The AppNL-G-F mouse retina is a site for preclinical Alzheimer’s disease diagnosis and research

In this study, we report the results of a comprehensive phenotyping of the retina of the AppNL-G-F mouse. We demonstrate that soluble Aβ accumulation is present in the retina of these mice early in life and progresses to Aβ plaque formation by midlife. This rising Aβ burden coincides with local microglia reactivity, astrogliosis, and abnormalities in retinal vein morphology. Electrophysiological recordings revealed signs of neuronal dysfunction yet no overt neurodegeneration was observed and visual performance outcomes were unaffected in the AppNL-G-F mouse. Furthermore, we show that hyperspectral imaging can be used to quantify retinal Aβ, underscoring its potential as a biomarker for AD diagnosis and monitoring. These findings suggest that the AppNL-G-F retina mimics the early, preclinical stages of AD, and, together with retinal imaging techniques, offers unique opportunities for drug discovery and fundamental research into preclinical AD. Electronic supplementary material The online version of this article (10.1186/s40478-020-01102-5) contains supplementary material, which is available to authorized users.


Introduction
Alzheimer's disease (AD) is the number one neurodegenerative disorder and cause of dementia. Accumulation of amyloid-beta (Aβ) plaques and hyperphosphorylated Tau in tangles are two hallmarks of AD, and are believed to lead to neurotoxic inflammation, neuronal dysfunction, and eventually neurodegeneration. Despite decades of research, the AD field is still struggling with finding techniques for adequate diagnosis and disease monitoring, rational treatment strategies and valid research models.
AD is supposed to start 20 years before the first cognitive symptoms appear [20,39,51], and it is nowadays believed that one should focus on this time window for diagnosis and future treatment [19,47]. In the search for new biomarkers and animal models that allow evaluation of this preclinical phase of AD, increasing attention has gone to the retina. Indeed, given that the retina is an integral part of the central nervous system, cellular and molecular mechanisms underlying neurodegeneration are conserved, and pathological processes occurring in the retina may be an indicator of those processes ongoing in other parts of the central nervous system [3,11,26]. In addition, state-of-the-art technologies for ocular imaging (e.g. optical coherence tomography and confocal scanning laser ophthalmoscopy) allow these processes to be visualised at a resolution of at least an order of a magnitude higher than conventional brain imaging techniques, without the need for invasive, costly procedures or tracers, and in a reproducible and quantifiable manner [11].
One emerging technique in the field of retinal biomarkers for AD, is retinal hyperspectral imaging. This noninvasive imaging spectroscopy technique relies on the wavelength-dependent effect of Aβ on light scatter, to measure the reflectance spectrum of a locus over a range of wavelengths and convert this into a hyperspectral score of which the magnitude correlates to the amount of Aβ. The exact pathological correlate of this hyperspectral signature is still unclear, although previous in vitro studies and theoretical modeling suggested that soluble Aβ 42 is being measured [29,30]. The latter would be a particular advantage, as converging evidence suggests that, in the preclinical phase of AD, soluble oligomeric Aβ accumulates to a threshold that eventually leads to symptomatic AD. Rather than intermediates to generation of Aβ plaques, these Aβ oligomers are nowadays regarded as being the most pathogenic form of Aβ. Aβ oligomers initiate neuropathological processes such as neuroinflammation, gliosis, degradation of synapses, impaired axonal transport, and oxidative damage, thereby leading to an enlarging pool of neurons with limited functionality and ultimately driving the diseased brain to the tipping point where actual neurodegeneration ensues [2,7,15,52].
