Inconsistencies and Controversies Surrounding the Amyloid Hypothesis of Alzheimer's Disease

Using the amino acid sequence corresponding to Aβ [9], the major constituent of amyloid plaques in AD, a precursor gene cDNA to Aβ (the amyloid precursor protei n, APP) was sequenced and mapped to chromosome 21 [12]. This finding had compelling implications in view of the observation many individuals with trisomy 21 (Down’s Syndrome) reach the neuropathogical criteria for AD by age 40 [13]. Such Down’s individuals would have a triple copy of APP, and therefore it was reasoned, excess Aβ production. Since Aβ is the main component of plaques seen in AD, it is presumed in turn that excess Aβ is the cause of AD in Down’s syndrome. Surprisingly, the fact that not all people with Down’s syndrome develop AD, despite plaques and increased Aβ expression, did not receive significant attention [13]. This observation may have quelled consideration of Aβ as the sole risk factor for AD.

Next, studies of familial EOAD uncovered genetic links between the APP gene and AD [14]. APP is processed into smaller peptide fragments, one of which is Aβ, via cleavage by α-, β- and γ-secretases (Figure 2). Importantly, EOAD-linked point mutations were identified not only in APP itself but also in presenilin-1 (PSEN1) and presenilin-2 (PSEN2) [15],[16] the key catalytic subunits of γ-secretase, known to cleave APP (Figure 2). No known AD-causing mutations are present in the gene encoding the β-secretase gene, beta-site APP cleaving enzyme 1 (BACE1).

The genetic mutations are reasoned to cause AD through aberrant processing of APP, leading to either increased levels of Aβ or an increased production of the 42 and 43 amino acid forms of Aβ (Aβ42/Aβ43) over the 40 amino acid form of Aβ (Aβ40). It is argued this triggers aggregation of Aβ [17]. The discovery that transgenic mice expressing familial human APP and PSEN mutations recapitulate many, but not all, of the features of the human disease [18] further established the link between aberrant Aβ production and the AD phenotype. This latter discovery, perhaps more than any other, tied the field to the amyloid hypothesis for the next decades.

The conclusions of the aforementioned studies were grounded in an unquestioned assumption that Aβ, rather than altered expression of APP or its products, causes AD pathology. The assumption arose because Aβ was the key component of plaques and because Aβ caused neurotoxicity in healthy cells [19]. Further, hyperphosphorylation of tau, thought to be downstream of Aβ, was seen as a critical mediator of the neurotoxic effects of Aβ [20] placing Aβ at the top of the pathological chain of AD events. A cycle thus began to develop early whereby studies were designed and then interpreted on the basis of the hypothesis that Aβ caused AD pathology, rather than being critically evaluated in the context of a range of possible interpretations.

Further, given the impact discoveries of mutations in APP, PSEN1 and PSEN2 have had in driving the amyloid theory, it is notable that, while these mutations account for the majority of EOAD cases, EOAD only comprises less than 5% of all AD cases [21]. In fact, the majority of AD cases are sporadic, idiopathic LOAD. It seems in retrospect presumptive to have extrapolated a role for Aβ in all AD based on the genetic evidence suggesting a role for altered APP processing in EOAD.

In general, the risk genes identified for LOAD are subtle, with no direct genetic association to the APP gene or its processing enzymes. The most well-known genetic link to LOAD is the apolipoprotein genotype E4 (APOE4) [22]. Recently another strong risk gene for LOAD was identified, a variant of the triggering receptor expressed on myeloid cells 2 gene (TREM2), implicating excessive innate immunity in Alzheimer’s pathogenesis [23]. Although these two mutations have been the strongest to date, many more have been associated with LOAD. Most of these genetic risk factors have been interpreted through the lens of the amyloid hypothesis, mainly by considering their modulatory effects on Aβ, though other interpretations are equally valid, an issue we discuss further in later sections.

