The vexing complexity of the amyloidogenic pathway

Abstract

The role of the amyloidogenic pathway in the etiology of Alzheimer's disease (AD), particularly the common sporadic late onset forms of the disease, is controversial. To some degree, this is a consequence of the failure of drug and therapeutic antibody trials based either on targeting the proteases in this pathway or its amyloid end products. Here, we explore the formidable complexity of the biochemistry and cell biology associated with this pathway. For example, we review evidence that the immediate precursor of amyloid‐β, the C99 domain of the amyloid precursor protein (APP), may itself be toxic. We also review important new results that appear to finally establish a direct genetic link between mutations in APP and the sporadic forms of AD. Based on the complexity of amyloidogenesis, it seems possible that a major contributor to the failure of related drug trials is that we have an incomplete understanding of this pathway and how it is linked to Alzheimer's pathogenesis. If so, this highlights a need for further characterization of this pathway, not its abandonment.

Overview of Alzheimer's Disease

Alzheimer's disease (AD) is a progressive neurodegenerative disorder with no cure or disease‐modifying therapies. There are both rare familial (inherited) forms (FAD or early onset AD [EOAD]) and much more common sporadic (SAD or late onset AD [LOAD]) forms of this disorder. There are presently roughly 50 million AD patients worldwide—5.7 million in the United States alone, with the number expected to increase two to fourfold by mid‐century, depending on the country—see 2018 Alzheimer's Association Alzheimer's Disease Fact and Figures: https://www.alz.org/media/HomeOffice/Facts%20and%20Figures/facts‐and‐figures.pdf). The cost to the U.S. economy of compensated and uncompensated care of AD patients is currently on the order of half a trillion dollars. 10% of men and 20% of women are expected to succumb to AD, which is inevitably fatal. Currently approved pharmacological treatments are limited to marginally effective management of early stage disease symptoms. LOAD patients typically survive for 4–8 years following initial diagnosis, but there is considerable heterogeneity in disease progression. Classically, patients with Alzheimer's initially present with short‐term memory problems and other mental deficits (some outlined in Fig. 1), with the disease progressing to expunge all memory, both short and long term. AD results in significant reductions in brain mass and, eventually, loss of organ functions and increased susceptibility to secondary disorders such as pneumonia. Brain tissues in patients who have died from AD or its complications usually exhibit the presence of extracellular amyloid plaques under the conditions of histological staining with Congo red, as viewed under a polarizing microscope. Except for the age of onset (usually much earlier for FAD than for LOAD), the symptoms and progression are similar for FAD and LOAD. In addition to amyloid‐β deposition, other cellular and biochemical changes in brain tissue occur in AD, including hyper‐phosphorylation of the tau protein and consequent formation of intraneuronal neurofibrillary tangles, increased glutamatergic signaling, neuroinflammation, and decreases in cholinergic signaling.1, 2, 3

Figure 1.

Self‐diagnosis by the corresponding author's father of his early symptoms for what proved to be late onset AD. A former mechanical engineer and Korean War era US Air Force cartographer, he succumbed to this disease 10 years later in 2017, at the age of 87.

The notion that the production and neurotoxicity of the amyloid‐β (Aβ) polypeptides is central to the etiology and pathology of AD—the “amyloid hypothesis” or “amyloid cascade hypothesis”—has been enormously influential over the past 30 years.4 However, in part due to the failure of a series of well‐publicized clinical trials designed around drugs or antibodies that target either Aβ aggregates or the biochemical pathway responsible for amyloid production, the “amyloidogenic pathway,” considerable skepticism has developed that amyloidogenesis is truly central to most forms of AD, particularly for the common sporadic late‐onset forms of this fell disorder.5, 6, 7 The genetic linkage of the amyloidogenic pathway to FAD is unquestioned.8, 9, 10, 11, 12, 13 However, while a number of gene variations have now been documented to be risk factors for the sporadic LOAD form of the disease,14, 15, 16 it has been a puzzling fact that the growing roster of risk factors has not included variations impacting genes that have an obvious role in the amyloidogenic pathway. Moreover, amyloid plaques are found in the brains of many who never develop AD symptoms.17 Conversely, there are rare AD patients who have few amyloid plaques or neurofibrillary tangles in their brains.18 These factors are probably part of the reason why the fraction of the US National Institute of Aging grant budget devoted to Aβ‐related research has decreased from 27% in 2007 to 18% in 2017.19 Doubts about the centrality of the amyloidogenic pathway in AD persist even though LOAD can now be diagnosed in living patients based on the detection of Aβ in the cerebrospinal fluid (CSF) and of Aβ plaques in the brain (Fig. 2).2, 3

Figure 2.

“Jack plot” illustrating the appearance with aging of biomarkers that now be be measured and used as predictors for the development of AD. The earliest predictor is the detection of long form Aβ in CSF, as collected via a spinal tap. This is followed by the detection of amyloid plaques in the brain through the binding of certain 18F‐labeled compounds and PET. The next predictor is the elevation of the tau protein in CSF and then the detection by magnetic resonance imaging of changes in brain morphology, along with detection by PET of problems with energy metabolism in the brain. Only after these biomarkers appear do early signs of memory loss and other relatively minor problems arise as “mild cognitive impairment”, which finally progresses to full blown AD. While not discussed elsewhere in this review, we suggest a significant additional layer of complexity for the amyloidogenic pathway as a target for AD therapeutics is the fact that Aβ production, deposition, and accumulation takes place over a period of decades. The conundrum of how to target a decades‐long process with a potential therapeutic in a drug trial of short duration represents yet another vexing problem. Figure reproduced with permission from The Lancet Neurology (Elsevier).3

It is not our goal here to conduct a comprehensive review of the evidence for or against a central role for Aβ production in the pathogenesis of AD, excellent examples of which are published elsewhere.20, 21 Rather, we provide some examples of how very complex the amyloidogenic pathway is. We argue that this complexity represents a confounding factor in efforts to treat or prevent AD by targeting either the biochemistry of this pathway or the toxic oligomer/aggregate forms of Aβ. We also highlight a recent breakthrough study that appears to, at last, establish a genetic link between the amyloidogenic pathway and LOAD.