A multitude of studies has evaluated the retinal phenotype of AD mouse models, and revealed retinal abnormalities that mirror observations in human AD subjects, including elevated Aβ levels, Aβ deposits co-localizing with sites of apoptosis, tauopathy, neurodegeneration of ganglion and amacrine cells, microgliosis and astrogliosis, abnormal electrophysiology, as well as tight junction attenuation and vessel malformations [4,10,16,24,35]. However, results are often conflicting, and interpretation of their biological relevance is obscured by the fact that these transgenic AD mice all rely on ectopic overexpression of familial AD-associated genes (mutations). Here, we address this shortcoming with a comprehensive retinal phenotyping study of the App NL-G-F knock-in mouse, which is characterized by physiological expression of humanized mouse amyloid precursor protein (APP), containing the Swedish, Beyreuther/Iberian and Arctic mutations, in disease-relevant central nervous system regions and cell types [32,40]. For this first-time retinal phenotyping study of the App NL-G-F mouse, we heavily relied on in vivo techniques to assess the possibility of longitudinal disease monitoring via retinal imaging and electrophysiology, besides the gold standard post mortem assays. Additionally, we evaluated the use of hyperspectral imaging to quantify retinal Aβ burden, based on the hypothesis that it detects soluble Aβ 42 and therefore may offer a way to study preclinical AD. Altogether, this approach represents a novel tool box, with the retina of the App NL-G-F mouse as a model and retinal imaging techniques as unique read-outs, for both fundamental and drug discovery research in the preclinical phase of AD.

Animals
B6.129S5-Apptm3.1Tcs (here referred to as App NL-G-F ) [32,40], App NL-G-F x CX 3 CR-1 GFP/+ (generated via crossing of App NL-G-F and B6.129P2-Cx3cr1tm1Litt (The Jackson Laboratory)), B6;C3Tg (APPswe,PSEN1dE9)85Dbo/ Mmjax (here referred to as APP/PS1) mouse lines and corresponding wild type (WT) controls were used. The App NL-G-F mouse carries an App gene construct, which contains a humanized Aβ region and includes the Swedish "NL", Iberian "F", and Arctic "G" mutations. The APP/ PS1 mouse contains transgenes for both APP and presenilin 1 (PSEN1 or PS1), under the control of the mouse prion promotor, with chimeric mouse/human APP with the Swedish mutation, and human PS1 lacking exon 9. The App NL-G-F model was chosen as it avoids potential artifacts introduced by APP overexpression by using a knock-in approach to express APP at endogenous levels and with appropriate cell-type and temporal specificity. The APP/PS1 overexpression mouse was included in this study as a reference AD model, given that its retinal phenotype is well-described, e.g. [9,12,22,29,33,45,54]. Mice were bred under standard laboratory conditions. App NL-G-F and WT female mice were sacrificed at 3, 6, 9, 12 and 18 months of age (± 1 month). Male mice were also included for the 18-month age group.

Aβ extraction and ELISA detection
For soluble Aβ extraction, two retinas from one mouse were pooled and homogenized in lysis buffer (50 mM NaCl, 0.4% diethanolamine, and protease inhibitor cocktail (Roche)). Samples were spun down and supernatant was centrifuged for 50 min at 100.000×g. Supernatant containing the soluble Aβ fraction was collected. From the pellet, insoluble Aβ was extracted with GuHCl (6 M GuHCl/50 mM Tris-HCl) by sonication, incubation at 25 °C for 60 min and centrifugation for 20 min at 170.000 x g. Aβ levels were determined via ELISA Meso Scale Discovery 96-well plates coated with antibodies for Aβ 40 JRFcAβ40/28 and Aβ 42 JRFcAβ42/26. JRFAβN/25 (human specific N-terminal epitope) labeled with sulfo-TAG was used as the detection antibody. Antibodies were provided by Dr. Marc Mercken (Janssen Pharmaceutica).

Microglia numbers and morphological parameters
were analyzed on maximum intensity projection images of retinal wholemounts, as described previously [5]. 2. Macroglia reactivity was quantified on retinal sections by manually counting GFAP-immunopositive fibers at the inner plexiform-inner nuclear layer border [43]. GFAP + fibers were counted over a length of 300 µm, on 20 images per mouse (two peripheral and two central images per section; five sections per mouse, including the central section containing the optic nerve head, and the sections located 240 and 480 µm anterior/posterior). For S100B, the immunopositive area was measured in the nerve fiber/ganglion cell layer over a length of 300 µm, on 20 images per mouse (cfr. above) [18]. 3. Retinal vasculature was analyzed on retinal wholemounts, on maximum intensity projections from the upper vasculature. Vessel density and branching complexity was analyzed with AngioTool software [56] on a binarized image. Retinal vessel diameter was measured manually at a distance of 300-400 µm from the optic nerve head; 5 measurements were averaged per vessel. 4. RBPMS + retinal ganglion cells were counted on retinal wholemounts, using a validated automated counting method [27]. 5. Melanopsin + retinal ganglion cell numbers were counted manually on wholemount images [50]. 6. ChAT + cells were counted in 3D, using the 'spots' module of Imaris software (version 7.4.2) on z-stack pictures of the wholemounts.

