Moving beyond amyloid and tau to capture the biological heterogeneity of Alzheimer’s disease

Abstract

Alzheimer’s disease manifests along a spectrum of cognitive deficits and levels of neuropathology. Genetic studies support a heterogeneous disease mechanism, with around 70 associated loci to date, implicating several biological processes that mediate risk for AD. Despite this heterogeneity, most experimental systems for testing new therapeutics are not designed to capture the genetically complex drivers of AD risk. In this review, we first provide an overview of those aspects of AD that are largely stereotyped and those that are heterogeneous, and we review the evidence supporting the concept that different subtypes of AD are important to consider in the design of agents for the prevention and treatment of the disease. Then, we dive into the multifaceted biological domains implicated to date in AD risk, highlighting studies of the diverse genetic drivers of disease. Finally, we explore recent efforts to identify biological subtypes of AD, with an emphasis on the experimental systems and datasets available to support progress in this area.

Genetic and pathological heterogeneity in Alzheimer’s disease: implications for research and treatment strategies

Alzheimer’s disease (AD) affects about 50 million people worldwide. With weak symptomatic treatments currently available and disease-modifying treatments emerging slowly, understanding the underlying molecular causes of this complex syndrome is imperative. Clinicians and pathologists have long recognized that late-onset (typical) AD (LOAD) manifests along a spectrum of cognitive deficits and levels of neuropathology. Although AD is somewhat stereotyped in terms of both the symptoms of cognitive dysfunction with age and the patterns of pathological aggregation of Aβ and tau in the brain, Alzheimer’s dementia is heterogeneous in many other ways. There exist a range of ages of onset and rates of progression, differences in the cognitive domains that are first affected, and a wide range in both the abundance and variety of neuropathological lesions. Genetic studies of LOAD support a heterogeneous etiology, with around 70 associated loci to date [1], and studies of the genes at these loci implicate several biological processes mediating risk for AD. Despite this heterogeneity, the majority of experimental systems for finding and testing new therapeutics are not designed to capture the genetically complex drivers of AD risk. In turn, past clinical trials may have suffered from an inability to predict which individuals are most likely to respond to the treatment under investigation based upon the drivers of disease in specific individuals.

In this review, we first provide an overview of those aspects of AD that are largely stereotyped and those that are heterogeneous. We next review the evidence supporting the concept that different subtypes of AD are important to consider in the design of agents for the prevention and treatment of the disease. Then, we delve into the multifaceted biological domains implicated in AD risk, highlighting studies of the genetic drivers of AD risk and resilience. Finally, we explore recent efforts to identify subtypes of AD, with an emphasis on the experimental systems and datasets available to support this approach to AD heterogeneity.

Stereotyped aspects of Alzheimer’s disease

In general, people likely to be diagnosed with AD first see a clinician due to concerns about decline in memory and orientation to time and place. While the clinical signs and symptoms of Alzheimer’s disease commonly include a multi-domain amnesiac dementia, there are numerous diseases besides AD that contribute to late-life dementia. It is estimated that between 10% and 30% of those clinically diagnosed with Alzheimer’s dementia do not in fact have the required abundance of pathognomonic neuropathological lesions in their brain that would unequivocally confirm the diagnosis [2].

Alzheimer’s disease is defined neuropathologically by the presence in the brain of extracellular amyloid plaques containing amyloid β-protein (Aβ) and intracellular neurofibrillary tangles containing tau. It is the presence and regional distribution of these hallmark lesions that define AD as a specific brain disease and can distinguish it from other forms of dementia. Aβ is generated by the sequential cleavages of its precursor, APP, first by β- and then by γ-secretase. The precise sites of cleavage of APP by γ-secretase are variable, and longer forms of Aβ produced by γ-cleavage accumulate abnormally in the brain, forming the fibrous plaques that represent one of the two diagnostic features of AD. Tau is a microtubule binding protein with multiple functions that is expressed highly in all neurons. Elevation in the degree of phosphorylation of tau can cause it to dissociate from microtubules and gradually accumulate into insoluble neurofibrillary tangles, the second pathological hallmark of AD. Longitudinal brain imaging studies support findings from molecular genetics that extracellular accumulation of Aβ largely precedes and can drive abnormal phosphorylation and aggregation of tau in neurons. Additional pathological findings in AD that suggest impairment in particular biological domains include the abnormal pruning of synapses, the accumulation of lipid droplets, and the enlargement and dysfunction of endolysosomes in the brain.

The centrality of Aβ and tau to AD biology was recently codified in the NIA-AA (National Institute on Aging - Alzheimer’s Association) AD Framework criteria for the disease [3]. This research framework expanded the definition of AD to include detecting and quantifying certain biomarkers of AD neuropathology in living people. Thus, a definitive diagnosis which previously required postmortem histopathology can now be made in living people. In this context, relevant biomarkers include biofluid and imaging measurements of Aβ, tau, and neurodegeneration (“ATN”). Not only the presence but also the temporal order of appearance of positive biomarkers is common across most AD cases, with substantial accumulation of Aβ appearing first, followed by widespread tau deposits and evidence of neurodegeneration. Together, these common features are central to AD, but there may be multiple molecular roads that can lead to these profiles.

