Neurons from individual early Alzheimer’s disease patients reflect their clinical vulnerability

Establishing preclinical models of Alzheimer’s disease that predict clinical outcomes remains a critically important, yet to date not fully realised, goal. Models derived from human cells offer considerable advantages over non-human models, including the potential to reflect some of the inter-individual differences that are apparent in patients. Here we report an approach using induced pluripotent stem cell-derived cortical neurons from people with early symptomatic Alzheimer’s disease where we sought a match between individual disease characteristics in cells with analogous characteristics in the people from whom they were derived. We show that the response to amyloid-β burden in life, as measured by cognitive decline and brain activity levels, varies between individuals and this vulnerability rating correlates with the individual cellular vulnerability to extrinsic amyloid-β in vitro as measured by synapse loss and function. Our findings indicate that patient induced pluripotent stem cell-derived cortical neurons not only present key aspects of Alzheimer’s disease pathology, but also reflect key aspects of the clinical phenotypes of the same patients. Cellular models that reflect an individual’s in-life clinical vulnerability thus represent a tractable method of Alzheimer’s disease modelling using clinical data in combination with cellular phenotypes.


Introduction
Alzheimer's disease is the most common age-related neurodegenerative disease and cause of dementia, estimated to affect close to 50 million people in 2015 worldwide with cases predicted to almost double every 20 years 1

. Autosomal dominant mutations in the Amyloid Precursor
Protein gene (APP) or genes encoding the APP proteolytic enzymes Presenilins 1 and 2 (PSEN1, PSEN2) are causative of early onset familial Alzheimer's disease. Largely based on insights from familial Alzheimer's disease, amyloid-β (Aβ) generation, metabolism or clearance is thought to underlie the pathogenesis of late onset forms of sporadic Alzheimer's disease. However, it is also apparent that whilst amyloid-related features predict clinical outcomes, this relationship shows very considerable inter-individual variation 2 . Some individuals show evidence of extensive amyloid pathology yet little apparent clinical impairment, and others have a relatively low amyloid burden in the context of moderately advanced dementia. Transgenic rodent models utilising human familial Alzheimer's disease gene mutations 3 have been extensively used to model various aspects of APP/Aβ pathobiology but have not proved useful in exploring the mechanisms whereby this pathobiology affects disease pathogenesis and, as a consequence, we have no effective preclinical model of sporadic Alzheimer's disease.
The advent of induced pluripotent stem cell (iPSC) technologies 4 now makes it possible to derive patient-specific cell lines capable of differentiating into various cell types and thereby human cellular models of disease. Although familial Alzheimer's disease iPSC-derived cells exhibit pathological APP-related phenotypes in vitro, sporadic Alzheimer's disease iPSCderived cells typically do not share the same phenotypes [5][6][7] . Recently however, iPSC-derived neurons were shown to display features in vitro that reflect analogous features from post-mortem material from the same individuals 8 . This has provided evidence on the feasibility of using individual cell models of disease to explore pathogenic mechanisms.

Deep and Frequent Phenotyping (DFP) pilot cohort clinical data
The DFP pilot study protocol was previously published 9 , and a subset of the clinical outcomes, namely Mini Mental State Examination (MMSE) scores and global magnetoencephalography (MEG) readout, and participants was used for this iPSC study listed in Table 1. See Supplementary Materials for more details.

Patient iPSC-derived cortical neuronal culture
The patient iPSC lines were reprogrammed from peripheral blood mononuclear cells using Sendai virus. Subsequently, the iPSCs were differentiated into cortical neurons driven by Ngn2 expression and plated in co-culture with primary rat astrocytes except for collecting conditioned media for the quantification of secreted Aβ (iPSC-derived neuronal monoculture).
See Supplementary Materials for more details on the generation of iPSC, differentiation into cortical neurons and quantification of secreted Aβ.