Optomotor response test
Visual performance was tested using the virtual reality optomotor test (Optomotry, Cerebral Mechanics), as previously described [43]. Briefly, the mouse was place on a platform in the testing arena and, using the staircase method, the maximal spatial frequency of moving black-white stripes for which the optomotor reflex was still present, was determined.

Electroretinogram
Following overnight dark adaptation, electroretinograms (Celeris, Diagnosys) were recorded for anesthetized mice. Lens electrodes with integrated stimulators were placed on the cornea after pupil dilation (0.5% tropicamide, Tropicol, Théa and 15% phenylephrine, Théa). Eyes were alternately stimulated, and the electrode on the contralateral eye was used as a reference. Full-field flash electroretinogram was recorded at increasing flash intensities of 0.003, 0.01, 0.1, 1, 2.5 and 7.5 cd*s/m 2 . For each intensity, 3 flashes were averaged. Inter-flash time increased from 5 s for the first two flash intensities, 10 s for the next two intensities to 30 and 45 s for the two last intensities. The scotopic threshold response was measured at 5*10 −6 cd*s/m 2 and 50 flashes were averaged. Analysis was performed with Espion v6.59.9 software (Diagnosys), as previously described [43]. The scotopic threshold response amplitude was defined as the peak of the curve. Using a band pass filter (75-300 Hz), the oscillatory potentials on the rising part of the b-wave were retrieved and the area under the curve was calculated. The latency of the oscillatory potentials was defined for OP2 [43].

Hyperspectral imaging
A hyperspectral visible and near-infrared snapscan camera (150 bands, 470-900 nm, Imec; HSI Snapscan software (version 1.2.0)) was mounted on a Leica DM6 microscope with LED light source (400-700 nm). The snapscan camera acquires the hyperspectral image (2048 × 1536 pixels) by the movement of the linescan hyperspectral imaging sensor with on-chip filter technology on a miniaturized translation platform inside the camera body. Filters for each spectral band cover five adjacent pixel rows. The camera was used in stop-motion mode in steps of three pixel rows to reduce cross-talk between adjacent spectral bands and with HDR enabled (three frames) to reduce signal-to-noise ratio. Acquisition was set to cover only the 70 visible light spectral bands (471-691 nm). For hyperspectral imaging, retinal wholemounts were stained with DAPI, prior to mounting on a glass slide using a drop of saline and a cover slip. The transmission spectra of central (1 mm from the optic nerve head) and peripheral (2 mm from the optic nerve head) regions of each retinal quadrant were acquired using a 40x objective with the ganglion cell layer nuclei used as the focal plane. Spectral analysis was performed using a modification of published methods [13] (Additional file 1: supplementary Fig. 1). In brief, each frame of the recorded hyperspectral image was spatially resized to 20% of its original size to smooth out acquisition noise. Each spectrum was normalized by the average spectral value of the three longest wavelengths (685, 687 and 691 nm). Pixels with spectral values above 1.5 and below zero were considered as outliers and masked out from further analysis. Spectral data from all mice were stacked in a matrix and a corresponding class label (AD or wild type control [CO]) was given for each entry. The between class covariance spectrum was calculated. This spectrum corresponds to the average spectral difference between AD and CO groups. The hyperspectral score for a given spectral pixel was obtained by calculating the dot product between its spectral data and the between class covariance spectrum. The average spectral score for a given mouse was calculated using all pixels of all the HS images collected for that retina.

