Experimental modelling of Alzheimer’s disease for therapeutic screening.

Neurodegenerative diseases, including Alzheimer’s disease (AD), pose a significant and urgent challenge to healthcare systems worldwide. With an increasing life expectancy, these progressive age-related disorders are expected to rise exponentially. No cure currently exists for AD, and the aetiology remains poorly understood. Furthermore, AD drug development faces one of the highest failure rates. Thus, a review of the experimental modelling of the disease is crucial to understanding how the current disease models can be applied to gain useful results while also considering their limitations. Disease models include in vitro , in vivo , ex vivo , and in silico systems as well as clinical trials. These systems are important for testing potential therapeutics to advance drug development, in addition to modelling the pathology of the disease to gain a greater understanding of the cause and progression. This review will discuss the current experimental models employed for the study of AD with the aim of providing an overview of how they are used and discuss their benefits and drawbacks as model systems, as well as highlighting the potential future of the experimental modelling of AD.


Introduction
Experimental modelling of Alzheimer's disease (AD) has been crucial to the development of current knowledge on the pathogenesis of the disease, and in the testing of potential treatments. At present, numerous models of AD exist to simulate the pathological alterations associated with the disease in humans including cell, animal, and computational models 1 . While these experimental models continue to be useful in AD research, none are able to replicate the complete pathophysiology of AD and as a result, there has been considerable doubt cast over the reliability of the results obtained through the use of these models. Development of experimental models that better mimic the complexity of AD in humans continues. This review aims to summarise the experimental models employed for AD at present, and discuss their role in the drug development process by providing examples of therapeutics that have been studied in each model. This highlights the ways in which to best utilise these models to obtain appropriate and reliable insight into the potential of screened therapeutics while acknowledging the limitations of each model. Alzheimer's disease Alzheimer's disease is a major cause of death worldwide, with over 50 million people suffering from this debilitating neurodegenerative disease 2 . Alois Alzheimer first described the disease in 1906 and noted the characteristic senile plaques and neurofibrillary tangles in patients' brains that continue to be synonymous with the disorder today 3 . As one of the most prevalent causes of death and the most common cause of dementia, AD is accountable for a vast social and economic burden. AD is age-related, and causes increasingly incapacitating symptoms as the disease advances including significant memory loss, confusion, language disturbances, and behavioural changes.
Despite its exponential prevalence in correlation with the rising global life expectancy and its devastating effects, AD remains incurable and drug development faces one of the highest failure rates in any therapeutic area. Only four drugs are clinically available for the treatment of AD in the UK and these drugs aim to mitigate symptoms only, with no disease-modifying effects. This disappointing situation did not change for almost two decades from 2003 when memantine was approved 4 (Figure 1). Since then, only around 50 drug candidates have passed Phase II trials and only one has succeeded Phase III. With so little progress in this area despite extensive research, it is crucial to review the drug discovery process for AD. At the core of the issue is the lack of understanding regarding the exact origin of AD. However, key hallmarks of the disease have been the subject of significant research, as well as the ways in which these hallmarks can be modelled for experimental therapeutic screening.

Figure 1
Key dates in AD drug development. The years listed are correct for the US Food and Drug Administration (FDA). At present, aducanumab has not been approved by the European Medicines Agency (EMA). 3 Development of AD Numerous hallmarks of AD have been identified since the first discovery of the disease, and these have been applied as targets for the development of AD therapeutics. Amyloid plaques and neurofibrillary tangles were the initial hallmarks of the disease, and remain the major targets for AD drug development. More recently, inflammation has also emerged as a key feature of AD pathology ( Figure 2). J o u r n a l P r e -p r o o f 4 Figure 2 Key hallmarks of AD: intraneuronal tau neurofibrillary tangles, extracellular amyloid plaques, and activated microglial cells. Activated microglia generate a rise in the production of pro-inflammatory cytokines and ROS, resulting in neuroinflammation.

Amyloid Hypothesis
The amyloid hypothesis has dominated AD research for the past two decades. This hypothesis postulates that the abnormal deposition of beta-amyloid (Aβ) proteins extracellularly in the brain is responsible for initiating the cascade of pathological alterations associated with AD 5 . The insoluble amyloid aggregates, generated via the proteolytic cleavage of amyloid precursor protein (APP), deposit around the neurons. Aβ aggregates induce neurotoxicity, however the exact relationship between amyloid deposition and the development of AD is still not fully understood.
Recently, the amyloid hypothesis has faced increasing controversy due to the lack of success in clinical trials of drugs that are aimed at counteracting amyloid aggregation 6 . While the drugs are reported to reduce plaque formation in vitro and in vivo, there has been a lack of positive results in patients in terms of improving cognitive function. Within the last few years, drugs which target the soluble neurotoxic amyloid oligomers rather than the plaques have demonstrated greater clinical efficacy. Evidence suggests that the oligomeric amyloid species may play a key role in triggering AD pathology 6 . Despite the previous failure of anti-amyloid treatments, significant evidence persists to demonstrate the clear importance of amyloid aggregation in AD pathology. Human biomarker studies have shown that plaque formation precedes other ADassociated changes including hyperphosphorylated tau deposition, neuron loss, J o u r n a l P r e -p r o o f 5 and cognitive decline 7 . Furthermore, familial AD (FAD) which is the hereditary form of the disease, responsible for a minority of AD cases, is associated with mutations in PSEN1, PSEN2, and APP genes which are all linked to the formation of abnormal amyloid plaque formation 8 . People with Down's syndrome also exhibit a genetic defect which is associated with a build-up of amyloid plaques, and consequently these individuals are at a greater risk of developing AD 9 . Carriers of the ApoE4 allele are pre-disposed to the development of the more common form of AD, sporadic or late-onset, as this allele reduces the rate of amyloid clearance in the brain which leads to a build-up of excess Aβ proteins 10 . Overall, it is clear that amyloid aggregation is a key marker of AD even at early stages in the development of the disease. Therefore, it remains an important target of AD therapeutics and a vital hallmark to replicate in experimental models of the disease.