Synapse imaging
Immunocytochemistry was conducted on neurons treated with various extrinsic Aβ insults on Day 80 of the neuronal differentiation. Antibodies against presynaptic Synapsin I/II, postsynaptic HOMER1, and dendritic MAP2 were used. The samples were imaged on an Opera Phenix automated microscope, and the synapses were quantified relative to the total MAP2+ area to derive synaptic density for all downstream analyses. See Supplementary Materials for more details.

Multi-electrode array (MEA) electrophysiology
The iPSCs were seeded directly onto the MEA plates for neuronal differentiation in co-culture with primary rat astrocytes. Baseline activities (2-

Statistical analyses
All statistical analyses were performed in GraphPad Prism 9.2.0. We reported Pearson's coefficient of correlation and two-tailed p-values for correlations by simple linear regression.
Kruskal-Wallis test was used for single-parameter comparisons amongst the patient lines.
Welch's t-test was used for the vulnerable-resilient group comparisons in the MEA experiment.

Data availability
Detailed raw data of the experiments are available from the corresponding authors upon reasonable request. The data from the DFP cohort can be requested via the Dementias Platform UK online portal (https://www.dementiasplatform.uk/research-hub/data-portal).

iPSC lines from a comprehensively-phenotyped cohort of early Alzheimer's disease patients
We set out to ask whether the heterogeneity of Alzheimer's disease patients could be accurately reflected in iPSC models by comparing clinical outcomes in vivo with patient-derived neuronal phenotypes in vitro. We asked specifically whether clinical vulnerability to Aβ burden in the brain can be reflected by Aβ-induced cellular vulnerability in neurons derived from the same patients. In this study, we tapped into the comprehensive clinical datasets of the DFP pilot cohort 9 (Table 1) from which we generated thirteen sporadic Alzheimer's disease iPSC lines and one familial Alzheimer's disease iPSC line (Patient #5) carrying an autosomal dominant APP mutation, to use in our experiments (Supplementary Table 1 and Supplementary Fig. 1).
Previously, the DFP study has highlighted the heterogeneity of disease and also very considerable inter-individual variation in the impact of that amyloid pathology 10 . This suggests a difference in vulnerability or resilience in the face of amyloid pathology that might reflect differences either in the hypothesised amyloid cascade or in factors that interact with that cascade. Here, we seek to investigate if the functional consequences in response to Aβ burden in the brains of Alzheimer's disease patients (instead of the accumulation of Aβ pathology per se 8 ) can be recapitulated in vitro using iPSC models derived from the same patients.

Aβ1-42
To understand if patient-derived iPSC models recapitulate the in-life clinical measures of their donors, we first differentiated all fourteen iPSC lines in parallel into cortical neurons in monoculture ( Supplementary Fig. 2a) and showed that Aβ1-42 levels in the conditioned media correlate significantly and negatively with the same pathological Aβ species in the CSF from the patient donors (Fig. 1a), a characteristic phenomenon of Alzheimer's disease patients thought to be due to the sequestration of Aβ1-42 in non-soluble cortical amyloid plaques 11 . Importantly, this relationship was not found for either Aβ1-38 or Aβ1-40 peptide comparisons and was not affected by the inclusion of the familial Alzheimer's disease line. However, the Aβ1-42 / Aβ1-40, and Aβ1-38 / Aβ1-40 ratios were significantly increased in Patient #5 harbouring an APP mutation compared to the other patient lines, consistent with previous observation from another study 12 . This result provides further evidence that patient-derived neurons reflect the pathological features in vivo of that patient. We next went on to examine patient-specific cellular vulnerability to Aβ in vitro.