Components of the Amyloidogenic Pathway

Our understanding of the role of the amyloidogenic pathway is closely linked to the genetics of the familial (inherited) form of AD. FAD is epidemiologically characterized by having an early onset (before 65 years of age) and nearly 100% penetrance throughout families that harbor pathogenic mutations or duplications in specific protein‐coding genes.8, 9 The first identified FAD gene was APP (chromosome 21; location: 21q21.3),10 which codes for the conserved, ubiquitously expressed single‐pass transmembrane amyloid‐β precursor protein (APP).11, 13, 22 Full‐length APP has 770 residues with a large ectodomain, a single transmembrane span, and a modest (ca. 45 residue) intracellular domain. The most common neuronal splice variant of APP encodes 695 residues. The role of the APP in the adult brain has not been fully elucidated; however, it may be important for proper neurological development during gestation, on top of playing roles in neuroplasticity, synaptic function, and homeostasis throughout life.23 The complete knock‐out of APP leads to small, slow, vacuous mice.24 Only when the ALP1 and ALP2 genes are also deleted is deletion of APP embryonic‐lethal.24 As of February 2019, some 18 different sites, all located in or proximal to the C‐terminal 99 residues of APP (C99), are known to be subject to mutations that cause FAD (see compilation at Alzforum.org/mutations/app).

One APP mutation is known to protect humans from AD: the “Icelandic mutation,” A673T.25, 26 Along these same lines is a fascinating case study of a Down syndrome (DS) patient who escaped from the usual early onset (30–40 years old) Alzheimer's neurodegeneration that normally accompanies DS.27, 28, 29 People with DS have an extra copy of chromosome 21 (trisomy‐21) and consequently, an extra copy of APP.12, 30 The presence of an extra copy of APP is thought to cause EOAD by increasing total expression of the APP protein, resulting in increased total Aβ.27, 31 A DS patient was recently identified who exhibited incomplete trisomy of chromosome 21 in which the APP locus was not duplicated, but most other genes of that chromosome were.31 This individual, who was studied extensively between ages 66 and 72, presented with the mild intellectual disability characteristic of DS. However, he did not exhibit the cognitive decline that normally accompanies DS, nor was dementia evident and only low levels of Aβ deposition were detected by amyloid positron emission tomography (PET) imaging. While these findings are not without dispute,32 this patient seems to provide a compelling example of how APP is linked to Alzheimer's.27, 31

Also causative of FAD is pathogenic dominant familial mutation in either PSEN1 33 and PSEN2,34 which code for the two isoforms of presenilin, the catalytic subunit of the heterotetrameric γ‐secretase protease. γ‐secretase cleaves a wide range of single span membrane proteins in their transmembrane domains,35, 36, 37, 38 the most studied of which are APP and the Notch receptor, the latter being a master regulator of cell and tissue differentiation and development.

The amyloidogenic processing of APP to Aβ involves cleavage by the β‐secretase (BACE1) and γ‐secretase proteases that successively hydrolyze APP in the juxtamembrane extracellular domain and then in the transmembrane domains, respectively (Fig. 3).39 BACE1 cleavage releases the large soluble extracellular N‐terminal fragment, sAPPβ, leaving the 99 residue transmembrane C‐terminal fragment of APP behind: C99 (also known as βAPP‐CTF). Liberated sAPPβ may have important roles such as stimulating increased axonal outgrowth, decreased cell adhesion, neural differentiation of human stem cells, and stimulation of GABA receptors to alter synaptic transmission.40, 41

Figure 3.

Canonical processing of the amyloid precursor protein. Full‐length APP (either neuronal 751‐ or 695‐residue isoform) is shown in the middle. On the right (red arrows) is the amyloidogenic proteolytic cascade that is initiated by β‐secretase and generates sAPPβ, AICD, and Aβ peptides. On the left (blue arrows) is the nonamyloidogenic proteolysis pathway that is initiated by α‐secretase and generates sAPPα, AICD, and Aα peptides. In both pathways, γ‐secretase cleavage is involved, as shown in light green.

The C99 domain of the APP has a single transmembrane span and a propensity both for homodimerization and cholesterol binding.42, 43, 44 In the amyloidogenic pathway, C99 is engaged by the heterotetrameric γ‐secretase complex, which binds C99 monomers and initially cuts C99 at either of two possible “epsilon cleavage” sites in the membrane near the cytosol, resulting the in the release of the amyloid intracellular domain (AICD). The function of AICD is controversial, with some evidence suggesting that it plays a role in regulating transcription of certain genes.45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59 Intriguingly, some of the these genes encode proteins clearly related to the amyloidogenic pathway51, 52, 55, 59 or to tau hyperphosphorylation and fibrillization.52, 56, 57 However, at least some of these studies have been challenged.60, 61 Painstaking research has shown that the remaining and still membrane‐anchored N‐terminus of C99 is not then released by γ‐secretase, but rather is subjected to processive cleavage in which tripeptides or tetrapeptides are successively released from the C‐terminal end of the transmembrane domain (TMD) domain until all that is left of C99 is either the long form of Aβ (mostly Aβ₄₂) or the short form (mostly Aβ₄₀) (Fig. 4).62 Aβ then dissociates from γ‐secretase and potentially from the membrane. Recent evidence suggests that β‐ and γ‐secretases sometimes form a complex in brain tissue that processively catalyzes the sequential cleavage of full‐length APP to Aβ.63, 64

Figure 4.

