Identification of genes regulating stimulus-dependent synaptic assembly in Drosophila using an automated synapse quantification system

ABSTRACT

Neural activity-dependent synaptic plasticity is an important physiological phenomenon underlying environmental adaptation, memory and learning. However, its molecular basis, especially in presynaptic neurons, is not well understood. Previous studies have shown that the number of presynaptic active zones in the Drosophila melanogaster photoreceptor R8 is reversibly changed in an activity-dependent manner. During reversible synaptic changes, both synaptic disassembly and assembly processes were observed. Although we have established a paradigm for screening molecules involved in synaptic stability and several genes have been identified, genes involved in stimulus-dependent synaptic assembly are still elusive. Therefore, the aim of this study was to identify genes regulating stimulus-dependent synaptic assembly in Drosophila using an automated synapse quantification system. To this end, we performed RNAi screening against 300 memory-defective, synapse-related or transmembrane molecules in photoreceptor R8 neurons. Candidate genes were narrowed down to 27 genes in the first screen using presynaptic protein aggregation as a sign of synaptic disassembly. In the second screen, we directly quantified the decreasing synapse number using a GFP-tagged presynaptic protein marker. We utilized custom-made image analysis software, which automatically locates synapses and counts their number along individual R8 axons, and identified cirl as a candidate gene responsible for synaptic assembly. Finally, we present a new model of stimulus-dependent synaptic assembly through the interaction of cirl and its possible ligand, ten-a. This study demonstrates the feasibility of using the automated synapse quantification system to explore activity-dependent synaptic plasticity in Drosophila R8 photoreceptors in order to identify molecules involved in stimulus-dependent synaptic assembly.

INTRODUCTION

The ability of the neuronal system to respond and adapt to changing environmental stimulation, or so-called activity-dependent synaptic plasticity, partially depends on the regulation of neurotransmission efficiency. One of the major ways to control neurotransmission efficiency is to regulate the efficiency of neurotransmitter release in presynaptic neurons (Alabi and Tsien, 2012). Synaptic vesicle release is performed in structures on the membrane of presynaptic neurons called the active zone (AZ) (Südhof, 2012). The synaptic vesicle release frequency correlates with AZ component proteins’ properties (Matz et al., 2010; Lazarevic et al., 2011; Davydova et al., 2014). However, the molecular mechanisms underlying the activity-dependent regulation of AZ component proteins remain underexplored. In particular, few studies have analyzed activity-dependent changes in the protein level of the AZ components. One of the reasons is the difficulty in accurately analyzing the AZ in target neurons due to the complexity of neuronal morphology and the high density of synapses. For instance, regarding 3D analysis, it becomes even more difficult because the time and mental burden of analysis is extremely large.

We previously found that the AZ number in adult Drosophila photoreceptor neurons changes in a neural activity-dependent manner (Sugie et al., 2015). The Drosophila compound eye comprises about 800 ommatidia, each of which contains eight types of photoreceptors (R1–R8). The Drosophila visual system comprises four optic ganglia: lamina, medulla, lobula and lobula plate. Photoreceptors R7 and R8 project axons to the medulla M6 and M3 layers, respectively (Clandinin and Zipursky, 2002; Hakeda-Suzuki et al., 2011, 2017; Takemura et al., 2013; Shimozono et al., 2019). In photoreceptor R8, round-shaped signals are observed on the axon using the STaR method by GFP labeling of bruchpilot (brp), an AZ component (Chen et al., 2014; Sugie et al., 2015; Araki et al., 2020). The number of Brp signals in photoreceptors was consistent with the number of synapses counted by electron microscopy (Chen et al., 2014). When flies were kept for 72 h under light (LL), dark (DD) and light–dark (LD) conditions, the number of Brp signals in photoreceptor R8 decreased in the LL condition compared to the DD and LD conditions (Sugie et al., 2015). In addition, the number of reduced AZs recovered when the LL condition was returned to the DD condition. Thus, the AZ seems to be disassembled and assembled under LL and DD conditions, respectively. The unchanged number of synapses in the LD condition may be due to competition between stimulus-dependent disassembly and assembly of synapses. We hypothesized that some signals are involved in disassembly and assembly (Sugie et al., 2015), and five genes were identified as being involved in stimulus-dependent disassembly: npc2a, beat-VI, obst-b, swim and CG34370 (Araki et al., 2020). Meanwhile, because the genes involved in synaptic assembly remained elusive, we set out to identify them using R8 photoreceptor synapses.

We performed a two-step RNAi screen against 300 genes expressed in the adult Drosophila brain, including memory-defective, synapse-related and transmembrane molecules. In the first screening, we utilized the aggregation-like signal from Brp-short-mCherry that correlates with synaptic number as an indicator of synaptic stability and obtained 27 candidate genes. In the second screening, we used the STaR system for direct quantification of the R8 synaptic number. To quantify the R8 synaptic number, we developed machine learning-based image analysis software named Synapse Quantifier, which automatically locates synapses in a 3D confocal image and counts the number of synapses along individual R8 axons. This enabled us to identify calcium-independent receptor for α-latrotoxin (cirl) as a gene involved in synaptic assembly signaling. Our data suggest that a presynaptic cirl–postsynaptic ten-a interaction mediates stimulus-dependent synaptic assembly.

