“Patient-specific Alzheimer-like pathology in trisomy 21 cerebral organoids reveals BACE2 as a gene-dose-sensitive AD-suppressor in human brain”

A population of >6 million people worldwide at high risk of Alzheimer’s disease (AD) are those with Down Syndrome (DS, caused by trisomy 21 (T21)), 70% of whom develop dementia during lifetime, caused by an extra copy of β-amyloid-(Aβ)-precursor-protein gene. We report AD-like pathology in cerebral organoids grown in vitro from non-invasively sampled strands of hair from 71% of DS donors. The pathology consisted of extracellular diffuse and fibrillar Aβ deposits, hyperphosphorylated/pathologically conformed Tau, and premature neuronal loss. Presence/absence of AD-like pathology was donor-specific (reproducible between individual organoids/iPSC lines/experiments). Pathology could be triggered in pathology-negative T21 organoids by CRISPR/Cas9-mediated elimination of the third copy of chromosome-21-gene BACE2, but prevented by combined chemical β and γ-secretase inhibition. We found that T21-organoids secrete increased proportions of Aβ-preventing (Aβ1-19) and Aβ-degradation products (Aβ1-20 and Aβ1-34). We show these profiles mirror in cerebrospinal fluid of people with DS. We demonstrate that this protective mechanism is mediated by BACE2-trisomy and cross-inhibited by clinically trialled BACE1-inhibitors. Combined, our data prove the physiological role of BACE2 as a dose-sensitive AD-suppressor gene, potentially explaining the dementia delay in ∼30% of people with DS. We also show that DS cerebral organoids could be explored as pre-morbid AD-risk population detector and a system for hypothesis-free drug screens as well as identification of natural suppressor genes for neurodegenerative diseases.


Introduction
Production 1-3 , and degradation 4 of β-amyloid peptides (Aβ) are among the central processes in the pathogenesis of Alzheimer's disease (AD). The canonical Aβ peptide is produced after sequential cleavage of the β-amyloid precursor protein (APP) by β-secretase and γ-secretase, generating a peptide that most often begins 99 amino acids (aa) from the C-terminus of APP with Asp1 and contains the next 37-42 aa of the APP sequence, generating a range of peptides (Aβ1- 37, 38, 39, 40 and 42). The longer of these peptides can be detected in toxic amyloid aggregates in the brain, associated with AD and other neurodegenerative disorders 5 . As APP gene is located on human chromosome 21, people with Down Syndrome (DS, caused by trisomy 21 (T21)) are born with one extra copy of this gene, which increases their risk of developing AD. Non-DS (euploid) people inheriting triplication of the APP gene alone (DupAPP) develop AD symptoms by age 60 with 100% penetrance. Paradoxically, only ~70% of people with DS develop clinical dementia by age 60, suggesting the presence of other unknown chromosome 21-located genes that modulate the age of dementia onset 6,7 . A number of secretases participate in the physiological cleavage of APP 1,8 , generating various peptides involved in neuronal pathology.
BACE1 is the main β-secretase in the brain 9 , while the expression and function of its homologue BACE2 (encoded by a chromosome 21 gene) remain less clear 10,11 . At least 3 different activities of BACE2 were recorded with regards to APP processing: as an auxiliary β-secretase (proamyloidogenic), as a θ-secretase (degrading the β-CTF and preventing the formation of Aβ), and as Aβ-degrading protease (AβDP) (degrading synthetic Aβ-peptides at extremely acidic pH). It remains unclear which of these activities reflect the role of BACE2 in AD. The potential activity of BACE2 as an anti-amyloidogenic θ-secretase can be predicted from studies on a variety of transfected cell lines that overexpress APP, and artificially manipulate the dose of BACE2 [12][13][14][15] .

Trisomy 21 (but not DupAPP) skews the ratios of Aβ non-amyloidogenic peptides
We compared organoids from isogenic iPSC lines, derived from the same individual with DS, mosaic for T21 and normal disomy 21 (D21) cells 18 . Cerebral organoids were derived following a standard protocol 19 , and shown to contain neurons expressing markers of all 6 layers of the human cortex ( Supplementary Fig. 1) and no significant difference in the proportions of neurons and astrocytes between the D21 and T21 organoids ( Supplementary Fig. 2). The integrity and copy number of the iPSC lines were validated at the point of starting the organoid differentiation, for chromosome 21 ( Supplementary Fig. 3), and the whole genome (available on request). T21/D21 status was further verified by interphase FISH on mature organoid slices, ( Supplementary Fig. 4a). The C-terminal region of APP can be processed by the sequential action of different proteases to produce a range of protein fragments and peptide species, including Aβ ( Supplementary Fig. 5). Aβ peptide profiles were analysed from organoidconditioned media (CM) whereby each CM sample was taken from a 6cm dish culturing a pool of 12-16 organoids derived from one iPSC line, in total: n=15 CM samples for Exp1 (3 trisomic isogenic lines, 2 disomic isogenic lines, 3 timepoints each), n=12 CM samples for Exp2 (2 trisomic isogenic lines, 2 disomic isogenic lines, 3 timepoints each) and n=20 CM samples for Exp3 (1 trisomic isogenic line, 1 disomic isogenic line, 1 DupAPP line, 1 line each for two different unrelated DS individuals, 4 timepoints each). CM was collected at a timepoints between days 100-137 of culturing and analysed using immunoprecipitation in combination with mass spectrometry (IP-MS) 20 . Please see "Methods" and "Supplementary Data" sections for more detailed explanations, and statistical controls used for individual iPSC line-to-line comparisons.
We observed that T21 organoids produce statistically highly significant increases in all four of these ratios, relative to isogenic D21, or non-isogenic DupAPP organoids (Fig. 1b). Therefore, we conclude that T21 causes these effects in our organoid system. The D21 ratios were not significantly different to DupAPP, suggesting that the third copy of genes other than APP causes these effects, though this needs to be tested on a larger number of individuals.
As the peptide profiling data strongly favour the hypothesis of a genetic-dose-sensitive antiamyloidogenic action of BACE2, we sought to zoom in on the BACE2 genetic locus in a systematic SNP-array analysis of 554 individuals recruited through the LonDownS Consortium who had undergone detailed assessment for dementia 22,23 ; 93 single nucleotide polymorphisms (SNPs) located within the BACE2 locus +/-50kb, were genotyped, and dementia age-of-onset determined, as described in Methods. We detect two new BACE2 SNPs (purple, Supplementary   Fig. 7) correlating with age of dementia onset in the DS cohort of the LonDownS Consortium, located in close proximity to a previously reported SNP (red, Supplementary Fig. 7) 24 . All of these 3 SNPs cluster in <4kbp segment, which is fully contained within a 12kbp deletion (blue line, Supplementary Fig. 7) that caused a de novo EOAD in a euploid patient 25 ( Supplementary   Fig. 7). These data corroborate the notion that subtle genotypic variation in BACE2 levels may play an important role in affecting the age of dementia onset in both DS and non-DS individuals.