Competing processive γ‐secretase cleavage reaction pathways for the C99 protein. This figure is reproduced with permission from Neurochemical Research (Springer).62

Aβ represents a polypeptide that contains half of what was originally a transmembrane domain. It is therefore unsurprising that it retains considerable affinity for membranes and has a high propensity to form aggregates in solution—particularly the longer Aβ₄₂ form.65 Conventionally, it has generally been thought that release of Aβ occurs mainly at the outer leaflet of the plasma membrane, with the peptide being released into the extracellular milieu. Aβ then forms toxic oligomers that go on to form cross‐beta fibrils that eventually become entangled with other molecules to form mature amyloid plaques in brain tissue.66, 67 The toxicity of extracellular Aβ may be related to its ability to self‐assemble at the membrane surface to form ion channels68, 69, 70 and/or to stimulate any one or more of a number of candidate receptors at the cell surface.71, 72, 73 One or more of these events are thought to then trigger ill‐defined pathways that lead to hyperphosphorylation and fibrillization of tau, ultimately resulting in intraneuronal neurofibrillary tangles and cell death.74 The exact nature of the physiological toxic Aβ oligomers is a matter of mystery. Numerous different forms of Aβ oligomers have been examined under laboratory conditions,21 but establishing which ones, if any, are pathophysiologically relevant has been elusive. While the toxicity of amyloid plaques was once questioned, there now seems to be little question that they are toxic because they trigger chronic inflammation in the brain.75, 76 They also likely serve as a reservoir of toxic Aβ oligomers, which can dissociate from the plaques to re‐enter solution.77, 78

Competing with the amyloidogenic pathway for processing of APP is a major “nonamyloidogenic” pathway that is initiated when the extracellular matrix‐metalloprotease α‐secretase (usually ADAM10, but sometimes ADAM9 or ADAM 17) cleaves full‐length APP,39, 79, 80 as is illustrated in the left half of Figure 3. Full‐length APP is clipped by α‐secretase at an extracellular site closer to the membrane than the β‐secretase cleavage site to release the transmembrane C83 fragment, which is then further processed by γ‐secretase in a manner analogous to C99 processing. The resulting APP fragments are the large extracellular sAPPα ectodomain, AICD, and P3 (a nonamyloidogenic fragment of Aβ, sometimes referred to as Aα). α‐secretase competes with β‐secretase for the initial cleavage of APP. Any factor that shifts the balance toward β‐secretase cleavage is expected to increase the total Aβ generated, potentially contributing to the pathogenesis in AD.81 Conversely, activation of α‐secretase cleavage is expected to reduce Aβ production. α‐Secretase also has other important substrates and is believed to be involved in maintaining synaptic health, neurogenesis, and neuronal homeostasis.81, 82 A fascinating emerging area of biology is how ADAM proteases are modulated by members of the iRhom (noncatalytic rhomboid homologs) and tetraspanin families of multispan membrane proteins.83

One confounding factor in efforts to develop agents that reduce amyloid‐β levels that act by inhibiting γ‐secretase and β‐secretase or by activating α‐secretase is the fact that all three of these proteases have multiple substrates besides APP, some of which are of great importance in their own right.37, 84, 85 For example, both γ‐secretase and α‐secretase play critical signaling roles when they cleave the Notch receptor, a master regulator of cell development. Inhibition of γ‐secretase activity is associated with severe toxicity when Notch cleavage is inhibited, disrupting Notch signaling.86, 87, 88 This has led to a search for “γ‐secretase modulator” compounds that do not alter Notch receptor cleavage, but that perturb γ‐secretase cleavage of C99 to tip the Aβ42:40 production ratio toward production of the shorter and less toxic forms.89, 90, 91 Whether compounds of this class will eventually be therapeutically useful remains to be seen.

Finally, it should be noted that in addition to the classical amyloidogenic and nonamyloidogenic pathways summarized above, there appear to be yet other proteolytic cleavage events that release other APP fragments encompassing all or part of Aβ.92, 93, 94, 95, 96 In some cases, products may be due to rare alternative cleavage events by one of the canonical secretases, but sometimes other proteases are involved. Whether some of these alternative APP‐derived peptides might contribute to AD is not yet clear.

