HEK293T cells were seeded at a density of 7.0 × 10⁶ cells per disk 24 h prior to transfection. The transfection involved employing a combination of an adenoviral helper plasmid, a Rep/Cap plasmid, and a plasmid mixture comprising DIO-Flpo and FDIO-EYFP or DIO-Flpo-Flpo and FDIO-EYFP-EYFP in specific ratios, delivered using PEI Pro (Cwbio, CW9309M). After a 72-h incubation period, the cells were harvested and collected by centrifugation. Viral particles were liberated through the application of a high-salt lysis buffer, followed by five cycles of freezing and thawing. To eliminate genomic DNA and residual plasmids, Benzonase (Sigma, 9025-65-4) was added prior to the precipitation of proteins, including AAVs, with polyethylene glycol (PEG, Aladdin, P103734) for 1 h. The precipitated PEG pellet was resuspended overnight, and AAVs were subsequently purified using an iodixanol gradient to eliminate contaminating proteins and empty capsids. The virus-laden iodixanol fraction was concentrated and underwent buffer exchanged through Amicon Ultra-15 filtration tubes (Merck, UFC910024). The concentrated AAV solution was sterile, filtrated, aliquoted, and stored at −80 °C for later use. The viral titer is determined using real-time quantitative PCR (qPCR) and adjusted to 5.0 × 10¹² vg/mL.

The nomenclature for dual AAV vectors adheres to the format: AAV2.X (where X represents the serotype) + Gene copy (single/double indicating the gene copy number) + ratio (the mixing ratio of Cre-dependent Flpo plasmid to Flpo-dependent enhanced yellow fluorescent protein). For instance, AAV2.2-single-1/100 indicates that this dual AAV vector was produced by co-packaging DIO-Flpo and FDIO-EYFP at a ratio of 1/100 with the AAV2.2 serotype.

Experiments were conducted using 6–8 weeks old PV-Cre mice (weighing ∼20 g). All animal procedures were performed in accordance with Animal Ethics Committee of Wuhan University (MRI2024-LAC010). Mice were anesthetized (1.25% avertin, 20 μL/g i.p.), and then intravitreal injection was performed using a microsyringe (Hamilton, #65). The injection volume was 1.5 μL per eye, and the injection site was located 0.5 mm posterior to the limbus, mouse was injected into the right eye only for fMOST. After the injection, mice were kept in a warm environment until fully awake.

Five weeks after injection, mice were deeply anesthetized and then underwent cardiac perfusion with ice-cold 0.1 M phosphate-buffered saline (PBS), followed by fixation with 4% paraformaldehyde (PFA). Ocular globes were carefully extracted and further fixed in 4% paraformaldehyde for 45 min before dissecting the intact retina for retinal flat-mount preparation. The optic nerve was transferred into a sucrose solution for gradient dehydration (at 4 °C). Subsequently, frozen sections were prepared at a thickness of 30 μm. Both retinal flat mounts and sections were washed with PBS and blocked overnight in a solution containing 4% BSAT at 4 °C. Immunofluorescence staining was performed using anti-GFP antibody (Abcam, ab13970) and anti-ChAT antibody (Millipore, AB144P), with all steps conducted at 4 °C. Images were acquired using laser scanning confocal microscopes (Zeiss LSM 880 and Leica Stellaris 5 WLL) and a high-resolution rapid fluorescence microscope (Leica Thunder).

The mice were deeply anesthetized and subjected to cardiac perfusion using ice-cold 0.1 M PBS followed by 4% PFA. Post-extraction from the cranial cavity, the brain and eyes were fixed in 4% PFA at 4 °C for 24 h, followed by a 24-h rinse in PBS. The brain and eyes were then prepared for resin embedding.

The samples underwent a graded dehydration process through a sequential series of ethanol solutions at concentrations of 50%, 75%, 95%, and 100%, with each step lasting 2 h. Following dehydration, the specimens were infiltrated with LR-White resin in increasing concentrations (50%, 70%, 85%, and 100%), each for 2 h. They were then immersed overnight in 100% LR-White resin overnight for 36 h at 4 °C, with the resin solution refreshed after 12 h. The final polymerization of the brains was conducted in a vacuum oven set at 38 °C for 24 h. The 100% LR-White resin was prepared by combining 100 g of LR-White with 0.24 g of ABVN as a polymerization initiator.

The LR-White resin-embedded brains were imaged using the fMOST system (Wuhan OE-Bio Co., Ltd.). Each brain was mounted with the fMOST automated data acquisition system, which comprises a 473 nm laser, a 40× water immersion objective (Olympus, N2667700), and a time delay and integration charge-coupled device (TDI-CCD) for signal detection. Semi-thin coronal sections of 2 μm thickness were sequentially imaged from anterior to posterior, with the data acquisition phase extending over 4–5 days per sample. This meticulous process generated approximately 6,500 coronal sections, facilitating the creation of a comprehensive brain dataset. The raw images, with a resolution of 0.35 μm × 0.35 μm × 2 μm, were processed using specialized software for stripe stitching and brightness normalization. The processed images were then used for 3D reconstruction using Imaris software. The resulting brain dataset was downsampled and aligned to the Allen Reference Brain Atlas, using the Common Coordinate Framework version 3 (CCFv3) for standardized spatial orientation.

