Alzheimer’s disease (AD) is a chronic neurodegenerative disorder causing memory loss and impairments in cognitive and behavioral functioning. AD is the most common form of dementia, which affects around 55 million people worldwide. The disease is characterized by depositions of β-amyloid (Aβ) into plaques, hyperphosphorylated tau (ptau) forming neurofibrillary tangles (NFTs), dystrophic neurites, synaptic loss and atrophy of neurons and brain regions. All these changes ultimately trigger the cognitive decline and behavioral symptoms of AD [1]. Dominantly inherited familial AD (fAD) has been linked to mutations in the genes amyloid precursor protein (APP) and presenilin 1 and 2 (PS1 and PS2), which play an important role in generation of Aβ aggregates. fAD accounts for less than 1% of all cases, while the far more common sporadic and late-onset AD (LOAD) has most likely a multifactorial aetiology including genetics, lifestyle, level of education and unknown external factors [2,3,4]. In the absence of effective treatments, interventions focusing on delaying the onset of AD by activation of the brain have gained more and more attention [5, 6]. Lifestyle factors like cognitive, social and physical activities might have the potential to postpone AD [7,8,9,10].

It is widely assumed that the neuropathological hallmarks of AD are causally related to cognitive decline in AD patients. This holds in particular for NFTs [11]. However, several studies have indicated a disjunction between the degree of AD pathology and its clinical manifestations. Early studies discovered that some individuals, characterized as cognitively normal, had advanced AD pathology at post-mortem examination [12, 13], which was later substantiated in cognitively intact individuals with longitudinal assessments of cognition and post-mortem assessments of pathology [14,15,16,17]. These findings led to the hypothesis that some individuals might have a “reserve” that allows them to cope with neuropathology and remain cognitively intact. Furthermore, epidemiological evidence showed a reduced risk of dementia in individuals with higher educational or occupational attainment [18, 19], IQ [20] and participation in leisure activities [21]. It was hypothesized that certain individuals exhibit progression of AD pathology while their lifelong experiences allowed them to cope better with disease-related changes.

The first aim of this review is to discuss the definitions of reserve to AD in the context of the epidemiological and experimental evidence. While it is hypothesized that the underlying mechanisms might be rooted in structural and functional brain mechanisms, more and more data on the cellular and molecular substrate of a reserve in AD have been published. Therefore, the second aim is to review the emerging cellular and molecular evidence of reserve. Insight into the cellular and molecular mechanisms that govern resilience to AD may be the starting point for the development of better treatments for AD and is therefore an important field of research.

In its most essential form, a reserve allows an individual to remain cognitively intact despite the presence of extensive AD pathology. Over the years numerous terms have emerged to describe this phenomenon, such as reserve, resilience, non-demented Alzheimer neuropathology (NDAN), clinically silent AD, pre-symptomatic AD, asymptomatic AD or preclinical AD. All of these definitions have been used to describe individuals with AD pathology that are cognitively normal, but whereas preclinical AD has been primarily used in living subjects that are biomarker positive but not necessarily resilient, the former definitions have all been used for individuals at autopsy. Importantly, it has been proposed that resilience is different from resistance, which refers to the absence or lower level of AD or comorbid neuropathology relative to the expected frequency or severity based on age, genetics or other characteristics [22]. For example, as levels of Aβ plaques and ptau have been shown to increase with age, individuals at advanced ages, like centenarians, are expected to have high levels of AD pathology. However, among cognitively intact centenarians, there are individuals without any Aβ plaques in the brain, which are resistant to developing Aβ plaques. There are also cognitively intact centenarians with similar amounts of AD pathology as an demented AD patient, which are resilient [23, 24]. Researchers often attempt to describe the phenomenon of reserve from their own perspective. For example, NDAN is used by researchers using post-mortem tissue, in which cognition is not always longitudinally assessed, making it impossible to make claims about maintenance of cognition. In the field, the different terminologies mentioned here are often used interchangeably, making it difficult to align relevant mechanisms that have been found so far. To develop potential consensus definitions, a recent workgroup has subdivided reserve into three concepts: cognitive reserve (CR), brain reserve (BR) and brain maintenance (BM) [25].