Amyloid‐β Is Not Always Generated at the Cell Surface and Does Not Always End Up in the Extracellular Milieu

Another source of complexity in the amyloidogenic pathway lies in the fact that there are two forms of γ‐secretase: one with presenilin 1 (PSEN1) and another with presenilin 2 (PSEN2) as the catalytic substrate. Familial AD mutations are found in both PSEN isoforms, but are much more common in PSEN1. One of the other subunits of the γ‐secretase complex, Aph1, also has two isoforms, Aph1a and Aph1b. Moreover, Aph1a is found in two common splice variant forms: Aph1aL and Aph1aS. Some of the different forms of γ‐secretase, based on the various possible combinations of subunits, appear to have different catalytic properties.97

In terms of the cellular location of amyloid formation, conventional thinking is that most PSEN1 cleavage of C99 occurs at the plasma membrane, leading to direct release of Aβ into the extracellular milieu.98 PSEN2 cleavage is generally thought to occur mostly in endosomes, with Aβ being released into the lumen.98 Whether this endosome‐luminal population of Aβ is secreted from cells via exocytosis, traffics on to lysosomes, or has a different trafficking itinerary is not well established. However, endocytosis of extracellular Aβ and subsequent lysosomal membrane “leakiness” (into the cytosol) was first reported over two decades ago.99, 100, 101 There are also many reports of C99 cleavage taking place in other organelles,102, 103, 104 with the resulting Aβ being reported in the lumen of these compartments, as well as in the cytosol.105, 106, 107, 108, 109 The trafficking pathways that lead to the presence of Aβ in some of these locations are by no means always clear. It has not been ruled out that some of the toxicity of Aβ is generated by its intracellular population. Could it be that it is intracellular Aβ that triggers hyperphosphorylation of tau and formation of neurofibrillary tangles? We decline to weigh in on this question but note hypotheses have been published that raise this possibility.110, 111, 112, 113

It should also be pointed out that not all extracellular amyloids are deposited in amyloid plaques located in the extracellular milieu of brain tissue. In the case of cerebral amyloid angiopathy (CAA), central nervous system amyloid deposits are formed in the walls of cortical and leptomeningeal arteries, arterioles, and, sometimes, capillaries and veins. CAA pathogenesis, like AD, is often associated with age. Similar to AD, accumulation of Aβ is thought to cause CAA and therapeutics are limited.114 Indeed, CAA often accompanies AD.115

While it may be unrelated to Alzheimer's, we point out a remarkable study documenting the accumulation of Aβ in the placentae of women afflicted with preeclampsia, raising the possibility that this often tragic condition is also linked to amyloidogenesis.116

The Mechanisms by Which FAD Mutations Promote the Etiology of AD May Be More Complex than Previously Appreciated

Inherited point mutations at any one of 18 residues of APP result in early onset FAD.8, 117 Nearly all the AD‐related pathogenic single nucleotide polymorphisms cluster around the α‐, β‐, or γ‐secretase cleavage sites located in the transmembrane and extracellular juxtamembrane domain (TM/JM) of APP‐C99.

It has generally been thought that there are several mechanisms by which FAD mutant forms of APP promote AD pathogenesis. These include increasing the Aβ42:40 production ratio, increasing the aggregation propensity of the released Aβ, and/or increasing total Aβ production.118, 119 Yet another mechanism seems to be reflected by the K687N mutation adjacent to the α‐secretase site, which results in decreased nonamyloidogenic cleavage by that protease, likely tipping the balance of APP processing toward increased β‐secretase cleavage, as well as altering the biophysical properties of the consequent mutant form of Aβ.120 Along the same lines, the K670N/M671L (Swedish) APP double‐mutant, in which the mutation sites are immediately N‐terminal to the β‐secretase cleavage site, is associated with increased β‐secretase cleavage and increased total Aβ production.121, 122, 123, 124, 125, 126, 127 In some cases, a single mutation is associated with more than one disease phenotype, such as the D694N (Iowa) mutant that is associated with both FAD and CAA. This mutation is thought to alter the biophysical properties of the Aβ peptides, resulting in increased fibrillization and toxicity.128, 129, 130, 131 As noted previously, the AD‐protective Icelandic APP A673T mutation proximal to the β‐secretase cleavage site is associated with defective β‐secretase cleavage and lower levels of Aβ.25, 26, 132

The mutations mentioned above are all located in the extracellular/juxtamembrane region of the APP, however most FAD‐associated mutations are at sites in the transmembrane domain, where it has long been thought that they cause FAD by altering γ‐secretase processing to elevate in the Aβ42:40 production ratio.133, 134, 135, 136, 137 However, in a recent cell‐based study, Xu et al. reported that while most FAD‐associated APP mutations at sites located in the cytosolic end of the APP TM domain are indeed associated with an increased Aβ42:40 production ratio, these mutants are also partially resistant to cleavage, thereby reducing total Aβ cleavage.138, 139 This is a surprising result, but is consistent with the recently determined landmark structure of the complex of the APP C83 domain and presenilin 1,140 which supports the notion that some APP mutations are likely to reduce the affinity of γ‐secretase for C99, which would lower physiological Aβ production. Is it possible that that reduced cleavage of APP‐C99 could also contribute to the etiology of AD? If so, then we might expect to see that the other class of FAD mutations—those that alter the sequences of PSEN1 and PSEN2—might sometimes be γ‐secretase loss of function mutations.