Heterogeneity in neuropathological burden and cognitive trajectory with age

Until recently, studies of Alzheimer’s disease in model systems, including human-derived cell culture models and rodent models expressing pathogenic transgenes, relied on comparisons between “AD” and “not AD”. However, Alzheimer’s disease/Alzheimer’s dementia are not homogenous in their development. Multiple clinicopathological subtypes of AD have been identified based upon the distribution of tau accumulation and atrophy and/or clinical presentation: typical (tau accumulation and atrophy in both hippocampus and association cortex), limbic-predominant, hippocampal-sparing, primary progressive aphasia and minimal atrophy. Longitudinal studies in community-dwelling populations undergoing deep phenotypic characterization are especially informative in defining the heterogeneity in neuropathological and cognitive outcomes with age. One of the longest running and most deeply studied cohorts of aging are the Religious Orders Study (ROS) and Memory and Aging Project (MAP) of Rush University Medical Center, which enroll individuals without signs of dementia at the age of 65 or above [4]. The participants are followed longitudinally with annual quantitative cognitive testing. Following death, their brain tissue is analyzed to acquire quantitative measurements of plaques, tangles, Lewy bodies, TDP43-positive inclusions and vascular lesions [5, 6]. In addition, a variety of - omics profiles are obtained in a majority of individuals that include bulk and single nucleus RNAseq, proteomic profiling, and metabolic profiling [7, 8]. The age at death ranges between 65 and 108 years, with a mean of 90. Prior to death, 44% (759/1746) received a diagnosis of Alzheimer’s dementia. After postmortem analyses, 64% (1122/1746) had received a neuropathological diagnosis of Alzheimer’s disease. Visualization of the data acquired from these cohorts regarding cognitive trajectory and neuropathological burden underscores the heterogeneity during brain aging and in AD in both cognitive trajectory and neuropathological burden (Figure 1A–D). Beyond the heterogeneity observed in these visualizations, a recent in-depth neuropathological study of these cohorts revealed a striking variety of mixed cognition-related neuropathological phenotypes present across the aging population, including assessments of Lewy bodies, TDP43 inclusions, hippocampal sclerosis, and strokes [5, 9, 10]. Organizing that data to look only at the subset of individuals who received both clinical and pathological diagnoses of AD reveals that only 24% lacked other histopathologies listed above, and that the brain pathologies in the remaining population were highly heterogeneous (Figure 1E). These data argue strongly for the utility of experimental systems that can capture the molecular underpinnings of this phenotypic diversity.

Figure 1. Heterogeneity in cognitive trajectory and neuropathological burden in the aging human population.

Data collected in the ROS and MAP cohorts for deceased individuals ⁱⁱ are shown to visualize the heterogeneity within the clinal diagnostic categories of not cognitively impaired (NCI) vs. Alzheimer’s dementia (AD) for each of the traits shown. Each dot represents data from one ROSMAP participant. The panels display age at death (A), as well as a range of quantitative measures: amyloid plaque deposition (B), neurofibrillary tangles (C), and cognitive trajectory with age (slope of changes in cognitive scores over age), (D).. E) Mixed pathologies are common within the population of individuals with AD. Shown is a Venn diagram depicting the percent of individuals with each of the listed brain pathologies ⁱⁱ. F) Percentages of the individuals in each of the groups, color coded as in E. To be included in the analysis, individuals had to have both a clinical diagnosis of Alzheimer’s dementia and a postmortem pathological diagnosis of Alzheimer’s disease, as well as determinations for each of the pathologies listed. Created with nVenn [178].

Genetic heterogeneity underlying risk for Alzheimer’s disease

While the majority of individuals with AD have a disease onset in late life, AD can have an onset as early as the 20s, 30s or 40s. The most common form of early-onset AD is trisomy 21 (T21), which causes Down syndrome (DS). Virtually all individuals with T21 develop AD pathology consistent with an AD diagnosis before 40 years of age and nearly half of DS adults have an onset of dementia before 55 years of age (reviewed in [11, 12]). AD in T21 is caused by elevation of the amyloid precursor protein (APP), the gene for which is located on chromosome 21. Increased expression of APP is the major factor contributing to the increase in Aβ plaques in individuals with T21, with Aβ deposition having been observed in DS individuals as young as 12 years old [13, 14]. There are rare cases of partial trisomy 21 (translocation DS) that do not include the APP locus, and these individuals still have the DS phenotype but do not develop AD [15]. Rare micro-duplications of just the APP locus also cause early-onset AD [16–19]. In addition to T21 and APP duplication, many missense mutations in APP or in PSEN1 or PSEN2 (encoding presenilins, the catalytic core of γ-secretase, which cleaves APP to generate Aβ) also cause an early-onset, familial form of AD (EOAD; fAD). These mutations either increase Aβ aggregation directly, increase Aβ levels overall, or selectively raise the relative levels of Aβ42, a longer, hydrophobic peptide more prone to aggregation. These genetic and neuropathological findings place Aβ in an etiological role in AD pathogenesis in individuals with these EOAD mutations.

The vast majority of patients with Alzheimer’s dementia have the “sporadic,” late-onset form of the disease (LOAD), with an onset of symptoms after the mid-sixties. The most significant genetic risk factors identified for AD are variants in APOE. Risk-conferring haplotypes encode APOE epsilon 4 (APOE4); and APOE4 is considered to be the strongest genetic risk factor for late-onset AD. Importantly, however, it should be noted that associations of APOE4 with AD neuropathology and clinical outcome are heavily influenced by ancestry [20]. Recent studies have shown that APOE4 has weaker effects in Blacks than Whites in the United States [21, 22] and that the effect size of APOE4 on AD risk is reduced in African ancestry [23, 24]. Recent genome-wide association studies (GWAS) for LOAD, which incorporated data from over one million subjects, identified over 70 distinct loci that have been associated with LOAD through GWAS [1]. Among these, rare coding variants have been identified in a subset including ABCA7, ABI3, AKAP9, BIN1, CLU, NCK2, NOTCH3, PLCG2, SORL1, TREM2, and UNC5C (reviewed in [25]). Together, these >70 genes document the influence of multiple biological pathways in the pathogenesis of AD (Figure 2). This suggests a potential for different disease-modifying therapeutic interventions depending upon the specific drivers of disease in each individual. It is important to note that there is evidence that several of these LOAD genetic risk factors operate at least in part through the Aβ production and clearance pathways and thus might benefit from amyloid-lowering therapeutics. Below, we describe evidence for the impact of different biological domains in AD risk, which is primarily rooted in genetic findings.