Amyloid accumulates in the retinas of App NL-G-F mice, with Aβ plaque formation at old age
Retinal amyloid burden was studied in App NL-G-F mice of various ages via ELISA. Soluble Aβ 42 was detected at 3 months, with stable levels until 18 months, yet a steep increase at 24 months (p = 0.0006) (Fig. 1a). Insoluble Aβ 42 , which is typically aggregated in amyloid plaques, was only clearly elevated at 12 months and dramatically increased at 24 months (p < 0.0001) (Fig. 1a). In contrast, and in line with the combined effect of the Swedish mutation that increases total Aβ production and the Iberian mutation that increases the Aβ 42 /Aβ 40 ratio, a steady level of soluble Aβ 40 accumulation was present from 3 to 24 months, albeit two orders of magnitude lower than Aβ 42 (Fig. 1b). Insoluble Aβ 40 (Fig. 1b) as well as soluble/ insoluble Aβ 38 and Aβ 36 levels (data not shown) were below the detection limit of the assay. For comparison, in the APP/PS1 overexpression mouse model, at 18 months of age (when retinal plaques are present [9,12,22,33,54]), we detected the accumulation of soluble Aβ 42 (comparable levels to App NL-G-F ), soluble Aβ 40 (higher levels than App NL-G-F ), and insoluble Aβ 42 (lower levels than App NL-G-F ) (Fig. 1c).
In accordance with these results, Aβ plaques were detected via 6E10 and 82E1 immunohistochemistry on retinal wholemounts of App NL-G-F mice of 18 months of age ( Fig. 1d-i). Morphologically, these resembled classical brain plaques, with a dense core and radiating fibrillary arms (Fig. 1k). Sizes typically ranged from 5 to 20 µm. Aβ plaque density was higher in the peripheral than in the central retina, while no overt differences between the retinal quadrants were seen (Fig. 1j). Plaques mainly localized in the inner plexiform layer (Fig. 1f ), often closely associated with microglia (Figs. 1l, 4d-g), ChAT + strata with cholinergic amacrine cell synapses (Fig. 1m), and PKCα + strata with bipolar cell terminals (Fig. 1n). In addition, some were found around the blood vessels in the nerve fiber layer, in close association with pericytes (Fig. 1o) and endothelial cells (Fig. 1p).
In all these cases, the Aβ deposits were extracellularly. Intracellular 6E10 staining was only seen in microglia (Fig. 1l), probably due to phagocytosis of Aβ plaques; and in RBPMS + and melanopsin + ganglion cells (data not shown), likely reflecting endogenous APP expression. Plaques did not associate with ganglion cells (RBPMS and melanopsin), bipolar (Chx10) and amacrine (ChAT) cell somata, nor astrocytes and radial Müller glia (GFAP) (data not shown). Oligomeric Aβ was visualized with an A11 immunostaining on retinal sections and was most abundant in all strata of the inner plexiform layer-comprising amacrine and bipolar cell presynaptic and ganglion cell postsynaptic terminals (Fig. 1q-u). Oligomeric Aβ was also observed extracellularly in the ganglion cell layer (Fig. 1t) and intracellularly in Chx10 + bipolar cells (Fig. 1u). A progressive accumulation over time-in 3-to 18-to 24-month-old mice -, increasingly affecting the inner plexiform layer and bipolar cells, was observed ( Fig. 1q-s). No Aβ staining, with 6E10 nor A11, was observed in the outer retina. These data show that, while soluble Aβ 42 (and Aβ 40 ) is present in the retinas of App NL-G-F mice early in life, retinal Aβ plaques only appear from 18 months of age onwards. Given this relatively slow progression of amyloidopathy, this retinal mouse model may be used to study the conversion from early to late stages of AD pathogenesis, as well as the effect of soluble Aβ on retinal structure and function. For the characterization of the latter, in addition to longitudinal studies, we will largely focus on ages 3 and 18 months in the rest of this study, as these represent disease stages without and with Aβ plaques in the retina, respectively.

