Exercise and hippocampal memory systems

Abstract

No medications prevent or reverse age-related cognitive decline. Physical activity (PA) enhances memory in rodents, but findings are mixed in human studies. As a result, exercise guidelines specific for brain health are absent. Here, we re-examine results from human studies, and suggest the use of more sensitive tasks to evaluate PA effects on age-related changes in the hippocampus, such as relational memory and mnemonic discrimination. We discuss recent advances from rodent and human studies into the underlying mechanisms at both the central and peripheral levels, including neurotrophins and myokines that could contribute to improved memory. Finally, we suggest guidelines for future research to help expedite well-founded PA recommendations for the public.

The elephant in the room: can we recommend exercise for memory?

The growing aging population together with an increase in sedentary lifestyle is an urgent public health concern, particularly after the fourth and fifth decades of life. If the pace of increased life expectancy [1] and inactivity [2] continues, an unprecedented percentage of future generations may live with cognitive impairment. By delaying cognitive decline, we may be able to compress our time with dementia or perhaps escape it altogether. Given the health benefits of physical activity (PA, see Glossary), and its wide accessibility at any age, determining how PA may sustain cognition during aging would have enormous societal and economic impact.

Human and animal studies show PA benefits brain function and cognition, and could counteract age-related cognitive decline [3]. However, a recent National Academies report [4] affirms that while PA appears promising, evidence from human studies is inconclusive towards preventing cognitive decline and insufficient to recommend exercise for reducing the risk of dementia. Their resolution was based on the inconsistency of cognitive benefits in randomized controlled trials (RCTs), and the lack of correspondence between cognitive change, and changes in intervention-induced biomarkers of brain dysfunction. The report called for more intervention studies with long-term follow-ups assessing clinical outcomes.

While valuable, this approach will require decades of lead time before a new recommendation can be considered. In this review, we re-examine evidence that has accumulated from animal models and human PA studies to unravel the mixed results from RCTs. We focus on cross-species approaches to identify possible changes in the hippocampus, a brain area crucial for memory formation [5, 6]. Determining how PA counteracts the trajectory of early memory decline will also provide prevention strategies for cognitive aging and Alzheimer’s disease (AD). Specifically, rodent models are utilized to evaluate components of human episodic memory vulnerable to aging, such as relational memory, pattern separation, and spatial navigation, as well as the underlying cellular and molecular mechanisms. By examining the effects of PA on memory function across different species and timescales, we hope to elucidate the process of change and identify relevant biomarkers of PA effectiveness for long-term brain health during human aging.

The mixed effects of PA on human cognitive function

The National Academies [4] and the National Academy of Medicine [7] reports point out that long-term prospective studies consistently show that PA reduces cognitive decline and dementia risk [8, 9]. Yet, several well-powered RCTs such as the LIFE [10] and MAX [11] trials failed to demonstrate PA benefits on cognition in older adults. However, in these studies cardiorespiratory fitness was not measured [e.g., 12]. This means aerobic training intensity could not be optimally personalized to starting fitness levels, and it is difficult to precisely verify fitness adaptations at a group or individual level. Indeed, fitness change may be a critical indicator of whether cognitive changes are expected, and the wide variation in fitness response can overshadow training group effects (see SI I-III and Table 1). Another source of discrepancy between results from observational and experimental designs may be a lack of consistent and precise cognitive outcome measures. For example, eleven meta-analyses of RCTs with primarily older adults used 9 unique terms for overlapping memory constructs (e.g., memory, short-term memory and verbal memory immediate; see SI IV). Therefore, relevant information may be masked when grouping across tasks with different levels of sensitivity and precision at different phases of aging and disease progression.

Table 1. Comparison of aerobic exercise training paradigms across species, animal models (yellow) and human (blue).