Figure 2. Diverse combinations of genetic factors converge on interrelated but varied biological domains to impact risk and resilience to AD.

(A) Over 70 different loci have been identified through GWAS of LOAD, and there are likely more rare genetic mutations and variants that have not yet been discovered. Across the human population, different combinations of risk and resilience factors converge on different biological domains that have been implicated in AD pathogenesis. Alterations of different combinations of these biological domains in turn impact the neuropathological and cognitive trajectory of individuals. (B) Experimental studies have revealed that many of these biological domains are highly interrelated (for example, endolysosomal dysfunction impacts APP metabolism). Therefore, rather than a single linear road to AD, we conceptualize several different but interrelated molecular pathways that lead to AD in different individuals. The diversity of “roads to AD” manifests in part in the heterogeneity in neuropathological and cognitive outcomes with age. Figure created with BioRender.com.

Genetic risk for Alzheimer’s disease impacts multiple biological domains

We posit that an array of known and unknown genetic variants associated with risk for and resilience to AD interact to have differential impacts upon sets of biological domains. The disruption of one or more of these biological domains in turn influences neuropathological and cognitive phenotypes (Figure 2). For each person, the molecular road from genetic variants and environmental factors to the accumulation of amyloid plaques and tau tangles in the brain and ultimately to dementia may be unique. In turn, disruption of certain biological domains may be the primary driver of AD pathogenesis in individual people. In this section, we describe biological domains impacted in AD, and the genetic evidence supporting their pathogenic influence in mediating risk and resilience to AD. It is important to stress that these biological domains are intimately linked to one another, and disruption of one can lead to disruption of others (e.g. dysfunction in the endolysosomal pathway leads to accumulation of Aβ; dysregulated microglial activation by Aβ can lead to synaptic vulnerability).

APP metabolism: Aβ generation and clearance

As stated above, some of the strongest pieces of evidence for a causal role for Aβ in AD comes from genetic studies of early onset AD (EOAD). Throughout life, APP proteolytic processing in mammals occurs in two simultaneous ways: amyloidogenic and non-amyloidogenic pathways. In the amyloidogenic or Aβ-generating pathway, APP undergoes consecutive cleavages by β- and γ-secretase to generate varying lengths of Aβ peptides (reviewed in [26]). Presenilin (encoded by PSEN1 and PSEN2) is the catalytic subunit of the γ-secretase complex [27, 28] and mediates an unusual intramembrane proteolysis of many polypeptide substrates, including APP. Studies on autosomal dominant missense mutations in EOAD have revealed that mutations in APP, PSEN1, PSEN2 affect the proteolytic processing and the cleavage sites in APP to increase production of total Aβ or longer Aβ peptides (Aβ42 relative to Aβ40, 39, 38, 37) [29–31]. In addition to these genotype-to-phenotype analyses of EOAD, decades of studies in in vitro and in vivo model systems have shown neurotoxic effects of elevated Aβ42 (for more detailed descriptions of these studies, please see other reviews focused upon this topic [32, 33]).

It is less clear whether elevated Aβ generation is playing a role in LOAD cases. Recent work from our group studying neurons derived from induced pluripotent stem cell (iPSC) lines generated from the ROS and MAP cohorts (which include participants with LOAD) has shown that the ratios of long (Aβ42) to short (Aβ37) Aβ generated by neurons from 50 different humans (both non-diseased and “sporadic” AD) are significantly associated with cognitive decline in the donors from whom they were cultured, and that a polygenic risk score for AD is significantly associated with the ratio of Aβ42:37 generated in these neurons [34]. These findings provide evidence that the profile of Aβ peptides generated by neurons is contributing in part to the risk for LOAD, and that Aβ generation may be a biological domain driving disease pathogenesis in genetic subsets of individuals with LOAD [34].

Experimental studies of certain genes conferring risk for LOAD support a casual role for Aβ clearance in LOAD. Aβ clearance involves contributions from many cell types. Aβ is eliminated from the brain via transport across the blood-brain barrier (BBB) into plasma, enzymatic degradation, removal though bulk flow of interstitial fluid and cerebrospinal fluid, and cellular uptake and degradation by glial cells and perhaps also neurons. APOE is an apolipoprotein, and its most well-established function is as a cholesterol transporter. In AD, APOE has been shown to function in part as a mediator of Aβ clearance into astrocytes [35–37]. Data from APOE-humanized mice show that APOE haplotypes variably comprising the ancestral allele (ε3), the AD protective allele (ε2) and the AD risk-conferring allele (ε4) differentially affect Aβ clearance in the brain such that clearance is greatest in mice expressing ε2 followed by ε3 followed by ε4 [37]. Other LOAD risk genes such as PICALM and CLU are also known to mediate Aβ clearance by playing roles in its transport across the BBB [38, 39]. Taken together, these studies support a role for both Aβ generation and clearance downstream of genetic risk factors for LOAD in at least a subset of individuals.

Tau proteostasis/cytoskeletal integrity

Mutations in MAPT, the gene encoding tau, do not result in AD but rather cause other forms of dementia such as frontotemporal dementia (FTD). However, accumulation of excessively phosphorylated wild-type (wt) tau is a hallmark of AD, and PET studies have shown that the regional pattern of accumulation of tau in the brain can follow local Aβ accumulation and shows a strong association with cognitive decline in individuals developing AD [40–42]. Given tau’s strong association with cognitive impairment, it has been postulated that complex gain -and/or loss-of-function mechanisms in the brain may be triggered by tau aggregation and accumulation in neuronal perikarya and neurites. Thus, proteostasis of tau is a biological domain clearly impacted in LOAD.