Hyperspectral imaging detects early retinal accumulation of Aβ in App NL-G-F mice
Aiming to provide a proof-of-concept for the use of hyperspectral imaging for the early detection of retinal Aβ accumulation and for monitoring its progression, we next performed ex vivo hyperspectral imaging of retinal wholemounts of 3-and 18-month-old App NL-G-F mice and age-matched WT controls. For both ages, the average spectra of the App NL-G-F mice were clearly distinct from WT. In young mice, this separation was most marked in the 475-500 nm range (Fig. 2a), while in old mice the spectral separation became even more apparent, extending to 600 nm (Fig. 2b). A hyperspectral  score was calculated for each mouse from the ex vivo retinal reflectance data, using a modification of methods developed for in vivo hyperspectral retinal imaging [13]. Hyperspectral scores were significantly higher in App NL-G-F mice compared to age-matched WT mice at both ages (p = 0.0014 at 3 months, p = 0.0002 at 18 months) (Fig. 2d-e). Furthermore, the difference in the mean hyperspectral score of App NL-G-F versus age-matched WT mice was also higher at 18 months than at 3 months, in line with the progressive accumulation of Aβ seen in the ELISA. Hyperspectral scores were calculated for each retinal quadrant in both central and peripheral locations (Fig. 2j). In keeping with the immunohistochemistry findings, which demonstrated a predominately peripheral distribution of Aβ, hyperspectral scores were higher in the peripheral retina than in the central retina ( Fig. 2g-h).
Remarkably, 18-month-old APP/PS1 mice had retinal hyperspectral signatures that closely mimics those of 18-month-old App NL-G-F mice (Fig. 2c, f, i). Given that these mice have similar levels of soluble Aβ 42 , but markedly different levels of soluble Aβ 40 and insoluble Aβ 42 , it is possible that soluble Aβ 42 is most contributory to the hyperspectral signature, as previously suggested by Vince and More [29].

Neuronal dysfunction yet no degeneration in the inner retina of App NL-G-F mice
Many studies have shown neurotoxic effects of soluble Aβ, and that functional deficits may precede neurodegeneration. Therefore, first electroretinography was performed with App NL-G-F and WT mice at different time points. Overall, no abnormalities were observed at early ages (Additional file 1: supplementary Fig. 2). At 18 months, however, the latency of both the oscillatory potentials and the b-wave, a measure for amacrine and bipolar cell function, respectively, showed significant changes (p < 0.0001 for the effect of genotype) in the App NL-G-F mice relative to age-matched WT mice ( Fig. 3d-g). Furthermore, also at 18 months of age, the latency of the a-wave was affected at the lowest light intensity (p < 0.0001), suggesting compromised photoreceptor function (Fig. 3b, c).
Next, we assessed potential neurodegeneration in the App NL-G-F retina via a longitudinal in vivo optical coherence tomography experiment. No differences in the overall thickness of the retina nor the nerve fiber, ganglion cell, inner plexiform or inner nuclear layers were identified at any age (Fig. 3h). Of note, no neurodegeneration was observed in the brains of the App NL-G-F mice either [40]. A small decrease in the thickness of the outer nuclear layer was found in App NL-G-F mice, from 12 months (p = 0.0016 for 12 months, p < 0.0001 for 18 months) of age onwards (Fig. 3i). Although these findings corroborate the electroretinogram results, we are unsure how to interpret their biological relevance given that this thinning is limited and occurs in the outer retina-where plaques nor oligomeric Aβ were seen. Loss or dysfunction of the retinal pigment epithelium due to amyloidopathy in these support cells may underlie this photoreceptor degeneration [6,34], or, alternatively, the AD phenotype of the App NL-G-F mice may lead to accelerated aging and therefore earlier age-related deterioration/ loss of rods [21]. This should be further investigated in follow-up studies.