Tau Hypothesis
Hyperphosphorylated tau fibrils aggregate as intraneuronal neurofibrillary tangles in the brains of AD patients. In healthy brains, tau is a phosphoprotein that promotes the assembly of tubulin into microtubules and stabilises this structure. Normal tau is highly soluble, whereas tau oligomers formed by hyperphosphorylated tau are insoluble and can self-assemble into neurofibrillary tangles 11 . Hyperphosphorylated tau is associated with numerous neurodegenerative diseases including Pick disease, dementia pugilistica, and fronto-temporal dementia with Parkinsonism linked to chromosome 17 12 . In such tauopathies, the presence of abnormally hyperphosphorylated tau in the neocortex is linked to dementia. The level of total tau in AD brains is four to eightfold greater than in normal aged brains, and this rise is exclusively in the form of aberrantly hyperphosphorylated tau 13 . Tau in the form of neurofibrillary tangles does not function as typical tau proteins in healthy brains, and appears to be inert. However, hyperphosphorylated tau occurring in the cytosol and not polymerised into tangles can induce toxic effects by inhibiting the assembly of tubulin and disrupting microtubule structures.
It can aggregate with normal tau into oligomers and consequently self-assemble into tangles, and it can also sequester other microtubule-associated proteins into amorphous aggregates 14 . It has been postulated that this disruption of microtubules and sequestering of microtubuleassociated proteins by the cytosolic hyperphosphorylated tau is the trigger for neurodegeneration and cognitive decline, and the aggregation of hyperphosphorylated tau into neurofibrillary tangles is likely a self-defence mechanism induced by the affected neuron 15 . As a result, inhibiting aberrant tau hyperphosphorylation is a key therapeutic route for the treatment of AD. Furthermore, accurately modelling this tauopathy is important for screening potential AD drugs.

Inflammation
While the former two hallmarks of AD have been well-established since the discovery of the disease by Alois Alzheimer 3 , a third key feature of AD has emerged J o u r n a l P r e -p r o o f 6 within the last two decades [16][17] . The brains of AD patients have been found to exhibit chronic inflammation due to a sustained immune response. The presence of elevated markers of inflammation is not exclusive to AD, and is now associated with numerous neurodegenerative diseases including Parkinson's (PD), multiple sclerosis (MS), and amyotrophic lateral sclerosis (ALS) 18 . Neuroinflammation is initially caused by neuronal loss and other AD pathologies as an acute neuroprotective response, however this becomes detrimental and exacerbates the severity of the disease as the immune response persists. As depicted in Figure 2, activated microglia disrupt the equilibrium of anti-inflammatory and proinflammatory signalling towards the latter and release a variety of toxic products, including numerous cytokines (e.g. interleukins, tumour necrosis factors) and reactive oxygen species (ROS) 19 . Chronic neuroinflammation is attributed to the exacerbation of amyloid and tau pathologies. Reactive microgliosis, whereby there is sustained activation of microglia as part of the inflammatory response, stimulates amyloid aggregation and chronically produces pro-inflammatory cytokines which damage neurons 20 . Cytokines, in particular interleukin-6, reportedly stimulate the hyperphosphorylation of tau by activating protein kinases (namely, CDK5) 21 . Furthermore, interleukin-1 enhances acetylcholinesterase (AChE) expression and activity which results in cholinergic dysfunction and the loss of cholinergic neurons 22 . Overall, the importance of inflammation in AD is evident and further study of its role in AD models is crucial to the development of anti-inflammatory therapeutics which have the potential to slow or delay the progression of the disease. 4 Drug discovery process The drug discovery process for AD is a time-consuming, arduous, and costly procedure ( Figure 3). It encompasses several stages including research and development, preclinical studies, clinical trials, and a final review and approval by the regulatory body: the Food and Drug Administration (FDA) in USA, and the European Medicines Agency (EMA) in European Union. Each stage also involves numerous steps and processes to focus in on the lead that will be optimised and taken to clinical trials, from the initial vast library of compounds. Following the identification and optimisation of a lead, the compound must undergo preclinical studies including in vitro and in vivo models, as well as toxicity studies 23 . The subsequent clinical trials will be discussed further in section 12, but notably this stage poses the greatest hurdle in the drug development process with the highest cost both financially and in terms of duration. The failure rate for diseasemodifying AD therapeutics in clinical trials is currently 100%, and the number of agents reaching clinical trials for the treatment of AD is around 97% lower than that for cancer 24 . This striking disparity is largely attributed to the higher success rate of cancer trials, which thereby attracts more funding and subsequently leads to the development of further therapies. Finally, following the clinical trials, successful drugs are passed to the appropriate regulatory body to be approved. This process includes the review of evidence substantiating the drug's safety and efficacy 23 . When a drug is approved, it can then be manufactured and prescribed 7 to patients. However, the regulatory body continues to monitor the product's safety in the marketplace.

