in APL neurons are organized more locally, allowing them to modulate microcircuits composed of KCs, DANs and MBONs ⁴⁷. Collectively, the spatial scale of AZ organizations depends on cell types, reflecting the function of individual MB neurons.

We observed that Brp::rGFP intensity correlates strongly with the AZ size in KC subtypes, in contrast to DPM and APL neurons (Fig. 3). Previous studies using localization microscopy showed that the molecular density of synaptic proteins at AZs can be dynamically adjusted in motor neurons, potentially modulating the synaptic vesicle release ^41,42,48. The variable Brp concentrations (i.e. the ratio of the Brp::rGFP intensity to the AZ size) in DPM and APL neurons may represent AZs with distinct Brp densities and therefore different functional modes (Fig. 3) ⁴¹. Notably, a recent study successfully predicted the neurotransmitter type of neurons using the ultrastructure of dense projections as one of the parameters ⁴⁹, suggesting neurotransmitter-specific AZ structures. Interestingly, both the DPM and APL neurons are reported to release multiple neurotransmitters ^{27,31,50–53}. Distinct Brp concentrations at AZs might thus support particular release modes or neurotransmitter types.

In KCs, the spatial analysis of local AZ configuration revealed strong correlations between Brp::rGFP intensities at neighboring AZs as well as AZ “hot spots”, regions with high Brp localization and AZ densities (Fig. 6). These results suggest previously unidentified, highly localized sub-compartmental AZ structures, which are likely organized across individual KCs. Presynaptic terminals of a single KC are found to be highly heterogeneous in terms of their plasticity upon dopamine modulation and associative learning ^44,45. A recent study found the sub-compartmentally heterogeneous activities of dopamine terminals explain memory specificity ²⁴. Such locally distinct dopamine input may be a source of the AZ hot spot formation in KC terminals. This model is consistent with the localization of two opposing dopamine receptors Dop1R1 and Dop2R to AZs ⁵⁴. Furthermore, we showed that the local configuration undergoes structural plasticity upon associative learning that involves presynaptic dopaminergic modulation ^55–60 (Fig. 7). Since cAMP signaling plays a crucial role in the AZ structural plasticity especially during associative memory formation ^39,48,61, it may underlie learning-induced plasticity in sub-compartmental AZ configuration ^17,18.

Materials and Methods

Animals

Flies were maintained on standard cornmeal food at 25 °C under a 12:12 h light-dark cycle for all experiments. All flies used for experiments are 3-7 days old adult males. Flies were transferred to fresh food vials after hatching and flipped every two days before experiments. The GAL4-UAS system was used to express transgenes, and balancers were removed for animals used in all experiments. Fly strains and resources used in this study are as follow: MB008B-GAL4 (BDSC 68291) ²³, MB009B-GAL4 (BDSC 68292) ²³; MB370B-GAL4 (BDSC 68319) ²³, MB504B-GAL4 (BDSC 68329) ²³ , VT64246-GAL4 (VDRC 204311) ⁶², GH146-GAL4 (BDSC 30026) ⁶³, R58E02-GAL4 (BDSC 41347) ⁶⁴, R13F02-GAL4 (BDSC 48571) ⁶⁴ , UAS-CD4::tdTomato (BDSC 35841, by Dr. Yhu Nung Jan & Chun Han), ⁵⁹brp::GFP₁₁ ³⁹, UAS-GFP_1-10 ³⁵.

Single-odor aversive conditioning

Adult flies at 3-7 days after eclosion were kept in a 25 °C incubator before and after conditioning. Approximately 50 flies were put in a training tube that is set inside the climate box for 2 minutes before the conditioning. Temperature in the climate box was set at 24 °C and the relative humidity was set to 70%. For the paired group, 2% 4-methycyclohexanol (4-MCH, Sigma-Aldrich, diluted in paraffin oil, Sigma-Aldrich) was presented in a cup with the diameter of 5 mm, together with 12 pulses of electric shocks at 90 V for 1 minute inside the training tube. The unpaired group received the same 1-minute electric shock first, followed by 1-minute rest, and then the same 1-minute 4-MCH presentation. Flies were dissected at 3 min, 20 min, 90 min, 270 min and 1 day after conditioning. Experiments were performed multiple times on different days for a single group.

Sample preparation

Animals used in experiments involving comparisons between control and experimental groups were dissected on the same day. Data were collected from multiple batches of experiments performed on different days. Flies were anesthetized by ice and placed on ice before dissection. The dissection was performed in cold PBS to collect the brain without the ventral nerve cord. The brains were subsequently fixed by 2% paraformaldehyde for 1 h and then washed 3×20 min by 0.1% PBT (0.1% Triton X-100 in PBS) in a PCR tube, typically 5-6 brains in one tube. Samples were mounted on microscope slides using mounting medium SeeDB2S ⁶⁵.

Confocal microscopy image acquisition

The image acquisition was performed using an Olympus FV1200 confocal microscope platform equipped with both PMT and GaAsP high sensitivity detectors and a 60×/1.42 NA oil immersion objective (PLAPON60XO, Olympus). The Kalman filter was turned on to 2× averaging while scanning. For Brp::rGFP samples in all GAL4 types, scanning speed was set at 2 µs/pixel with a voxel size of lateral 0.079 × 0.079 × axial 0.370 µm. For experimental point spread function (PSF) imaging, SeeDB2S immersed beads were scanned in a setting that is 60×/1.42 NA oil immersion objective; 473 nm laser power: 2.0%; 559 nm laser power: 1.5%; voxel size: 0.079 µm × 0.079 µm × 0.370 µm; scanning speed: 4 µs/pixel, to produce multiple bead images.