RESULTS

Overall flow of the screening strategy for genes that regulate synaptic assembly

To identify genes that regulate synaptic assembly, we performed a genetic screen of the Drosophila visual system (Fig. 1A). In the LD condition, in which a 12-h LL and 12-h DD cycle was repeated, the number of R8 photoreceptor synapses in the Drosophila visual system was maintained (about the same number as in the DD condition) (Sugie et al., 2015). Since synaptic disassembly in the LL condition and assembly in the DD condition are competitive, the balance of synaptic stability seems to be barely maintained in the LD condition (Fig. 1B) (Sugie et al., 2015; Kawamura et al., 2021). In this sensitized background, we hypothesized that the synaptic number would decrease when the genes that regulate synapse assembly were knocked down by RNAi (Fig. 1B).

Fig. 1.

Screening strategy. (A) Illustration of the Drosophila optic lobe, which consists of the retina (red), lamina, medulla, lobula and lobula plate. Photoreceptor R8 projects axons to the medulla M3 layer and makes synapses along the M1–M3 layers; however, no synapses are formed around the M2 layer (arrow). (B) Strategy for screening genes that regulate synaptic assembly in the LD condition. In the LD condition, synapses are disassembled and assembled repeatedly without changing the number of synapses in wild type. We hypothesized that the number of synapses would decrease when synaptic assembly genes were knocked down in the LD condition. (C) Flow of the screening.

We designed a plan for RNAi screening to identify genes that regulate synaptic assembly by expressing Gal4 specifically in the R8 photoreceptor, and knocking down candidate genes using the Gal4/UAS system (Brand and Perrimon, 1993). Figure 1C shows the screening workflow. We selected 300 genes, including membrane protein genes, memory-defective genes and synapse-related genes, that are expressed higher than ‘moderately high’ in the fly head (see Materials and Methods). To screen the 300 genes, we decided to use a marker that could easily detect a decrease in the synaptic number for the first screening since quantifying the synaptic number for 300 genes was substantially time-consuming. Forced expression of Brp-short-mCherry, a fusion protein of a fragment of Brp (Brp-short) and mCherry, in target neurons allows us to visualize presynapses in a cell-specific manner (Fouquet et al., 2009). Brp-short-mCherry signals in photoreceptor R8 exhibited a reduced number of synaptic signals on axons under LL conditions, and aggregate-like signals were generated in the upper medulla (Fig. 2A–2D) (Sugie et al., 2015; Araki et al., 2020). Furthermore, because the decrease in the synaptic number and the aggregate formation of Brp-short-mCherry are correlated, we concluded that the aggregate is likely due to the accumulation of destabilized Brp along the axon, and that the proportion of R8 axons displaying an aggregation signal can be used as an indicator of synaptic disassembly (Fig. 2E–2G).

Fig. 2.

Result of the first screening for genes that regulate synaptic assembly. (A–C) R8 synapses labeled by Brp-short-mCherry. (A) Representative image and illustration of an axon with and without an aggregation signal. The arrowhead marks an aggregation signal. Scale bars, 2 μm. (B, C) Representative images of the 3 days LD (B) and LL (C) conditions. Scale bars, 20 μm. Arrowheads represent aggregation signals. Dashed lines show the boundary of the M1 layer. (D) Quantification of the proportion of axons with an aggregation signal. 3 days LD: n = 139 axons; 3 days LL: n = 72. (E, F) R8 synapses visualized with Brp-short-mCherry and the STaR method. Merged image of axon with no aggregation (E), Brp-short-mCherry (E’), Brp-GFP (E’’). Merged image of axon with aggregation (F), Brp-short-mCherry (F’), Brp-GFP (F’’). (G) Quantification of the number of Brp-GFP puncta. Axons with no aggregation: n = 18; axons with aggregation: n = 24. Scale bars, 2 μm. (H) Result of the first screening with Brp-short-mCherry. (I–N) Confocal images of R8 synapses visualized by Brp-short-mCherry. Control (I), arr overexpression (OE) (J), sgg OE (K), letm1 RNAi (L), cype RNAi (M), cirl RNAi (N). Arrowheads represent aggregation signals. Scale bars, 10 μm. The proportions of axons with aggregation signals and the number of Brp-GFP puncta were analyzed with the Chi-square test and unpaired t-test, respectively. * P ≤ 0.05, *** P ≤ 0.001.

We selected 27 genes that showed a high aggregate formation rate with their RNAi. We next developed automatic image analysis software to quantify the synaptic number using the STaR method by directly counting the Brp-GFP signal in R8 (Chen et al., 2014). Using the software, we selected several candidate genes that showed a decreased synaptic number by RNAi. Because cell death itself decreases the synaptic number, we analyzed the cell membrane morphology of the candidate RNAi lines and excluded neurodegenerative genes from the candidates. Finally, we found one gene that regulates synaptic assembly (Fig. 1C).

First screening for synaptic assembly signals

The 300 RNAi lines of the selected genes were obtained from the Bloomington Drosophila Stock Center. Each RNAi line was crossed with senseless-FLP and GMR-FsF-Gal4 (Chen et al., 2014), which induce high RNAi expression, specifically in photoreceptor R8. In addition, we combined temperature-sensitive Gal80 (Gal80[ts]) (McGuire et al., 2004) to allow an adequate level of Gal4 to be expressed in a temperature-dependent manner, resulting in approximately 20% of axons forming Brp-short-mCherry aggregates in the wild type at 25 ℃ in the LD conditions (Fig. 2D, 2H, 2I). We first checked whether the aggregation formation rate could reflect synaptic stability when the synaptic assembly or disassembly genes were expressed. Previous studies have shown that arrow (arr) and shaggy (sgg) are synaptic assembly and disassembly genes, respectively (Sugie et al., 2015; Kawamura et al., 2021). The aggregation formation rate decreased in arr overexpression (14.14%) and increased in sgg overexpression (52.43%) compared with the control (20.14%) (Fig. 2H–2K), indicating that the proportion of axons with an aggregation signal matched the synaptic stability level. We performed RNAi screening by quantifying the rate of axons showing Brp-short-mCherry aggregation. Since the rate of aggregation in the wild type under LL conditions is about 45% (Fig. 2D), we set 45% as the baseline for synaptic destabilization. We selected 27 genes that showed a rate of axon aggregation higher than 45% (Fig. 2H, 2L–2N, Supplementary Table S1).