The Mechanisms by Which Presenilin Mutations Promote the Etiology of AD May Also Be More Complex than Previously Appreciated

Presenilin‐1 and its presenilin‐2 isozymes are multispan membrane proteins that are ubiquitously expressed throughout the body, including the nervous system.35, 141 In general, γ‐secretase complexes have a diverse range of biological functions that involve regulating synaptic plasticity, learning and memory, as well as neuronal survival and homeostasis.35, 38, 141 The presenilin‐1 protein (PSEN1; coded by the PSEN1 gene) harbors over 180 documented human FAD‐causing missense mutations affecting about 130 amino acid sites (Alzforum.org). There are at least 39 FAD mutations in presenillin‐2 (PSEN2; the product of the PSEN2 gene), which impacts 17 different sites. Together, mutations in PSEN1 and PSEN2 account for 90% of the known autosomal dominant FAD gene mutations.36, 141, 142 It is interesting to note that frameshift or nonsense mutations in the PSEN1 gene have not yet been linked to AD, but instead result in a skin disorder, acne inversa, most likely because of a resulting reduction in Notch cleavage.143, 144

γ‐secretase is composed of a heterotetrameric (1:1:1:1) complex of presenilin, anterior pharynx‐defective 1 (APH1), presenilin enhancer 2 (PEN2), and nicastrin (NCT).35 There is also evidence that γ‐secretase may sometimes function as a multimer of heterotetramers.145 The primary biochemical role of γ‐secretase is to catalyze the intramembrane proteolysis of Type 1 (N‐terminal extracellular single span) transmembrane proteins such as APP‐C99; however, at least 90 protein substrates (such as Notch receptors, integrins, cadherins, etc) have been reported,35, 36, 37 earning it the pseudonym “the proteasome of the membrane.” As noted earlier and illustrated in Figure 4, γ‐secretase cleavage of C99 involves competing multistep cleavage pathways that lead primarily to Aβ42 or Aβ40.

In recent years, several studies have suggested that many FAD mutations in PSEN1 are partial or total loss‐of‐function (LOF) mutations. Cacquevel et al.146 noticed a drastic LOF of purified γ‐secretase containing FAD‐mutant forms of PSEN1 (PSEN1^Mut), with other groups describing similar LOF phenomena for FAD‐associated PSEN mutants.147, 148 Sun et al. later published findings based on studying purified γ‐secretases harboring pathogenic PSEN1 mutants and assaying C99 cleavage in vitro, as well as in a cell‐based assay.149 In a wide range of biochemical conditions, they found that about 90% of tested FAD‐associated PSEN1^Mut‐containing γ‐secretases exhibited catalytic LOF in the presence of PSEN1^WT‐containing γ‐secretase and did so in a dominant negative manner. Indeed, PSEN1^Mut γ‐secretases behaved analogously to pharmacological antagonists of wild type (WT) γ‐secretase activity, such that titration of PSEN1^Mut γ‐secretase into solution containing PSEN1^WT γ‐secretase induced LOF of WT at ratios as low as 2:1 (mutant:WT γ‐secretase).145 This interesting result can be rationalized based on direct dominant negative interactions between WT and mutant forms of γ‐secretase (see additional discussion below).58, 63, 145

For most FAD mutants that induced partial LOF, it has also usually been seen that the Aβ42:40 production ratios for the residual activity were significantly increased (although a decreased Aβ42:40 ratio was observed for 10% of mutants). One might expect that reduced total Aβ production by γ‐secretase would be protective against AD, although the increased Aβ42:40 for the residual activity could conceivably be the dominant effect, promoting disease pathogenesis. However, arguing against this possibility was the observation that the age of onset of FAD from mutation to mutation did not inversely correlate with the corresponding change in the Aβ42:40 ratio.149

Before further discussing whether γ‐secretase LOF could sometimes contribute to the development of AD, we highlight additional evidence that FAD mutations in PSEN1 may often induce LOF. Figure 5 illustrates the locations of disease mutations sites in the structure of the protease (PSEN1) in complex with one of its substrates, the APP‐C83 protein (C83 being the product of α‐secretase cleavage of full‐length APP, see Fig. 3).140 We make two observations about the distribution of the FAD mutants in this structure. First, the disease mutations are most highly localized around the substrate‐binding interface. It is not hard to imagine how some of these mutations might interfere with substrate recognition or the catalytic mechanism (in either case leading to reduced catalytic activity) and/or that they alter the distribution of cleavage sites leading to an increased Aβ42:40 ratio. However, mutations sites are also well distributed throughout entire presenilin structure, with many sites not located near the substrate interface. This widespread distribution of disease mutation sites is akin to what is seen for retinitis pigmentosa (RP) mutations in rhodopsin (see fig. 23 in Reference 150) and for a number of other disease‐linked membrane proteins, including the cystic fibrosis transmembrane regulator protein (see fig. 3 in Reference 151). What is known for RP/rhodopsin and several other disease mutation/membrane protein relationships is that there are at least two (nonexclusive) mechanisms by which mutations can result in disease pathogenesis. Some mutations are located at ligand or substrate binding sites and directly induce dysfunction of the protein. However, other mutations—particularly those located distal from the functional site—cause LOF by inducing misfolding. The misfolded protein is often degraded by endoplasmic reticulum protein folding quality control, although the misfolded protein in some cases escapes degradation to form toxic aggregates that may further contribute to disease pathogenesis. The fact that the distribution of presenilin mutations conform to what is a now a well‐established pattern for a number of other disease‐linked membrane proteins150 is fully consistent with the notion that some presenilin mutations result in LOF due to misfolding of the protein, with FAD being promoted by the resulting loss of catalytic function and/or by toxicity of the misfolded presenilin.152 The dominant negative nature of these mutations possibly may be explained by a model in which the folding‐defective γ‐secretase is still able to form higher ordered oligomers with WT, with the consequence being that both mutant and WT are targeted for degradation when the mutant form is recognized as folding defective by endoplasmic reticulum (ER) quality control.145 A similar phenomenon has previously been documented for peripheral myelin protein 22, a tetraspan membrane protein for which dominant mutations that induce misfolding of one allele result in mistrafficking of both the mutant protein and associated WT PMP22.153 To summarize, the structure of the presenilin/C83 complex is consistent with the notion that mutation‐induced loss of γ‐ secretase function may be a factor that colludes with an increased Aβ42:40 production ratio to trigger pathogenesis of FAD.