Note ↑ and ↓ indicate an increase and a decrease of each factor, respectively, induced by PA. Most of the studies include hippocampal structures, and observed PA effects on cognitive function and biological markers. Abbreviations: AD, Alzheimer’s disease; ADAS-cog, Alzheimer’s Disease Assessment Scale-Cognitive subscale; Aβo, Amyloid-β oligomers; BDNF, Brain-derived neurotrophic factor; CA1, Cornu ammonis 1; CA3, Cornu ammonis 3; CBV, Cerebral blood volume; CSF, Cerebrospinal fluid; CTSB, Cathepsin B; DG, Dentate gyrus; ERRα, Estrogen-related receptor alpha; fEPSP, Field excitatory postsynaptic potential; FNDC5, Fibronectin type III domain-containing protein 5; HRR, Heart rate reserve; IGF-I, Insulin-like growth factor-I; LBD, Lewy body dementia; MCI, Mild cognitive impairment; MMSE, Mini-Mental State Examination; MWM, Morris water maze; PGC-1α, Peroxisome proliferator-activated receptor gamma coactivator 1-alpha; VO2 max, Maximal oxygen consumption.

	SUBJECTS	CONTROL	EXERCISE REGIMEN	REGION EVALUATED	FINDINGS	REF
CROSS-SPECIES REPORTS	• Male PGC-1α null mice  • 5, 6 and 13 weeks old	• Housed without wheels  • 30 days	• Voluntary wheel running  • 30 days	• Hippocampus	• FNDC5 regulates BDNF  • ↑ FNDC5 through increase of PGC-1α, ERRα	129
	• Male young adults  • 25-30 years old	• No exercise  • 12 weeks  • n=4	• Cycle ergometer  • 12 weeks, 3 days/week, 4 times × 4 min/day  >90% peak aerobic capacity  • Treadmill walking  12 weeks, 2 days/Week, 45 min/session  70% peak aerobic capacity  • n=6	• Plasma	• ↑ Irisin in plasma	130
	• Male C57BL/6 mice, Swiss mice and AD mouse model (APP/PS1 ΔE9)  • 2.5-3 months old	• No exercise  • 5 weeks  • n=100	• Swimming  • 5 weeks, 5 days/week, 1 hour/day  • n=101	• Hippocampus	•↑ Novel object recognition  •↑ Contextual fear conditioning  •↑ Synaptic plasticity (fEPSP)  •↑ Hippocampal FNDC5 mRNA, FNDC5/irisin and BDNF	131
	•Ex vivo cortical slices (n=5)  •16-66 years old (60% female)	• n=11	• N/A	• Hippocampus  • Cortex  • CSF	•↓ Hippocampal irisin (protein) in late AD  •↓ Irisin CSF in AD and LBD  •↓ FNDC5 mRNA and FNDC5/i risin by Aβo  • Recombinant irisin stimulate cAMP-PKA-CREB pathway
	• Postmortem brain tissue  • 68-100 years old (>50% female)  • Control  • Early AD (n=7)  • Late AD (n=7)
	• Male C57BL/6 mice  • 4 weeks old	• Sedentary  • 4 weeks  • n=32	• Voluntary wheel running  • 4 weeks  • n=32	• Dentate gyrus  • Frontal cortex  • Plasma	•↑ Muscle CTSB (mRNA and protein)  •↑ Plasma CTSB  •↑ Hippocampal CTSB (mRNA)  •↑ Behavioral performance in MWM (probe)	132
	• Young adults  • 19-34 years old (56% female)	• Treadmill walking  • 16 weeks, 2 sessions/week, 10-25 m in/session  • 50% maximum heart rate  • n=23	• Treadmill running  • 16 weeks, 3 sessions/week, 45-75 min/session  • 70-90% maximum heart rate  • n=20	• Hippocampus  • Plasma	•↑ CTSB plasma  • Positive correlation between CTSB level and late complex-object recall score
	• Male C57BL/6 mice  • 7 weeks old	• Sedentary  • 2 weeks  • n=23	• Voluntary wheel running  • 2 weeks  • n=23	• Entorhinal cortex  • Dentate gyrus  • CA1  • CA3	•↑ DG CBV  • Correlation between DG CBV and neurogenesis	64
	• Young adults  • 21-45 years old (82% female)	• Within-subject design	• Aerobic training  • 12 weeks, 4 sessions/week, 40 min/session  • n=11	• Entorhinal cortex  • Dentate gyrus  • CA1  Subiculum	•↑ DG CBV  • Correlation between DG CBV and VO2 max, and cognition  •↑ 1st-trial learning of new declarative memories
SEPARATE REPORTS IN HUMAN	• Older adults with MCI (49% female)  • 65-95 years old	• Attending 2 education classesabout health promotion  • 6 months  • n=25 amnestic MCI  • n=25 MCI	• Multicomponent exercise (Aerobic exercise, muscle strength training, postural balance retraining, dual-task training)  • 6 months, 2 days/week, 90 min/day  • 60% maximum heart rate  • n=25 amnestic MCI  • n=25 MCI	• Medial temporal lobe including entorhinal cortex  • Whole brain cortical atrophy	•↑ Group × time interaction in MMSE and logical memory I in aMCI group  •↓ Whole brain cortical atrophy  • Association between low total cholesterol levels before the intervention and an improvement of logical memory 1  • Significant relationship between a higher level of BDNF and improved ADAS-cog performance	165
	• Older adults (62% female)  • 55-80 years old	• Stretching and toning  • 48 weeks, 3 sessions/week, 40 min/session  • n=60	• Walking on indoor track  • 48 weeks, 3 sessions/week, 40 min/session  • 50-75% HRR  • n=60	• Hippocampus  • Caudate nucleus  • Thalamus	•↑ Anterior hippocampal volume  • Correlation between increased VO2 max and an increase of hippocampal volume  • Relationship between changes in BDNF and in hippocampal volume	66