Non-amyloidogenic Aβ peptide ratios mirror between T21 organoids and DS-CSF
In order to assess if the peptide ratio differences from Fig. 1b have any relevance in vivo, we analysed the Aβ-peptide profiles immunoprecipitated from human cerebrospinal fluid (CSF). We have previously produced IP-MS data on CSF from people with DS and age-matched controls 26 .
We repeated the calculations shown for organoids in Fig. 1b, on IP-MS results from CSF samples from DS (n=17) and age-matched euploid people (n=12). All four relative ratio calculations showed an increase in peptide ratios in CSF from people with DS, compared to agematched euploid controls, of which three comparisons were statistically highly significant (Fig.   1c). This suggests that in DS brains, the third copy of BACE2 skews the anti-amyloidogenic processing significantly towards BACE2-cleavages, relative to other anti-amyloidogenic enzymes cleaving at the α-site. Importantly, these CSF results validate the in vivo relevance of the peptide ratios obtained using CM from iPSC-derived cerebral organoids (comparison of Fig.   1b and Fig. 1c).

Aβ-degrading activity of BACE2 is cross-inhibited by clinically trialed BACE1-inhibitors
Chemical inhibition of BACE1 remains an attractive therapeutic strategy for AD. As BACE2 is a homologous protein, most inhibitors tested in clinical trials also cross-inhibit the (proamyloidogenic) β-secretase activity of BACE2, which has been proven as the cause of several unwanted side-effects, such as skin pigmentation changes. As our data suggest that the opposite, Aβ-degrading, activity of BACE2 plays an important role, we designed a new FRET-based in vitro assay, in which efficient AβDP-cutting after Aβ aa34 by BACE2 at pH=3.5 could be measured (Fig. 2), while zero activity by BACE1 was detectable under same conditions (Supplementary Information). We demonstrated that at least two BACE1 inhibitor compounds (of which one recently used in clinical trials) inhibit the AβDP (Aβ-clearance) activity of BACE2 in a dose-dependent manner (Fig. 2). This has, to our knowledge, so far not been shown, and could provide an additional explanation for the failure of some BACE1-inhibitor clinical trials, and should be taken in consideration when testing new inhibitors.