Figure 5.

Locations of FAD mutation sites in the presenilin 1 cryo‐EM structure.140 Known FAD disease mutation sites are highlighted in green (WT residue side chains are shown). The magenta protein is a substrate, APP C83, bound to the active site, with the C‐terminal end of its TM domain unraveled in preparation for cleavage.

The possibility that loss of γ‐secretase catalytic function could contribute to AD pathogenesis was further suggested by results of clinical trials for γ‐secretase inhibitors Semagecestat154, 155 and Avagecestat,156 which despite achieving the desired significant reduction in total Aβ, exhibited a lack of efficacy coupled to toxicity and worsening of AD symptoms. This prompted the cessation of the related drug trials.5, 6, 157 In general, nearly every disease‐modifying drug candidate that targets the amyloidogenic pathway has thus far been unsuccessful due to issues with toxicity, lack of efficacy, or both.6

It seems possible that γ‐secretase LOF may contribute to some phenotypes of FAD. If so, there must be associated mechanisms contributing to disease pathogenesis in some patients that are not yet well recognized. These etiological mechanisms would possibly act in collusion with an increased Aβ42:40 production ratio. In the next sections, we discuss possible mechanisms by which γ‐secretase LOF could contribute to the etiology of AD.

Is full length APP Sometimes Toxic?

As noted above, the relationship between APP processing and FAD appears to be more complex than simply being based on total Aβ levels and the Aβ42:40 ratio. Consistent with this thinking are findings from studies of transgenic mice expressing a BRI‐Aβ fusion protein that was designed to be processed such that extracellular Aβ40 and Aβ42 are generated independent of either full‐length APP or its C99 fragment. Interestingly, the mice developed amyloid deposits resembling those seen in humans; however, at all ages tested, the mice did not show cognitive or behavioral deficits.158 Similarly, Hamm et al. reported that early (pre‐plaque) neurodegenerative symptoms observed in TgCRND8 AD mouse models were Aβ independent, as indicated by the promotion of AD phenotypes using β‐ or γ‐secretase inhibitors.159 Accumulation of full‐length APP in mitochondria has been observed in several AD patients but not in age‐matched controls.160 Along these same lines, full‐length APP was found to aggregate within dystrophic and degenerative neurons (but not glial cells or astrocytes) in humans with AD, leading some groups to posit that accumulation of full‐length APP precedes Aβ accumulation as a contributing factor to disease progression.161 The concept that full‐length APP could actively contribute to AD, independent of its role in amyloidogenesis is suggested by the results summarized above, but the evidence is not yet strong. Moreover, the lack of β‐secretase mutations as a cause or risk factor for AD is unsupportive of this idea. On the other hand, support for a possible role for the APP C99 domain as a toxic agent in some forms of AD is stronger and is consistent with evidence that many FAD mutations induce full or partial γ‐secretase LOF, leading to accumulation of APP‐C99 and/or APP‐C83, which could be toxic under some conditions.162 Additional evidence for such toxicity is presented in the following section.

Is C99 Sometimes Toxic?

The transgenic mouse strain (3xTg: APPswe, M146V presenilin 1, TauP301L) is associated with enhanced C99 accumulation and the absence of tau hyperphosphorylation. In a recent study, 3xTg was compared to the corresponding 2xTg strain lacking the presenilin mutation, such that amyloid‐β production is unimpeded. It was seen that the 3xTg mice developed more severe AD‐like symptoms including apathetic behavior (an early behavioral symptom of AD), decreased long‐term potentiation, and decreased spontaneous locomotor activity.163 In the same 3xTg mouse strain, it was found that C99 accumulation was the primary contributor to hippocampal lesions.164 It has also been reported that C99 accumulation in J20 AD‐model mice is associated with alterations in the early brain network, a process that could be reversed by treatment with β‐secretase inhibitors.165 In a familial Danish dementia mouse model, in which a deficiency in the protein BRI2 resulted in increased total APP levels, only pharmacological inhibition of β‐secretase processing of APP (and not γ‐secretase) rescued memory and deficits in long term potentiation (LTP), implicating C99 as the pathogenic agent.166 Lauritzen et al. demonstrated in two separate mouse models that C99 accumulation induced defects in lysosomal‐autophagic function, another early hallmark of AD neuropathology, and that the defects were not observed in β‐secretase‐inhibitor treated mice.167 In okadaic acid‐treated mice (in which tau hyperphosphorylation is induced), total levels of C99 increase and accumulate primarily in axons, suggesting C99 redistribution (on top of accumulation) may also play a role in AD‐related pathogenesis.168