To better understand the effects of aerobic PA on memory, we summarize effects from the meta-analysis reports at the task level. We focused specifically on neuropsychological tests predicting cognitive impairment and dementia such as word list and story recall, and a more general decline in visuospatial memory in complex figure tasks. We also included tasks assessing constituent processes in episodic memory such as relational memory and pattern separation, as well as wayfinding (Box 1). These processes may identify early memory dysfunctions, as they rely on the hippocampal circuitry and deteriorate early in AD [13-15]. We counted the number of results consistent with a hypothesis in favor of PA (improved) or against the hypothesis (declined), relative to null effects [e.g., 8]. Positive results for cognitively normal middle-aged and older adults were proportionally highest in relational memory tasks, and there were no negative effects for any task (Figure 1A, Key Figure). Evidence that PA improves middle-aged and older adults’ performance on relational memory tasks is consistent with data from cross-sectional studies with older adults [16-18], and with both item and spatial relational tasks (Figure 1B) in pre-adolescents [19-23] and young adults [24-26]. Albeit, some studies with smaller sample sizes of young adults observed null effects [16, 17, 27, 28], which may be due to lack of statistical power. Although, PA has shown improved pattern separation and wayfinding (Box 1) in animal models and young adults, it appears that no RCTs in middle-aged or older adults have focused on these tasks, which represents an important future direction.

Box 1: PA effects on spatial memory and wayfinding.