Patient iPSC-derived neurons demonstrate a spectrum of synaptic vulnerability to A insults
Dysregulation, and eventually loss, of synapses is one of the earliest pathological phenotypes of Alzheimer's disease and leads to cognitive decline and memory loss 13,14 . Electrophysiology, in particular MEG, is thought to be a surrogate of synaptic dysregulation and loss and hence provides an opportunity to explore whether the individual impact of Alzheimer's disease pathology on synaptic health in people in vivo is reflected in their cells in vitro. We therefore sought to investigate synaptic vulnerability to Aβ insults in vitro; iPSC lines were again differentiated in parallel into cortical neurons, this time plated in co-culture with rat cortical astrocytes ( Supplementary Fig. 2a, 2b). We then treated the neurons with a range of extrinsic Aβ insults listed in Table 2.
All three exogenous A treatments resulted in decreased synaptic density in all patient-derived cortical neurons relative to control treatments. However, the different patient lines showed different levels of impact of Aβ insults on synapse loss, allowing us to rank lines from the most resilient to the most vulnerable ( Fig. 1b and Supplementary Fig. 2c, 3). Notably, cellular vulnerability in the patient carrying the familial Alzheimer's disease APP mutation that generated the most endogenous A1-42 was within the range, but was relatively resilient to the impact of exogenous A insults. All neurons displayed functional activity by firing action potentials on Day 80 of neuronal differentiation ( Supplementary Fig. 2d). The synapse loss datasets demonstrated good reproducibility over three repeat independent iPSC differentiations.
By comparing the extent of synapse loss between differentiation repeats, we confirmed that the specific levels of vulnerability in each line of iPSC-derived neurons in response to Aβ insults remained consistent across all differentiation repeats ( Fig. 1c and Supplementary Fig. 4a, 4c).
Importantly, similar patient line-specific vulnerability measured by synapse loss was also consistent across the different Aβ insults used, especially between Aβ1-42 and Aβ25-35 oligomers where there is a significant and positive correlation ( Supplementary Fig. 4b). A positive correlation was also observed across differentiation repeats when the neurons were treated with Alzheimer's disease brain homogenate ( Supplementary Fig. 4a, 4c). The synapse loss data indicated that the degree of synapse loss due to the exposure to extrinsic Aβ in functional cortical neurons is patient-specific, cell-autonomous, and reproducible across insults and differentiation repeats.

Synaptic vulnerability to Aβ insults in vitro reflects clinical vulnerability to Aβ burden in vivo
Next, we explored if the levels of synaptic vulnerability to Aβ insults in the patient-derived neurons in vitro was a reflection of the individual's response to amyloid in life as measured using electrophysiological measures of synaptic activity and measures of cognitive decline, the ultimate clinical manifestation of synaptic dysfunction. While in the in vitro experiments the cells were exposed to the same amount of A insult, in vivo the individuals showed a range of amyloid burden. Global MEG recordings and cognitive decline measured by MMSE score loss rate (Table 1)  We then selected the three most vulnerable together with the three most resilient patient lines and investigated whether their electrophysiological activities were also differentially affected based on their synaptic vulnerability in vitro. As for the synaptic loss measures, the neurons derived from the most vulnerable patient lines exhibited greater reductions of firing and burst rates caused by the exposure to A1-42 oligomers as compared to the most resilient patient lines (Fig. 2b). The scrambled A1-42 peptide control did not elicit any change in the levels of neuronal activities ( Supplementary Fig. 6). Additionally, the differences in synapse loss in the patient-derived neurons could not be explained by their APOE variants ( Supplementary Fig. 7) nor by the single case of an APP mutation carrier who scored as both relatively resilient to amyloid in vivo and to A in vitro suggesting that the resilience/vulnerability to A is not driven either by the most significant genetic variant associated with sporadic Alzheimer's disease or by mutations in the APP gene itself.
In conclusion, we show here that neurons derived from Alzheimer's disease patients retain the same individual vulnerability to Aβ that the person from whom they were derived, showed using both biomarkers and clinical measures that reflect the synaptic phenotypes measured in vitro.