The CR hypothesis postulates that individuals with a reserve process cognitive tasks in a more efficient manner. It is hypothesized that individuals with CR have increased adaptability to cope with neuropathology through specific processes, like increased functional efficiency of brain networks. BR relates to anatomical differences like a higher number of neurons, synapses or other structural brain resources by which individuals thus can withstand more atrophy or synaptic loss before clinical manifestations occur. Finally, the definition of BM has emerged to illustrate the notion of preservation of brain morphology or absence of neuropathological change over time [26]. In BM, processes like neurogenesis, repair mechanisms or removal of pathology by glial cells might play an active role.

Definitions of CR, BM and BR have been proposed as a general framework, which can be used to harmonize alternative definitions or be seen as equivalent to definitions used by other researchers. Thus far, CR, BM and BR have primarily been used in clinical studies focusing on imaging and fluid biomarkers. Lifelong experiences, such as years of education, IQ, social interactions, complexity of occupation, leisure activities and socioeconomic status, are often used as proxies for reserve as they are associated with a later onset of AD and with processes that can be linked to the proposed definitions. For example, more years of education has been linked to larger brain volumes [27, 28], related to BR. However, the use of such proxies often results in a biased measure and affects outcome measures independent of reserve. For example, more years of education has been directly linked to better health [29], and individuals with higher education levels may perform better in cognitive status questionnaires. Furthermore, the role of these proxies in resilience remains unclear as it was recently demonstrated that IQ explains more variation in rate of cognitive decline than years of education [30], suggesting that IQ might be a better proxy for resilience than years of education. To overcome these issues, more recent studies have used multiple proxies to combine lifelong exposures that are related to resilience into a composite score [31,32,33]. In addition, others have implemented more direct measures related to resilience and neural mechanisms by using functional magnetic resonance imaging (fMRI) or changes in fluid biomarkers to identify alterations in network activity or relevant brain changes, respectively. This approach has been labeled the “neural implementation” of CR [25].

Researchers focusing on molecular and cellular mechanisms often use the definitions resilience or resistance, which, according to the proposed framework, could be attributed to CR or BM, respectively. The notion of resilience has also been recapitulated in animal models by identifying animals with better cognition than expected based on age or pathology [34, 35]. Repetitive cognitive training or enriched environment improved cognition in animal models of AD [36, 37], underlining the idea that increased plasticity through cognitive training contributes to resilience. The influence of positive novel experiences (such as enriched environment) or negative experiences (stress, early life adversity) on cognition has been well established in rodents [38,39,40,41]. Researchers often attempt to characterize mechanisms that have been identified in animal studies as either CR or BM. For example, in young and aged rats, which were either cognitively impaired or unimpaired based on behavioral tests, genes related to age were correlated with cognition to identify genes related to CR [42]. The authors concluded that upregulated genes, counteracting aging stressors that impair cognition, such as neuroinflammation and oxidative stress, and downregulated genes, related to nervous system development, reflect adaptive changes in the circuit to preserve cognition. However, when focusing on such complex molecular or cellular mechanisms, it becomes difficult to relate the mechanism back to the proposed definitions. For example, neuroinflammation in the ageing brain can both be linked to exacerbation of neuropathology or to synapse pruning [43], and thus influences both CR and BM.

When focusing on the molecular and cellular mechanisms of resilience, an analogy can be made to what has been called the cellular phase of AD [44]. In this concept there is a complex cellular phase, in which cells respond to Aβ and ptau aggregations. In this phase, which possibly could last for decades, there is a gradual shift from the initial reversible physiological reactions to pathology to irreversible compensation mechanisms, which could be independently of Aβ and ptau. This would consequently disturb the brain homeostasis and lead to clinical symptoms. In resilience, cellular reactions to pathology might not develop into irreversible changes, resulting in maintenance of brain homeostasis. The absence of some of these irreversible cellular changes can be linked back to CR and BM, such as loss of synaptic inputs evolving into alterations in connective patters or the initial clearance of pathology evolving into clearance dysfunction. The proposed definitions are well suited to describe changes on the macroscopic level (e.g. changes in brain volume or circuit changes) or microscopic level (changes in the amount of pathology or the amount of synapses). However, the possible upstream molecular mechanisms orchestrating the observed changes in CR and BM are often complex, rooted in both types of reserve and not fully elucidated. Upstream effectors such as transcription factors might have such a broad effect that it will influence both resilience or resistance, or CR and BM, making it impossible to distinguish between the two. Hence, when focusing on CR and BM, there is a possibility that the downstream effects of the cellular and molecular mechanisms are observed, but not the fundamental mechanism itself.