Figure 3
Drug development process from research and development to clinical trials and final review. The typical total duration and cost associated with the process is 9-16 years and around $2 billion.

Natural and synthetic compounds as AD therapeutics
Traditionally, synthetic single-target therapeutics have been designed and implemented for the treatment of neurodegenerative diseases. This includes small molecule inhibitors against targets such as cholinesterases and amyloid aggregation. With the advance of computational simulations and in silico studies, synthetic drug design has become progressively simpler. Predictions on pharmacokinetic and pharmacodynamic properties can be made rapidly and with increasing accuracy 25 . Large libraries of compounds can be narrowed down to a manageable number of structures with promising activity, which are then synthesised and evaluated. This can save immense costs and time. Since the beginning of the century, a multi-target approach to drug development has gained attention due to the lack of disease-modifying effects observed with the administration of single-target therapeutics in patients [26][27][28][29] .
This approach involves the generation of synthetic hybrid compounds with the capacity to counteract multiple targets of complex diseases such as AD simultaneously. This effect is expected to slow or prevent the progression of the disease, and such multi-target agents have demonstrated promising results in experimental models. However, no therapeutics of this type for AD have passed clinical trials so far.
Semi-synthetic drugs, or synthetic drugs based on natural scaffolds, constitute the majority of clinically approved AD therapeutics: donepezil (selective AChE inhibitor), rivastigmine (non-selective cholinesterase inhibitor), and memantine (NMDA receptor antagonist) 30 . Galantamine is the exception to the other semisynthetic approved AD drugs. It is derived from plants, specifically from the Amaryllidaceae family 30 . Natural products have become increasingly popular due to the widely held conception that 'natural' means safe. While there are some J o u r n a l P r e -p r o o f 8 reports of fewer side effects, natural agents can still induce toxic effects 31 . Furthermore, the conversion of natural products into therapies faces several challenges including difficulty isolating the active agent(s), limited efficacy, and poor bioavailability.
Nevertheless, animal and plant-based products have exhibited potential as therapeutics including multi-target activity and synergistic effects between active agents within an extract 31 .

6
Current AD therapeutics At present, only four drugs are clinically available for the treatment of AD in the UK ( Figure 4). Of these, three are acetylcholinesterase (AChE) inhibitors while the other is an antagonist of the N-methyl-D-aspartate receptor (NMDAR) 32 . AChE inhibitors, including donepezil, rivastigmine, and galantamine, are typically prescribed for mild to moderate AD cases whereas the NMDAR antagonist, memantine, is for severe cases. The AChE inhibitors have differing modes of action, but with the same core aim of preventing cognitive decline associated with the loss of cholinergic neurons. While donepezil and rivastigmine function to prevent the degradation of acetylcholine (ACh, a neurotransmitter) by inhibiting the activity of AChE, galantamine exerts a similar effect via an alternative mechanism by inducing increased levels of ACh through the stimulation of preand post-synaptic nicotinic receptors 33 . Memantine interacts with NMDARs to block the effects of glutamate, a neurotransmitter which exerts excessive stimulation on neurons causing excitotoxicity and preventing normal neurotransmission 34 . Although these drugs alleviate symptoms of AD, they are unable to slow or prevent the progression of the disease. AChE inhibitors are also associated with adverse gastrointestinal effects 33 . Therefore, the need for a disease-modifying or curative agent for this disease remains incessantly urgent. Although numerous agents are entered into clinical trials every year, not since 2003 has there been a novel drug approved for the treatment of AD in the UK 4 . The disappointing failure rate in AD trials has brought about a shift in research focus, namely the development of drugs with alternative targets to the typical anti-amyloid agents 35 . As the amyloid hypothesis has been challenged in recent years due to a lack of positive results in human testing, the number of agents entering clinical trials targeting tau and inflammation have increased. Furthermore, combination therapies and multi-target drugs have also gained attention 35 . This approach has been driven by the fact that modulation of a single target of complex, multifactorial diseases such as AD is not sufficient to yield the desired disease-modifying efficacy. Nevertheless, anti-amyloid agents constitute the majority of AD drugs in clinical trials ( Figure 5). However, these trials are now directed at patients in early or preclinical stages of AD. A potential justification for the high failure rates of AD drugs, in particular the anti-amyloid agents, is that the patients recruited for trials were often in late-stages of the disease with symptoms so severe that any disease-modifying effects would be unlikely 36 . With the recent approval of aducanumab by the FDA (discussed further in section 11), there is renewed hope in the field particularly for agents which can target the neurotoxic, soluble oligomeric form of amyloid 6 .
With the above considerations in mind, the drug development process has continued. At present, there are around 70 AD drugs in clinical trials (based on a clinicaltrials.gov search of drug trials that are currently active). There are 11 agents in Phase 1 trials, 43 in Phase 2, and 13 in Phase 3. Figure 5 below displays the major targets of the agents in each phase of the trials. Most of the agents were small molecule therapeutics (59%), followed by antibodies (26%). The remaining drugs were combination therapeutics or DNA/RNA based (5% each), and supplement/dietary (3%) or hormones (2%).