Detecting synapses through machine learning-based image processing

To narrow down the candidate genes even more, we undertook a second screen. We tried to analyze the synaptic number by directly counting the semiendogenous Brp-GFP signal (Chen et al., 2014) and screened the 27 candidate genes that showed a decreased synaptic number.

To perform the screening, we developed a “Synapse Quantifier,” image analysis software for synapse quantification, by modifying the convolutional neural network (CNN)-based system described previously (Han et al., 2020). Although Imaris (Bitplane) is widely used to quantify confocal images of the synapse, no software automatically calculates the number of synapses per axon in photoreceptor R8. The CNN was trained for 12 × 12 pixel images of 800 synapses and 800 backgrounds (Fig. 3A). Images of synapses fused with adjacent synapses, as well as isolated synapses, were included in the training images so that fused synapses could be appropriately recognized. In addition, images of areas between synapses were trained as background images. The CNN calculates the score according to similarity to the trained synapses (Fig. 3B).

Fig. 3.

Synapse detection algorithm for R8 synapses. (A) Examples of 12 × 12 pixel cropped synapse images classified as “synapse” and “background” used to train the CNN. (B) Schematics of the CNN. (C) Synapse detection in a 2D plane. The xy-plane is scanned by a 12 × 12 pixel window with a specified stride and analyzed by the CNN. Adjacent synapse images of distance below the overlap threshold are regarded as images of the same synapse. Among them, the synapse image with the highest score is analyzed. (D) Synapse detection in a 3D space is performed by repeating the same process along the z-axis.

Scanning throughout the region of interest using a 12 × 12 pixel window, each image was classified by the trained CNN. We set the “overlap threshold” as the threshold of the distance to distinguish neighboring frames that recognize the same synapse. When a single synapse was recognized multiple times, only the synapse with the highest score within the distance defined by the overlap threshold was considered (see Materials and Methods) (Fig. 3C and 3D). Thus, the locations of all synapses were identified in a 3D confocal image using this system.

Counting synaptic number along individual R8 axons

We needed to directly quantify the number of synapses per R8 axon to identify genes that affect synaptic number. The Synapse Quantifier includes an algorithm that automatically classifies synapses in a 3D space according to their relative distance and counts the synaptic number in each group.

Although the R8 axons are aligned in parallel, the R8 synapses are not evenly distributed along them. A gap exists between the proximal and distal groups of synapses along a single R8 axon at the M2 layer (Fig. 1A, Fig. 4A, black arrows) (Sugie et al., 2015). If we classify the synapses simply by considering their relative distances, the proximal and distal synapses are classified into two different groups. To solve this problem, we applied a co-efficient σ when calculating the distance between synapses so that the distance along the y-axis was underestimated (Fig. 4B). In addition, the coordinates of the synapses should be manually rotated so that the R8 axons become parallel to the y-axis (Fig. 4C). By adjusting σ and coordinates of the rotation angle appropriately, the proximal and distal synapses of the same R8 axon are regarded as the same group, and the synaptic number along individual R8 axons is counted semiautomatically (Fig. 4D).

Fig. 4.

Grouping and counting the synapses. (A–C) Flow of synapse detection. (A) Synapse coordinates detected by the CNN-based algorithm. Arrows indicate the gap between proximal and distal groups. (B) Grouping of the synapses into four groups according to their mutual distance. The y-axis distance is corrected according to the coefficient σ. (C) Rotation of the synapse coordinates and grouping of the synapses into two groups. (D) The appearance of Synapse Quantifier. (E, F) Representative images of R8 synapses labeled with Brp-GFP. (E’, F’) R8 synapses detected by Synapse Quantifier. (G) Comparison of the number of R8 synapses between manual and automatic counting. R is the correlation coefficient (n = 33 axons).

The plot of synapse coordinates detected using the Synapse Quantifier showed a distribution very similar to the original synapse image, indicating that the synapses were successfully detected (Fig. 4E and 4F). Furthermore, the number of R8 synapses counted using the Synapse Quantifier was compared with that of manual quantification (Fig. 4G). In more than 75% of cases, the error rate was less than 10%. Moreover, the correlation coefficient was about 0.7. These data demonstrate that the Synapse Quantifier can detect and count synapses adequately.

Results of the second screen with the Synapse Quantifier

Using the Synapse Quantifier, we screened the RNAi lines that showed a decreased number of synapses visualized by the Brp-GFP marker (Chen et al., 2014). Of the 27 genes that showed the disassembled synapse phenotype in the first screening, RNAi of 19 genes resulted in a decreased mean number of synapses after 3 days of LD (Fig. 5A, Supplementary Table S2). The reduction was significant for three genes (letm1, cype, cirl) (Fig. 5A–5F), which we selected for further analyses.

Fig. 5.