Tau is a microtubule-binding protein that functions in part to stabilize and facilitate the assembly of the microtubule cytoskeleton in neurons. Structurally, the tau protein is roughly divided into two major parts: (i) the tau projection domain in the N-terminal region and (ii) the repeat domains responsible for tau’s binding to tubulin in the C-terminal region. The function of the N-terminal region remains incompletely defined, but some studies have suggested that it can regulate tau polymerization, the cellular localization of tau [43], and microtubule stabilization [44]. Tau post-translational modifications also regulate tau proteostasis. Tau phosphorylation, for example, reduces its microtubule-binding ability, which contributes to both its aggregation and disruption of the cytoskeleton. MAPT knock-out mice are viable and fertile, and brain development is not grossly altered. Nonetheless, several neuronal phenotypes have been reported despite the likely functional redundancy of tau with other microtubule-binding proteins [45–49].

Tau degradative enzymes, the ubiquitin-proteasome system (UPS), and the autophagy-lysosome pathway (ALP) each contribute to the turnover of tau, which in turn is important for maintaining neuronal health (reviewed in [50, 51]). A gain-of-toxic-function of aggregated tau likely plays a central role in tau-induced neurotoxicity [52]. Accumulation of pathologically aggregated tau has been shown to affect brain cells in numerous ways including: impairment of the protein degradation machinery [53–59], disruption of synaptic function [60], induction of mitochondrial dysfunction and oxidative injury [61], alteration of nucleocytoplasmic transport [62], dysregulation of RNA-binding proteins [63], and promotion of genomic instability [64, 65]. Thus, certain LOAD genetic variants that act to reduce tau turnover would likely contribute to AD risk. In accord, our group showed that polygenic risk for AD was associated with protein levels of various proteasome components across a cohort of iPSC-derived neurons from LOAD cases [34].

Endolysosomal dysfunction and retromer trafficking

GWAS of LOAD have identified several genes important in the process of endolysosomal transport, including BIN1, CD2AP, PICALM, and SORL1. Analyses of postmortem brain tissue, stem cell derived neurons and organoids all support the hypothesis that endosomes are enlarged and altered in AD neurons [66–69]. Based on these data, it has been suggested that an “endolysosomal traffic jam” may be a central cellular event in AD pathogenesis [70, 71].

Among the known LOAD risk loci, SORL1 is best known as a neuronal sorting receptor, and both SNPs and coding mutations in SORL1 have been identified in numerous studies to be associated with LOAD [72–75]. SORL1 mediates intracellular trafficking of APP (and many other substrates), thereby regulating the cleavage of APP and generation of Aβ [76]. Recent studies using iPSC-derived neurons have shown that reduction or loss of SORL1 leads to impairment of endosomal trafficking, autophagy, and endosomal recycling [77–80]. PICALM encodes a clathrin-binding protein and mediates trafficking between the trans-Golgi network and endosomes [81, 82]. BIN1 regulates trafficking-related functions such as clathrin-mediated endocytosis and synaptic vesicle recycling, and loss of BIN1 results in increased size of early endosomes and impaired trafficking of β-secretase in a manner that enhances Aβ generation [83–86]. CD2AP encodes an adaptor protein that has a role in endocytosis and lysosomal trafficking [87]. CD2AP levels affect the degradation rate of APP, with reduced CD2AP levels resulting in a reduction in Aβ [88] and thus accumulation of uncleaved APP in early endosomes [89]. These studies suggest a convergence downstream of variation in these LOAD-associated genes on endolysosomal dysfunction. In this regard, Aβ is normally generated in large part in recycling endosomes, which have the mildly acidic pH necessary for the two aspartyl proteases which generate Aβ throughout life (BACE1 and Presenilin/γ-secretase) to be protonated and thus be catalytically active.

Immune function/microglial activation

Many of the loci identified by GWAS in LOAD are associated with innate immune function, with several genes being highly expressed within microglia, monocytes and or macrophages. These genes include ABI3, CD33, CR1, HLA-DRB5-DRB1, INPP5D, MEF2C, MS4A4A, MS4A6A, PLCG2, SORL1, SPI1 and TREM2 [90–92], among others (Figure 4). Of these genes, TREM2, PLCG2, ABI3, and SORL1, have received the most experimental attention, as rare coding variants in these genes have been identified and linked to AD [93–95].

Figure 4. A variety of experimental systems are necessary to capture different aspects of disease pathogenesis.

Each experimental system has its advantages and disadvantages. Shown are three examples of genetic risk/protective factors for AD, and the direct and indirect molecular consequences of these variants on biological domains. For the identification of the proximal, cis effects of specific genetic variants on RNA and protein levels, having homogenous, robust cultures of a single cell type can be of the highest utility for two reasons: 1) reductionist systems are generally more robust and reproducible due to the simplified nature of the system and 2) the cis-effects observed are less likely be confounded by the expression of other, downstream consequences of the variant due to the presence of other cell types. For the same reasons, these same monocultures are useful for capturing the direct, cell autonomous consequences of genetic variants on a particular biological domain. However, more complex systems such as co-cultures of multiple human cell types or organoids may be needed for capturing the downstream, indirect effects on secondary biological domains if interactions between multiple cell types are involved. Animal models are necessary for capturing system-level dysfunction that impacts upon cognition and behavior. However, animal models are less useful for studying the direct effects of genetic variants associated with risk for human diseases, as it is important to interrogate these effects in the context of the human genome. Finally, no experimental system is able to precisely model Alzheimer’s dementia, and ultimately the determination of causality for dementia can only come from human clinical trials that target that biological domain. Three examples are provided to outline models that were proposed based upon data from human cellular and animal experimental systems. Example #1 summarizes findings from a number of studies using human iPSC models and animal models, which have shown that SORL1 coding variants (including truncation variants) can lead to reduced SORL1 protein levels and impaired retromer transport, which in turn affects the accumulation of Aβ and phosphorylated tau (reviewed in [180]). The second example summarizes studies from human cells that revealed that a LOAD-associated variant in CD33 affects CD33 splicing [181], reduced cell surface CD33 levels [182, 183] and induced an impairment of Aβ clearance [182, 184]. Finally, the third example highlights the results of a recent study that showed that high polygenic risk score across a large cohort of human iPSC derived neurons was associated with elevation of longer Aβ peptides, which in turn induced a reduction in protein phosphatase 1 (PP1) level and altered tau proteostasis [34].