Figure 5
Targets of AD drugs currently in clinical trials. These data were taken from clinicaltrials.gov, and the search was focused on drugs in clinical trials for Alzheimer's disease that are currently active ('active, not recruiting').

7
Experimental models of AD Experimental models are critical for elucidating the fundamental mechanisms underlying AD, as well as evaluating novel therapeutics. Typically, in vitro and in vivo models (e.g. cell and rodent models respectively) are employed prior to clinical trials on human patients 37 . Ex vivo models (e.g. rodent brain slices) and, more recently, in silico models (e.g. virtual ligand screening) have also been developed to further aid in modelling AD.

General advantages and disadvantages of current AD models
Experimental models are vital for toxicity studies prior to human trials. Currently, a minimum of 2 mammalian species are required for preclinical toxicity studies. Any toxic effects would typically be established in initial in vitro and in vivo studies, and attempts would be made to reduce these adverse effects prior to mammalian and subsequent human testing 28 . These studies provide important information about projected safe and tolerable dosage ranges, as well as the pharmacokinetic profile of the drug. As mentioned above, experimental models are also useful for deciphering AD pathology by simulating the changes observed in humans during the disease. In vivo models can provide information regarding the complex pathogenesis of AD and reproduce the progressive nature of the disease as seen in patients. In in vitro models, in-depth cellular studies can be performed to establish the mechanisms that generate the hallmarks of AD 37 . Unfortunately, none of the current experimental models can reproduce the complexity of the disease as observed in human patients. Poor translation of positive preclinical results to patient trial outcomes has been attributed to the lack of accurate disease modelling 37 . Therefore, research is ongoing to produce an experimental model that can better represent AD development. Furthermore, it is increasingly common practice to employ several AD models in preclinical studies that replicate different features of the disease to achieve a more reliable indication of the potential effects in humans. Although an accurate representation of the human condition during AD remains unavailable at present, the importance of experimental modelling is indisputable in terms of advancing current knowledge of AD development and testing novel therapies. 8