When investigating molecular effects to such an extent, their link to phenotype is sometimes difficult to establish. Studies focusing on molecular effects of Aβ and tau on individual cells often refer to cellular resilience, in which specific cell types are not affected by pathology. For instance, in post-mortem tissue, different excitatory neuron subtypes were found to be more resilient to tangle formation [45], or altered levels of proteins such as mitofusin 2 (MFN2) or RAR Related Orphan Receptor B (RORB) were associated with resilience or vulnerability to tangles [46, 47]. These resilient cellular subtypes or mechanisms cannot always be linked back to phenotypic traits, such as cognition. Furthermore, the challenge with these studies is to translate a cellular view of resilience back to an overall view of resilience in an organism. It is often unknown how molecular changes in specific cells, often studied in only one brain area, influence a potential brain-wide phenomenon. Currently, molecular and cellular mechanisms are often attributed to resilience or resistance due to correlations with cognition or the amount of pathology. However, causal evidence of these mechanisms on behavioral phenotype (cognition) or biochemical phenotype (pathology) is often incomplete. Studies elucidating if these complex cellular and molecular responses can be untangled into separate mechanisms related to CR and BM are required to further understand how they contribute to cognition.

While it is evident that some individuals exhibit resilience to AD, a clear cellular or molecular substrate is lacking. We propose to use resilience as a general term when focusing on complex cellular and molecular substrates of reserve in AD, and cellular resilience when there is resilience on a cellular level, for example neuronal subtypes without tangle formation, which cannot be traced back to cognition. Only when both causal evidence and effects on phenotype are present, molecular and cellular processes can be attributed to CR or BM, or resilience and resistance, respectively. Possible mechanisms related to resilience rooted in cellular and molecular mechanisms are often derived from translational approaches in animal models or from human post-mortem studies, which will be highlighted in the next sections of this review.

Most evidence for resilience to AD has been obtained in large-scale longitudinal cohorts by investigating if there is a better cognitive performance as could be expected based on the amount of pathology, inferred from the analyses of blood or cerebrospinal fluid (CSF), of brain imaging or of post-mortem neuropathology. More recently, animal models have been used to discover potential mechanisms of resilience. While each of these approaches have their advantages and disadvantages, all are required to help understand the molecular basis of resilience.