Finally, based on the localization of plaques in the nerve fiber and inner plexiform layers, and in order to confirm and complement the optical coherence tomography data, we next examined neuronal subpopulations in the inner retinas of 18-month-old animals. We quantified: (1) ganglion cells, using the pan-ganglion cell marker RBPMS (Fig. 3l); (2) melanopsin + ganglion cells, given Full electroretinogram data is given in Supplementary Fig. 2. h, i Analysis of the thickness of the retinal layers, as measured by optical coherence tomography, shows no differences between App NL-G-F and WT mice from 2 to 18 months of age (h), with the exception of a thinning of the outer nuclear layer (i). Two-way ANOVA with Sidak's multiple comparisons test (F 1,106 = 32.14 for the effect of genotype, p = 0.0016 and p < 0.0001 for 12 and 18 months, respectively); n = 8-9. j, k Representative optical coherence tomography images of WT (j) and App NL-G-F (k) retinas. Dotted lines delineate the outer nuclear layer. l Quantification of RBPMS + retinal ganglion cell density reveals ganglion cell loss in 18-month-old App NL-G-F compared to WT mice. Unpaired two-tailed t test (t 14  that this subtype appears to be selectively vulnerable to Aβ pathology [31] (Fig. 3m); and (3) ChAT + cholingeric amacrine cells (Fig. 3n). While a small loss of RBPMS + ganglion cells was observed (6.26 ± 2.28%, p = 0.0158), melanopsin + and ChAT + neuronal populations did not show any sign of degeneration in 18-month-old App NL-G-F mice. In line with all of the above observations, and in keeping with the relatively minor loss of ganglion cells and neural plasticity leading to reversible/inducible changes of retinal ganglion cell electrical activity during a critical period of dysfunction preceding death [37], 18-month-old App NL-G-F mice behaved normally in the optomotor response test for visual acuity (Fig. 3o) and contrast sensitivity (data not shown). Altogether, these results indicate that the majority of inner retinal neurons in the App NL-G-F mouse survive progressive amyloidosis up till 18 months. Nevertheless, although no cell death was observed at this age, amacrine and bipolar cell function may be used as a sensitive readout for AD pathology in the retina of App NL-G-F mice.

Local microglia reactivity, astrogliosis and vascular changes in the retina of aged App NL-G-F mice
Because of the strong implication of microgliosis and astrogliosis in AD pathogenesis, and as neurotoxic inflammation may precede neurodegeneration, we next examined the retinal glial cells. First, microgliosis was investigated in entire Iba1-labeled retinal wholemounts of App NL-G-F mice aged between 3 and 18 months. Overall quantification of microglial density (Fig. 4a), soma size (Fig. 4b) and roundness (data not shown) did not reveal population-wide differences between the two genotypes at any age. Analysis of 18-month-old App NL-G-F x CX 3 CR-1 GFP/+ mice retinas, with GFP-labeled microglia, confirmed these findings (data not shown). However, a detailed investigation of these retinas, co-labeled for 6E10, revealed local microglia reactivity around Aβ plaques, as was previously shown in the brains of these mice [40] and in the retinas of other AD mouse models (3xTgAD, APP/PS1) [10,36,41]. Indeed, microglia clustered around Aβ plaques, and these cells showed less ramified, thicker processes and larger somas (p = 0.0112) compared to microglia in regions without plaques (Fig. 4c-g) or in wild type CX 3 CR-1 GFP/+ mice. Second, Müller glia and astrocyte reactivity were assessed on transverse sections of the retina. Longitudinal assessment of Müller glia reactivity, by counting the number of GFAP-positive radial fibers, revealed higher reactivity in 18-month-old App NL-G-F versus WT mice (p = 0.0432) (Fig. 4h-j). Further substantiating these findings, and in line with the brain phenotype [40], analysis of the S100B immunopositive area, which is specific for astrocytes, showed astrogliosis in the retina of 18-month-old App NL-G-F mice (p = 0.0057) (Fig. 4k-m).
Looking at non-neuronal cell types, besides glial cells, also vasculature may undergo pathological alterations in the context of AD. Changes in retinal blood vessel function and morphology, including branching complexity, vessel density and vessel diameter, are being explored as potential early biomarkers for AD [8,38,53,55]. Given that a subset of plaques was found closely associated with the retinal vasculature, we next assessed vessel morphology on retinal wholemounts stained with isolectin B4 (Fig. 5). Vessel density (Fig. 5a) and branching complexity (Fig. 5b) were unaltered in App NL-G-F versus WT mice at 18 months of age. However, diameters of the retinal venules (Fig. 5c), yet not of the retinal arterioles (Fig. 5d), were smaller in App NL-G-F mice (p = 0.0235). Of note, Shi et al. recently reported pericyte loss accompanying retinal vascular amyloidosis, hence the observed reduction in diameter may be a manifestation of this [44].