Synapse loss leads to a loss of dendritic mass [24] and, crucially, may precede, and indeed drive, neuron loss in a range of conditions [25]. Deficits in synaptic plasticity are measurable at just one month of age in mouse models of AD [26] and synapse loss is evident during early stages of the human disease [27]. Elegant research has revealed the number of neocortical synapses to be a better correlate of cognition than both Aβ plaques and NFTs [28] and a greater loss of synapses than neurons is evident in human AD brains [29]. These observations place synapses at the forefront of understanding AD pathogenesis. It has therefore been suggested that AD is primarily a synaptic disorder [30],[31]. In a mouse model of AD, synapse loss arises from over-elimination of synapses, rather than a failure of synapse formation [32].

The aforementioned studies of Aβ at synapses are difficult to interpret. Aβ plaque deposition can occur without associated synapse loss [43], and conversely synapse and dendritic tree loss can occur in areas without Aβ deposition, although synapse loss does usually appear exacerbated near Aβ plaques [44]. Thus it would be prudent to treat the suggestion that Aβ plaques have a primary causative role in synapse destruction in human LOAD with caution.

As for Aβ oligomers however, while many reports identifying Aβ oligomers as triggers of synaptic degeneration support the amyloid hypothesis, technical restrictions limit interpretations of these results, and their relevance to the human disease is unclear, as we discuss below. For example, are pathological effects in mice resulting from injection or over-expression of Aβ relevant to the human condition?

Furthermore, synaptic gene dysregulation in early AD can occur independently of alterations in the expression of APP and regulators of APP metabolism [45]. Finally, genetic studies suggesting a role for APP and its processing in familial EOAD may have been incorrectly extrapolated to LOAD. Thus, as with plaques, it is conceivable that Aβ oligomers play a role, but it would also be prudent to treat the suggestion Aβ oligomers play a primary or sole causative role in synapse destruction with a degree of caution.

Later we discuss our view that understanding AD requires first understanding the complex biology of the multicellular synapses [46], the role of glia in synapse removal, and the means by which these cells can be driven towards excess synapse removal and/or destruction.

The Food and Drug Administration (FDA) has over the years approved five drugs for AD; Donepzil, Galantamine, Memantine, Rivastigimine and Tacrine. It is notable that each of these are unrelated to the amyloid hypothesis and were not tested in transgenic AD mice before being used in the clinic [66].

Meanwhile, many anti-amyloid treatments that were tested in mice have completed, or are undergoing, extensive clinical trials in humans. We summarise the most high profile of these drugs in Table 1. They are divided into those directly targeting Aβ by active and passive immunization, those targeting inhibition or modulation of the γ-secretase APP cleaving enzyme (Figure 2), presenilin, and those targeting the APP β-secretase cleavage enzyme BACE1.

So far, anti-Aβ treatments have broadly failed to meet their primary clinical endpoints and some major phase 3 trials were halted early. None of the tested treatments have produced a discernible functional recovery, or altered the course of disease. In fact alarmingly some, specifically inhibitors of γ-secretase, lead to an increased decline in cognition (Table 1). With each successive failure the validity and foundations of the amyloid hypothesis, on which these drugs have been based, is called increasingly into question. Haste to run Phase 3 trials without Phase 2 success, and similar criticisms, have recently been made of this commercially-driven enterprise [67].

Why is the hypothesis failing clinically? Some suggest the disease is not being targeted early enough [68], noting that in animal models anti-Aβ approaches clear hyperphosphorylated tau aggregates when given to young, but not old, animals [69] and, also, detailed analysis of recent trials have shown hints of treatment benefit in individuals treated early in disease [57].

Planned human intervention studies aim to address this issue in two ways. The DIAN [70] and the API Colombia study [71] use anti-Aβ antibody treatments in presymptomatic individuals at risk for familial EOAD. If these trials succeed, the results will provide evidence for a degree of Aβ involvement in EOAD. They will not necessarily prove Aβ causality in all EOAD, nor will they provide information on the role of Aβ in LOAD.

The Anti-Aβ asymptomatic Alzheimer’s disease trial [72] meanwhile tests the effect of starting anti-Aβ treatment at the pre-symptomatic stage, in individuals predicted to develop LOAD on the basis of brain amyloid accumulation as measured by positron emission tomography (PET) imaging. This will effectively test the hypothesis that anti-Aβ treatments provide cognitive benefits when given earlier in sporadic AD.