Result of the second screening with Synapse Quantifier. (A) Result of the second screening. Quantification of the number of R8 Brp-GFP puncta per axon, performed by Synapse Quantifier (stride was set to 3). We selected the three genes (letm1, cype and cirl) whose synapses were significantly reduced by RNAi as candidate genes. (B–E) Representative image of R8 synapses visualized by Brp-GFP. Control (B), letm1 RNAi (C), cype RNAi (D), cirl RNAi (E). Scale bars, 2 μm. (F) Quantification of the number of Brp-GFP puncta by Synapse Quantifier (stride was set to 2). Control (n = 73 axons), letm1 RNAi (n = 44), cype RNAi (n = 56), cirl RNAi (n = 46). (G–K) Morphology of the cell membrane. Control#1 (UAS-myr-RFP (II)) (G), Control#2 (UAS-myr-RFP (III)) (H), letm1 RNAi (I), cype RNAi (J), cirl RNAi (K). Scale bars, 2 μm. (L) Quantification of the proportion of axons with neurodegeneration. Control#1 (n = 185 axons), Control#2 (n = 103), letm1 RNAi (n = 67), cype RNAi (n = 73), cirl RNAi (n = 102). The number of Brp-GFP puncta and the proportion of axons with neurodegeneration were analyzed with the unpaired t-test and Chi-square test, respectively. n.s. P > 0.05, ** P ≤ 0.01, *** P ≤ 0.001, **** P ≤ 0.0001.

We also visualized the morphology of the R8 cell membranes to determine whether the synaptic number decreased due to neurodegeneration (Fig. 5G–5K). We did this by expressing RFP with a membrane transition signal (myr-RFP). We quantified the proportion of axons that showed morphological abnormality (Fig. 5L). letm1 and cype knockdown resulted in morphological abnormalities, which are characteristics of axonal degeneration (Fig. 5I and 5J, white arrows, and 5L). Therefore, letm1 and cype were excluded from further analysis, and cirl remained as the final candidate.

Transsynaptic interaction of Cirl and Ten-a may work in stimulus-dependent synaptic assembly

To examine possible off-target effects of the cirl RNAi line, we analyzed R8 synapses of the cirl knockout mutant. Because the cirl mutant (Scholz et al., 2015) showed a significant decrease in synaptic number to a similar extent to the cirl RNAi (Fig. 6A–6C), we concluded that there were negligible off-target effects. Furthermore, we analyzed the number of R8 synapses of cirl RNAi on days 0 and 3 after eclosion. While control flies showed no significant change in the R8 synaptic number, cirl RNAi showed a significant decrease on day 3 (Fig. 6D), suggesting that Cirl functions to stabilize/assemble synapses in the adult stage.

Fig. 6.

Pre- and postsynaptic knockdown of cirl, ten-a and ten-m. (A, B) Representative images of R8 synapses in the 3 days LD condition. Control (A) and cirl mutant (B). Scale bars, 2 μm. (C) Quantification of the number of Brp-GFP puncta by Synapse Quantifier. Control (n = 61 axons), cirl mutant (n = 61). (D) Quantification of the number of Brp-GFP puncta by Synapse Quantifier. Control 0 days (n = 53 axons), Control 3 days LD (n = 72), cirl RNAi 0 days (n = 72), cirl RNAi 3 days LD (n = 46). (E–H) Representative images of R8 synapses with R8-specific knockdown in the 3 days LD condition. R8-specific knockdown was performed with an R8-specific Gal4 driver (senseless-FLP & GMR-FsF-Gal4). Control (E), cirl RNAi (F), ten-a RNAi (G), ten-m RNAi (H). Scale bars, 2 μm. (I) Quantification of the number of Brp-GFP puncta by Synapse Quantifier. Control (n = 26 axons), cirl RNAi (n = 46), ten-a RNAi (n = 31), ten-m RNAi (n = 47). (J–M) Representative images of R8 synapses with postsynaptic neuron-specific knockdown in the 3 days LD condition. Postsynaptic neuron-specific knockdown was performed with ort-gal4. Control (J), cirl RNAi (K), ten-a RNAi (L), ten-m RNAi (M). Scale bars, 2 μm. (N) Quantification of the number of Brp-GFP puncta by Synapse Quantifier. Control (n = 46 axons), cirl RNAi (n = 48), ten-a RNAi (n = 39), ten-m RNAi (n = 47). The numbers of Brp-GFP puncta were analyzed with the unpaired t-test. n.s. P > 0.05, * P ≤ 0.05, *** P ≤ 0.001, **** P ≤ 0.0001.

cirl is the unique Drosophila homolog of latrophilin. Mammalian latrophilin 1 (LPHN1) was identified by its characteristic role in binding to α-latrotoxin, a component of black widow spider venom that induces a rapid increase in synaptic vesicle exocytosis (Krasnoperov et al., 1996; Davletov et al., 1998). In addition, LPHN1 can interact with Lasso to perform axon guidance and synaptogenesis (Ushkaryov et al., 2019). This suggests that an interaction between Cirl and the Drosophila Lasso homologs, Ten-a and Ten-m, can mediate synapse assembly by Cirl. Single-cell RNAseq data in the adult stage show that ten-a and ten-m are expressed in R8 and its secondary neurons (Dm9, Mi1, Mi4, Mi15, R7, L1, and Tm20) (Davis et al., 2020). Therefore, we analyzed the synaptic number following knockdown of cirl, ten-a and ten-m, specifically in the R8 photoreceptor cells or their secondary neurons. Because photoreceptor R8 is histaminergic, we performed RNAi experiments in cells expressing the histamine receptor Ora transientless (ort) using ort-Gal4 for secondary neuron-specific knockdown. We found that the synaptic number was reduced when cirl was knocked down in the R8 photoreceptor (Fig. 6E–6I) or when ten-a was knocked down specifically in secondary neurons (Fig. 6J–6N), suggesting that an interaction between Cirl and Ten-a, bridging the presynaptic and postsynaptic membranes, respectively, mediates stimulus-dependent synapse assembly.