The data were processed using ImageJ and Aivia software. The images and fluorescence intensities were processed or analyzed by ImageJ. To ensure more reasonable normalization of fluorescence signal intensity, we first standardized the imaging parameters for capturing fluorescent images to avoid signal intensity deviations caused by differences in exposure time, gain, etc. All images were obtained under the same microscope settings. Then we standardized the method for fluorescence signal quantification:

Finally, apply the normalization formula: subtract the average gray value of the background (background mean gray value) from the average gray value of the target region to complete the fluorescence intensity correction.

The number of fluorescent cells in retinal flat mounts refers to the total count of all EYFP-positive cells in each whole retina, for each group, we analyzed the retinas from five different mice. For each retina, we selected multiple EYFP-positive RGC somata as regions of interest and calculated the mean fluorescence intensity per retina; these per-retina values were then used for statistical analysis. Statistical analysis was performed using GraphPad Prism 9.0. One-way analysis of variance (ANOVA) was used for comparisons between groups, followed by Tukey’s post-hoc test for multiple comparisons. Data are presented as mean ± standard deviation (SD), and differences with p < 0.05 were considered statistically significant (ns = not significant, *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001).

We employed PV-Cre mice and a dual AAV system for specifically and sparsely targeting of PV⁺ RGCs. The dual AAV system is consisted of two vectors: a Cre-dependent Flpo plasmid and Flpo-dependent enhanced yellow fluorescent protein (EYFP) plasmid. We obtained four dual AAV vectors products (AAV2.2-single-1/100, AAV2.2-single-1/300, AAV2.2-single-1/600, AAV2.2-single-1/1000) by mixing the Cre-dependent Flpo plasmid and Flpo-dependent EYFP plasmid at four different ratios (1/100, 1/300, 1/600 and 1/1000) during the AAV production. We intravitreally injected these four products into the retina of Cre mice and collected the retina and optic nerve 5 weeks postinjection (Figure 1A). In the WT control mice, EYFP expression in neither retinal cell bodies nor the optic nerve was observed. Meanwhile, the results showed that as the mixing ratio increases from 1/100 to 1/1000, the number of labeled PV⁺ cells progressively decreased (Figure 1B). The number of fluorescent cells in retinal flat mounts is the total count of all EYFP-positive cells in each whole retina. Specifically, the product of 1/100 ratio labeled approximately 542.0 ± 39.2 cells and the product of 1/300 labeled approximately 144.0 ± 47.6 cells, but it is still insufficient to observe the complete morphology of individual cells. The product of 1/600 ratios labeled 35.4 ± 5.8 cells, while the product of 1/1000 ratio resulted in up to 4 labeled cells. Surprisingly, in one retina of 1:1000 ratio group, only a single PV⁺ RGC was labeled (Figure 1C). Accordingly, the number of EYFP-positive optic nerves showed a similar trend to that of EYFP-positive somata, ∼7–8 axons were labeled in 1/100 group, while only 1 axon was labeled in 1/1000 group (Figure 1D). There was no significant difference in the fluorescence intensity of individual retina soma in PV-Cre mice across the different ratio groups (Figure 1E). Thus, we established a novel strategy for sparse and high-brightness labeling of specific types of retinal cells by combining transgenic mice with a dual-plasmid AAV viral system.

Next, we systematically optimized the copy number parameters of the target gene in the core plasmid rAAV-EF1a-DIO-Flpo-Flpo-WPRE-hGH-pA and rAAV-nEF1a-FDIO-EYFP-EYFP-WPRE-hGH-pA, within the payload capacity of AAV2.2. By increasing the copy number of Flpo and EYFP within these plasmids, we successfully elevated the expression levels of the EYFP protein, as reflected in the enhanced fluorescence intensity seen in labeled cells (Figure 2A). Interestingly, at a plasmid ratio of 1/600, there was no significant difference observed in the number of EYFP-positive cells or the labeled axons when comparing double-copy with single-copy groups (Figures 2B, D). Nonetheless, the average fluorescence intensity per cell exhibited a significant difference between these groups. Specifically, the mean signal intensity in the single-copy group was 70.2 ± 10.9, whereas the double-copy groups’ intensity was significantly higher 94.1 ± 11.4 (p < 0.01) (Figure 2C). This optimization paves the way for more detailed studies on the morphology of RGCs and their projections to downstream brain regions, enhancing our understanding of neural pathways.