Cellular studies have also provided evidence for toxicity of C99. Pera et al. showed that C99 build‐up in the mitochondria‐associated endoplasmic reticulum membranes (MAMs) of AD model cells resulted in altered lipid metabolism/composition in the MAMs and in mitochondrial membranes.108 These changes, unobserved in wild‐type cells, interfered with the assembly and activity of mitochondrial respiratory complexes, offering a possible explanation for the bioenergetic shortfalls associated with AD. Primary cells from DS patients showed endocytic abnormalities that were reversible by reducing expression of either full‐length APP or β‐secretase.169 The group also found that γ‐secretase inhibition was sufficient to induce endocytic defects in healthy fibroblasts and worsened endosomal pathology in DS fibroblasts, suggesting that the abnormalities were derived at least in part from C99 accumulation but not Aβ.169 Neuronally differentiated induced pluripotent stem cells derived from two FAD and two LOAD patients (which displayed elevated levels of the AD biomarkers Aβ, phosphorylated‐tau, and active GSK‐3β) exhibited reductions in phosphorylated‐tau and active GSK‐3β after treatment with β‐secretase inhibitors but not γ‐secretase inhibitors, suggesting that AD‐promoting GSK‐3β activation and tau hyperphosphorylation are more closely linked to C99 production than to Aβ.170

While it is premature to declare that toxicity from accumulation of C99 contributes directly to AD pathogenesis, the studies summarized above suggest that C99 does appear sometimes to be toxic. Further investigation will be required to rigorously test whether FAD mutation‐induced γ‐secretase LOF results in C99 accumulation in brain tissue and, if so, whether this accumulation contributes to the etiology of the related phenotypes of FAD. If this was proved to be the case, it would fundamentally alter our understanding of the relationship of the amyloidogenic pathway to AD, providing new insight into the failures of previous drug trials and perhaps pointing to promising future therapeutic strategies.

The First Direct Genetic Evidence for the Role of the Amyloidogenic Pathway in Late‐Onset AD

While it is clear that genetically dominant mutations in APP, PSEN1, or PSEN2 cause FAD, one of the mysteries of the amyloid hypothesis is why incompletely penetrant mutations in these proteins have not been discovered as risk factors for sporadic LOAD. This is despite the fact that variations in many other protein‐coding genes have been identified as risk factors for LOAD (e.g. APOE4, TREM2, and COBL).14, 15, 16 Of course, for most cases of LOAD, it may be that a combination of nongenetic factors leads to accumulation of toxic forms of Aβ. Such factors could include chemical modification of Aβ (e.g. by reactive oxygen species or ROS by‐products), decreased transport of Aβ out of the brain, or decreased proteolytic cleavage of Aβ. Moreover, a number of the known genetic risk factors for LOAD do have linkages to the amyloid pathway, likely being involved in APP trafficking (e.g. SORL1,171 Aβ transport [e.g. APOE172 or PICALM173]), inflammatory response to formation of amyloid plaques (TREM2174 or ABCA7175), or combinations of several pathways (e.g. BIN1176).

In what may eventually prove to be a major breakthrough in AD research, new data have been published by the Jerold Chun group at the Burnham Institute that describe a nonclassical genetic mechanism that brings mutations in APP into play as risk factors for LOAD.177, 178 This work was based on the fact that genomic heterogeneity exists within complex multicellular organisms such as humans, a phenomenon termed genomic mosaicism. The authors uncovered a novel phenomenon by which APP‐specific gene recombination events lead to the expression of genomically integrated mutant APP genes, and the hyper‐accumulation of these mutant APP genes correlates with disease in sporadic LOAD patient‐derived neurons. Both healthy‐ and LOAD‐derived neurons underwent APP‐specific gene recombination events, resulting in new mutant genomic APP sequences (referred to as gencDNAs). Many of these gencDNAs were transcribed into mRNAs and presumably translated into the mutant APP proteins. While this phenomenon was observed even in age‐matched nondiseased controls, neurons from LOAD patients contained a roughly 10X greater number of unique APP gencDNA sequences, and it was demonstrated that APP gencDNA accumulation correlated with LOAD progression in J20 AD‐model mice. We summarize this fascinating work here in Figure 6. This study was made possible by the donation of brains from both control and LOAD patients.

Figure 6.

Model for gencDNA formation and expression in neurons based on work of the Chun lab.177, 178 Illustrated by blue arrows is the canonical gene expression program for neuronal APP splice variants APP‐751 and APP‐695. The gencDNA creation pathway, illustrated by red arrows, is considered to be initiated by APP reverse transcription followed by translocation and recombination back into chromosomal DNA following DNA damage. Newly formed APP gencDNAs can be expressed and fed into the typical gene expression program in blue.

For diseases such as LOAD, where a single cause of disease is difficult to pinpoint, a directed interrogation of the diseased‐cell transcripts can be a useful way to fingerprint the disease biological state. The prefrontal cortices (an area of the brain involved in cognition, personality, and behavior) for six non‐AD individuals (ages 80–94; three females, three males) and for seven with verified LOAD (ages 72–88; six females, one male) were sectioned and prepared for an intensive genomic and transcriptomic analysis. Using flow cytometry to separate neurons from other brain cells, the Chun lab was able to analyze the transcriptomic state of small populations (50 at a time) of primary neurons from control and LOAD brain tissue. Using the reverse transcriptase‐based polymerase chain reaction (RT‐PCR) and Southern blot, they found transcripts that represented not only the APP‐751 and APP‐695 neuronal splice variants, but also lower molecular weight transcripts that were verified to be APP mutants using ³²P antisense probes. The smaller, mutant transcripts were present both in healthy and LOAD neurons, however, the disease neurons expressed a roughly 10X greater fraction of unique reads, suggesting this process is out of balance in LOAD. They then sequenced the transcripts.