AβDP product (Aβx-34) colocalises with BACE2 in human brain and organoid neurons
As in vitro experiments showed that BACE2 can very efficiently cleave the Aβ34-site in the FRET peptide (Fig. 2) and synthetic Aβ1-40 peptide in solution at an acidic pH 12 , we sought to visualize if the presence of the substrate (Aβ1-40), enzyme (BACE2) and one of the products of this reaction (Aβ1-34) can be detected in our organoids, in a sub-cellular compartment known to be acidic. Firstly, by immunofluorescence (I.F.) using pan-anti-Aβ (4G8), anti-BACE2, or neoepitope-specific antibodies against Aβx-40 and Aβx-34 27 , we detected significantly higher signals (normalized to pan-neuronal marker) in T21 organoid neurons, compared to isogenic D21 ones ( Supplementary Fig. 4b-d). Pearson's coefficient showed a high level of colocalisation (>0.55) of both the main substrate (Aβx-40) and its putative degradation product (Aβx-34) with BACE2 in neurons of cerebral organoids, in LAMP2+ compartment (known to be a subset of lyzosomes, therefore low pH vesicles) ( Fig. 3 & Supplementary Fig. 8). In comparison, the Pearson's coefficient for BACE1 with Aβx-34 was only 0.16 ( Fig. 3 & Supplementary Fig. 8), and its pattern of sub-cellular localization was different to BACE2 (high colocalization with Rab7 and Sortilin, much lower with LAMP2). Using I.F. on human brain sections, a similar highly significant difference was observed (Fig. 4a, b): Aβx-34 colocalised with BACE2 (0.52 (±0.034SEM)) as opposed to BACE1 (0.01 (±0.021SEM)). The colocalised signal of Aβx-34 and BACE2 was seen in 3 categories of objects ( Fig. 4), in all analysed samples: 4 individual DS-AD brains ( Fig. 4a-c), 5 euploid sporadic AD subjects (example in Supplementary Fig. 14a, for complete list of brain samples see Supplementary Table 1) and (in the fine vesicle compartment only) in 5 non-demented control euploid subjects' neurons (age 42-84), as well as DS brain from a 28yr old with no plaques or dementia, (examples in Fig. 4d, for complete list of brain samples see Supplementary Table 1). Lambda scanning and Sudan black B stainings were independently used to subtract the autofluorescence of lipofuscin granules ( Supplementary Fig. 15f, g). This has proven that the fine vesicular pattern and large amorphous extra-cellular aggregates are not autofluorescent lipofuscin granules, but real colocalisations of BACE2 and Aβx-34 ( Supplementary Fig. 15). Colocalised signals of Aβx-34 and BACE2 were particularly strong in areas surrounding neuritic plaques (Fig. 4a-c).
As AβDP cleavage by BACE2 is efficient only at low pH, we sought to analyse in more detail the BACE2 and Aβx-34 co-localisation in highly acidic cellular compartments. For this reason, we costained lysosome markers LAMP1 or LAMP2 with Aβx-34. Additionally, macroautophagic vacuoles containing Aβ were shown to accumulate in AD distended neurites 28 , which is why we also stained with the macro-autophagosome marker LC3A. As we further found that Aβx-34 did not colocalise with LAMP1 or LC3A, but colocalised strongly with LAMP2 (Fig. 3, Supplementary Figure 8 and Supplementary Information), we tested colocalisation with the components of an alternative autophagy pathway: chaperone-mediated autophagy (CMA), and found a very high level of colocalisation (Fig. 3).
This pinpoints the triplication of BACE2 as a likely cause of specific anti-amyloidogenic T21 effects we observed in Fig. 1a. Furthermore, we used two different dyes to detect any presence of amyloid deposits (the traditional Thioflavine S, and a newer, more sensitive dye AmyloGlo 29 ) in organoid sections. Remarkably, elimination of the third BACE2 copy caused the T21 organoids (that had not shown any overt amyloid deposits at 100DIV, see T21C5 in Supplementary Fig. 11, top row) to develop extremely early AD-plaque like deposits (AmyloGlo+ and Thioflavine S+) in the cortical part of the organoid by 48DIV ( Supplementary Fig. 11, middle row), that progressed aggressively and became much stronger and denser by 96 DIV, accompanied by massive cell death ( Supplementary Fig. 11, bottom row, Supplementary Fig. 12).
In order to prove that extracellular deposits staining positively with amyloid dyes really are related to hyperproduction of Aβ amyloidogenic peptides, we cultured T21C5∆7 organoids in media containing high concentrations of β and γ secretase inhibitors. Early T21C5 and T21C5∆7 organoids were treated with a combination of β-secretase inhibitor IV and compound E (γ secretase inhibitor XII) (Supplementary Table 2) from 20DIV to 41DIV (Fig. 6). Amyloid-like deposits were readily detected with AmyloGlo in the untreated and vehicle only treated T21C5∆7 organoids (Fig. 6b), but were completely absent from T21C5∆7 organoids treated with β and γ secretase inhibitors. Inhibitor treatment also significantly reduced the number of neurons expressing pathologically conformed Tau (TG3-positive cells) in the T21C5∆7 compared to untreated controls (Fig. 6c). No AmyloGlo positive aggregates or TG3-positive cells were detected in T21C5 organoids under any treatment conditions at DIV41 (Fig. 6a, c) and were also absent in the same organoids at DIV100 (Fig. 7 g, l, Supplementary Fig. 11). Also, no obvious deleterious effects of the inhibitors, or vehicle control, could be seen in early unedited T21C5 organoids.
Further histo-pathological verification showed that elimination of one copy of BACE2 triggered progressive accumulation of extracellular deposits that co-stain with Thioflavine S and antibodies against Aβ, both 4G8 and neo-epitope specific Aβx-40&Aβx-42. The antibody signal intensity in colocalisations with Thioflavine S drastically increased upon pre-treatment with 87% formic acid (Fig. 7a-d), proving that the deposits contain insoluble Aβ material. This is further corroborated by the isolation of fibrillary material from the detergent-insoluble fraction of the CRISPR-edited organoid. When viewed by Transmission Electron Microscopy (TEM) the filaments found exhibited a straight morphology of <10nm diameter ( Supplementary Fig. 13a), closely resembling fibrils grown in vitro from synthetic Aβ1-40 peptide (Supplementary Fig.   13c). Furthermore, neuritic plaque-like features were detected by IHC co-staining with Gallyas in CRISPR-edited organoids (Fig. 7m, n), but not their unedited T21 control (Fig. 7l). Human brain from an AD patient is shown for comparison stained with Gallyas (Fig. 7k). Tau pathology was also observed by IHC using the hyper-phosphorylated Tau antibody AT8 (Fig. 7e, f), and by I.F. for conformationally altered Tau (TG3, Fig. 7g-j). The relative increase in the amount of conformationally altered (pathological) Tau in CRISPR-edited organoids T21C5Δ7, compared to unedited T21 control organoids, was also independently confirmed by immunoblotting using TG3 antibody. As shown in Fig. 7o, the protein material isolated from T21C5Δ7 organoids produced significantly more TG3 signal than unedited controls, albeit having a weaker signal with the general 3R-Tau antibody (consistent with the observed neuronal loss, Supplementary   Fig. 12).