DISCUSSION

Cirl is a new molecular component in stimulus-dependent synaptic assembly

The AZ structures of photoreceptor R8 in Drosophila are disassembled by prolonged exposure to light and assembled when light exposure is stopped (Sugie et al., 2015). We hypothesized that some molecules are involved in synaptic disassembly and assembly in a stimulus- (or neuronal activity-)dependent manner, and we have previously identified several molecules responsible for disassembly (Sugie et al., 2015; Araki et al., 2020). However, the molecules involved in stimulus-dependent synaptic assembly remained unclear. In this study, we set up a screening strategy to identify genes that function in synaptic assembly. By utilizing the aggregation-like signal from Brp-short-mCherry as a simple indicator of synaptic disassembly, we identified 27 candidate genes from 300 initially selected genes. We successfully selected candidate genes affecting synaptic number from those initial candidate genes, and cirl was finally identified as a gene involved in stimulus-dependent synapse assembly.

Cirl is one of the adhesion GPCRs (GPCRs with binding properties) and is involved in mechanosensory neuron function by interacting with TRP channels in Drosophila (Scholz et al., 2015). Mammalian Lphn1, a homolog of cirl, can interact with Lasso and induce synaptogenesis (Ushkaryov et al., 2019). Therefore, we hypothesized that ten-a and ten-m, which are lasso homologs, are also involved in synaptic assembly. We found a significant reduction in the number of R8 synapses after knocking down cirl or ten-a in R8 or postsynaptic neurons, respectively, in the 3-day LD condition, suggesting their positive role in stimulus-dependent synaptic assembly. In Drosophila, because ten-a and ten-m are required for synaptic partner matching and synapse formation in the olfactory system (Hong et al., 2012; Mosca et al., 2012; Mosca and Luo, 2014), exploring the possible contributions of cirl and ten-a/-m in stimulus-dependent synaptic assembly in the olfactory system is highly desirable. Our data suggest a new model for the interaction of cirl and ten-a in the synaptic assembly of photoreceptor R8. A previous study has shown that microtubule stability acts on synaptic stability (Sugie et al., 2015), suggesting that cirl is responsible for regulating signaling, resulting in microtubule stabilization. Single-cell RNAseq data from the adult stage show that cirl and ten-a are expressed in both photoreceptor R8 and its postsynaptic partners (Davis et al., 2020). Among the postsynaptic partners, Mi1 neurons and photoreceptor R7 show a high expression level of ten-a (Davis et al., 2020). Since stimulus-dependent synaptic assembly occurs mainly around the M1–M2 layer of R8 axons (Sugie et al., 2015; Bai and Suzuki, 2022), where Mi1 neurons have many postsynaptic connections with R8 (Kind et al., 2021), cirl–ten-a interaction may affect R8–Mi1 synaptic connections.

Synapse Quantifier is a new tool for synapse analysis in 3D

Some anatomical studies have reported synapses at the neuromuscular junction (NMJ) in the larval stage of Drosophila (reviewed by Bai and Suzuki, 2020; Chou et al., 2020). The NMJ has a very simple structure, with a sparse and small number of sensory neurons and a low density of synapses. Therefore, antibody staining can effectively visualize synaptic structures and components. However, the CNS in the adult brain is tightly packed with various neurons, making it difficult to analyze the synapses of neurons of interest. Recent advances in synaptic visualization tools have enabled high-resolution analysis of the pre- and postsynaptic components in the CNS (Duhart and Mosca, 2022). However, the complexity of CNS neuronal morphology and the high density of synapses still make it difficult to analyze synapse images visualized by confocal microscopy.

CNN-based machine learning is a powerful approach that is widely used for image recognition and classification (Xu et al., 2019; Diez-Hermano et al., 2020; Han et al., 2020). In this study, we developed a custom-made program for synapse image analysis by combining CNN and an algorithm that locates and classifies synapses according to their relative distance. Initially, similarity to the trained synapse images calculated by the CNN was used to exclude overlapping synapse images and to appropriately detect fused synapses. Subsequently, the grouping algorithm semiautomatically counts the synaptic number along individual R8 axons. Importantly, the algorithm does not require any reference image that visualizes the R8 axon shape. Because synapses are appropriately grouped and counted using a single-channel synapse image, no modification in biological experiments is necessary for expanding the scope of application. Similarly, the combination of CNN and appropriate algorithms will extend the potential of CNN-based machine learning in numerous biological studies.

Theoretically, the images used to train the CNN should be obtained in a microscopic system identical to the system used to analyze the experimental images. For example, in this study, the synapse images used for training were obtained using the Nikon C2 system, and the screening image data were obtained using the same system. However, because the images of individual synapses are relatively simple and the CNN tends to obscure the details of the images, the CNN trained using a Nikon C2 may be used to analyze images obtained using another microscopic system. Indeed, we confirmed that we can also successfully detect and count synapses in confocal images obtained using the Olympus FV system (Supplementary Fig. S1). Although collecting multiple training data is laborious, the expansion of the CNN provided in this study to other microscopic systems will markedly reduce the effort required when the Synapse Quantifier is applied to different studies.