Another prominent suggested reason for clinical failures of anti-Aβ drugs in particular are that the agents used initially were not properly validated and were flawed [68]. A recent study has shown the monoclonal anti-Aβ antibodies, solanezumab and crenezumab, fail to target human Aβ as effectively as they target over-expressed human Aβ in mouse models [73]. The possibility was also countenanced that only amyloid plaques, potentially functionally inert [74], rather than soluble Aβ oligomers were targeted in early trials. Furthermore monotherapies may not be capable of effectively reducing Aβ plaque load. A double pronged approach to reduce Aβ by both active immunisation and inhibition of β-secretase has effectively cleared plaques in mice [75]. However, as reviewed recently [67], therapeutic approaches targeting plaque and approaches targeting soluble Aβ have both now been tested in humans, with equally negative outcomes.

Whilst the latter conclusions suggest that anti-Aβ treatments may be failing because they poorly target Aβ in human tissue, the conclusion does not disprove an alternative view for the failure of clinical trials, namely that Aβ is not responsible for all AD. Indeed for some the failure of clinical trials was by no means a surprise [76] and the validation that bapineuzumab does effectively bind Aβ in human tissue [73], but did not provide recovery in clinical trials, only provides support for this view. In the following sections we elaborate on the notion the central focus on Aβ is, on the available evidence, unwarranted.

Niels Bohr

Oligomers are seen in the brain tissue of AD mouse models, although this does not correlate with cognitive decline [102]. Nevertheless, in one study oligomers were found in post-mortem human AD brains but not cognitively normal controls [40]. Furthermore, oligomers appeared to differentiate AD, dementia without AD pathology and cognitively normal patients in another [103]. This evidence is much cited in support of a role for oligomers in AD, but the data fail to define clearly which manifestation of many possible Aβ oligomers are toxic, or if Aβ oligomers are responsible for toxicity [104].

Debate also continues over the nature of amyloid assemblies in vivo, with studies reporting various assemblies showing different toxic effects (reviewed in [105]). Moreover, at present it is only possible to study Aβ oligomers secreted from in vitro cultures, or extracted from post-mortem brain tissue [105],[106]. Accordingly, some have inferred current evidence for Aβ oligomerisation may simply be an artifactual consequence of detection techniques such as sodium dodecyl sulphate (SDS) polyacrylamide gel electrophoresis (PAGE) [107]. Hence it is not yet clear if Aβ oligomers are present in the original tissue, or rather arise due to experimental manipulations.

For example, SDS-PAGE can detect Aβ oligomers in human brain homogenates, yet surface-enhanced laser desorption/ionization time-of-flight MS (SELDI-TOF MS), which requires less manipulation of samples prior to analysis, fails to detect dimeric Aβ in human brain homogenates [107]. Both SDS-PAGE and SELDI-TOF MS could detect Aβ monomers from human brain homogenate. As the authors note, this suggest an over-reliance on low-resolution techniques, such as immunoblotting, may have distorted our understanding of Aβ biochemistry in the in vivo human brain. Clearly, our understanding of the true biochemistry and presence of Aβ monomers, dimers and higher order oligomers in the human brain in vivo is limited. Unfortunately, data that brought the amyloid hypothesis to its current, Aβ oligomer-based state, including the aspects concerning synaptotoxicity, arose from material driven by such low-resolution techniques.

The similarities between EOAD and LOAD pathology are integral to the amyloid hypothesis. Yet, both LOAD and familial EOAD are highly heterogeneous. They exhibit (1) different ages of onset [112], (2) differing temporal progressions [113], (3) different and varying cognitive symptoms [114] and (4) dissimilar pathological presentations [115].

Furthermore, a third of patients with EOAD show non-memory symptoms, whilst in LOAD only 6% have non-memory symptoms [116]. In twins environmental influences play a part in the timing of onset and on the levels of pathological markers at the end stage [117]. This heterogeneity has been recognised in recent AD diagnostic guidelines from the National Institute on Aging (NIA) and Alzheimer’s Association (AA) [118].