Alzheimer's disease-like pathology develops reproducibly in unedited cerebral organoids from 71% of DS donors, and it is donor-specific
Our data in Figs. 5, 6, 7 show that severing the BACE2 dose by a third, using CRISPR/Cas9, might tip the balance against the anti-amyloidogenic activity, and provoke AD-like pathology.
Our data in Fig.1 suggest that anti-amyloidogenic activity of BACE2 is gene-dose dependent, and its level varies between individuals, with SNP allelic differences in BACE2 gene correlating with age of dementia onset. We therefore hypothesized that organoids grown from some people with DS may develop AD-like pathology without any CRISPR-Cas9 intervention. We then tested this hypothesis using iPSC lines from 6 different individuals with DS, and one DupAPP patient (Table 1). We detected amyloid-like aggregates (both diffuse and compact in appearance) in 5/7 unedited iPSC-derived organoids from people with DS, and one with DupAPP (Fig. 8).
The two donors whose iPSC-organoids did not show pathology are (i) the T21 iPSC from our isogenic model (whose clinical status is unknown) and (ii) QM-DS6, a donor who remains free from dementia symptoms at age 37 (Table 1). Organoids from another 5 DS donors, and one DupAPP patient, (all diagnosed with clinical dementia) all showed presence of diffuse and compact amyloid-like deposits (Fig. 8) as well as presence of neuritic plaque-like features (focal hyper-phosphorylated tau (AT8+), conformationally altered tau (TG3+) and filamentous Tau (AT100+)) within neuropil neurites within plaque-like circular foci ( Fig. 9a-n). This was corroborated by Gallyas intra-neuronal positivity ( Fig. 9o-t). Similarly as for T21C5Δ7, we were able to isolate fibrillary material from the detergent-insoluble fraction of QM-DupAPP organoid ( Supplementary Fig. 13b), that on TEM resembled fibrils grown in vitro from synthetic Aβ1-40 peptide ( Supplementary Fig. 13c). Most importantly: tested individual organoids from one donor (from multiple iPSC lines and multiple independent experiments) either all did (DupAPP, QM-DS1-5), or all did not (isogenic T21, QM-DS6) show AD-like pathology (Table 1), proving the pathology is donor dependent. This open possibilities of developing assays for pre-therapy riskstratification and individualized drug-response quantitation.