Discussion
In this study, we demonstrate for the first time that cellular vulnerability to Aβ insults in vitro reflects clinical vulnerability to Aβ burden in vivo, specifically in people living with Alzheimer's disease dementia, by establishing the correlation between synapse loss in individual Alzheimer's disease patient-derived neurons and their clinical outcomes. This was further supported by neurons from the more vulnerable group of patients exhibiting more deleterious response to extrinsic Aβ insults as measured by their levels of neuronal activity as compared to the resilient group. This approach of integrating clinical in-life data with disease modelling in the laboratory presents a tractable method of Alzheimer's disease modelling with iPSCs.
Decline in cognition estimated from time since onset and current cognitive score, and 'brain activity' assessed using MEG were selected as clinical outcomes likely to be reflections of synaptic health and so broadly analogous to the synaptic loss data we measured in vitro. We report here that the amount of cognitive decline as a function of amyloid burden correlates with more severe Aβ-driven synapse loss and loss of synaptic function, as measured using MEA electrophysiology, in the patient-derived neurons. Although it has been known that synapse loss correlates with cognitive decline in Alzheimer's disease 14,15 , and that MEG identifies neurophysiological changes that are specific to Alzheimer's disease, it remains unclear how different brain MEG signals change at different stages of Alzheimer's disease progression 16,17 .
Interestingly, we find a clear correlation between greater brain activity levels measured by MEG correlating with more severe Aβ-driven synapse loss in the patient-derived neurons. This apparently counterintuitive observation is in line with a considerable amount of evidence for hyperexcitability in the early phases of Alzheimer's disease. Neurons exhibit hyperactivation particularly during the mild cognitive impairment stage before hypoactivation as disease progresses 18,19 , and hyperexcitability leading to seizure activity is increased in Alzheimer's disease, perhaps as a function of amyloid related pathology 20 . Our findings substantiate the role of hyperexcitability in early Alzheimer's disease and provide a model with which to explore such therapeutics discovery.
It has recently been shown that several measures of secreted Aβ peptides in iPSC-derived cortical neurons from Alzheimer's disease patients reflect the extent of Aβ neuropathology of their donors 8 . We extend that previous work on post-mortem, end of life, neuropathological findings to in-life, early in disease, clinical measurements by showing that the levels of Aβ1-42 secreted from patient-derived neurons correlate with the levels of the same pathological Aβ species in the patient CSF samples (Fig. 1a). However, we have now shown that not only is there a correlation between cellular phenotypes and analogous phenotypes in post-mortem brain and in patients, but that the functional consequences of those phenotypesthe response to Aβ as well as the amount of Aβare preserved in the cells. Crucially, the inclusion of the familial Alzheimer's disease case did not affect the cellular-clinical correlation in vulnerability to Aβ. The neurons from the familial Alzheimer's disease individual in fact belong to one of the more resilient patient lines in vitro even though this individual has the greatest Aβ burden measured by amyloid PET within this cohort, further reinforcing our interpretation that the iPSC models specifically reflect the vulnerability to Aβ measured by clinical outcomes instead of the levels of Aβ accumulation in the brain.
In conclusion, we reveal that cellular vulnerability reflects clinical vulnerability to Aβ in

Ethics statement
The DFP cohort study was approved by the London Central Research Ethics Committee, 14          Welch's t-test was used for statistical analysis. Box plots (centre line, median; box limits, interquartile range; whiskers, data range; points, all data points) showing the percentage of synapse loss caused by Aβ 1-42 oligomers, Aβ 25-35 oligomers and Alzheimer's disease brain homogenate with patients distinguished by their APOE variant genotypes. n = 12 (e4-), 20 (e3/e4) and 9 (e4/e4) independent neuronal differentiation repeats per patient line.

Supplementary materials and methods
All reagents were purchased from Sigma unless stated otherwise. All iPSC-derived neuronal culture incubated at 37°C and 5% CO2.

Deep and Frequent Phenotyping cohort pilot study and clinical data
The Deep and Frequent Phenotyping (DFP) cohort pilot study protocol was previously  counted per recording length.