Important to note also is that although essentially 100% of individuals with Down’s syndrome have neuritic fibrillar plaques and NFTs by the fifth decade, the onset of dementia is highly variable, with only 70% becoming demented by their 70s, but with most maintaining their baseline cognitive abilities through their 40s and into older ages [13].

Regional aggregation of Aβ may differ in familial and sporadic cases. Remarkably, amyloid deposition is actually greater in some regions in sporadic AD cases than in early onset cases with presenilin mutations [119], indicating not all cases of AD follow the same distinct pattern of amyloid deposition as suggested in 2002 [120]. This observation alone raises doubt that the clinical phenotype of AD is solely related to Aβ deposition.

Furthermore, up to 50% of AD cases have mixed pathologies with other neurodegenerative conditions. For instance, α-synuclein deposition (otherwise seen in Lewy bodies), is a common co-morbidity with amyloid deposition, with more than 50% of AD patients also exhibiting α-synuclein accumulation [121].

Many consider disease heterogeneity as the manifestation of human genetic variation and environmental factors influencing progression of an Aβ or a tau-driven disease. Animal models provide support for this view, since they show homogeneous disease phenotypes, where variation can be introduced through environmental actions such as exercise and environmental enrichment and through mouse strain genetic background, providing a model for the amyloid hypothesis. However, as we shall next discuss, animal models may not truly reflect either LOAD, or even all cases of EOAD. A real possibility, instead, is that human AD may be heterogeneous in presentation because the causes of AD may be heterogeneous, causing the diversity of symptoms that characterise the disease.

Almost all mouse models of AD are engineered to over-express human APP to such an extreme extent that animals show pathology within months of birth. The treatments in current human testing have usually been shown to alter this pathology before being developed for the clinic. However, this approach has yet to produce a result that has translated to a positive human clinical outcome [66].

Little evidence indicates that APP is overexpressed in the human AD brain [122]. Indeed, total Aβ may be reduced [123]. More worrying, most mouse models do not show substantial neuronal loss, despite the presence of large depositions of amyloid [124]. Further, in contrast to human AD, where synapse loss is integral, mice show a highly variable presentation. Some mice show increased synaptic density in specific brain regions, while most models show reduced synaptic density.

Thus, while providing reasonable models for assessing the ability of a treatment to remove Aβ in vivo, and for investigating the relationship of Aβ to other features of the disease, such as its inflammatory components [125], the reality is that removing an overexpressed Aβ molecule in these mice may not be relevant to removing an under expressed, but aggregated molecule, from the human brain. Additionally, a number of questions remain as to the relevance of these transgenic mice to human AD:

Given the limitations of the mouse models, several groups have attempted to investigate alternatives. This includes other species that may better recapitulate AD pathology including rats, octodon degu, chicks, dogs, guinea pigs, rabbits, dolphins and non-human primates, although these are much more expensive to investigate and some still rely on APP over-expression.

Notable exceptions to the over-expressing mouse models include the senescence accelerated mouse model (SAMP8) [140] and the anti-NGF mouse [141]. These latter models replicate several features of AD without relying on human familial mutations. Others have used mice with inducible neuronal loss to replicate the patterns of loss seen in human AD [142]. Stimulation of inflammation also recapitulates many AD features in mice, including increased levels of cleaved APP fragments, altered tau phosphorylation [143] and declining motor and cognitive skills.

Recently a more relevant mouse model was created in which humanized Aβ, with human AD-causing mutations, was inserted into endogenous mouse APP[144]. These mice showed Aβ pathology, neuroinflammation and memory impairment, although there was an absence of tau pathology. This study supports a role for mutant APP (but not necessarily for Aβ per se) in some familial forms of disease. It does not however show Aβ causality in more common sporadic forms of disease. It is also prudent to recall that the majority of familial cases of AD are linked to mutations in presenilin genes, rather than mutations in APP, which are rare [145].