Discussion
Several human brain studies show detectable expression and β-secretase activity of BACE2, though at much lower levels than that of BACE1 [30][31][32][33] . Chemical inhibition of β-secretase activity is an attractive therapeutic approach aimed at reducing the production of Aβ [34][35][36] . Complete knock-out of BACE1 abolished all β-secretase activity in mouse neurons, while leaving some degree of β-secretase activity in astrocytes 37 . This activity was abolished by the complete knockout of both BACE1 and BACE2, leading to a hypothesis that a BACE2-driven β-secretase activity in astrocytes may contribute to accelerate the Aβ-production and AD-pathology in DS 37 .
In human brain, the β-secretase activity of BACE1 correlated positively with the amount of Aβ, whereas the β-secretase activity of BACE2 did not 30 . On the other hand, SNPs at the BACE2 locus (and not BACE1) correlate with the age of onset of dementia in people with DS 24 , as well as sporadic LOAD in euploid people in the Finnish population 38 , and a recent report showed that a de novo intronic deletion within one allele of BACE2 caused EOAD in a 50 year old euploid person 25 .
All of the above data (and new data we show in Supplementary Fig. 7) implicate that a single allele alteration in the genetic dose of BACE2 is capable of affecting the risk of AD-dementia, but do not resolve the question whether BACE2 per se acts predominantly as an accelerator, or a suppressor of AD pathology. The answer to this question requires clarification, as most chemical inhibitors used in clinical trials have dual activity against BACE1 and BACE2 35,39 .
The increased ratios of 1-20&1-34 (BACE2-AβDP) to the amyloidogenic and α-site products are among our most consistent and robust observations in T21 organoid CM and DS-CSF ( Fig. 1b-c). The 1-34 generating cleavage can only occur after the cuts by both β-and γ-secretases have released Aβ, because the hidden transmembrane site between aa34 and aa35 is inaccessible to any proteolytic enzymes until the soluble Aβ (1-37 to 1-42) molecules are released from the membrane 12,13 . Therefore, the Aβ1-34 species can only be a product of an AβDP activity (a catabolic degradation or clearance of an already made Aβ1-37 to 1-42 peptides). Besides BACE2, the only enzymes with potential to cleave the peptide bond Leu34 -Met35 are BACE1 12,14 , and extracellular matrix (ECM) metalloproteinases (MMP2 and MMP9) 40 , since no other Aβ degrading enzymes (neither IDE, nor NEP, nor ECE) are known to cleave at this site 41 .
BACE1 action is unlikely to cause the increased ratios we observe, as BACE1 can only generate this cut in solution at very high enzyme concentration and after prolonged incubation 12 .
To further corroborate this point, we designed a novel FRET-assay and established the conditions in which BACE2 can efficiently cleave at Aβ34 site ( We also demonstrated that two BACE1 inhibitors (β-Secretase Inhibitor IV -CAS 797035-11-1 (Calbiochem, originally a Merck compound)), and LY2886721 (Eli Lilly compound recently used in clinical trials) both inhibit the AβDP activity of BACE2 in vitro, while the γ-secretase inhibitor (DAPT) had no effect.
This suggests that the AβDP activity (cutting the peptide bond Leu34 -Met35) has a different enzymatic preference, conditions, and pH, as compared to the classical β-secretase cleavage that both BACE1 and BACE2 are capable of. As FRET assays cleaving this classical (before Asp1) site are generally used to measure the BACE1 inhibitors' selectivity for BACE1 or BACE2, our data suggest that the degree of selectivity for any given inhibitor calculated this way, does not necessarily reflect whether the same selectivity would apply for their cross-inhibition of the Leu34 -Met35 site cleavage (AβDP) activity. Interestingly, the presence of the Aβx-34 degradation product, both alone 27 and co-localising with BACE2 ( Fig. 4) show elevated levels in cells and extracellular aggregates immediately surrounding neuritic plaques, suggesting BACE2 degradation of not only newly produced Aβ, but also of Aβ that is released and re-deposited (from and to) existing deposits. A recent report on widespread somatic changes in individual neurons suggests an additional mechanism for the production of toxic Aβ species, including products that do not require secretase cleavage 42 , underscoring the importance of efficient Aβ degrading mechanisms that protect from AD, such as the one exerted by BACE2 that we describe here. A recent mouse model has shown that introducing a third dose of chromosome 21 to a mouse that several hundred fold over-expresses Aβ40 and 42 worsens the amyloid plaque load, and this correlates with an unexpected decrease in the Aβ40/42 ratio 43 . This unfavourable ratio effect (the cause of which is unknown) is expected to worsen the plaque load and AD pathology, and a mere 1.5x increase of Bace2 dose in this mouse model has no chance in protecting the mouse against a >100x overload of Aβ. In another mouse model, where transgenic BACE2 was artificially overexpressed together with transgenic wtAPP, it actually decreased Aβ40 and 42 to the wt mouse control levels, and the presence of BACE2 transgene reversed behavioural pathologies seen in TgAPP mouse 44 .
This indicates that a balance of doses of APP and BACE2 affects levels of soluble Aβ40 and 42, and their oligomerization and aggregation as a consequence. Our results in Figs 5, 6 and 7 further corroborate that a significant disturbance of this balance by a reduction in BACE2 copy number is sufficient to cause an early AD-like pathology in T21 cerebral organoids. We did not see any amyloid plaque-like structures at >100DIV organoids from three independent T21 iPSC lines (or normal disomic lines) of our isogenic system (Supplementary Figs. 1,2,4,8,11,14,Figs. 3,7,8). Surprisingly, CRISPR/Cas9 elimination of the third copy of BACE2 in the same T21 line caused widespread AmyloGlo+ deposits at 41DIV, and widespread neuritic plaque-like structures with profound neuron loss ( Supplementary Fig. 11, 12) and Tau pathology at 96DIV (Figs. 6, 7). Our data in Figs.1 and Supplementary Fig. 7 suggest that anti-amyloidogenic activity of BACE2 is gene-dose dependent, and its level varies between individuals, with SNP allelic differences in BACE2 correlating with age of dementia onset. We therefore hypothesized that organoids grown from some people with DS may develop AD-like pathology without any CRISPR-Cas9 intervention. Diffuse amyloid plaque-like appearance with Tau pathology was recently reported in 110 days old cerebral organoids from only a single DS-hiPSC line 45 so far.
We subsequently analysed iPSC-derived organoids at approximately the same cell culture age from a total of 7 different individuals with DS and one with DupAPP. We found flagrant ADlike pathological changes in 5/7 DS tested (71%), as well as the one DupAPP. Very interestingly, when this assessment was repeated in independent experiments, and when individual organoids from a single experiment were compared, it was a black/white picture: either they all had ADlike pathology, or none did, driven solely by the genotype of the donor (Table 1). Our data, though not conclusive, are illustrative of the stratifying potential of this technology. For example, the cerebral organoids from individual QM-DS3 showed the worst AD-like pathology with fibrillary amyloid deposits ( Fig. 8f, i, j, Table 1), and this individual was diagnosed with dementia at age 37. In contrast, organoids from individual QM-DS6 showed no pathology (Fig.   8b, Table 1), and this individual was also dementia-free at age 37. This opens up possibilities for finding correlations with clinical parameters, for which a much larger number of individuals would have to be tested.
To confirm that the AmyloGlo deposits were in fact aggregated β-amyloid containing material, early organoids were treated with a combination of β-secretase inhibitor IV (βI-IV) and gamma secretase inhibitor XII (Compound E) (Fig. 6 a, b). The combination of these inhibitors should prevent any production of Aβ, and therefore eliminate AmyloGlo positivity. After treatment for 21 days, the inhibitor treatment did indeed prevent the formation of plaque-like deposits within T21C5Δ7 organoids, confirming that such deposits are comprised of β-amyloid. The same treatment conditions also significantly reduced the number of TG3-positive cells in T21C5Δ7 organoids (Fig. 6c), highlighting the ability to modulate both amyloid and tau pathology in the cerebral organoid system. This also demonstrates the feasibility of using this AD-like organoid pathology in future hypothesis-free drug screens for chemical compounds that may prevent/inhibit amyloid production or aggregation.
In view of our results, it becomes inviting to hypothesize that triplication of BACE2 may be the cause of the delayed onset of dementia in 30% of people with DS compared to DupAPP 7 , and (because of the predicted abundance of BACE2 mRNA in endothelial cells) also the cause of a significantly lower degree of cerebral amyloid angiopathy (CAA) in the brains of people with DS compared to those of DupAPP 46 . Our organoid system is not informative in this regard, as we could not detect any endothelial cells in our organoids (not shown). This, however, is also an advantage, as it allows uncovering the mechanisms that are specific to neurons in the absence of endothelial or blood cell derived tissue components.
In neurons, a recent report also found that an increased APP dose may act (through an unknown mechanism) as a transcriptional repressor of several chromosome 21 genes, including BACE2 47 .
This observation needs further verification and mechanistic explanation, but if true, it would imply that the protective effect of the third copy of BACE2 in DS that we observe is actually quenched by the third copy of APP, which opens up possibilities of chemically intervening to inhibit this transcriptional repression and potentially unleash a much greater degree of BACE2 protection. An integration of the two observations (the one in 47 and the one in our report) suggests this could be exploited as an additional new protective/therapeutic strategy for AD in general.
We found, surprisingly, an equally high or higher level of colocalisation of Aβx-34 with LAMP2A, as with the general LAMP2 (Fig. 3). The high level of colocalisation with LAMP2A and absence of colocalisation with either LC3A or LAMP1 (Fig. 3), suggest that AβDP activity of BACE2 that generates Aβ34 is not related to classical lysosomal degradation or macroautophagy, but rather could be related to a CMA-like process 48,49 . The only published study that linked CMA with APP processing 50 found a motif that satisfies the criteria for a CMArecognition KFERQ motif at the very C-terminus of APP (KFFEQ), and this paper demonstrated that C99 (β-CTF) can bind HSC70. However, paradoxically, when this motif is deleted from the β-CTF, the binding to HSC70 is not abolished, but rather increased, suggesting the presence of another, alternative CMA-recognition motif within the β-CTF peptide 50 . The association of the AβDP x-34 product with LAMP2A/CMA compartment is a provocative new observation that requires further studies.
In conclusion, we found that relative levels of specific non-amyloidogenic and AβDP (Aβclearance) products are higher in T21 organoids and DS-CSF, and they respond to the dose of BACE2 (and not APP). We also demonstrated that BACE2-AβDP activity generating one of these products can be cross-inhibited in solution by recently clinically tested BACE1-inhibitors.
All components of the AβDP degradation reaction (hitherto only demonstrated in solution in vitro): the main substrate (Aβx-40), the enzyme (BACE2), and its putative degradation product (Aβx-34), we found highly colocalised in discrete intracellular vesicles in human brain neurons, (and not astrocytes), suggesting that at least some of the AβDP activity generating Aβx-34 takes place intra-neuronally and physiologically during lifetime, before the onset of AD pathology, in both normal and DS brains. Furthermore, we directly demonstrated that the third copy of BACE2 protected T21-hiPSC organoids from early AD-like amyloid plaque pathology, therefore proving the physiological role of BACE2 as an AD-suppressor gene. The BACE2's θ-secretase antiamyloidogenic cleavage and the AβDP degradation actions could both be contributing to an overall AD-suppressive effect. Regardless of the contribution of each of these modes of action, our combined data suggest that increasing the action of BACE2 could be exploited as a therapeutic/protective strategy to delay the onset of AD, whereas cross-inhibition of BACE2-AβDP activity by BACE1-inhibitors would have the unwanted worsening effects on disease progression. We also show that cerebral organoids from genome-unedited iPSCs could be explored as a system for pre-morbid detection of high-risk population for AD, as well as for identification of natural dose-sensitive AD-suppressor genes.