In addition to detecting the normal neuronal APP splice variants (APP‐751 or APP‐695, which contain all 18 APP exons except for 8 or 7–8, respectively), the group identified several unique APP sequences that map back to noncanonical inter‐exon junctions (IEJs). Over‐represented among these sequences were APP variants that joined exons from early in the APP locus (exons 1, 2, 3, or 6) to those toward the end of the locus (exons 11, 14, 16, 17, or 18), skipping a large a portion of the central exons. Using RNA in situ hybridization, the group detected the transcripts in the cytoplasm, indicating they were exported out of the nucleus as mature mRNAs. They also demonstrated that non‐neurons from the same individuals did not express mutant APP transcripts, nor were mutant PSEN1 transcripts expressed. A dangling question was whether these strange APP transcripts reflected a malfunction in gene splicing or if expression was genomic of nature. To verify whether the transcripts found in their previous experiment were expressed from the WT APP locus, Chun and colleagues developed a DNA in situ hybridization (DISH) strategy that would target only genomic loci with IEJs (absent of intron sequences in between them). They identified several mutant genomic IEJs (gencDNA) in both healthy and LOAD neurons that were absent in other cell types; however, APP gencDNA sequences were greatly overrepresented in LOAD neurons. Using fluorescently labeled DISH probes and sequence‐specific restriction enzyme digests, it was verified that the gencDNA sequences were indeed chromosomal and that they did not colocalize with the WT APP locus.

Sequencing of the gencDNAs revealed a pattern of insertions, deletions, and missense mutations that clustered around the 3′ (exons 1, 2, and 3) and 5′ (exons 16, 17, and 18) ends of the APP open reading frame. Exons 16, 17, and 18 include the sequence for the APP C‐terminal C99 domain. It was found that some LOAD cell populations also contained gencDNAs that encoded mutations corresponding to known FAD APP mutations, which were totally absent in nondiseased neurons. It was also verified that some mutant APP gencDNAs were expressed into proteins. Moreover, expression of one such construct caused significant reduction of cell viability in culture. This discovery points to a somatic mechanism by which mutant forms of APP are generated and can come into play in the etiology of LOAD.

Although it was not tested, it seems reasonable that some of the identified mutant APP gencDNAs present in LOAD neurons would be toxic via the same mechanisms by which inherited mutations in APP promote FAD. For example, one identified gencDNA (D3/16) coded for an APP variant that harbored three previously known FAD pathogenic point mutations: A673V (known to increase relative β‐secretase cleavage of APP and increase aggregation potential of Aβ), as well as V715M (French) and V717F (Indiana), which are thought to be pathogenic due to increasing the Aβ42:Aβ40 production ratio. If APP gencDNAs containing known FAD mutations are expressed into the corresponding protein, it would help to explain why the hallmarks of LOAD are so similar to FAD. This concept is further supported by lack of any detected FAD mutant APP gencDNA in neurons from healthy brain tissue. To summarize, the evidence points to a newly discovered mechanism by which the APP gene is selectively retrotranscribed (inferred by the lack of genomic introns in the mutant genes and detected PSEN1 gencDNA sequences) and integrated into the genome by homologous recombination.

How does gencDNA form in situ? Chun et al. hypothesized that the APP recombination event was mediated by endogenous reverse transcriptase (RT) activity coupled to DNA damage. To test this, gencDNA formation in Chinese hamster ovary (CHO) cells was induced and the relative levels of endogenous RT activity was measured in brain samples, in CHO cells, and following purification of RT in vitro. It was shown that the formation of novel APP gencDNAs could be induced using hydrogen peroxide (H₂O₂), which introduces single‐ and double‐stranded DNA breaks. In contrast, gencDNA did not form when the cells were treated with abacavir and azidothymidine, commonly prescribed RT inhibitors used to treat HIV infection. While the RT activity was observed in the brain cells was modest (ca. 100,000X less activity than in CHO cells), that it would be detected at all points to the possibility that this process could happen in neurons in vivo. To further validate this, a J20 sporadic AD model mouse (which develops amyloid plaques in an aging‐dependent manner) was examined, confirming that the number of neuronal (contrary to other brain cell types) APP gencDNA loci also increases with age, correlating with the development of the AD‐like state in the mouse model.

The question remains as to where the “endogenous” RT activity found in neurons comes from. RT enzymes are a common tool of viruses for integrating their genes into the host genome, however, the brain is an immune‐privileged organ with several barriers and potent immune responses to infection. Nonetheless, certain viruses are able to enter the central nervous system (such as alpha‐herpes, rabies, HIV, and polioviruses) through various routes.179, 180 The authors postulated that DNA damaging events linked to AD, such as head injuries, may further assist gencDNA formation in the context of active RT. Could it be that APP recombination is what connects familial AD with some phenotypes of LOAD? While this study was conducted with only a small set of human subjects, if confirmed in future studies, it may point to new direction for disease‐modifying therapeutics that act through very different mechanisms than previously investigated. It also stands to reason that genomic restructuring events such as these could promote the progression or onset of neurodegenerative disorders beyond sporadic AD.

Final Perspectives

We have provided a selective overview of the complexity of the amyloidogenic pathway. We have also highlighted a couple of recent studies that provide fresh evidence that the amyloidogenic pathway is not a red herring in the pursuit of potential Alzheimer's therapeutics. Indeed, the studies featured in this review suggest that it is partly the complexity of this pathway that has confounded early efforts to target it or its amyloid products in order to prevent or treat AD. We conclude by respectfully suggesting that continued research on the amyloidogenic pathway is well merited as the only route to eventually ascertaining whether this pathway is druggable for preventing or treating AD.

Conflict of Interest

The authors declare that they have no conflicts of interest related to this work.