The most accurate method to diagnose AD pathology and comorbidities is by post-mortem neuropathological analysis. It is essential to perform a neuropathologic diagnosis on post-mortem tissue, as common co-morbid neuropathological changes, for instance related to hippocampal sclerosis or early Parkinson’s disease, also influence cognition. To illustrate this point, Montine et al. [48] have defined apparent resilience, referring to specific lesion types measured in vivo with positron emission tomography (PET) scans without consideration of common co-morbidities, and essential resilience, which can only be determined through post-mortem neuropathologic assessment. An individual with apparent resilience might be cognitively intact due to the absence of common comorbid pathology, while a demented individual with similar amounts of AD pathology might have these comorbidities. The authors demonstrated that most cases of apparent resilience were in fact also classified as resistant to co-morbid disease, i.e. resilient as they remain cognitively intact despite presence of AD pathology and resistant to pathological comorbidities as these were lower than expected based on age and levels of AD pathology. Whereas clinical and pathological data are both crucial in determining whether a donor is resilient to AD pathology, often these data are incomplete or not adequately described in post-mortem studies. Unfortunately, most longitudinal cohort studies that test cognitive capabilities over time often have long post-mortem delays, while other studies or brain banks have more specialized neuropathological protocols [49], but have limited clinical information or lack longitudinal measurements of cognition. Another important caveat is that different pathological AD staging systems are used resulting in different subpopulations that are being studied, which is often the result of the scarcity of true resilient donors. Whereas some studies strictly classified resilient individuals with only high amounts of AD pathology [Braak VI [50] and CERAD ≥ 2 [51], most researchers have pooled cognitive intact individuals with Braak stages III-VI [52]. Importantly, different glial subtypes were recently demonstrated in the visual cortex in post-mortem tissue of resilient versus AD patients at Braak stages III-IV [53]. This study demonstrates that cellular and molecular changes can happen before the appearance of tangle formation in the occipital cortex. It remains however uncertain if donors with lower amounts of AD pathology would remain cognitively intact if pathology would progress further. Some donors might remain cognitively intact while others would progress to dementia, which might confound the comparison of AD-related effects between resilient and AD-patients (Fig. 1). While the use of post-mortem tissue allows the implementation of omics on a cellular and molecular level, several confounders that generally lead to unwanted variability must be taken into consideration, including post-mortem delay, agonal state, fixation methods or medication. In addition, due to the absence of a clear biomarker for resilience, the relation between measures for cognition and for pathology are used to identify resilience. As it has been estimated that up to 40% of variance in cognition is not explained by AD pathology or risk factors [54], other parameters than pathology and cognition likely play a role. Once a clear biological substrate for resilience has been established, these other factors, which are likely independent of Aβ and ptau, should become more apparent. Finally, it is important to realize that post-mortem tissue provides a snapshot of the molecular mechanisms and pathology at time of death, which might deviate from a longitudinal assessment. With the development of novel PET and fluid biomarkers for specific features such as Aβ and ptau burden, neuronal damage, α-synuclein, or glial markers, resilience and resistance can be estimated more accurately and longitudinally in living subjects and possibly validate molecular and cellular changes initially discovered in post-mortem tissue. As investigating if resilience mechanisms are causal or simply correlate with resilience is currently not possible in human subjects, others have tried to study the concept of resilience in animal models.

Representation of cognitive aging in the presence of AD neuropathology in AD-patients and resilient individuals. Simplified overview of the AD neuropathological burden versus cognitive functioning during aging or disease progression. A Schematic of a resilient and AD brain with the neuropathological hallmarks of AD, neurofibrillary tangles (green arrows) and amyloid-beta plaques (brown arrows) and atrophic brain regions, as can be seen by the shrinkage of gyri and sulci and larger volume of the vessels, in AD. B Schematic of the progression of cognition and AD pathology over time. Plaques are present years before the onset of the disease and increase over time (brown line). The onset of phosphorylated tau (green line) follows and corresponds better with clinical progression in AD (red line). In later stages of the disease, comorbid pathology is often found in AD, such as TAR DNA-binding protein 43 (TPD-43) inclusions, α-synuclein (α-syn) or hippocampal sclerosis (purple line). It is hypothesized that resilient donors have a similar progression of AD pathology but are able to stay cognitively intact for a longer period (blue line). Multiple studies have also shown that in fact resilient donors are also to a certain extend resistant, as they have reduced amounts of AD pathology and comorbid pathology (dashed lines). Importantly, when low to intermediate amounts of AD pathology are present in the brain, it is difficult to differentiate resilient donors and those who would progress to dementia. Thus, in that timeframe it is difficult to assign resilient donors, which is here depicted as low likelihood of AD based on the amount of pathology

To date, it remains difficult to model the concept of resilience in animal models. One model that tries to recapitulate the idea of the positive effects of lifestyle factors on cognition is enriched environment (EE). Several studies have shown beneficial effects after EE, including improved cognition and a reduction of the amount of both Aβ and ptau [55,56,57]. Similar to the human situation, it remains unclear which lifestyle factors or factors associated with EE, such as social interaction, cognitive stimulation, novelty or exercise, contribute to these beneficial effects. Importantly, while exercise in rodents has shown robust effects on cognition, EE without exercise has also shown to protect from age-related cognitive decline [58]. Some researchers have used a different approach by identifying learners and non-learners in aged animal populations or in AD animal models. These studies found a role for neurogenesis by showing activity dependent activation of newborn neurons in aged animals and the activation of phospholipase A2 (PLA2G4E) in the Tg2576 model by comparing learners to non-learners [35]. Another strategy that has recently been put forward is trying to reflect the genetic diversity of humans in the 5xFAD model by crossing it with BXD mice, which are a series of recombinant inbred strains [59]. By using such an approach, different transcriptional networks enriched in astrocyte and microglia markers have been identified as drivers of resilience in AD [60]. Interestingly, these networks showed a considerable overlap with previously identified networks in the human brain [61], underlining that genetics play an important role in resilience. In addition, animal models provide an excellent opportunity to investigate cellular resilience. Mechanisms that allow cell types or synaptic components to become resilient to Aβ or ptau can be identified in well-controlled environments. For instance, in the APP/PS1 model, it was shown that Aβ-induced deficits in learning and memory and synaptic plasticity depend on α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid (AMPA) receptor subunit GluA3, rendering synapses without GluA3 resilient to Aβ [62].