Human subjects -clinical assessment
Human subjects were participants in the "The London Down Syndrome Consortium  Table 1.

Primary hair follicle keratinocyte sampling
Upon specific informed consent, three to six individual strands of hair were non-invasively plucked from the scalp hair of donor subjects, and placed in transport medium [DMEM (Sigma D5546), 2mM glutamine (Sigma G7513), 1x Pen/Strep (Sigma, P4333), 10% foetal calf serum].

Reprogramming of primary keratinocytes
Primary keratinocyte cultures were expanded to 70% confluency, electroporated with plasmids encoding reprogramming factors in episomal vectors (non-integrational reprogramming), and  (Life Technologies) supplemented with penicillin/streptomycin. Passaging was carried out using ReLESR and 10 μM ROCK inhibitor was included in culture media for 24 hours after passaging.
Cerebral organoids. Cerebral organoids were generated following the standard protocol with the following changes 18 . iPSC lines were first transitioned into feeder free conditions using either mTESR1 or E8 media with geltrex. To form embryoid bodies (EBs), hiPSCs were washed once with PBS, then incubated with Gentle Cell Dissociation Solution (Stemcell Technologies) for 4mins. This solution was then removed and accutase added and incubated for a further 4mins. mTESR1/E8 medium at double the volume of accutase was added to the cells and a single cell suspension generated by titruating. Cells were centrifuged to remove accutase and then resuspended in hESC medium supplemented with 4ng/ml FGF2 and 50micromolar ROCK inhibitor. 9000 cells were used to form a single EB in each well using either a V shaped ultra low attachment 96 well plate (Corning). Specifically, iPSCs were allowed to form embryoid bodies (EBs) in suspension by culturing for 6 days in hESC medium with low FGF, in non-adherent culture dishes. After 5-7 days, EBs were transferred into a 24 well ultra low attachment plate for neural induction. Neural induction was achieved by culturing for further 5-7 days in DMEM-F12 supplemented with 1% of each: N2, GlutaMAX and MEM-NEAA, plus 1μg/ml heparin.
Neurally induced EBs showing neuroectodermal "clearing" in brightlight microscopy were embedded in matrigel droplets, and transferred to 6cm dishes containing organoid differentiation medium-A, (for 4-5 days), followed by organoid differentiation medium+A 18 . Organoid maturation was carried out with 12-16 organoids per 6cm dish on an orbital shaker at 37°C, 5% CO2. Aliquots of conditioned medium (CM) were collected from mature organoids (100-137 days old from day of EB formation), 3-4 days after feeding (to allow time for cells to secrete products into the culture media). Three completely independent experiments were carried out each time starting from undifferentiated iPSC stage, and CM was collected at 3-4 timepoints in each experiment. CM was immediately frozen and stored at -80°C.
For inhibitor treatment, organoids were treated from 20DIV (6 days after embedding in matrigel) to 41DIV. βI-IV and Compound E were added freshly to the media before use at final concentrations of 2.5μM and 6nM respectively. Media was replaced every 3-4 days during treatment. DMSO of the same volume was used as a vehicle only control.