Importantly, convergent signaling pathways have been described linking subsets of myeloid LOAD risk factors to one another mechanistically (for example, TREM2 and PLCG2 [96, 97]; CD33 and INPP5D [98, 99]; TREM2 and INPP5D [100]). Broadly, analyses of experimental systems modeling these LOAD coding variants implicate dysregulation of immune-related pathways and the phagocytic capacity of myeloid cells (reviewed in [101]) as additional biological domains affected in LOAD. Aberrant phagocytosis (e.g., as shown for CD33 [102]) can have multiple detrimental consequences, including a reduction in the clearance of apoptotic cells and other cellular debris, including Aβ aggregates, and the abnormal pruning of certain synapses (discussed in the next section). In addition to altered phagocytosis, altered functioning of these LOAD-associated microglial genes can lead to the release of cytokines and other factors that can in turn have detrimental effects on other brain cell types. For example, reduction in INPP5D levels in iPSC-derived human microglia induces sublethal activation of the inflammasome and release of IL-1β and IL-18 [99]. The genetic landscape of LOAD clearly implicates this biological domain as having a causal influence on disease risk and/or progression and emphasizes the need to unravel the specific role of microglia in disease progression.

Synaptic vulnerability

Loss of synapses is quantitatively correlated with cognitive decline in LOAD (reviewed in [103]). Vulnerability of synapses in AD may be influenced by genetic risk factors acting autonomously in neurons, resulting in a destabilization of synapses and/or in microglial cells to induce abnormal synapse pruning. During normal brain development, microglia utilize the complement system to tag and prune excess (less functionally active) neural synapses and refine neural networks [104, 105]. Microglia have been shown to be involved in engulfing synapses in early stages of AD through the use of the classic complement system. Dysregulation of complement signaling can lead to an increased phagocytosis of synapses and loss of synapses. Intriguingly, variants at the complement receptor 1 (CR1) locus were initially identified to be associated with LOAD through GWAS [1, 90, 106, 107], supporting the concept that abnormal microglia-mediated synapse destruction could be one of the molecular roads leading to LOAD. Data from APP mutant transgenic mice revealed that complement binding and phagocytosis by microglia results in early synaptic loss in transgenic mice with plaque deposition [108]. Evidence also exists in model systems that protein turnover at presynaptic sites is impaired in the presence of elevated Aβ levels, which also may contribute directly to synaptic vulnerability and synapse pruning [109–111]. Further, astrocytes play a critical role in synaptic function and also play an indirect role in complement-mediated synapse elimination. Thus, dysregulation on biological domains in several cell types, including microglia and astrocytes, may converge on synaptic vulnerability.

BBB integrity/vascular function

Blood-brain barrier (BBB) dysfunction has been proposed to play a causal role in some forms of cognitive decline and dementia (reviewed in [112]). Accumulation of Aβ is clearly a common link between AD and cerebral amyloid angiopathy (CAA), both of which can result in dementia but perhaps with differences in the specific mechanism of brain injury (reviewed in [113]). CAA is defined by fibrillar Aβ accumulation in the walls of cerebral and meningeal vessels, which in turn can affect BBB integrity [114–117]. In many cases of EOAD and LOAD, CAA also occurs, and mouse models of AD also show evidence of breakdown and dysfunction of the BBB [37, 38, 118–120], Importantly, a causal role of BBB dysfunction in LOAD pathogenesis has been suggested by genetic studies that have identified associations with several genes known to be expressed by and function in BBB cells (endothelial cells, pericytes, astrocytic end feet), including APOE, CLU, and PICALM (Figure 3). APOE can bind to Aβ and help mediate its transport across the BBB via low density lipoprotein receptors. Data from transgenic mice expressing human APOE isoforms showed that compared to mice with APOE2 and APOE3, mice with APOE4 show reduced Aβ clearance across the BBB [37, 38, 121]. CLU encodes the apolipoprotein clusterin, also known as APOJ. Like APOE, APOJ can bind to Aβ and appears to be the major carrier protein of Aβ in human plasma and CSF [122]. PICALM, another gene associated with LOAD via GWAS, is expressed in pericytes and endothelial cells that (among other cell types) comprise the BBB, and it has been shown to mediate Aβ transcytosis across the BBB [39].

Figure 3. LOAD risk genes are expressed strongly in a variety of cell types in the brain.

Risk-conferring and protective genetic factors for AD are expressed in a variety of cell types, underscoring the idea that there are multiple molecular roads to AD. Shown are examples of iPSC-derived brain cell types (images courtesy of the Young-Pearse lab, for illustration purposes). LOAD GWAS hits expressed in each cell type are noted within the boxes. Expression was determined by probing a large database of single cell RNAseq data obtained from over 400 samples of ROSMAP human brain tissue [179] (see also ⁱⁱⁱ ). Data from all participants were aggregated and pseudobulk scores calculated to determine if the listed gene was in the upper (green) or lower (black) quartile of expression relative (abundance) to all genes detected in that cell type. If a LOAD GWAS gene is not listed under a given cell type, then it was not detected in that cell type in the snRNAseq dataset.