In rodents, PA improves hippocampus-dependent spatial memory in paradigms including the Morris water maze, the Y-maze, and the radial arm maze [89]. These maze tasks test a type of spatial navigation called wayfinding, which is based on coding your position in space and relative to landmarks to build a mental map for finding your way in space. PA also promotes performance in spatial memory tasks with low motor demand, such as contextual fear conditioning and spatial pattern separation [see review 88]. It would be important to see this translate to aging humans because one of the first signs of AD related cognitive decline is difficulty with navigation via wayfinding. However, few studies have examined PA or fitness benefits on wayfinding in older adults. One RCT examined a more general form of spatial working memory as an outcome with older adults. This trial found that aerobic exercise led to 2% increases in hippocampal volume, which were associated with greater gains in fitness and improvements in memory performance [66]. Similar to relational memory, promising trends have been reported for navigation processes in younger age groups. A study with middle-age adults showed preliminary evidence that fitness may enhance activation in regions important for wayfinding during spatial learning in a virtual maze, though these benefits were not seen in performance [136]. Another study, with adolescent males, showed that greater fitness was related to faster learning, but not recall of objects in a virtual environment [28]. With respect to mechanisms, the difference in learning versus recall may be a clue to early disease processes that would be good to target with PA. A study with older adults presenting a preclinical AD biomarker profile (e.g., cerebrospinal fluid Aβ42 levels below 500 pg/ml) but without other cognitive symptoms, showed this group was slower to acquire new memories during wayfinding, but showed preserved recall and recognition [137]. This suggests some components of spatial navigation may be more sensitive than others to early effects of pathology on spatial memory. Further, more studies across species could examine the slope and shape of learning curves as common metrics for reporting PA effects on learning and memory with more contact to computational and process models for learning mechanisms.

Figure 1, Key Figure: Human PA and fitness affect hippocampal memories and networks.

(A) Bar graph shows the proportion of statistically significant effects in favor of improved memory for aerobic compared to control groups in training studies with older adults. Ratios indicate absolute number of positive effects relative to all studies. See SI IV-V for more information. (B) In non-spatial relational tasks, letters symbolize items, which could take the form of words, faces, scenes, etc. Participants are tested for memory of relations compared to recombined or novel pairs. The spatial relational reconstruction task requires retrieval of object-location pairs and inter-relations. A signature error involves swapping item identity with correct locations [162]. Mnemonic discrimination tasks test memory precision. Participants incidentally encode objects and are tested on memory for repeats (targets), similar objects (lures), or new objects (foils) [39, 163, 164]. (C) Association networks modified by aging and PA. Network links and bar graphs are colored according to network. Links show connections within each network. Only the default network includes aspects of the hippocampal and medial temporal lobe memory system, as approximated in the oval inset. The figure to the right of the inset illustrates a different view of the relative positioning of medial temporal lobe structures (ph, parahippocampus; erc, entorhinal cortex; prc, perirhinal cortex). The dentate gyrus and CA sub-regions are coiled together along the long axis of the hippocampus. Young>Old functional connectome indicates the links that were stronger for young adults [76], and the bar graph shows the percentage of links within each network in favor of young adults. The functional connectome below shows links that were stronger with greater fitness, and the bar graph shows the percentage of links within each network having a positive association with fitness. Of all the networks with connections that are weaker for older adults, the hippocampal-cortical default network had the greatest percentage (15.38%) that were stronger with greater fitness [76].

PA improves pattern separation in rodents and humans

The ability to discriminate among ambiguous or similar experiences is a crucial feature of episodic memory, and therefore pattern separation is regarded as a critical function of the hippocampal network. It is considered that in the hippocampal network, the feed-forward projections from a smaller number of entorhinal cortex neurons, project via the perforant pathway, onto a larger population of dentate gyrus cells, providing an anatomical basis for the concept of pattern separation. Information is then processed from the dentate gyrus to area CA3 for pattern completion and then to area CA1 for encoding, forming the ‘tri-synaptic hippocampal circuit’ [29]. The dentate gyrus also affects spatial or episodic memories through pattern separation [30, 31]. This theory has been supported by testing rodents with lesions in select hippocampal subfields (area CA1, area CA3 or dentate gyrus) on tasks requiring discrimination among similar stimuli. For example, animals with dentate gyrus damage have deficits in their ability to distinguish between two similar objects, with one covering a baited food well, whereas CA1-lesioned rats can [32]. In addition, dorsal dentate gyrus-lesioned rats, but not dorsal CA1- and CA3-lesioned rats, are impaired in detecting novel object locations [33]. More recently, following a neurotoxic dentate gyrus lesion, rats could not learn new scene stimuli but were able to retrieve learned familiar scenes [34]. Newly-formed scene memories in the lesion group were also easily disrupted by presenting similar versions of the new scene. Thus, the integrity of the dentate gyrus is critical for pattern separation.