In vitro models of AD
In vitro models of AD allow the study of pathological changes at a cellular level (Table 1). These models have the advantage of strictly controlled environmental conditions, in addition to lower costs and simpler maintenance and handling compared to in vivo models 37 . Studies can also be carried out with shorter timescales, and preliminary efficacy and pharmacodynamic experiments can be performed on these cell models 38 . Although initial toxicity studies can be carried out, these models cannot provide reliable pharmacokinetic data due to their simplicity. For the purposes of this report, in vitro models will include 2D and 3D cell culture and induced pluripotent stem cells (iPSCs) while tissue models and primary cultures will be discussed later as ex vivo models. HBMEC -human brain microvascular endothelial cell BCEC -brain capillary endothelial cell RBE4 -rat brain endothelial cell SH- Both studies tested the effects of the natural extracts on Aβ1-42-induced cytotoxicity, in addition to LDH release and MDA. While the differentiated cells were pre-treated for 12 hours with extracts followed by exposure to 100 μM Aβ1-42, the undifferentiated cells were treated with extracts in the presence of 0.5 μM Aβ1-42 for 24 hours. A reduction in cell viability to 35% was reported in differentiated PC-12 cells, whereas a viability of 63% was reported for the undifferentiated cells at their respective stressor concentrations and conditions. In both cases, an increase of around 150% LDH leakage was reported. However, for MDA levels, an increase of around 25% was observed in undifferentiated cells while the differentiated cells only demonstrated a 10% increase. Tong et al 55 tested the therapeutic effects of shikonin, isolated from the traditional Chinese herb Lithospermum erythrorhizon, which is used for wound healing and various allergic conditions. Yang et al 47 investigated the neuroprotective effects of various phenylethanoid glycosides derived from Herba Cistanche -a traditional Chinese herbal medicine for treating kidney disorders.
These natural products demonstrated antioxidant and anti-apoptotic properties as well as significant neuroprotective effects.
Natural products, including traditional Chinese herbal medicines, have also been tested for their potential as AD therapeutics in SH-SY5Y cells. Chang and Teng 45 tested β-asarone, a major component of Acorus tatarinowii Schott, in undifferentiated SH-SY5Y cells stressed with Aβ25-35. The authors found that βasarone was able to protect against inflammation and autophagy induced by Aβ25- 35. A similar methodology was employed by Li et al 56 for testing a different natural product -trichostatin A, which is produced by Streptomyces hygroscopicus. Trichostatin A is an established reversible inhibitor of histone deacetylases, and demonstrated antioxidant and anti-autophagy activity in an undifferentiated SH-SH5Y model stressed with Aβ25-35. Like PC-12, SH-SY5Y cells are also regularly used in a differentiated form. Krishtal et al 57 compared the effects of Aβ1-42 on undifferentiated and RA/BDNF-differentiated SH-SY5Y cells (retinoic acid with brain-derived neurotrophic factor). The authors reported that undifferentiated cells cannot be used as a reliable model for the toxic effects of native Aβ since they exhibited a low sensitivity to Aβ1-42. However, only 48-hour and 72-hour timepoints were tested. At 48-hours, the viability of undifferentiated cells decreased to 84% yet there was no significant reduction in viability at 72-hours. However, this contrasted with the results seen for differentiated cells where no significant decrease was observed at 48-hours, but viability was significantly reduced to 57% at 72-hours. On the other hand, in a subsequent publication 58 , the authors reported that the same conditions resulted in a viability of 57% at 48hours in undifferentiated cells rather than the previously reported 84%. Various inducers of differentiation have been employed for SH-SY5Y experiments. In Krishtal et al 58 , undifferentiated cells were compared to cells differentiated with N(6),2'-O-dibutyryladenosine 3':5' cyclic monophosphate (dbcAMP), retinoic acid (RA) with brain-derived neurotrophic factor (BDNF), and RA with tetradecanoylphorbol acetate (TPA).
The authors observed that dbcAMPdifferentiated cells had a significantly increased susceptibility to the toxic effects of Aβ1-42. Cells treated with RA/BDNF were also more sensitive to Aβ1-42, but only at lower concentrations (10 μM Aβ1-42). In contrast, RA/TPA-differentiation induced a high resistance to Aβ1-42-induced neurotoxicity. Different differentiation-inducing agents also result in various phenotypes. For example, dbcAMP treatment induces a noradrenergic phenotype 58 , and RA with TPA stimulates a dopaminergic phenotype, while RA alone induces a cholinergic phenotype and RA with BDNF further enhances the cholinergic markers 44 . Therefore, it is vital to consider the desired phenotype for studies, and whether it is more appropriate to employ an undifferentiated or differentiated cell line as part of the experimental design.
Similarly, iPSCs can be employed in an undifferentiated or differentiated form. iPSCs are artificial stem cells derived from somatic cells, and can be used to generate any specialised cell type. The major benefit of iPSCs is that comparisons can be made between healthy and diseased human patients, leading to the potential for personalised medicine. Furthermore, these comparisons can aid in the identification of disease-associated markers which can enhance current knowledge on the pathogenesis of AD while revealing novel potential therapeutic targets. Li et al 52 generated iPSCs from cells isolated from the blood of an AD patient with a presenilin 1 mutation and from a cognitively normal individual, and they observed that both Aβ and p-Tau levels were elevated by over 2-fold in the diseased iPSC-derived neurons compared to the control. Upon treatment with the BACE-1 inhibitor, LY-2886721, the levels of Aβ and p-Tau were significantly reduced. Once isolated from the patients, iPSCs can be differentiated into the cell type of interest. Li et al 52  The authors demonstrated that this 3D system could be used to complement 2D in vitro studies for testing the therapeutic effects of potential drugs, while allowing the evaluation of their mechanism of action at a cellular level.

9
In vivo models of AD Both transgenic and non-transgenic animal models of AD have been developed to simulate the pathological changes associated with the human disease. Most commonly, mammalian models such as mice and rats are employed for AD studies, however the use of non-mammals including C. elegans (Caenorhabditis elegans) and fruit flies (Drosophila melanogaster) is advantageous as they are subject to less stringent ethical standards and incur lower costs 1 ( Table 2). In general, animal models allow in depth studies into AD pathogenesis, and can reproduce the major disease hallmarks 37 . Animal models are also crucial for safety assessments of novel therapeutics as their complex systems provide a better reflection of human pharmacokinetics and therefore improved toxicity predictability, in comparison to cell models. However, the complexity of animal models results in a lack of control on experimental conditions 38 . Furthermore, transgenic models are limited in their ability to accurately reflect the human condition as sporadic AD cases are associated with age rather than genetic mutations 61 . With compounding evidence pointing to the multifactorial nature of AD, disease models with only a single, often artificial, cause are not able to reproduce the complete human pathology. Higher costs and strict ethical standards are also associated with animal models, compared to in vitro models. 9.1 Therapeutics tested in in vivo models of AD C. elegans, Drosophilia fruit flies, and zebrafish are increasingly employed as in vivo models due to their low costs and relative ease of maintenance. AD-like phenotypes are commonly induced in these models via transgenic methods, however they can also be chemically induced. Capatina et al 71 treated zebrafish with scopolamine to stimulate memory impairment and oxidative stress. Pretreatment with an extract of Rosmarinus officinalis reportedly reduced oxidative stress as evidenced through the analysis of oxidative stress and lipid peroxidation markers (superoxide dismutase, catalase, glutathione peroxidase, malondialdehyde). Levels of acetylcholinesterase were also found to be regulated following treatment with the extract. Spatial memory in the zebrafish was assessed using the Y-maze, with a significant improvement observed in locomotion pattern and memory in the Rosmarinus officinalis extract-treated animals. Yuen et al 72 employed a more common transgenic model of C. elegans which expresses human Aβ1-42 in body-wall muscle cells causing paralysis. Danshen, a traditional Chinese medicine obtained from Salvia miltiorrhiza, was able to reduce the toxicity of Aβ1-42 in the nematodes as demonstrated by the delay of the onset of paralysis in treated worms. However, no significant reduction in Aβ1-42 levels was detected although the extract was able to prevent Aβ1-42 aggregation in vitro. Treatment with Danshen extract was shown to significantly reduce ROS levels, therefore the authors postulated that the delay in paralysis in treated worms may be due to the protection against Aβ1-42-induced toxicity via ROS inhibition.
In general, rodents (i.e. mouse and rat) are the most popular in vivo models for AD. These models are employed most frequently due to their relatively low maintenance costs, and ease of genetic manipulation and breeding. Furthermore, the nervous systems of rodents are similar to that of humans and their behaviours are complex which allows for the study of AD-relevant cognitive impairment in these models. An AD-like disease state can be induced in these models using a variety of methods; including transgenic, chemically/mechanically induced, and spontaneous. However, no one model can completely emulate the complex pathology of human AD. Table 3 provides a summary of the common rodent models for AD, with the corresponding phenotype of each model relevant to AD. Aβ plaques, tau tangles, inflammation, cognitive impairment, learning and memory deficits [86] J o u r n a l P r e -p r o o f Numerous experiments are performed in rodents to assess the disease state of the animal, the phenotypic relevance to human AD, and for therapeutic screening. Behavioural tests which study the cognitive function of the rodents are commonly employed for AD experiments, as these are relevant to the major AD symptom of memory impairment. Table 4 below lists the common behavioural tests employed in AD studies with rodent models. Table 4 Common rodent behavioural tests for AD.