Lipid homeostasis

Disruption of lipid homeostasis has been identified in AD, and its association is supported by the identification of several LOAD genetic risk factors encoding apolipoproteins and lipid transporters. For example, APOE and CLU are associated with AD, and these lipoproteins serve key physiological roles as lipid carriers. In addition, GWAS have identified two members of the ABCA family (ABCA1 and ABCA7) as risk genes for LOAD [123, 124]. ABCA (ATP-binding cassette subfamily A) transporters play a role in maintaining cholesterol homeostasis in the brain. Both ABCA1 and ABCA7 are responsible for loading cholesterol onto the apolipoproteins [125]. Loss of ABCA1 results in a decrease in APOE protein levels and APOE lipidation [126, 127].

Furthermore, ABCA1 deficiency increased Aβ deposition in an AD mouse model [128, 129]. LOAD SNPs at the ABCA7 locus were recently shown to be associated with altered ceramide metabolism and cognition, and ABCA7 knock out mice showed similar dysregulation in ceramide metabolism [130, 131]. SORL1 also affects apolipoprotein levels and lipid homeostasis, with loss of SORL1 in neurons resulting in a reduction in CLU and APOE protein levels and an accumulation of lipid droplets (LD) [80]. Recent studies have highlighted the important role that microglial lipid homeostasis plays in AD. LD accumulate in microglia with age, and these LD-accumulating microglia show defects in phagocytosis and secrete pro-inflammatory cytokines [132]. Importantly, in Tau P301S transgenic mice, the inclusion of an APOE4 transgene induces lipid accumulation in microglial lysosomes, which in turn affects lysosomal function [133]. Further, APOE4-mediated microglial lipid dyshomeostasis results in disruption of neuronal network activity [134]. These studies suggest the importance of lipid transporters and regulation of lipid homeostasis as a potential pathogenic driver for at least a subset of AD cases.

RNA splicing

Proteomic and RNA profiling of postmortem brain tissue in large well-studied cohorts have revealed strong associations between AD and RNA splicing and stability factors. In postmortem AD brain tissue, core spliceosome components are enriched in insoluble protein fractions and closely associated with neurofibrillary tangles [135–137], although these observations were made at the end of the disease process. Cytoplasmic mislocalization of spliceosome components and aberrant mRNA splicing have also been associated with neurofibrillary tangles in AD brain tissue ([135, 137, 138]. Analyses of large human aging cohorts reveal hundreds of aberrant pre-mRNA splicing events in AD brains, and perturbation of tau may be sufficient to cause similar splicing errors in human neuronal culture and fly models [139–141]. To date, LOAD genetic studies have not clearly implicated specific genetic variants in spliceosome components or regulators in AD pathogenesis. On the other hand, mutations of RNA-binding proteins implicated in pre-mRNA splicing, including TDP-43 and FUS, have been identified as causal in frontotemporal dementia (FTD) and amyotrophic lateral sclerosis (ALS) [142–148]. These findings in human brain tissue coupled with studies in mouse models, suggest that loss-of-function in spliceosome components can lead to pre-mRNA splicing errors and ultimately neurodegeneration.

Mitochondrial and metabolic homeostasis

Aging is the greatest risk factor for AD. The quality and activity of mitochondria have been shown to decline during normal aging in humans. Neurons have a high metabolic demand, making them especially vulnerable to these age-related reductions in mitochondrial function. Evidence of cerebral mitochondrial dysfunction and glucose hypometabolism early during AD progression have been reported (reviewed in [149]), along with evidence of elevated oxidative stress such as increased mutation of mitochondrial DNA and damage to membranes. Mitochondria in postmortem AD brain show altered morphology as well as a reduction in mitochondrial proteins (reviewed in [,149],[8, 150]). It is unclear whether the reported findings of mitochondrial impairment in AD represent primary mechanisms driving risk for AD and/or whether mitochondrial dysfunction is occurring secondary to earlier impairment of other biological domains, such as those that result in elevations of Aβ, aggregated tau, microglial dysfunction and/or disruption of other biological domains described above. Regardless, mitochondrial homeostasis is clearly yet another biological domain affected in AD, and a reduction in mitochondrial function is likely to contribute to AD risk or progression in at least a subset of individuals.

Potential convergence of dysregulated biological domains on Aβ plaque and tau tangle accumulation, synaptic loss and dementia

A key question that arises when considering the consequences of genetic risk factors for AD on biological processes is the following: Are the conserved neuropathologies in AD (Aβ plaques and tau tangles) the central pathways required for development of Alzheimer’s dementia, with the other genes and biological domains modifying progression or risk of AD, or can these other biological processes cause AD by themselves? Alzheimer’s disease is defined by the presence of plaques and tangles, and thus disruption of these other biological domains alone would not be expected to cause AD per se without also affecting the accumulation of Aβ and tau. However, other types of dementia can occur in the absence of plaques and tangles, and therefore it is possible that disruption of these other biological domains could impact upon synapse loss and/or dementia directly. A useful thought exercise is to consider the following: if there were aged humans who no longer expressed the APP or MAPT gene, could these individuals develop dementia due solely to disruption of one or more of the “other” biological domains (apart from Aβ accumulation or tau proteostasis) outlined in Figure 2? We hypothesize that it would indeed be possible. This hypothesis is reinformed by consideration of other types of dementia that occur in the absence of plaques and tangles, and a variety of animal models that show dementia-like cognitive phenotypes with disruption of some of the genes and biological domains outlined in Figure 2.

Experimental approaches for studying subtypes of AD

In this section, we highlight some of the experimental approaches used to identify and study AD pathogenesis at a molecular level. These are classified into three major categories: -omics level studies of human postmortem brain tissue, studies of rodent models, and human iPSC-based experimental systems.