We are intrigued by the highly relevant modelling of AD based on other risk factors of disease. For example, diabetic mice develop many similar features to AD mice [146] and a mouse model of chronic heart failure shows alterations in the metabolism of cerebral Aβ and cognitive impairments [147]. These mouse models show it is not necessary to have familial AD mutations, nor do they need to have the aggregating form of amyloid, to re-create several features of disease. However in general, at this stage, the genetic mouse models hold front and center stage in AD studies. Studies based on these need to be increasingly treated with caution and consideration given to the use of different models.

It is widely accepted that Aβ, when injected or over-expressed in substantial excess, can cause pathology in rodents, but what is the scenario by which genetic mutations in the human cause AD? APP trafficking, function and cleavage is complex and highly controlled. We show in this section that mutations in APP can cause changes to a range of APP cleavage products, all of which could affect synaptic function. Meanwhile, presenilin mutations impact the cleavage of numerous proteins also with synaptic functions. Thus, even in EOAD caused by mutations in APP or presenilin, Aβ may not be the sole basis of disease. In sum, a valid, but inconvenient interpretation of AD genetics is that the aberrant processing of APP, or of other proteins cleaved by presenilin containing enzymes, is the key contributing factor in familial AD, rather than solely aberrant production of Aβ that contributes to histopathology. Furthermore, as we first discuss below, given the numerous genes and processes implicated in AD, it seems highly unlikely that any single gene such as APP alone will account for this disease in the majority of AD cases.

Genome-wide association studies (GWAS) have identified a number of genetic risk factors for AD. To date, the most emphatic demonstration that numerous genes contribute to the risk of AD has come from a meta-analysis of four GWAS data sets consisting of 17,008 AD cases and 37,154 controls from 15 countries [148], which implicated 11 new regions of the genome as risk factors for AD. The findings reinforce the importance of the innate immune response and inflammation (HLA-DRB5/DRB1, INPP5D, MEF2C) already implied by previous work (CR1, TREM2). Also reinforced is the importance of cell migration (PTK2B), lipid transport and endocytosis (SORL1). New hypotheses on AD pathogenesis have also emerged related to genetic mutations in molecules associated with hippocampal synapse function (MEF2C, PTK2B), the cytoskeleton and axonal transport (CELF1, NME8, CASS4) as well as myeloid and microglial cell functions (INPP5D). While efforts are often made to relate new data to the amyloid hypothesis [149], many of these mutations are not conveniently placed within it.

Furthermore, new research showing the molecular signatures of AD vs. normal aging indicates that the molecular phenotype of AD is highly complex, with a variety of transcriptional changes differentiating AD from aging. Transcriptional profiles for neuroinflammatory and lipid metabolism genes in particular are altered early in disease in this dataset [150].

GWAS and ageing data is therefore increasingly consistent with a view that Aβ (or more likely altered APP production, function and cleavage) exists somewhere within a highly complex disease framework that is yet to be understood. It is unclear whether numerous mechanisms converge on a single primary pathway, or, if AD will need to be redefined as a host of diseases manifesting ultimately as memory loss, resulting from synapse loss and neurodegeneration. The latter view, that memory becomes problematic when brain function is disrupted, has simple appeal, but is a nightmare from a therapeutic perspective.

There is a widespread assumption that all the genetic links to AD effectively modify the cleavage of Aβ to produce more of the longer forms, Aβ42 and Aβ43, or increase the ratio of longer Aβ peptides compared to shorter ones and in turn that this is causative of AD. Certainly, evidence for increased levels of Aβ42, or for increases in the Aβ42:Aβ40 ratio as a result of APP, PSEN1 and PSEN2 mutations has been found in both in vivo and in vitro studies [151]-[153]. However the conclusion is not warranted in view of the full data set.

In a study examining the effects of eight FAD PSEN1 mutations on Aβ production, most of the mutants produce no change in the Aβ42:40 ratio [154]. Furthermore, family members with the same FAD mutations exhibit heterogeneity in their clinical and neuropathological phenotypes [155]. These results are supported by studies showing heterogeneous effects of FAD PSEN mutations on the Aβ42:40 ratio depending on the mutation [156]-[158], but are contradicted by reports of universal increases in Aβ42 as a result of FAD mutations [153],[159]. Nevertheless as it stands, it seems unlikely that FAD mutations lead to the same phenotypic amyloid cleavage, resulting in increased Aβ42 and/or increases in the Aβ42:40 ratio.