(See figure on next page.) Fig. 4 Locally reactivated microglia and macrogliosis in the retinas of aged App NL-G-F mice. a, b Quantification of microglial density (a) and soma area (b) on Iba1-labeled retinal wholemounts of 3-to 18-month-old App NL-G-F and WT mice reveals no differences between genotypes when analyzed for the entire retina. An effect of aging was observed for microglia density (Two-way ANOVA, F 4,33 = 2.976, p = 0.0334); n = 3-6. c Morphometric analysis of plaque-associated microglia versus microglia distant from Aβ plaques in 18-month-old App NL-G-F x CX 3 CR-1 GFP/+ retinas reveals that plaque-associated microglia have a larger soma, which is indicative of their activation. Unpaired two-tailed t-test (t 63 = 2.614, p = 0.0112); n = 2-40 cells from 6 mice. d-g Representative images of immunostaining for Aβ (6E10) on retinal wholemounts of 18-month-old CX3CR-1GFP/ + (d-e) and App NL-G-F x CX 3 CR-1 GFP/+ mice (f-g), illustrating that green fluorescent microglia surrounding Aβ plaques display morphological alterations typical for reactive microglia, with thicker and less ramified processes and a larger soma. h Counting the number of GFAP + radial fibers on retinal sections of App NL-G-F and WT mice shows that macrogliosis increases with age, and significantly differs between genotypes at 18 months. Two-way ANOVA with Sidak's multiple comparisons test (

Discussion
In this study, we investigated the extent to which ADassociated disease processes are present in the retina of the App NL-G-F knock-in mouse, and whether they can be monitored via in vivo retinal imaging and electrophysiology. First, we demonstrated that soluble Aβ (mostly Aβ 42 ) accumulation is present early in life in these mice yet retinal Aβ plaques are only observable at 12 months of age and older. This rising Aβ burden in the retina coincides with local microglia reactivity, astrogliosis, and abnormalities in retinal vein morphology, and leads to compromised electrophysiological function of selected retinal neurons. No overt neurodegeneration was seen, however, and visual performance outcomes were intact. Altogether, these findings suggest that the App NL-G-F retina resemble changes that occur in the brain in preclinical AD, during which Aβ accumulation instigates the earliest detrimental effects on astrocytes, microglia and vasculature, poses a stress on neuronal functionality but does not lead to overt functional impairment and neuronal death yet. Second, the data presented in this study suggest that there is an identifiable pre-plaque stage in AD in which only soluble forms of Aβ are present, followed by a phase in which soluble, oligomeric forms of Aβ are likely in equilibrium with plaques. The App NL-G-F retina and reported retinal imaging techniques offer unique opportunities for further fundamental and preclinical research into this sequence of biological events over the course of preclinical AD. Hence, the App NL-G-F mouse is not only the first AD model to accumulate Aβ without phenotypes related to APP overexpression and to more closely model human pathophysiology, its retina also is the first model organ to study preclinical AD. Combined with the available non-invasive, high-resolution imaging techniques and functional assessments of the retina, this creates a toolbox for researchers to shed new light on Aβ proteostasis and novel disease-modifying therapies focused thereon. The rather mild phenotype is in stark contrast to observations made in the brains of App NL-G-F mice. Indeed, previous studies have shown that (sub)cortical Aβ plaques appear at 2-4 months of age, with astrogliosis and microglial reactivity at 6-9 months, and cognitive impairment as early as 6 months [40]. The later formation of Aβ plaques in the App NL-G-F retina could be due to lower expression levels of Aβ compared to the brain, as shown for other AD mouse models [1], or to different clearance rates. However, importantly, it confers a unique strength to the App NL-G-F retina as a model for AD research, as it creates a time window to investigate the earliest phases of AD pathogenesis and is closer to the slow disease progression seen in humans. A second asset of the App NL-G-F retina is that, by diverging from many other available AD models, its phenotype shows striking similarities to the neuropathological changes seen in retinas of human AD patients. This includes the shape and size of the retinal plaques, which we found to be 5-10 times smaller than their counterparts in the brain [22,23]; the localization of these plaques to the inner retinal layers-where endogenous APP expression has been observed [17,25]-and associated with retinal blood vessels; as well as the microglial reactivity, astrocytosis and changes in retinal vessel morphology. Of note, the manifestations of retinal vascular amyloidosis are of particular interest in light of the recent findings by Shi et al., who reported pericyte loss accompanying retinal vascular amyloidosis in the