Large-scale studies of human brain tissue

Large scale multi-omics studies of postmortem brain tissue can highlight particular pathways and proteins that are associated with degrees of pathology, cognitive decline, and clinical AD diagnosis [8, 151–153]. These studies are highly valuable in providing data from the actual target organ affected in the disease. Data sets are currently available encompassing hundreds of samples of aged brain tissue from individuals with and without AD and dementia across several cohorts using bulk and single nuclei RNAseq, proteomic profiling, epigenetic profiling and metabolomic profiling through the AMP-AD Knowledge Portalⁱ [154]. These datasets have helped identify RNA and protein modules of co-expression that are altered in AD, and integration with quantitative neuropathological measurements has allowed researchers to disentangle associations between these networks and AD pathology and cognitive decline and resilience. These resources reveal a plethora of possible factors that may influence AD risk and progression. To validate and test mechanistic hypotheses generated from these studies, establishing experimental systems that measure biological processes implicated in LOAD are critically important.

Late-onset Alzheimer’s disease mouse models

Multiple early-onset AD mouse models have been generated by overexpressing mutant human APP, PSEN1 or MAPT, to study the molecular mechanisms in the brain and test therapeutic interventions, but so far, these approaches have not shown strong translatability for rescuing the cognitive impairments associated with LOAD. A major contributor to this problem is likely that most of these mouse lines only model limited aspects of disease (for example, Aβ accumulation). Improved mouse models more closely relevant to LOAD are clearly needed. Among the first such efforts established valuable sets of humanized APOE replacement mouse lines that harbor each of the three APOE alleles [155, 156]. These lines were crossed to humanized MAPT and APP/PSEN1 models, and the offspring have yielded important insights into the role of APOE in AD pathogenesis (reviewed in [157]). A similar approach underway for advancing mouse models of LOAD involves the generation of new lines that harbor novel combinations of LOAD genetic risk variants [158, 159]. The resultant consortium, called Model Organism Development and Evaluation of LOAD (MODEL-AD) aims to build new LOAD mouse models that can successfully predict the effectiveness of novel therapeutic approaches by modeling interactions between multiple genetic risk variants [160–162]. Applying CRISPR/Cas9 technology and a Cre/lox system, MODEL-AD researchers have generated mouse models carrying LOAD-associated risk variants in ABCA7, ABI3, ADAMTS4, APP, BIN1, CD2AP, CEACAM1, CLASP2, CR1, CR2, EPHA1, ERC2, IL1RAP, IL34 KIF21B, KLOTHO, MEOX2, MTHFR, MTMR4, PICALM, PILRA, PLCG2, PLXNB1, PTK2B, PTPRB, SCIMP, SLC6A17, SORL1, SNX1, and SPI1 in mice with pathogenic TREM2 (R47H) and APOE (ε4) variants. In addition to providing these lines to the scientific community, pathological phenotypes and transcriptomic profiles of the mouse brains generated in MODEL-AD also are available to the community. These unbiased approaches to analyzing these mouse models are important complements to the standard measures of cognition and plaque and tangle burden. Also in recent years, an ambitious project at the Jackson Labs introduced a novel AD mouse population by combining the well-established 5xFAD mouse model of AD with genetically diverse reference strains [163]. This collection of models revealed that diverse strains of mice show differential vulnerability to the introduced fAD mutations as to both accumulation of neuropathology and cognitive outcomes [163]. By incorporating genetic diversity, this resource provides a unique in vivo model for studies of LOAD pathogenesis.

Human iPSC-derived experimental systems

For the study of disease pathogenesis, having a human-based experimental system is necessary for capturing genetic variants in the context of the human genome. Recent developments in iPSC technology have afforded the opportunity for enhanced modeling of the cell biology responsible for neurological diseases by enabling in vitro analysis of human brain cells (reviewed in [164]). Disease mechanisms can be probed both by the genetic manipulation of iPSCs and by modifications of their cellular environment. Through these types of iPSC-based studies, the impact of variation in specific genes on the biological domains described above can be readily studied. Until recently, most iPSC studies focused on using these cells to study highly penetrant mutations in early-onset AD. These studies showed consequences of APP and PSEN1 mutations on Aβ generation and tau phosphorylation as well as other cellular phenotypes [67, 69, 165–170]. More recently, iPSC-based studies have supported a role for APOE in lipid homeostasis, synaptic integrity, and inflammation [171–174]; for SORL1 in endolysosomal dysfunction and retromer trafficking [77–79]; for PICALM in endocytosis in astrocytes [175] and in pericyte functioning at the BBB [39], and for BIN1 in neuronal endocytosis [86] and the regulation of the proinflammatory response in microglia [176]. A large-scale effort called the “iPSC Neurodegenerative Disease Initiative” (iNDI) is underway to generate a set of iPSC lines in a single genetic background for the study of over 100 variants associated with LOAD and AD related dementias (ADRDs) [177]. A set of three isogenic lines will be generated for each variant that includes homozygous mutation, heterozygous mutation, and a revertant line in which the mutation is edited back to the wild-type allele. These lines then will be differentiated to various brain cell fates and analyzed using a number of platforms that include the acquisition of transcriptomic and proteomic profiles [177]. Many of these lines are already available to the scientific community, and along with the data sets to be generated from these lines, this work will serve as a valuable resource for studying AD and ADRDs.

iPSC technology, beyond studies of the individually expressed genes, provides a unique opportunity to define the molecular underpinnings of person-specific variation in brain neuropathology and cognition. While GWAS have identified numerous SNPs associated with LOAD, the identified variants together explain less than half of the global genetic variance in LOAD susceptibility, and most individual variants do not measurably influence disease course. The ability to derive iPSCs from an array of well-characterized subjects allows for a closer examination of the mechanism of disease progression in particular subsets of people who likely have both known and unknown genetic influences increasing their LOAD risk. The power of iPSC technology lies in part in its ability to capture the genetics of an individual in a well-controlled and manipulable living experimental system. Through analyses of diverse brain cell types derived from iPSCs, the contribution of genetic variants to biological domain function and dysfunction can be isolated and studied in relevant cells. Such studies may lead to both the refinement of known pathways implicated in AD and to the identification of new pathways heretofore not associated with disease.