To elucidate the precise projection patterns of RGC to visual centers at single-cell resolution, we employed fMOST to acquire high-resolution three-dimensional data from the whole brains of transgenic mice subjected to anterograde labeling. Through image registration, neuron tracing, and three-dimensional reconstruction, we successfully reconstructed the complete and morphologically intact axon of a single RGC, along with its distinct projection pathways and terminal branches in all target areas, including the superior colliculus and lateral geniculate nucleus. We performed whole-brain imaging and data reconstruction of sparsely labeled PV⁺ RGC 5 weeks after virus injection to the right eye of PV-Cre mice (Figure 3A). The complete projection patterns of individual neurons were reconstructed (Figure 3B). Computational neuroanatomical analysis revealed that the neuron demonstrated origin locations and distinct projection patterns.

We found that PV⁺ RGCs primarily project to regions such as the superficial layers of the dorsal lateral geniculate nucleus (dLGN), the ventral lateral geniculate nucleus (vLGN) (Figure 3C) and the superior colliculus (SC) (Figure 3D). Supplementary Video 1 demonstrates the complete projection continuously traced from single-cell labeling in the retina to brain regions. We observed the overall trajectory of individual RGC axon upon exiting the optic chiasm. As the axon approach the midbrain and forebrain regions, it exhibits distinct target-specific branching behaviors, directing their projections to SC and LGN, respectively. Notably, we did not observe any RGC axon branches projecting to non-visual nuclei, consistent with their projection specificity. Within the LGN, axon terminals form highly intricate, densely branched arborizations (Figure 3C). Compared to the terminals in the SC, those in the LGN generally display greater complexity, with more branching points, forming a smaller but more compact structural volume (Figure 3D). In summary, our fMOST imaging data reveal a distinct pattern: individual RGC transmit information to both the SC and LGN in a parallel manner via target-specific axonal branching. However, the presynaptic structures formed in these primary visual centers differ significantly in spatial scale, morphological complexity, and subnuclear localization, likely reflecting their distinct functional requirements in processing motion perception (SC) versus pattern and detail vision (LGN-cortical pathway).

AAV2.2 vectors have been sufficient to meet the needs of studies involving sparse labeling and central projection tracing (Chen et al., 2010). To better delineate neuronal morphology, we selected AAV2.NN as a candidate, which exhibits stronger penetration of the inner limiting membrane in the retina and a larger payload capacity compared to AAV2.2 (Pavlou et al., 2021). We found that the RGC projections labeled by AAV2.NN were more distinct and denser under the same condition compared to that of AAV2.2 (Figure 4A). Whether the number or the fluorescence intensity of labeled cells, AAV2.NN is greater than AAV2.2 (Figures 4B, C). The number of axons also exhibits a corresponding trend (Figure 4D). Therefore, we performed morphological classification of PV⁺ RGCs labeled with AAV2.NN.

The classification of RGC subtypes is instrumental in understanding parallel processing of visual information at the retinal level. In optic nerve degenerative diseases like glaucoma, different RGC subtypes demonstrate varying susceptibility to damage (Denniss et al., 2025; Liu et al., 2023). Recognizing this selective vulnerability is essential for developing targeted neuroprotective therapies. Furthermore, ensuring that regenerating axons in optic nerve regeneration correctly reach their target brain regions relies heavily on their subtype identity.

Recent authoritative studies have classified mouse RGCs into more than 40–50 transcriptomic subtypes (Goetz et al., 2022; Sanes and Masland, 2015), highlighting the complexity and diversity of these cells. In the retina, there is a specific stratified correspondence between the dendrites of ON and OFF RGCs and the dendrites of starburst amacrine cells (SACs) within the inner plexiform layer (IPL) (Sun et al., 2015; Chen et al., 2021; Guadagni et al., 2016). The dendrites of ON RGCs are primarily distributed in the inner sublayer (sublamina b) of the IPL, located at approximately 40% depth (measured from the ganglion cell layer, GCL), overlapping with the dendritic band of ON-type SACs (Supplementary Video 2; Baden et al., 2016; Sanes and Masland, 2015). Conversely, the dendrites of OFF RGCs are concentrated in the outer sublayer (sublamina a) of the IPL, at a depth of about 77% (measured from the GCL), corresponding to the dendritic band of OFF-type SACs (Supplementary Video 3). The dendrites of ON-OFF cells extend into both the ON and OFF sublayers (Supplementary Video 4; Baden et al., 2016).

As shown in Figure 5, PV⁺ RGCs can be subdivided based on their dendritic positioning within SAC synapses into three types: ON, ON-OFF, and OFF (Yi et al., 2012). Morphologically, ON cells can be further classified into distinct subtypes: PV_on1, characterized by a large soma and a large dendritic field (large parasol-like), probably with fast conduction speed, primarily responsible for motion detection and luminance information (Goetz et al., 2022; Sanes and Masland, 2015); PV_on2, featuring a small soma and a small dendritic field (small parasol-like), which might have slower conduction speeds and is mainly involved in high spatial resolution and fine vision (Goetz et al., 2022); PV_on3, having a large soma with a small dendritic field (sunflower-like); and PV_on4, distinguished by dendrites exhibiting bilateral symmetry (butterfly-like). This detailed classification and understanding of RGC subtypes provide a foundation for exploring their functional roles and developing interventions that address specific vulnerabilities and regeneration strategies.