Newer evidence for the role of a longer form of Aβ, Aβ43, in disease pathogenesis may be the result of a familial AD-linked presenilin mutation [160]. The Aβ43:Aβ42 ratio is increased in mice harbouring this particular presenilin mutation, with no change in Aβ40 or Aβ42 levels [160]. We simply cannot draw conclusions at this stage about the role of any form of Aβ in disease. It would be prudent to include measurements of a variety of Aβ cleavage forms in disease, and determine the importance of qualitative versus quantitative changes in Aβ production over time, during disease [161].

Evidence from rare clinical case studies illustrates mutations in presenilin genes can be associated with neurodegeneration independently of amyloid plaque deposition. Presenilin mutations have been found in frontotemporal lobe dementia (FTD) without amyloid pathology [162], dementia with Lewy bodies [163], posterior cortical atrophy dementia [164] and atypical dementia [165]. However newer evidence suggests presenilin mutations may not be the true causes of all these amyloid-independent neurodegenerative states, as genetic defects in the progranulin (PGRN) gene can explain FTD, atypical phenotypes and parkinsonism, also associated with presenilin mutations [166],[167]. There is however at least one clinical case involving a point mutation in PSEN1, that is associated with the development of Pick’s disease with tauopathy, but without amyloid plaques [168].

Regardless, in AD, where amyloid plaques are present, there is good reason to suggest that the effects of presenilin mutations on APP biology may account only in part, if at all, for the AD pathology they associate with. In fact, there are several known functions for presenilin which could be impacted by FAD-linked presenilin mutations. These alternative functions include roles in macroautophagy, APP vesicle transport, cell survival, cleavage of a wide variety of possible substrates, and the importance of presenilin for synaptic function, which together have culminated in the presenilin hypothesis of AD [169]. There are also links between presenilin function and the innate immune system, which may account for the presence of neuroinflammation in EOAD, independently of amyloid [170]. The numerous effects FAD-linked presenilin mutations would have on all these processes could together account for AD independently of, or additional to, any effect on amyloid pathology.

A `presenilin hypothesis’ of AD has been articulated [169]. We further suggest that presenilin mutations fit best within a hypothesis that AD is a disease driven by synapse loss. Though the AD literature has largely focused on the role of PSEN1 in APP cleavage (Figure 2), presenilin mutations affect a range of proteins and therefore processes, particularly those involved in synaptic function, as summarised above. Given that presenilin appears important for cleaving proteins that are crucial at synapses, presenilin mutations would lead to synaptic dysfunction. It would also follow that drugs targeting presenilin in humans are destined to have profound detrimental effects on the brain with long-term use. This indeed was the result of recent clinical trials of presenilin antagonists, also termed γ-secretase inhibitors (Table 1).

Nevertheless γ-secretase inhibition continues to be pursued. Recent evidence using conditional presenilin KOs has suggested that presenilin function is not as critical in the adult brain as the developing brain, allowing the field to justify its ongoing use as a target for drug intervention (See [171] for a detailed review). This is in spite of the reservations that emerge from consideration of data reviewed above.

The APOE4 allele has been a known genetic risk factor for sporadic AD, and it remains the strongest known [22]. It has been linked to the amyloid hypothesis by indications it is involved in the clearance pathway of Aβ, with deficits causing a toxic Aβ accumulation and aggregation [172]. Meanwhile, an alternative avenue of enquiry shows that APOE4 has intimate connections with innate immunity, and this was reasoned to explain its broader relationship with inflammatory disease, not just AD. One observation is that APOE suppresses TNF secretion from inflammatory cells [173] (also see [174]). Importantly an APOE mimetic that suppresses TNF secretion has successfully treated experimental models of neurodegenerative disease, including traumatic brain injury [175], stroke [176] and AD [177]. As well as reducing behavioral deficits, in the study of AD, the APOE mimetic also reduced Aβ plaques and tau tangles [177]. The relationship of APOE4 to inflammation therefore opens a channel of enquiry directed to explain why stimulation of APOE expression in mice enhances normal Aβ clearance (both soluble oligomers and plaques) and reverses behavioural deficits. In line with the observation of the link between presenilin mutations and inflammation, the links between APOE4 and inflammation further point to inflammation as a major player in AD pathogenesis independently of Aβ.