human AD retina [44]. With one exception [14], this phenotype is thus far unstudied in AD mouse models and our data suggest that the App NL-G-F retina can be used to further explore this retinal manifestation of AD. Furthermore, the App NL-G-F retina also appears to be a research model for in-depth investigation of the synaptotoxic effects of Aβ, given the converging data on (soluble) Aβ accumulation in synapses in the inner plexiform layer and electrophysiological abnormalities of amacrine and bipolar cells. Notably, an explanation for the diverging and more severe phenotype of many other transgenic AD mouse models may be sought in the fact that those models are far removed from the human disease state, with non-endogenous promotors, overexpression of APP byproducts, PS1 mutation and/ or abnormal APP and Aβ expression levels. This study is pioneering in that it is the first-time retinal phenotyping study of the App NL-G-F knock-in mouse, which was created to overcome all of the aforementioned artefacts and currently still is unique in its kind. Finally, we have also showcased here the unique potential of the retina for in vivo monitoring of structural and functional changes in AD. Of the battery of tests used, hyperspectral imaging and electroretinograms proved to be sensitive read-outs to assess AD disease status. Results from our post mortem studies suggest that these may in future studies be complemented with confocal scanning ophthalmoscopy for microgliosis and optical coherence tomography angiography or fundus imaging for detailed retinal vascular assessments. Together, this creates a set of in vivo readouts for preclinical investigation that cannot be matched by any existing in vivo brain research methods. Notably, the fact that identical techniques and outcome measures can be used in mice and patients, confers potential translational value to this research. A better understanding of the time course of the pathological processes underlying preclinical AD, is an important prerequisite for the development of new therapies targeting this phase. Furthermore, retinal imaging may provide an accurate and convenient indication of the efficacy of therapies. Currently, Aβ status is evaluated via CSF assays or PET imaging, which is mainly indicative for the deposition of fibrillary forms of Aβ. As such, these do not always provide information about oligomeric Aβ, which may be the more relevant Aβ species to target [46,49]. Our data substantiate that hyperspectral imaging is a highly sensitive technique that can be used to quantify Aβ burden in the earliest stages of disease. This imaging spectroscopy technique is based on the fact that soluble Aβ aggregates-most likely Aβ 42 -interfere with traversing light, leading to Rayleigh light scattering and a characteristic hyperspectral signature, the magnitude of which appear to be proportional to the amount of Aβ present [29,30]. In the past year, following promising post mortem and in vivo studies in both animal and human retinas [28][29][30], results of the first clinical trial showing that in vivo retinal hyperspectral imaging can discriminate between Aβ PET-positive cases and controls emerged [13]. This, together with our data, suggests that hyperspectral imaging can be used to monitor retinal Aβ accumulation, with the potential to provide quantitative measures of Aβ oligomers. As such, hyperspectral imaging in the App NL-G-F retina may become a gateway to improved preclinical studies of the new generation of Aβ therapeutics. Finally, on the long run, this research may also lay the foundations for the rational use of this retinal imaging biomarker for AD diagnosis, disease monitoring, patient stratification and even population-wide screenings.

Conclusion
To conclude, decades of AD research have taught us that we should not try to unravel AD pathogenesis from a simple, neuron-centric perspective, or by assuming a linear series of events leading from Aβ accumulation to dementia. AD is to be understood as a complex pathological cascade, taking decades to gradually evolve from a clinically silent phase to dementia, that is driven by the proteopathic stress imposed by Aβ and Tau conformations, and feedback and feedforward responses of astrocytes, microglia, and vascular cells [48]. Research based on this new framework requires novel approaches, including mouse models that more closely mimic human pathophysiology and allow the study of the earliest, preclinical phases of AD. In this research, we have shown that the retina of App NL-G-F mice is apromising model for preclinical AD, in which Aβ oligomers accumulate and pathological cellular responses are elicited. Furthermore, we have added further evidence in support of the potential of hyperspectral imaging to quantify retinal Aβ burden and to serve as an AD biomarker. Altogether, the App NL-G-F retina and retinal imaging techniques are unique tools for both fundamental and drug discovery research in the preclinical phase of AD.