IP-MS.
CM from organoids was analysed by IP-MS, using a previously described method 20 . The team performing the MS was blinded to the genotypes in all experiments. In exp1, all three independent trisomic lines (T21C6, T21C5, and T21C13) were compared to two independent disomic lines (D21C3 and D21C7), whereas in exp2, two independent trisomic lines (T21C6 and T21C13) were compared to two independent disomic lines (D21C7 and D21C9).
In exp3, a T21C6 line was compared to the isogenic D21C9 line, and to hiPSC lines from 3 unrelated individuals: a DupAPP FEOAD patient (QM-DupAPP), and two unrelated adult people with DS (QM-DS1 and QM-DS2).
In all 3 experiments, IP-MS results for all iPSC lines that were used in a particular experiment are shown. IP-MS results were used to calculate the relative ratios of peptides and these ratios were taken as data points for the statistical comparisons.
IP-MS spectra were also obtained from the CSF samples of people with DS and age-matched normal controls. Peak ratios calculated as described above. The cohorts, methods and spectra behind these data were previously described 26 .

Fluorescence in situ Hybridization (FISH)
. FISH on organoid cryosections was performed as described 51 . Briefly, slides were rinsed in PBS, rehydrated in 10 mM sodium citrate buffer and incubated in the same buffer at 80⁰C for 20 min. Slides were cooled down and incubated in 2x Saline Sodium Citrate (SSC) for 5 min and in 50% formamide in 2x SSC for 1h. After incubation slides were covered with previously prepared hybridization chamber and incubated with 10 μL of   Fig. 14).
Western Blot. For western blots, whole cell lysates of CRISPR edited or unedited iPSCs (Fig.   5c) or organoids (Fig. 7o) were separated in a 10% acrylamide gel by SDS-PAGE and transferred to a nitrocellulose membrane according to the manufacturers protocols (Bio-Rad).
Following a 60min incubation in 5% non-fat milk in TBS-T the membrane was incubated with primary and secondary antibodies (Supplementary Tables 3, 4). For the stainings shown in Fig.   7o quantitations were done strictly on the same membranes re-stained using the antibodies shown. For the protein of interest (BACE2 or TG3), the signal was adjusted to corresponding βactin loading control for all samples. Such adjusted values for unedited C5 (wt) (n=4) were set to 1, and used to calculate the fold change for C5Δ7 (n=4) replicates, and the resulting fold-change values for pairs run on the same gel were averaged and analysed by student's t-test. Membrane stripping between stainings was carried out using Thermo-Fisher stripping solution, following manufacturer's instructions.
AmyloGlo and Thioflavine S staining. For AmyloGlo staining, OCT embedded slices were rinsed with PBS, and incubated in 70% ethanol for 5 min at RT, followed by washing with milli Q water for 2 min at RT. Slices were then incubated with AmyloGlo solution for 10 min in the dark at RT, followed by washing in 0.9% saline solution for 5 min at RT, and counterstaining with DRAQ5 for 10min at RT. Thioflavine S staining was performed as described 52 Supplementary Fig.   15a ), pre-incubation with BACE2 specific immunogenic peptide ( Supplementary Fig. 15b-e) and Lambda (λ) scan function on confocal microscope ( Supplementary Fig. 15f, g). Three different samples (DS-AD1, DS (28 yrs) pre-AD and euploid sporadic AD (73 yrs) after IHC were stained with 0.1% Sudan black B in 70% ethanol for 20 min at RT and analysed on confocal microscope with Aiyrscan. Sample DS-AD1 was stained with antibodies solution, 12 hrs pre-absorbed with BACE2 specific immunogenic peptide, and analysed on confocal microscope and slide scanner. Lambda scan records a series of individual images within a defined wavelength range (in our case from 630 nm to end of spectrum) and each image was detected at a specific emission wavelength, at 10 nm intervals. For lambda scan analysis, samples were stained with one primary antibody and labelled with far-red secondary antibody (647). As negative control, we used secondary antibody (647) alone and, as additional negative control, one sample was counterstained with DAPI only, without secondary antibody. As we used a far-red (647) antibody, we analysed expression from 630 nm to the end of spectrum at 10 nm intervals. Aβx-34 and BACE2 antibodies showed specific peaks, significantly over and above the autofluorescent signal, in all three specific ROI indicated in Fig. 4 Fig. 15).
After staining samples were rinsed in water, dehydrated in a graded series of ethanol, cleared in Histo-Clear and mounted with Histomount mounting medium. Samples were scanned by shown for the lack of space, data available on request). The genome integrity of the isogenic iPSC lines was previously published 18 (but was repeated here as described above). No additional rearrangements due to re-programming or passaging were observed.
BACE2 locus SNPs: The cohort of people with DS has been described in recent reports 22,23 . In brief, participants donated DNA samples and had detailed cognitive and clinical assessments to determine dementia status 54 . Age of dementia diagnosis was established and used in SNP analysis. BACE2 SNP genotyping for the LonDownS cohort was undertaken as previously  Supplementary Fig. 7) were nominally associated with AOO in the LonDownS cohort, but were not significant after correction for multiple testing.