In humans, studies using functional magnetic resonance imaging (fMRI) with the blood oxygen level-dependent (BOLD) signal show area CA3 and dentate gyrus activity is elevated during pattern separation performance [35-37] while CA1, subiculum and parahippocampal cortices including the entorhinal cortex showed a bias toward pattern completion [37]. Limits in fMRI spatial resolution make it difficult to reliably distinguish between the CA3 and dentate gyrus. Activity spanning both regions may reflect projections of pattern separated information from the dentate gyrus to CA3. A small but growing number of studies with humans suggest pattern separation, evaluated with mnemonic discrimination tasks (Figure 1B), is a critical process for evaluating PA effects on hippocampal function. Specifically, an intervention study with young adults showed that 6 weeks of high-intensity training, which increases fitness, improved discrimination [27]. However, the small sample and lack of control group limit the finding. Studies using a cross-sectional design, also suggest greater PA [38] and higher fitness [39] correlate with better discrimination. Further, this correlation appears strongest when fitness and circulating brain-derived neurotrophic factor (BDNF) are high [40]. Moreover, a single PA session improved object discrimination in young adults, further linked to changes in functional connectivity during the test phase between hippocampal (dentate gyrus/CA3) and cortical regions involved in recall, suggesting PA can induce rapid functional changes that affect performance, without structural alterations [41]. Yet, a voxel-based morphometry study in young adults also found higher fitness was associated with greater right entorhinal volume, which was related to mnemonic discrimination [42]. Thus, rapid functional changes appear in similar systems that are enhanced structurally with greater fitness, suggesting mechanisms linking rapid and accumulated effects could help understand how regular PA can improve memory. However, further studies in middle-aged and older adults are needed to elucidate these mechanisms.

In the dentate gyrus of the hippocampus, new neurons are born throughout adulthood in mammals, including humans, a process called adult neurogenesis [43, 44]. It has been suggested, that these new neurons play a key role in pattern separation. For example, mice with ablated hippocampal neurogenesis exhibited reduced spatial memory for similar, but not discrete, spatial locations in the radial arm maze task. These mice also showed performance deficits in the touchscreen-based pattern separation task [45]. Similarly, genetic suppression of adult neurogenesis impaired discrimination between similar contexts and disrupted normal population coding in CA3 [46]. Moreover, in transgenic mice in which tetanus toxin blocked the output of developmentally-born granule cells onto area CA3, while adult-born granule cells were intact, enhanced contextual discrimination in a highly similar context in comparison to control mice was observed, suggesting a key role of adult-born granule cells for pattern separation [47]. This concept is further supported by research showing that lesion of the lateral entorhinal cortex, a major input to adult-born granule cells results in deficient performance in the touchscreen task [48] (Figure 2A,B).

Figure 2: Effects of PA on hippocampal circuits and pattern separation in mice.

In mice voluntary wheel running increases the number of new hippocampal neurons and modifies their network. (A) Using viral vectors to trace the structures that project directly to adult-born neurons it was observed that running enhances entorhinal cortex innervation of adult-born dentate granule cells [58]. Control and running afferent circuitry are depicted in a 3D reconstruction of photomicrographs taken throughout the dorso-ventral extent of hippocampal-entorhinal cortex slices derived from young adult male mice. In control conditions, afferent traced cells (labeled with rabies virus expressing MCherry, red) that project directly to new neurons, are observed in the lateral entorhinal cortex (LEC) and perirhinal cortex (PRC), and only sparsely in the caudomedial entorhinal cortex (CEnt). Exercise increases the input from the CEnt and LEC onto the adult-born granule cells. Nuclei are stained with 4′,6-diamidino-2-phenylindole (DAPI), blue [58]. (B) LEC input to new neurons is important for pattern separation. (B, C) Mice are trained in the touchscreen to distinguish between two identical stimuli spaced apart (big) closely together (small) to evaluate pattern separation. The ability to differentiate which icon is associated with reward in the more challenging, small separation condition, is (B) reduced following LEC lesion [48], and (C) enhanced by running [53]. (D) Model showing the relative contribution of the regions that directly innervate adult-born granule cells under control and exercise conditions [58].