APP synthesis, trafficking and cleavage are complex and highly regulated processes (Figure 2). It is important to recognise that familial AD APP and presenilin mutations may not only impact Aβ production, but also the production of the other peptides produced from APP including sAPPα, sAPPβ, p3 and AICD, as well as the relative levels of full-length APP. Interestingly, overexpression of AICD can cause an AD-like phenotype [178], whilst increased cleavage of sAPPβ is associated with familial Danish Dementia with similar aetiology to AD [179]. Furthermore lowered levels of neurotrophic sAPPα are seen in AD, and mutations which inhibit the α-secretase enzyme ADAM10, which liberates sAPPα from its precursor, are found in the promoter region and coding sequence of some individuals with AD [180]. Depletion of sAPPα by inhibition of ADAM10 trafficking can bring about sporadic AD phenotypes [181], corroborating an independent role for APP cleavage products other than Aβ in bringing about disease phenotypes.

Whilst the functions of p3 and sAPPβ are little explored (which in itself is a remarkable reflection of the intense focus on Aβ at the expense of other cleavage products of APP), a wealth of evidence exists for physiological functions of sAPPα, Aβ, AICD and full-length APP [182] (Figure 2). These studies raise questions as to whether familial AD driven by presenilin and APP mutations is primarily a result of aberrant Aβ expression, or if it is in fact a result of altered APP cleavage, and the resultant effects of altered APP cleavage on sAPPα, sAPPβ, Aβ, AICD, p3 and full-length APP. This brief discussion does not even approach the possible physiological functions of C83, C99 (Figure 2), the different functions of the multiple isoforms of APP including APP_695, APP₇₅₁ and APP_770, or the highly homologous proteins to APP, APLP1 and APLP2, which are also physiologically expressed in the human brain and may serve redundant functions with APP proteins.

The field has pursued the idea that Aβ accumulation is the central cause of AD based in part on an amyloid-centric interpretation of the genetics. Yet as described above, numerous products of APP are affected by APP mutations and by presenilin mutations.

One valid conclusion, in view of the complex biology of APP function and cleavage, is that APP and presenilin mutations cause AD because they alter several cleavage products of APP, which each in turn contribute to AD. In this view of disease, alterations in Aβ expression would only be a part player in pathology or, perhaps even, a by-product, and an indicator of altered APP function and cleavage. This conclusion, if true, would predict that clinical trials of anti-Aβ drugs will fail even in some or all cases of EOAD.

Meanwhile, analysis of the effects of presenilin mutations does not lead to the conclusion that AD is caused by Aβ. The cleavage of a range of proteins is affected by presenilin mutations and many would affect the synapse. In fact as we illustrated, presenilin mutations do not always alter Aβ production as may be expected. Finally, many mutations and risk factors associated with AD may not relate to Aβ metabolism.

Yet, amyloid-centric interpretations continue to flourish. A recent study showed that a mutation, determined to modestly decrease Aβ levels of the course of lifespan, is preventative of AD [23]. This data was taken to suggest that a life-long reduction in Aβ reduces the risk of AD. However, the mutation also results in marginally (albeit non-significant) increases in levels of beneficial neurotrophic sAPPα. Regardless, an alternative interpretation is that a life-long change in APP function and cleavage could protect against AD independently of lowered Aβ production.

As suggested by Figure 3, the amyloid hypothesis is at least incomplete, and quite possibly largely incorrect. Therefore it follows that therapies targeting Aβ or APP processing may not treat LOAD, and possibly may not even work in some cases of EOAD. Given this conclusion, it is worthwhile to consider alternative possibilities. There are a number of theories in the literature that must be given serious consideration and ultimately integrated into a holistic view of disease. We will elaborate on just some of these, below.