Quantitative paralogous amplification-pyrosequencing
Quantitative paralogous amplification-pyrosequencing was carried out based on the published method 55 . This method takes advantage of the existence of identical sequences on chromosome 21 and one other autosome, allowing amplification of both loci with a single primer pair.
Paralogous sequence mismatches in amplified products from chromosome 21 (GABPA and ITSN) can be quantified relative to their paralogous regions on chromosome 7 and 5 respectively. As such, trisomic cells show a 60:40 ratio for the paralogous sequence, while disomic cells produce a 50:50 ratio. Primers used are listed below, and pyrosequencing was performed on the Pyromark Q48 machine (Qiagen) following standard procedures.
CRISPR/SpCas9-HF1Cas9 editing of the BACE2 locus. The guide-RNA (gRNA) targeting BACE2 Exon 3 was cloned into a vector containing the high fidelity SpCas9-HF1 56 and blasticidin S resistance gene. The complete plasmid was delivered via Lipofectamine3000 to a trisomic iPSC line T21C5 (full official name NIZEDSM1iT21-C5), which was described and characterized in a previous report 18 . Untransfected iPSCs were removed by treatment with blasticidin (2 μg/ml for 48h). Individual colonies were picked and further sub cloned by limiting dilution to achieve clonal cell lines. DNA was purified from individual clones, PCR amplified and sequenced by Sanger Sequencing. Sequences were analysed in Mutation Surveyor (V3.1.0) and "Tracking InDels by dEcomposition (TIDE)" (TIDE V 2.0.1, Desktop Genetics). TIDE analysis of the CRISPR-targeted clone 2.3.5 DNA sequence gave a score of 65% of the wt read remaining (not shown). The quality of the gRNA was assessed using two different prediction software platforms: CCTop online software 57 , and the MIT online platform (http://crispr.mit.edu/). The same two software platforms were used to predict the off-target sites.
Neither platform found any off-targets with 0, 1 or 2 mismatches. The top 10 CCTop-predicted sites were PCR amplified in both Δ7 and WT clones, then sequenced by Sanger Sequencing to rule out off target events. No differences in the sequence were found.
Protein isolation from Cortical Organoids. Organoids were collected at specified durations in culture (expressed as Days In Vitro (DIV)) and washed twice with ice-cold PBS. The samples were resuspended in ice-cold NP-40 Buffer (150mM NaCl, 1% NP-40, 50mM Tris pH8) containing EDTA free protease inhibitors (complete cocktail, Roche) and lysed using a 1ml tissue homogenizer (Fisher). Each sample was centrifuged at 10,000rpm for 10 minutes at 4˚C and the homogenates were stored at -80˚C. Protein concentration was determined using the bicinchoninic acid method (BSC, Pierce).

Detection of fibrillary material from organoids by transmission electron microscopy
(TEM). Organoids were lysed following the same procedure for protein extraction, however, samples were initially spun at 20,000g for 20 minutes at 4˚C. Following the first centrifugation, supernatants were removed and kept on ice. The remaining cell pellets were resuspended in 5x weight/volume buffer (10mM Tris-HCL pH7.5, 0.8M NaCl and 10% sucrose) 58 containing proteases inhibitor and spun at 20,000g for 20 minutes at 4˚C. An equal volume of supernatant 1 was added to the supernatant from the second centrifugation step. 1% N-lauroysarcosinate (weight/volume) was added and the samples were rocked at room temperature for one hour. The samples were ultra-centrifuged at 100,000g for one hour at 4˚C. The supernatant was decanted and the sarkosyl-insoluble pellet was resuspended in ice cold PBS prior to imaging. The samples were deposited on to glow-discharged 400 mesh formvar/carbon film-coated copper grids.
Negatively stained with a 2% aqueous (w/v) uranyl acetate solution and then immediately analysed at 100 kV using a JEOL TEM1010 equipped with a Gatan Orius camera.
TEM analysis of synthetic Aβ1-40 fibrils in vitro. Synthetic Aβ peptide powder (China peptides) was treated with 1,1,1,3,3,3-hexafluoro-2-propanol (HFIP) and lyophilized. The peptide was then dissolved in 20µL of 100 mM NaOH and then diluted with buffer. A 50µM stock of this monomeric Aβ peptide was grown at 37˚C shaking at 180 rpm for 48-60 hours before recording the TEM images. 4µL of extract was added to a 15 nm thick, lacey carbon on 300 mesh grid (glow-discharged) for 2 minutes followed by negative staining with 2% uranyl acetate for 1 minute and then air dried. The grids were then viewed under FEI T12, 120 kV Transmission electron microscope equipped with a 4K CCD camera (FEI) at 30000X magnification under low dose conditions.

Data availability
All data that support the findings described in this study are available within the manuscript and the related supplementary information, and from the corresponding authors upon reasonable       and after digestion with HpyCH4IV(cut), for the initial clone 2.5, and its colony-purified sub-clone 2.3.5 (renamed further below as "Δ7"). The 294bp fragment in 2.3.5 is reduced to 65% of the wt value (normalized to the 439bp band), and a de novo 255 bp fragment appears in CRSPR targeted line (red asterisk). c Western blot stained with anti-BACE2 antibody of the lysates of the iPSC line Δ7 compared to the wt T21C5 iPSC line.
Quantification of the total actin-normalised BACE2 signal showed a significant reduction in Δ7 compared to   Tau. β-actin was used as a loading control. Human brain tissue of a 75 year old is shown for comparison.
Comparison of the average values (n=4) for CRISPR-edited T21C5Δ7 showed a highly significant relative increase in TG3 compared to unedited (n=4) T21C5 organoids, as indicated in the graph, p=0.0127. Scale bar: 5μm.