MATERIALS AND METHODS

Drosophila strains and genetics

The stock numbers of RNAi lines used in this study and genotypes are listed in Supplementary Table S1 and S3, respectively. The following stocks were used: senseless-FLP, GMR-FsF-Gal4, Brp-FsF-GFP (Chen et al., 2014), UAS-Brp-short-mCherry (Fouquet et al., 2009), UAS-arr (Culi and Mann, 2003), UAS-sgg (Bourouis, 2002), tub-Gal80[ts] (II) (BDSC 7019), tub-Gal80[ts] (III) (BDSC 7017), UAS-myr-RFP (II) (BDSC 7118), UAS-myr-RFP (III) (BDSC 7119), ort-gal4 (BDSC 56515), ten-a RNAi (BDSC 29439), ten-m RNAi (BDSC 29390) and cirl ^KO (Scholz et al., 2015).

RNAi screening

For the RNAi screen, three categories of genes (transmembrane proteins, memory-defective and synapse-related) were selected from 1,854 genes expressed higher than ‘moderately high’ in the fly brain based on FlyBase RNAseq information. Deduced amino acid sequences of each gene were analyzed with THMMM and SignalP, and 205 genes with transmembrane domains were selected as transmembrane proteins. Seventy memory-defective genes and 130 synapse-related genes were selected from ‘phenotypic class’ and ‘tissue/cell affected’ data on FlyBase, respectively. Because some selected genes were included in more than one category, 378 non-duplicated genes were reselected. RNAi lines for 300 of these genes were available and were ordered from the Bloomington Drosophila Stock Center.

Light exposure

Flies were grown in a 12-h LD cycle environment at room temperature (25 ℃) with 50–60% humidity. For experiments in LD or LL conditions, LED panels (MLP-LSK2478DA4, Musashi Electric) were used and set to 2,000 lux on an illuminance meter.

Dissection and fixation

The experimental procedures for brain dissection and fixation were performed as described previously (Hakeda-Suzuki et al., 2017). Fly brains were dissected in 0.1% PBT (PBS containing Triton X-100) and then fixed with 4% paraformaldehyde at room temperature for 60 min. Samples were washed three times with 0.1% PBT and incubated at room temperature for 2 h. Finally, samples were washed once with PBS and filled with VECTASHIELD (Vector Laboratories).

Imaging and quantification

All confocal images were captured using a Nikon C2+ confocal microscope. Each brain image was taken at a depth of 30 to 50 μm from the surface; Brp-short-mCherry and Brp-GFP images were taken at Z steps of 1 μm and 0.5 μm, respectively. In Fig. 2, the proportion of axons with an aggregation of Brp-short-mCherry was quantified as the percentage of axons with labeled spots with a radius of 2.8 μm or larger in the spot function of Imaris (Bitplane). The number of Brp-GFP puncta in Fig. 2G quantified the number of labeled spots with a radius of 0.35 μm or larger in the spot function. The number of synapses in Fig. 4G and Supplementary Fig. S1C was determined by counting the number of Brp-GFP puncta signals from the 2D view of the Z-stack image. The number of synapses per axon predicted by the Synapse Quantifier was directly obtained from the 3D view mode.

Image quantification using Synapse Quantifier

GUI usage of Synapse Quantifier

The details of GUI of Synapse Quantifier are described in Supplementary Fig. S2 and Supplementary Table S4.

Training the CNN

Training data for the CNN were manually collected by cropping 12 × 12 pixel areas that contained synapse and background images using ImageJ (800 each; Fig. 3A). Images of synapses that were fused with adjacent synapses were included to detect fused synapses as well as isolated synapses. In addition, images between neighboring synapses (inter-synapses) were included as background images.

CNN-based machine learning was performed using a custom-made MATLAB script available online. The structure of the CNN is simplified from that used in our previous study (Han et al., 2020). The CNN consists of 11 layers (Fig. 3B). After layer 1 receives the original 12 × 12 pixel image, there are two sets of convolution, normalization and ReLU layers, connected by a max pooling layer between them. Following the fully connected layer in layer 9, the softmax function in layer 10 calculates the final classification results according to their similarity to synapses and backgrounds in layer 11. The accuracy of the trained CNN was more than 99.9%. The trained CNN is available online. In the following, the value in the node “synapse” in layer 11 is used as “score”, which represents the similarity to the trained synapse images.

Synapse detection using the CNN

The region of interest (ROI) specified in a 3D confocal image is scanned by a 12 × 12 pixel window. The default Stride parameter for scanning is 2 pixels, which was increased to 3 pixels for the second screening to speed up the analysis (Supplementary Fig. S2A #11). Each image is classified by the trained CNN and “score” is calculated. Since the scanning stride is smaller than the size of synapses, a single synapse is detected multiple times. A parameter, “overlap threshold”, is used to define the minimum distance between neighboring synapses (Supplementary Fig. S2A #6). If the distance between adjacent synapse images is smaller than the threshold, they are regarded as images of the same synapse. The image of the highest “score” within the area defined by “overlap threshold” is considered in the followings analyses. The removal of the overlap is performed first along the x-y plane (2D analysis; Fig. 3C) and then along the z-axis (3D analysis; Fig. 3D).

Grouping and counting synapses

The detected synapses are classified according to their relative distance. The threshold distance along the x-y and z axes are separately specified (XY radius and Z radius; Supplementary Fig. S2A #12, #13). However, there is a gap between proximal and distal groups of synapses along a single R8 axon (Fig. 1A). To classify the proximal and distal synapses in the same group, the coordinates of synapses are manually rotated using the rotation slider (Supplementary Fig. S2A #21). Here, the distance between synapses is underestimated along the y-axis according to the weight along the y-axis (“σ” in Fig. 4D; “Y weight” in Supplementary Fig. S2A #14). As a result, the proximal and distal synapses of the same R8 axon are regarded as the same group, and the result can be confirmed in the preview, in which the group number is color-coded (Supplementary Fig. S2A #17, S2B). The coordinates of all synapses and number of synapses in each group are summarized in an exported Excel file (Supplementary Fig. S2D, S2E).