PSF acquisition and image deconvolution

Sub-resolution fluorescent beads (Tetraspeck Microspheres 0.1 µm, Thermo Fisher Scientific, T7279) were imaged to generate the PSF. The bead solution was diluted with distilled water and sonicated several times to eliminate aggregation. PSF images acquired were processed using Amira software (Thermo Fisher Scientific) with the Extract Point Spread Function module. The PSFs extracted from each image were averaged to create a single PSF for later image deconvolution (Resizing voxel size: 0.079 µm × 0.079 µm × 0.370 µm; 32 pixels × 32 pixels × 21 slices of image size). Image deconvolution was performed by using the Richardson-Lucy iterative non-blind algorithm in a Fiji plugin DeconvolutionLab2 ⁶⁶, or by CLIJ2 GPU-based Richardson-Lucy deconvolution ⁶⁷. Images were deconvolved 20 times iteratively to improve image quality.

Brp clusters segmentation and data analysis

For Brp::rGFP cluster detection, the image stack size was adjusted to double the XY plate pixel number using Fiji to increase the detection precision. The images were subsequently processed using the 3D maxima finder and the 3D spot segmentation function of the 3D suite package in Fiji. 3D maxima were identified for each cluster and used as the starting point for pixel clustering ³⁶. An intensity threshold (threshold) and the noise tolerance were set to reduce noise and adjust the sensitivity of 3D maxima detection respectively. The threshold set the value of all pixels that have values below the threshold to NaN (not a number). The noise tolerance is set in the 3D maxima finder. To be noticed, the threshold is only applied to the image used to find 3D maxima. The 3D maxima finder generates the 3D maxima. The 3D segmentation function then requires the 3D maxima and the raw image, so that there is no intensity cut affecting the measurement of cluster volume (used deconvolved image). The local thresholding method for the 3D segmentation was set to Gaussian fit, with the “SD value” set to 2. Cluster volume was measured as the voxel count within the 3D ROI.

The above procedures may include a tiling and stitching step for images and data because of the limited computer power. Typically, an image was tiled into 25 sub-stacks with overlapped margins. After being processed separately, the data acquired from all tiled sub-stacks was stitched back with a deletion of overlapping data points to generate the complete detection result. Nearest neighbors were identified for each particle and spatial distances in between (NND) were calculated using Python. For each sample, mean NND was calculated separately. For a particular Brp::rGFP cluster, the number of surrounding clusters within r (r = 3 × mean NND) spherically including itself is referred as the “AZ density”. The “local intensity” of a particular Brp::rGFP cluster is the average of Brp::rGFP intensity of all clusters surrounding within r (r = 3 × mean NND) including itself. The plot function of Python was used to plot the color-coded 3D reconstructed maps.

F-score calculation

The 3D detection parameters were optimized by matching the 3D maxima detection results to the ground truths. Ground truths were defined by annotating AZs manually, according to the size, position and signal pattern of Brp puncta in the images. The F-score was used to assess detection precision. F-score is the harmonic mean of a system’s precision and recall value, calculated by the following formula: 2 × [(Precision × Recall) / (Precision + Recall)]. The precision is the number of true positives divided by the total number of identified positives, including incorrect identification. The recall is the number of true positives divided by the total number of true data points defined as the ground truth. F-scores shown in Fig. 1 were automatically calculated using a customized Fiji Marco that aligned 3D maxima with ground truth. Different combinations of noise threshold and noise tolerance were tested for the test images cropped from raw images.

Statistical analysis

Statistical analysis was performed using GraphPad Prism version 8, 9 and 10 or Python. For statistical comparisons, Kruskal-Wallis test with original False Discovery Rate method of Benjamini and Hochberg correction was applied expect as explicitly mentioned. A desired false discovery rate of 0.05. Pearson’s correlation coefficient (R) was calculated using Python.

Data and materials availability

Further information and requests on resources and reagents should be directed to and will be fulfilled by the corresponding author, Dr. Hiromu Tanimoto (hiromut@m.tohoku.ac.jp). This study did not generate any unique code. All data needed to evaluate the conclusions in the paper are present in the paper.

Acknowledgements

We thank all lab members of Tanimoto Lab at Tohoku University for valuable discussions. We thank Dr. Yuki Tsukada at Keio University for crucial advice on Fiji-based image analysis.

We thank Vienna Drosophila Resource Center and Bloomington Drosophila Stock Center for transgenic flies.

Additional information

Author contributions

Conceptualization: HW, HT

Methodology: HW, YM, NY, SE, SK

Investigation: HW, YM, SE

Visualization: HW, SE

Resources: SK Supervision: HW, NY, HT

Writing—original draft: HY, HT

Writing—review & editing: HW, SE, HT

Funding

Ministry of Education, Culture, Sports, Science and Technology (22H05481)

Ministry of Education, Culture, Sports, Science and Technology (22KK0106)

Ministry of Education, Culture, Sports, Science and Technology (20H00519)

Japan Science and Technology Agency (JPMJSP2114)

Tohoku University (Research Program 'Frontier Research in Duo)