Contextual memory
Fear conditioning Reference memory, hippocampaldependent associative learning Animal is exposed to aversive stimulus (mild shock) associated with a conditioned stimulus (tone). Freezing response associated with tone alone is measured [87] Passiveavoidance learning Reference memory, associative learning Animal learns to avoid mild aversive stimulus associated with entering desired compartment (darkness) [88] Spatial memory

Morris water maze
Reference memory, working memory, hippocampal spatial memory Animal must find stable platform in circular pool based on prior learned visual clues [89] Radial arm (water) maze Reference memory, working memory, spatial memory Animal placed in maze with several arms radiating from central platform and must guide themselves towards food reward (in water, maze is submerged and escape platform used in place of food reward) [90] Barnes maze Reference memory, working memory Animal placed on circular platform with several holes around circumference and must find escape box accessed through one of the holes [91] Working memory Y-maze/ T-maze Reference memory, working memory Animal placed in 3-arm maze and alternations (explorations of each arm) are recorded [92] Object recognition

Learning and recognition memory
Animal given different objects to explore then positions of objects are changed and some novel objects introduced to test recognition [93] Two widely used mouse models are the 3xTg and the 5xFAD models, which develop Aβ plaques at 6 months and 2 months respectively 37 . The 3xTg model exhibits behavioural symptoms at 4 months, while the corresponding age for 5xFAD mice is 2-4 months. The 3xTg mice overexpress transgenic APP and tau and display a progressive onset of symptoms, but the 5xFAD model overexpresses transgenic APP and develops a significantly more severe and rapid-onset disease with severe amyloid pathology 94 . Therefore, the 3xTg model is considered a more appropriate model of the age-related sporadic AD (SAD) while the 5xFAD mice are used to model familial AD (FAD). Esquerda-Canals et al 95 Table 4 for their study of 3xTg mice treated with an anti-Aβ antibody, including the Morris water maze and the object recognition test. In the Morris water maze, an improvement in spatial memory was observed in treated mice and an improvement in recognition memory was evident in the object recognition test, however no significant improvements were detected in exploratory behaviour or anxiety. The authors attributed this to a reduced clearance of Aβ in the amygdala compared to other brain regions. Despite a reduction in Aβ compared to the untreated mice, the amygdala remained the most affected region with Aβ following treatment which could explain the amelioration of hippocampal-dependent tasks but not those associated with the amygdala.
Based on the multi-target approach for drug development, Kupershmidt et al 96 generated M30 with the active group from rasagiline (a monoamine oxidase B (MAO-B) inhibitor) and an antioxidant-iron chelator moiety. The authors previously reported improved cognition following M30 treatment in the APP/PS1 model. In this study, Kupershmidt et al 96 employed an aging mouse model. An improvement in recognition memory was observed in M30-treated mice using the object recognition test with an increase in recognition index of around 2.5-fold. Furthermore, M30 could reportedly reduce cortical iron levels and Aβ deposition as well as inhibit MAO-B activity in aged mice by around 37%.
Following the above transgenic and spontaneous examples, an alternative method for inducing an AD phenotype in rodent models is through chemical or mechanical administration. Chemically-induced AD models are particularly common, and a variety of chemicals are available for this purpose. Aluminium chloride (AlCl3) is commonly employed as it induces an AD-like phenotype with cognitive impairments and increased acetylcholinesterase activity. Khalaf et al 97 100 , with an increase in escape latency of around 3-fold and 1.4-fold respectively. Zhang et al 100 also observed a negative effect on recognition memory in the object recognition test. Metformin reportedly improved spatial memory but had no effect on motor function, while silver nanoparticles prevented deficits in spatial and recognition memory. 10 Ex vivo models of AD Ex vivo models can combine the advantages of both in vitro and in vivo systems, through the direct investigation of intact affected tissues with the ability to control the extracellular environment 101 . Most commonly, primary cell and tissue cultures and brain slices are employed as ex vivo models taken from genetically modified AD rodents 102  By using primary cell culture rather than the animal model, detailed studies into cell physiology and effects of drug treatments can be carried out at a cellular level. As a result, Beggiato et al 103 were able to establish that palmitoylethanolamide (PEA) exerts its protective effects against neurodegeneration by counteracting reactive astrogliosis. Salau et al 104 also tested the neuroprotective effects of a natural product, but used primary tissue culture. Rat brain tissue was harvested, treated with vanillin, and subjected to Fe 2+ -induced neurotoxicity. By using primary tissue, the therapeutic effects of vanillin could be studied in a model which represents in vivo conditions whilst also allowing investigation of the mechanisms of the neuroprotective activity -in this case, vanillin could ameliorate oxidative imbalance and dysregulated metabolic pathways, elevate ATPase activity, and inhibit cholinergic enzymatic activities. Brain slices, for example from mice as reported by Kniewallner et al 102 , can be studied ex vivo to observe effects of stress and/or drug treatments on each cell and tissue type. Kniewallner et al 102 explored the effects of platelets isolated from AD mice on healthy mouse brain slices. They reported previous attempts to generate a similar in vivo model, however infused platelets did not enter the brains of the mice therefore this model was not successful. However, the authors also noted the drawbacks of the ex vivo model -specifically, that the model lacks blood flow and therefore the platelet localisation and adhesion to vessels may not reflect an in vivo condition. Human samples have also been used as ex vivo models; for example, post-mortem brain or tissue samples from AD patients. These samples provide direct insight into the disease pathology, but have limited accessibility 105 . Furthermore, the acquisition of appropriately matched controls can be challenging, and differing handling practices between various sources can affect comparability. As Scholefield et al 105 reported when studying post-mortem brain tissue with ex vivo rat brain tissue, the human samples can be highly variable depending on the methodology used in addition to which brain region is being tested. Platelets and lymphocytes, or induced pluripotent stem cells, have the benefit of ease of accessibility from AD patients 1 . While post-mortem human brain tissue provides the most direct insight into pathological changes, platelets and lymphocytes can allow the investigation of cellular pathological mechanisms and are not susceptible to rapid degradation as with post-mortem tissues.