It is suggested that a similar pathogenesis operates in AD as in Type 2 diabetes (T2D), but restricted to the brain, thus describing AD primarily as a result of cerebral insulin resistance [183]. Certainly cerebral insulin resistance is present in AD just as in T2D [184], and the appropriate alterations in post-insulin receptor intracellular signalling have been impressively demonstrated in fresh AD autopsy brains [185]. This idea is likely inseparable from the argument that AD is an inflammatory disease, since evidence that excessive TNF induces insulin resistance is biochemically precise [186] and, as has been reviewed [187], is widespread across many inflammatory diseases, infectious and sterile. Moreover an agent that inhibits TNF production [188] and another that controls insulin resistance [189] have both been shown to reverse AD in experimental models [190].

We and others have long proposed a role for neuroinflammation, driven by microglia and astrocytes, as a trigger for Alzheimer’s pathogenesis [46],[125],[187],[191]. The case for chronic inflammation, as classically defined, rather than Aβ, being the primary initiator of AD has a long history, with new evidence continuing to accumulate. From 1989 it has been reported that inflammatory cytokines are essential for the excess APP required for the amyloid hypothesis of AD [192], as well as up-regulating its cleavage to form Aβ [187]. In addition, parallel studies demonstrating that oligomeric Aβ influences synapses through inducing the inflammatory cytokine TNF [187],[193],[194] have been enlightening. A possible role for neuroinflammation in synapse pathology early in disease has now been acknowledged [195] and there is much evidence, from genetics and measuring indicatory of inflammation very early in AD, that it is in the right place at the right time to be causal, and likely to precede Aβ and tau pathologies [187]. More recently, the clinically approved specific anti-TNF agent, etanercept, is reported to prevent changes caused by administering Aβ to mice intracerebroventricularly [196].

The concept of hyperphosphorylated tau being a primary mediator of AD, like amyloid, has a long history, which continues to grow [197]. Much interest still exists in where tau sits in the pathogenesis of AD [198]. In our view, AD is sufficiently diverse that it is conceivable that the role of tau, and where it sits in AD pathology, could vary among individuals. If tau is a primary activator of disease in some cases, it is imperative that the reported harmlessness of phosphorylated tau to neurons during mammalian hibernation [199] be discussed in AD research circles. Furthermore, hyperphosphorylated tau can be considered another histological sign of cytokine activity [187].

The inflammatory hypothesis is an example of how amyloid and tau research can be integrated into a novel set of ideas, both expanding the amyloid hypothesis and including it. However, while we use the term `neuroinflammation’ throughout this text and elsewhere, we note that neuroinflammation is poorly defined. In its simplest form neuroinflammation is currently defined by altered glial cell morphology and excess pro-inflammatory cytokine release [46]. This must ultimately give way to a more complex and subtle view of glial function/dysfunction within the multicellular synapse [200].

Stepping back to consider the multitude of factors we have summarised above, a complex picture emerges that consistently points to synaptic dysfunction and loss as a major link between the diverse characteristics of the disease. We have recently pointed out the synapse needs to be re-defined and understood as a multicellular structure where glia play a critical role [46]. This allows us in turn to re-imagine AD.

Microglia and astrocytes are essential to normal synapse biology, including the removal [201],[202] and formation of synapses [203],[204], and maintenance of synaptic function [205],[206]. Disruptions in signalling between glia and synapses, which may involve several known cytokines such as TNF, could therefore drive the well-known synapse loss in AD, either independently of, or in conjunction with Aβ [46].

Clearly, understanding the physiological roles of microglia and astrocytes at synapses, as opposed to simply considering them as cells with key roles in innate immunity and `neuroinflammation’, is a critical avenue for future research. We suggest future research will reveal that the entire current concept of `neuroinflammation’ is poorly understood, defined and characterised. This concept will require profound rethinking before we can truly understand the role of glia in the unperturbed brain and in AD pathogenesis [46].