A potentially fruitful approach would be to combine iPSC technology with epidemiology to examine “person-specific” differences in a large collection of lines generated from humans with known clinical, pathologic, and genomic heterogeneity. This approach provides an excellent opportunity to capture AD heterogeneity in a manipulate experimental system. To this end, our group has recently reported the generation of iPSC lines from 53 deceased individuals in the ROS and MAP cohorts that span the genomic, clinical, and pathological spectrum of aging and AD [34]. We differentiated these iPSC lines to cortical neuronal fates (iNs) and obtained RNA and protein profiles of the iNs as well as quantitative measures of APP cleavage products and phosphorylated tau proteoforms. These iPSC-derived data showed significant concordance with data obtained from brain tissue of the same individuals. Through this approach, we identified multiple layers of association between the aged brain and human neuronal cultures that strongly support the utility of iPSC-derived systems to interrogate mechanisms of neurodegenerative disease. We found associations between measures of Aβ and tau in these iNs with rate of cognitive decline in individuals from whom they were derived. Further, we observed concordance between iNs and brain tissue in the expression of AD-associated RNA and protein modules. Through these analyses, the iN experimental system was shown to be valuable for both the identification and validation of protein phosphatase 1 (PP1) as a molecular link between LOAD polygenic risk score (PRS), Aβ, and tau [34]. Future studies in this and other cohorts of iPSC lines will facilitate the study of biological domains outlined in the previous sections in a variety of cell types. Perhaps more importantly, these experimental systems provide a unique pre-clinical system to test person-specific responsiveness to therapeutic interventions.

The known genetic risk factors for AD are expressed highly in several different cell types (Figure 3). One of the strengths of the iPSC system is that these cells are pluripotent, meaning that they can be directed to become any cell type in the body. Increasingly more complex experimental systems are being developed using iPSCs to generate microphysiological systems that recapitulate various aspects of brain biology. Each of these experimental systems -- monocultures of single cell types, multi-cellular co-cultures of different cell types, organoids, and brain-on-chip systems -- have their utility in studying AD pathogenesis (see Figure 4). As outlined above, the highly reductionist system of iPSC-derived monocultures (i.e., of pure neurons) can be markedly useful for capturing the processes immediately downstream of genetic variants; for example, the RNA and protein changes that are associated with genetic mutation or variant at a given locus. They also are useful for quantifying direct, cell-autonomous effects of genetically encoded influences on biological domains. However, more complex cellular systems such as co-cultures of multiple cell types, assembloids, and organoid models are helpful for capturing intercellular communication between cell types and/or indirect effects on systems-level processes (i.e. examining effects of genetic variants on synaptic pruning through co-culture of neurons and microglia). An obvious limitation of these human in vitro systems is that they do not allow for the assessment of behavioral and cognitive outcomes, which require in vivo animal models. No single experimental system, however, fully captures Alzheimer’s dementia, and one of the challenges facing the field is determining whether pathways identified using the forementioned experimental models are relevant to processes in the AD brain. While definitive validation may only come by engaging the relevant targets in clinical trials, one approach to raise confidence for disease relevance is the use of large-scale studies of human brain tissue to examine if concordant findings are observed.

Concluding Remarks

It is becoming increasingly appreciated that there are multiple molecular pathways that contribute to Alzheimer’s dementia and that a single therapeutic strategy may not be sufficient for the modification of disease trajectory across all patients. Studies are beginning to emerge that attempt to provide frameworks for subtyping AD based upon biomarker data, genetic data and/or RNAseq and proteomic profiles from the postmortem brain or systemic cells and fluids. New experimental systems that capture human genetic diversity and allow for the assessment of multiple biological domains impacted by known and unknown AD genetic risk and resilience factors may prove to be highly valuable for identifying and testing therapeutic strategies in diverse genetic backgrounds (see Outstanding Questions). With the emerging success of some anti-amyloid treatments and the continued development of tau-targeting therapies, it is important to consider additional strategies that target other biological domains affected in different subtypes of AD. Through the use of human cell-derived in vitro systems and additional approaches reviewed above, the development of subgroup-specific strategies for the prevention and treatment of AD may become a reality.

Outstanding Questions.

Could anti-amyloid and/or anti-tau therapies administered at a sufficiently early stage prevent the emergence of AD or stall its progression, at least in some cases?

What would be the optimal timepoints for administering therapeutic interventions that target different biological domains implicated in AD risk? Different mechanisms driving Alzheimer’s dementia may operate at different stages of the disease process, and the most efficacious intervention strategies may differ based on the disease stage.

Can patient iPSCs be used to stratify patients with MCI/AD based on the dysfunction of specific biological domains (for example, by quantifying Aβ generation, microglial reactivity and phagocytosis, the efficiency of endolysosomal trafficking, etc.)?

What would be the ways to determine the subtype or cellular stage of AD for each person, and to determine which treatment might be most suitable for a specific individual (i.e., “precision Alzheimerology”)? Can iPSC studies be used to identify combinations of genetic variants driving AD subtypes, such that a directed genome analysis can be used to subtype individuals?

Highlights.

Alzheimer’s disease (AD) is heterogeneous as regards the age of onset, rates of progression, and the order in which cognitive domains are affected. Further, a diverse combination of cognition-relevant neuropathological phenotypes occurs among people with AD.

Over 70 genetic loci have been associated with AD, and genes at these loci point to different biological processes and pathways in a variety of cell types that contribute to risk and resilience in AD.

Improved experimental systems that capture biological diversity are necessary to disentangle subtype-specific drivers of AD. Systems such as human iPSCs and new rodent models will be critical tools for elucidating molecular mechanisms leading to dementia, testing new therapeutic strategies, and determining subtype-specific responsiveness to interventions.

Matching the right treatment or combinations of treatments to the appropriate population is a key challenge for the next wave of therapeutic development for AD.