Clinical trials
As mentioned above in 'Drug discovery process' (section 5), clinical trials for AD therapeutics are expensive, time-consuming, and have a high failure rate. The current design for clinical trials involves three main phases ( Figure 6). First in human (FIH) Phase 1 studies employ a small group of volunteers (20-100 people) to test the safety and dosage of the drug. This study typically lasts a few months, and employs healthy participants. In some cases, including the testing of genotoxic drugs for terminal cancer, volunteers with the condition may be used. The objectives of Phase 1 clinical trials also include pharmacokinetics, i.e. ensuring that the drug can pass to the targeted area and remain in the body for a duration sufficient to exert its effect. Preliminary studies into its therapeutic capacity in humans may also be performed at this stage. Next, Phase 2 studies are carried out with the drugs that have succeeded in Phase 1. This stage is performed on a larger scale (up to several hundreds of volunteers with the condition) and can last around 2 years. The purpose of these studies is to investigate the efficacy and any side effects of the drug, as well as determining an optimum effective dose. Around 80% of drugs do not pass Phase 1 or 2, often due to toxicity or lack of efficacy. For those that do enter Phase 3, long-term studies (around several years) on the safety and efficacy in hundreds to thousands of participants are carried out. Any adverse side effects are monitored, and the effect of the drug is compared to existing treatments. Where a drug is successful in Phase 3, an application is submitted to the regulatory body after which the drug can be marketed. A final Phase 4 may occur at this stage where the long-term safety and efficacy is evaluated in patients who have been prescribed this medication.