Ultimately, an interdisciplinary approach is required to elucidate how educational attainment and lifestyle factors can lead to structural and functional changes on both a macroscopic, cellular and molecular level. More large-scale longitudinal community cohorts are required to adequately identify resilient donors with longitudinal cognitive testing and autopsy with a short post-mortem delay and extensive neuropathological characterization. One example is the longitudinal centenarian cohort [63]. Importantly, findings derived from post-mortem studies should be validated in vitro or in vivo to demonstrate causality and link them to either CR or BM, or resilience or resistance, respectively. To recognize the different strengths and limitations discussed above, the most important papers discussed in this review are summarized in Table 1, indicating how resilience was determined in each study.

Evidence for the existence of resilience to AD has mainly come from large community-based cohorts with longitudinal measures of cognition and neuropathological examinations at autopsy. These large cohorts like the Religious Orders Study (ROS), the Rush Memory and Aging Project (MAP), the Nun study or the Baltimore Longitudinal Study of Aging (BLSA) have all shown a discordance between the degree of post-mortem AD pathology and ante-mortem cognition [102,103,104]. Similar patterns have been observed in other large cohorts, like the 90 + study [105], the Honolulu-Asia Aging Study (HAAS) [106], and the Medical Research Council Cognitive Function and Ageing Study (CFAS) [107]. In addition, it has been estimated that around one-third of the community-dwelling elderly with intermediate and high levels of AD neuropathology remained cognitively unimpaired [52]. Thus, a considerable number of the elderly population shows a discrepancy between their cognition and AD pathology. Recently, case-reports have emerged of extreme resistant or resilient cases, in which individuals with fAD were able to maintain cognition up to three decades after the expected onset of symptoms. While these individuals were resilient towards Aβ, there was either regional resistance towards tau pathology [108], demonstrated by low levels of ptau in the frontal cortex and hippocampus [109], or resilience to ptau, demonstrated by the more extensive load throughout the brain [110]. Even though a final neuropathological diagnosis of AD is required in addition to the clinical symptoms, developments in the identification of CSF and neuroimaging biomarkers [111] have allowed monitoring the progression of AD in vivo. In line with reports that extensive Aβ deposits are found in the brain at autopsy in cognitively intact individuals, multiple studies have shown that 20–40% of cognitively intact individuals over the age of 65 have Aβ biomarkers (measured by PET Pittsburgh Compound-B (PiB) uptake) above the diagnostic threshold in at least one AD vulnerable brain region [112,113,114]. This has also been extended to different CSF biomarkers, where AD signatures of these markers were above diagnostic thresholds for cognitively intact elderly [90, 115].

It is crucial to recognize the potential impact of the many different types of amyloid deposits, as they differ in their associations with clinical symptoms. The most commonly described types are diffuse plaques, which are non-fibrillar and vary in morphology, and dense core deposits [116, 117], which contain fibrillar Aβ and correlate better with clinical severity [118]. Furthermore, dense core plaques can further be characterized by the presence of a neuritic component if the focal Aβ deposit contains tau-positive or dystrophic neurites. Neuritic plaques (NPs) correlate well with clinical severity and are thought to closely associate with neuronal loss or atrophy in AD [119]. For example, in a cohort of individuals with intact cognition, NPs were associated with lower scores on cognition tests [120]. Strikingly, reduced amounts of NPs have also been observed in the superior temporal sulcus in resilient individuals compared to AD patients [