Unfortunately, in humans, PA effects on adult neurogenesis cannot be assessed in vivo. However, studies in multiple mouse and rat strains have shown that voluntary wheel running and forced treadmill training increase adult neurogenesis in the dentate gyrus throughout life in both sexes [49] and in the presence of AD-like pathology (Box 2). Use of viral vectors to birth-date the adult-born granule cells showed that PA increased their dendritic arborization and cell body area at 7 days-old [50]. In 21-day-old adult-born neurons, PA increased motility of dendritic spines and accelerated their maturation in both young [51] and middle-aged mice [52]. This running-induced enhancement of adult-born neuron morphology and number is associated with improved stimulus [53] and object discrimination in mice [54] (Figure 2C). Consistently, in a Bax conditional knockout mouse, which specifically prevents programmed cell death of adult-born neurons, thereby upregulating adult neurogenesis, there was greater discrimination and rapid contextual encoding between two similar environments [55]. Altogether, these findings suggest adult hippocampal neurogenesis is important in pattern separation and can be facilitated by PA.

Box 2. PA effects on Alzheimer’s disease pathology.

Mice and rats do not naturally develop hallmark AD pathologies of amyloid plaques and tau neurofibrillary tangles. Most animal models express familial AD genes that lead to build-up of amyloid and tau proteins in young adulthood. PA can counteract these processes and enhance neural plasticity in many of the mouse models. For instance, in APPswe/PS1dE9 mice, PA increases synaptic plasticity [138], dendritic arborization [139], cell proliferation and neuronal differentiation [140]. PA also increased cell proliferation [141] and cell survival [142] in triple transgenic (3xTg) mice with both amyloid and tau [143]. Recently, it was also shown that in 5xTG AD mice running for three hours per day for four months increased neurogenesis and BDNF levels, and prevented age-related decline in mnemonic discrimination [144]. Other PA benefits on cognition have been reported across AD mouse models and behavioral assays, including spatial learning and memory [145], contextual memory [139], novel object [131] and recognition memory [146]. Studies suggest PA reduces Aβ and tau hyper-phosphorylation [140, 147], with high intensity running potentially having the most impact on reducing pathology via increased clearance [148]. Data from humans also suggest PA interacts with processes leading to pathology aggregation such as clearance or perhaps resilience to its presence. In humans, pathology is measured in vivo with positron emission tomography (PET) using tracers binding to amyloid or tau. No training studies have reported on amyloid or tau brain imaging outcomes. However, one intervention found greater fitness improvements in AD patients were related to increased hippocampal volume and episodic memory [149]. Without PET, proxy biomarkers of brain pathology can be measured from cerebrospinal fluid (CSF). Two small training studies failed to show changes in CSF Aβ (Aβ₄₂) with either amnestic impairment [150] or early AD [151]. In contrast, cross-sectional studies report favorable relations between sensor-based PA and CSF Aβ₄₂ [152]. Greater PA is also associated with lower risk of amyloid in the default network (PET PiB) associated with aging [153] or APOEe4 [154]. Thus, evidence supports PA benefits during periods of amyloid and tau cortical spreading. However, important questions remain about the extent to which PA at this stage is affecting clearance rather than mitigating its source. Just as valuable are preserved benefits on BDNF, neurogenesis, and morphology, as these could be mechanisms for promoting cognitive resilience even without pathology reversal per se.