Figure 6
Clinical trial process with 3 main phases, in addition to the number of participants typically used, the duration, and the key aims at each stage.
So, why are so many drugs failing AD clinical trials? This issue has been discussed in numerous reviews [106][107][108]27 , which have posed similar potential explanations for these failures. Issues have been identified with drug design as well as clinical trial design which could contribute to the widespread lack of efficacy observed in drugs aimed at AD (Table 5). With regards to drug design, the difficulty in identifying suitable therapeutic targets is attributed to the poor understanding of the mechanisms of the disease. Furthermore, poor drug delivery and penetration as a result of an inability to cross the blood brain barrier (BBB) is a prevalent reason for the failure of drug candidates in clinical trials that have otherwise demonstrated promise in early drug development. In terms of the clinical trial design, concerns have been raised regarding the length of the studies and the variability in clinical endpoints between trials 108 . Due to the progressive nature of the disease, extended durations may be required to detect any disease-modifying effects. Anderson et al 108 employed a clinical trial simulator to show that, even with a study which lasted 5 years, measurement variability between individuals results in difficulty identifying a treatment with 80% efficacy. However, increasing the trial duration would consequently incur higher costs and potentially result in a higher drop-out rate. The use of patients with mild to moderate stage AD has also been questioned, as the progression of the disease may be too advanced by this point for the drugs to be effective 27 . As a result, participants with earlier stages of the disease are being employed in trials and several prevention studies are also being performed. Biomarkers for amyloid and tau are currently used in clinical trials to identify participants at risk of developing AD. However, novel biomarker analyses are being investigated as a technique for monitoring target engagement and drug efficiency. In addition to amyloid and tau, biomarkers for several other AD targets such as oxidative stress and inflammation are increasingly employed as investigational agents during clinical trials 106 . A key case demonstrating an issue with clinical trial design is that of aducanumab. An interim futility analysis of aducanumab (an anti-amyloid monoclonal antibody developed by Biogen) deemed that the drug would not achieve statistically significant clinical effects by the end of the trial based on the data obtained so far 109 . In the 3-month period between the completion of the futility analysis and the announcement of drug futility, trial participants had the opportunity to complete the trial. Upon reanalysis of the data to include participants that continued during this 3-month period, significant clinical effects were found. Subsequently, Biogen applied for FDA approval for aducanumab 107 and it was granted in June 2021 110 . However, this decision has been met with great controversy -in particular, due to the fact that the FDA approved aducanumab against the recommendations of its expert advisory committee which had agreed that there was insufficient evidence of any clinical benefit to approve the drug 111 .  117 . In silico methods are typically used alongside traditional in vitro experiments to validate the results. By using in silico modelling for drug design, predictions can be made about pharmacokinetics as well as target affinity 30 . Therefore, large libraries of potential ligands can be screened to identify leads with the greatest predicted target affinity which can then be synthesised and tested. This saves considerable costs and time as only a selected number of ligands need to be synthesised following the virtual ligand screening. Possible improvements on the ligand structure to optimise affinity for the target can also be recommended using in silico modelling 118 . Based on the size of the active site on the target, side chains can be added or removed to enhance ligand binding interactions. Furthermore, based on the properties of residues within the active site, substituents can be altered on the ligand structure to form interactions with these areas (e.g. whether hydrophobic or hydrophilic substituents would be more appropriate). One example of disease modelling from Anastasio 119 demonstrated that cerebrovascular disease can contribute to amyloid dysregulation and, in turn, the progression of AD. By modelling the various elements which are associated with the amyloid regulatory pathway, it was possible to identify alternative therapeutic targets, and therefore recommend potential treatments. By developing this model further, the authors could make predictions on the response to pharmacological interventions, and were able to demonstrate the potential for oestrogen to significantly reduce amyloid levels. A more recent model from Madrasi et al 120 based on quantitative systems pharmacology (QSR) was developed to rationalise the lack of clinical efficacy of amyloid-modulating therapeutics.
With the growing availability of artificial intelligence (AI) and machine learning, these techniques have recently been applied to AD research in several capacities -for example, determining individual risk of AD, drug development, and in efforts to decipher the cause of AD. AI is capable of processing large datasets and analysing it with a high degree of accuracy. However, this relies significantly on the quality of the data input. A machine learning diagnostic platform to detect AD by analysing retinal images was reported by Wisely et al 121 . Rodriguez et al 122 applied AI to identify potential candidates for repurposing as AD therapeutics by studying differentially expressed genes in relation to disease progression then recommending potential treatments which have an affinity for the identified targets. Despite the clear benefits of AI including rapid processing and low error compared to human methods, these techniques remain extremely costly to implement which currently limits their application and regular use.
The generation of brain organoids from iPSCs is another example of an AD experimental model which is likely to be increasingly employed in the future. Brain organoids allow the study of brain development and the mechanisms of neurological and neurodegenerative disorders, in addition to the screening of therapeutic compounds [123][124] . Furthermore, by employing patient-derived iPSCs for the generation of brain organoids, personalised therapeutic strategies could be developed and novel insights into molecular and genetic disease mechanisms may be revealed 123 . While significant advances have been made in the last decade, a number of challenges exist with the use of brain organoids including the technical difficulty in culturing these models and the lack of reproducibility 124 . Due to the lack of immune and vascular systems, these models can be improved to enhance their physiological relevance 125 . As with current AD experimental models, brain organoids are currently not able to completely simulate the pathological features of the disease. However, with continued research, the brain organoid is a promising preclinical model that has the potential to bridge the translational gap between animal models and clinical trials. 13 Conclusion Experimental models of AD are important for both the advancement of the knowledge on disease pathogenesis as well as the development of novel therapeutics. At present, no experimental model can fully replicate the pathophysiology of human AD. The high failure rate of clinical trials for AD drugs indicates that there is an issue with the current systems for modelling the disease, as the positive results observed in these models often do not translate into clinical benefits. However, by acknowledging the limitations of each model, it is possible to continue gaining useful information on AD. Employing multiple experimental models which mimic various aspects of the disease in preclinical studies can provide a more representative depiction of the human condition. Furthermore, as the current models are adapted and new experimental models are generated, these systems continue to gain translational power and produce more reliable results. With the approval of the first novel therapeutic for AD in two decades, the future of drug development to combat this debilitating disease is increasingly hopeful.