DATA AND CODE AVAILABILITY

The MATLAB source codes (“eval_synapse_GUI.fig”, and “eval_synapse_GUI.m”) are available at GitHub (https://github.com/satouma7/SynapseQuantifier).

CONFLICTS OF INTEREST

The authors declare that we have no competing interests.

ACKNOWLEDGMENTS

We gratefully acknowledge Dr. Tobias Langenhan at Leipzig University for providing the cirl ^KO flies. We thank the Bloomington Drosophila Stock Center for providing fly stocks. We thank Enago (www.enago.jp) for the English language review. This work was supported by JSPS KAKENHI #21J12660 (J. O.), a Grant-in-Aid for Scientific Research (B) and a Grant-in-Aid for Scientific Research on Innovative Areas from MEXT (#16H06457, #21H05682 and #21H02483 (T. S.); #18K14835, #18J00367 and #21K15619 (Y. N.); #19K22592 and #21H02837 (A. S.); #21H02484, #22H05169, #22H05621 and #20H01823 (M. S.)), a Takeda Visionary Research Grant from the Takeda Science Foundation (T. S.), and the Japan Agency for Medical Research and Development (AMED) #JP22ek019484s (A. S.).

REFERENCES

•  Alabi,  A. A., and  Tsien,  R. W. (2012) Synaptic vesicle pools and dynamics. Cold Spring Harb. Perspect. Biol. 4, a013680.
•  Araki,  T.,  Osaka,  J.,  Kato,  Y.,  Shimozono,  M.,  Kawamura,  H.,  Iwanaga,  R.,  Hakeda-Suzuki,  S., and  Suzuki,  T. (2020) Systematic identification of genes regulating synaptic remodeling in the Drosophila visual system. Genes Genet. Syst. 95, 101–110.
•  Bai,  Y., and  Suzuki,  T. (2020) Activity-dependent synaptic plasticity in Drosophila melanogaster. Front. Physiol. 11, 161.
•  Bai,  Y., and  Suzuki,  T. (2022) Activity-dependent circuitry plasticity via the regulation of the histamine receptor level in the Drosophila visual system. Mol. Cell. Neurosci. 119, 103703.
•  Bourouis,  M. (2002) Targeted increase in shaggy activity levels blocks wingless signaling. Genesis 34, 99–102.
•  Brand,  A. H., and  Perrimon,  N. (1993) Targeted gene expression as a means of altering cell fates and generating dominant phenotypes. Development 118, 401–415.
•  Chen,  Y.,  Akin,  O.,  Nern,  A.,  Tsui,  C. Y. K.,  Pecot,  M. Y., and  Zipursky,  S. L. (2014) Cell-type-specific labeling of synapses in vivo through synaptic tagging with recombination. Neuron 81, 280–293.
•  Chou,  V. T.,  Johnson,  S. A., and  Van Vactor,  D. (2020) Synapse development and maturation at the drosophila neuromuscular junction. Neural Dev. 15, 11.
•  Clandinin,  T. R., and  Zipursky,  S. L. (2002) Making connections in the fly visual system. Neuron 35, 827–841.
•  Culi,  J., and  Mann,  R. S. (2003) Boca, an endoplasmic reticulum protein required for wingless signaling and trafficking of LDL receptor family members in Drosophila. Cell 112, 343–354.
•  Davis,  F. P.,  Nern,  A.,  Picard,  S.,  Reiser,  M. B.,  Rubin,  G. M.,  Eddy,  S. R., and  Henry,  G. L. (2020) A genetic, genomic, and computational resource for exploring neural circuit function. eLife 9, e50901.
•  Davletov,  B. A.,  Meunier,  F. A.,  Ashton,  A. C.,  Matsushita,  H.,  Hirst,  W. D.,  Lelianova,  V. G.,  Wilkin,  G. P.,  Dolly,  J. O., and  Ushkaryov,  Y. A. (1998) Vesicle exocytosis stimulated by α-latrotoxin is mediated by latrophilin and requires both external and stored Ca²⁺. EMBO J. 17, 3909–3920.
•  Davydova,  D.,  Marini,  C.,  King,  C.,  Klueva,  J.,  Bischof,  F.,  Romorini,  S.,  Montenegro-Venegas,  C.,  Heine,  M.,  Schneider,  R.,  Schröder,  M. S., et al. (2014) Bassoon specifically controls presynaptic P/Q-type Ca²⁺ channels via RIM-binding protein. Neuron 82, 181–194.
•  Diez-Hermano,  S.,  Ganfornina,  M. D.,  Vegas-Lozano,  E., and  Sanchez,  D. (2020) Machine learning representation of loss of eye regularity in a Drosophila Neurodegenerative model. Front. Neurosci. 14, 516.
•  Duhart,  J. C., and  Mosca,  T. J. (2022) Genetic regulation of central synapse formation and organization in Drosophila melanogaster. Genetics 221, iyac078.
•  Fouquet,  W.,  Owald,  D.,  Wichmann,  C.,  Mertel,  S.,  Depner,  H.,  Dyba,  M.,  Hallermann,  S.,  Kittel,  R. J.,  Eimer,  S., and  Sigrist,  S. J. (2009) Maturation of active zone assembly by Drosophila Bruchpilot. J. Cell Biol. 186, 129–145.