PA influences hippocampal memory by rewiring neuronal networks

From a functional perspective, increased dentate gyrus neuron production brings several benefits. It may contribute to maintenance of spine and synapse density, and help prevent hippocampal shrinkage in healthy aging [56]. In addition, although adult-born granule cells account for a small percentage of the total population of dentate gyrus cells, they have enhanced excitability, lower thresholds to induce long-term potentiation (LTP), and enhanced LTP compared to developmentally-born granule cells [57]. Moreover, PA modifies synaptic plasticity in adult-born neurons. Specifically, voluntary running increases short-term synaptic plasticity from the lateral entorhinal cortex onto adult-born neurons, evoked selectively by lateral perforant pathway stimulation [58]. PA also alters the network of adult-born neurons, augmenting the innervation from entorhinal cortex and areas important for spatial memory and theta rhythm generation, such as caudomedial entorhinal cortex, medial septum and supra- and medial mammillary nuclei [58] (Figure 2A,D). Indeed, evidence from animal models and preliminary results in humans show that physical movement induces synchronized neuronal firing at theta rhythm in hippocampal circuits [59, 60]. PA also modifies the expression of genes important for synaptic transmission in the lateral entorhinal cortex [61], a cortical area preferentially innervating adult-born granule cells [48, 62]. Thus, PA could coordinate the activity of neural ensembles to favor memory processes and enhance integration of adult-born neurons into the existing hippocampal-entorhinal circuitry that otherwise deteriorate with aging [63].

Mapping PA and fitness effects on human hippocampal networks

In humans, repeated non-invasive imaging is possible for tracking change throughout the brain. Some imaging methods can also be used in rodents to link imaging markers in humans to plausible biological substrates. In a first proof-of-concept study, exercise-induced increases in cerebral blood volume were observed with MRI in the dentate gyrus of young mice and middle-aged humans [64] (Table 1). More recently, a training study with older adults (N=40) [65] reported increased fitness was associated with improved hippocampal cerebral blood volume. Effects were strongest in the whole hippocampus, but cerebral blood volume changes accounted for increased anterior hippocampal volume [see also, 66, 67]. Thus, in agreement with animal models, exercise-induced changes in human hippocampal volume may couple with increased resting metabolic state or vascular density or both [see also, 68]. However, a study in young to middle-age adults found that anterior hippocampal volume changes from 6 weeks of training that improved fitness increased markers of myelination rather than cerebral blood volume [67]. Using multiple MRI modalities, they also found both fitness and hippocampal volume changes were absent after a 6-week break, further supporting the link between fitness and hippocampal adaptations, as well as the importance of long-term PA.

It is also worth emphasizing that PA effects on hippocampal volume are broader than maintenance or reversal of aging. Benefits have been shown in pre-adolescents [21], adolescents [28], college-age adults [25], cognitively unimpaired older adults [Table 1, 66, 68-70], older adults with mild cognitive impairment and AD [71, 72]. Thus, similar to animal models, PA effects on hippocampal structure are seen across the lifespan, suggesting an intrinsic adaptive benefit offering protection and resilience during aging.

In addition to regional structural growth and metabolic support, studies using neuroimaging with humans have also shown that moderate intensity PA [73-75] and fitness [25, 76-78] strengthen functional connectivity of a hippocampal-cortical brain network known as the default network (Figure 1C). Functional connectivity in this context refers to the extent to which the resting BOLD activity across network regions is coupled over time, typically quantified by the correlation strength of their timeseries. The default network spans cortical association regions including the medial and lateral surfaces of the temporal, parietal, and prefrontal cortices. While other association networks also show reduced functional connectivity with aging, the default network uniquely mirrors the spatial topography of structural connections of the medial temporal lobe and the cortical spread of AD pathology [79], and functional connectivity has been observed to predict risk of later cognitive impairment [e.g., 80]. For instance, Figure 1C illustrates association networks that we and others have shown to have reduced functional connectivity with aging [76, 81]. Although all the networks contain long-range cortical connections, more than any other network, the hippocampal-cortical default network is both affected by aging and protected with fitness. The extent to which alterations of the hippocampus are critical to this functional pattern needs to be examined. Especially given that most training effects on hippocampal structure and metabolism have been strongest in the anterior hippocampus, whereas the default network primarily includes the posterior hippocampus and parahippocampal cortex.