Meso Scale Discovery immunoassay of Aβ peptides
The iPSC derived neurons were grown as described previously without the Day 4 passage onto rat astrocytes until Day 40. Cell conditioned media was collected after 48 hours and stored at -80°C. Cells were washed once with PBS, and M-PER™ (Thermo) added for 20 min on ice.
Cell suspension was centrifuged at 14,000 g for 10 min at 4 o C. The supernatant was collected,

Human Alzheimer's disease brain homogenate extraction
The extraction protocol of human Alzheimer's disease brain homogenate was modified from a published method 6  mM NaHCO3, pH = 7.4) with a ratio of 1 g of tissue to 4 ml of aCSF supplemented with a panel of protease inhibitors (5 mM EDTA, 1 mM EGTA, 5 ug/ml leupeptin, 5 µg/ml aprotinin. 2 µg/ml pepstatin, 120 µg/ml Pefabloc and 5 mM NaF). The homogenisation was followed by centrifugation at 200,000 g for 110 min at 4°C and the supernatant was transferred into a Slide-A-Lyzer™ G2 Dialysis Cassettes 2K MWCO in 100 times volume of aCSF without protease inhibitors for 72 h. The aCSF was replaced every 24 h and the resultant aliquots were frozen at -80°C.
The iPSC-derived neurons were incubated with either 25% Alzheimer's disease brain homogenate (1:1 mixture from the two cortices) or aCSF without protease inhibitors as the treatment control in the cell culture medium (v/v) for 72 h at 37°C before paraformaldehyde fixation.

Immunocytochemistry
Adherent neurons were fixed in 4% paraformaldehyde for 5 min, followed by treating with 0.5% saponin in PBS for 20 min for permeabilisation. To block the samples, we treated the plates with 10% normal goat serum with 0.01% tween-20 in PBS for 30 min. Primary antibodies were then left incubating with the samples at 4°C overnight with 1% normal goat serum and 0.01% tween-20, before washing with PBS 3 times. Secondary antibodies were then applied in 1% normal goat serum and 0.01% tween-20 at room temperature for 1 h before washing for another four times. The primary antibodies we used were: Guinea pig anti-SYNAPSIN I/II (Synaptic Systems, 1:500), rabbit anti-HOMER1 (Synaptic Systems, 1:500),

High-content imaging and analysis
Synapse: The 96-well plates were imaged on the Perkin Elmer Opera Phenix high-content imager. We captured 15 images per well with a 43X objective at +1 µm focus level with the binning value of 1. We then analysed the image with the Harmony software v4.9 from Perkin Elmer with a customised pipeline. The MAP2-positive neurites were identified with 0.5 overall threshold as the region of interest and resized by expanding outwards by 5 px to cover synaptic signals which lay slightly above the MAP2 signals. Both presynaptic (SYNAPSIN I/II) and postsynaptic (HOMER1) signals were then identified with Method A of the "Find spots" function with threshold values of 0.17 and 0.14, respectively. We also filtered away the spots which were larger than 100 px 2 . Finally, the synapses were ascertained by finding HOMER1 signals in the vicinity of SYNAPSIN I/II signal regions which had been resized by expanding outwards by 5 px. The absolute number of synapses was then normalised to the total MAP2positive area to derive synaptic density which was used for all downstream analyses. All the values of synaptic density downregulation due to the Aβ extrinsic insults were then normalised to the corresponding treatment controls i.e., Aβ1-42 normalised to scrambled Aβ1-42, Aβ25-35 normalised to Aβ35-25 (reversed), and Alzheimer's disease brain homogenate normalised to aCSF.
Cortical markers: We captured 15 images at -1, 0 and +1 and µm focus levels per well with a 20X objective and binning value of 2. We analysed the images on the same Harmony software by first identifying human nuclei among the co-culture with rat astrocytes and filtering away the nuclei with circularity less than 0.6. The percentage of cortical marker-positive cells was calculated by selecting the human nuclei with cortical marker mean signal intensity greater than a threshold which was determined as the mean intensity across all human nuclei. Finally, we derived relative cortical marker expression by normalising the percentage of cortical marker-positive neurons to the geometric mean across all fourteen patient lines.