•  Hakeda-Suzuki,  S.,  Berger-Müller,  S.,  Tomasi,  T.,  Usui,  T.,  Horiuchi,  S.-Y.,  Uemura,  T., and  Suzuki,  T. (2011) Golden Goal collaborates with Flamingo in conferring synaptic-layer specificity in the visual system. Nat. Neurosci. 14, 314–323.
•  Hakeda-Suzuki,  S.,  Takechi,  H.,  Kawamura,  H., and  Suzuki,  T. (2017) Two receptor tyrosine phosphatases dictate the depth of axonal stabilizing layer in the visual system. eLife 6, e31812.
•  Han,  X.,  Wang,  M.,  Liu,  C.,  Trush,  O.,  Takayama,  R.,  Akiyama,  T.,  Naito,  T.,  Tomomizu,  T.,  Imamura,  K., and  Sato,  M. (2020) DWnt4 and DWnt10 regulate morphogenesis and arrangement of columnar units via Fz2/PCP signaling in the Drosophila brain. Cell Rep. 33, 108305.
•  Hong,  W.,  Mosca,  T. J., and  Luo,  L. (2012) Teneurins instruct synaptic partner matching in an olfactory map. Nature 484, 201–207.
•  Kawamura,  H.,  Hakeda-Suzuki,  S., and  Suzuki,  T. (2021) Activity-dependent endocytosis of Wingless regulates synaptic plasticity in the Drosophila visual system. Genes Genet. Syst. 95, 235–247.
•  Kind,  E.,  Longden,  K. D.,  Nern,  A.,  Zhao,  A.,  Sancer,  G.,  Flynn,  M. A.,  Laughland,  C. W.,  Gezahegn,  B.,  Ludwig,  H. D. F.,  Thomson,  A. G., et al. (2021) Synaptic targets of photoreceptors specialized to detect color and skylight polarization in Drosophila. eLife 10, e71858.
•  Krasnoperov,  V. G.,  Beavis,  R.,  Chepurny,  O. G.,  Little,  A. R.,  Plotnikov,  A. N., and  Petrenko,  A. G. (1996) The calcium-independent receptor of α-latrotoxin is not a neurexin. Biochem. Biophys. Res. Commun. 227, 868–875.
•  Lazarevic,  V.,  Schöne,  C.,  Heine,  M.,  Gundelfinger,  E. D., and  Fejtova,  A. (2011) Extensive remodeling of the presynaptic cytomatrix upon homeostatic adaptation to network activity silencing. J. Neurosci. 31, 10189–10200.
•  Matz,  J.,  Gilyan,  A.,  Kolar,  A.,  McCarvill,  T., and  Krueger,  S. R. (2010) Rapid structural alterations of the active zone lead to sustained changes in neurotransmitter release. Proc. Natl. Acad. Sci. USA 107, 8836–8841.
•  McGuire,  S. E.,  Mao,  Z., and  Davis,  R. L. (2004) Spatiotemporal gene expression targeting with the TARGET and gene-switch systems in Drosophila. Sci. STKE 2004, pl6.
•  Mosca,  T. J.,  Hong,  W.,  Dani,  V. S.,  Favaloro,  V., and  Luo,  L. (2012) Trans-synaptic Teneurin signalling in neuromuscular synapse organization and target choice. Nature 484, 237–241.
•  Mosca,  T. J., and  Luo,  L. (2014) Synaptic organization of the Drosophila antennal lobe and its regulation by the Teneurins. eLife 3, e03726.
•  Scholz,  N.,  Gehring,  J.,  Guan,  C.,  Ljaschenko,  D.,  Fischer,  R.,  Lakshmanan,  V.,  Kittel,  R. J., and  Langenhan,  T. (2015) The adhesion GPCR latrophilin/CIRL shapes mechanosensation. Cell Rep. 11, 866–874.
•  Shimozono,  M.,  Osaka,  J.,  Kato,  Y.,  Araki,  T.,  Kawamura,  H.,  Takechi,  H.,  Hakeda-Suzuki,  S., and  Suzuki,  T. (2019) Cell surface molecule, Klingon, mediates the refinement of synaptic specificity in the Drosophila visual system. Genes Cells 24, 496–510.
•  Südhof,  T. C. (2012) The presynaptic active zone. Neuron 75, 11–25.
•  Sugie,  A.,  Hakeda-Suzuki,  S.,  Suzuki,  E.,  Silies,  M.,  Shimozono,  M.,  Möhl,  C.,  Suzuki,  T., and  Tavosanis,  G. (2015) Molecular remodeling of the presynaptic active zone of Drosophila photoreceptors via activity-dependent feedback. Neuron 86, 711–725.
•  Takemura,  S.-Y.,  Bharioke,  A.,  Lu,  Z.,  Nern,  A.,  Vitaladevuni,  S.,  Rivlin,  P. K.,  Katz,  W. T.,  Olbris,  D. J.,  Plaza,  S. M.,  Winston,  P., et al. (2013) A visual motion detection circuit suggested by Drosophila connectomics. Nature 500, 175–181.
•  Ushkaryov,  Y. A.,  Lelianova,  V., and  Vysokov,  N. V. (2019) Catching latrophilin with lasso: a universal mechanism for axonal attraction and synapse formation. Front. Neurosci. 13, 257.
•  Xu,  C.,  Theisen,  E.,  Maloney,  R.,  Peng,  J.,  Santiago,  I.,  Yapp,  C.,  Werkhoven,  Z.,  Rumbaut,  E.,  Shum,  B.,  Tarnogorska,  D., et al. (2019) Control of synaptic specificity by establishing a relative preference for synaptic partners. Neuron 103, 865–877.e7.