Notably, from longitudinal studies, long-distance cortical functional connectivity changes have been observed after 1-year [73] and in relation to fitness [75, 76], but not after shorter training periods (e.g., 6 months) unless fitness is improved [78]. It is also worth emphasizing that fitness relations with default network functional connectivity most often include the prefrontal cortex [25, 76, 77], which may counteract age-related reductions in functional connectivity between posterior and prefrontal cortices [82] and could help explain broader cognitive benefits of aerobic exercise to executive function [73, 77]. The underlying mechanisms in such long-range network effects measured with fMRI are expected to be multidimensional, reflecting integrative effects on neurotransmitter function, neurotrophic support for long-term potentiation and synaptic plasticity, vascular perfusion and reactivity (see outstanding questions). Complementary methods at the cellular and molecular level are also critical to clarifying potential mechanisms and informing potential signaling pathways from exercising muscles to the brain.

Outstanding Questions.

• What experimental cognitive tasks are most sensitive to early dysfunction in hippocampal circuits?

• What signaling pathways increase central BDNF expression from PA, and can they be non-invasively and reliably measured in humans?

• Are there additional critical signaling pathways mediating or moderating effects of PA on hippocampal-dependent memory, such as stress response pathways including the hypothalamic-pituitary-adrenal axis and the sympatho-adrenal axis? How are they modified by PA dose (intensity, duration, etc), sex, age, medications, pathology?

• How can non-invasive imaging outcomes be added to animal studies to link changes in biological substrates to testable forms in humans?

• What experimental designs can examine more directly whether, for who, and how PA changes the trajectory of memory decline? For instance, a prospective design could identify “decliners” to enrich a training study, and high-density measurement at training onset and offset could clarify the time-course and durability of PA-dependent changes in hippocampal and memory outcomes.

• Will it be possible to harness peripheral factors as therapeutic interventions for aged or frail individuals who cannot exercise?

Cellular and molecular PA mechanisms

Trophic factors are closely associated with PA benefits for brain function [83]. BDNF plays an important role in synaptic plasticity, neurite outgrowth, neurogenesis and cell survival [84]. Knockdown or deficits in BDNF/tyrosine receptor kinase B (TrKB) signaling impairs memory function [85]. In rodents, PA upregulates hippocampal BDNF/TrkB mRNA and protein expression after both short (2-7 days) and long (1-8 months) periods, across a range of intensities, exercise and housing conditions [86, 87, for review, see 88]. A meta-analysis reported a large PA effect on BDNF mRNA (effect size=1.82, 13 studies, 17 effects) and protein (effect size=1.25, 10 studies) in the rodent hippocampus [89].

Conversely, blocking hippocampal TrkB receptors abolishes benefits of exercise on learning and memory [90] and selectively ablating TrkB in adult hippocampal progenitor cells precludes running-induced neurogenesis [91], suggesting that BDNF may mediate, in part, the beneficial effects of PA on memory processes.

Utilizing BDNF as a biomarker for PA effects on the brain across species is complicated by several factors. While the neurotrophin can easily be measured in both the rat and mouse brain, blood serum and plasma levels in mice are too low to detect [92, 93]. In rats, peripheral BDNF can be assayed but shows no change after acute running [94] or a decrease after several weeks of exercise in adult [95] and aged [96] subjects. In humans, BDNF cannot be measured non-invasively in the brain, but it can be measured in the blood. It has been estimated the brain contributes 70-80% of circulating plasma BDNF during exercise [97]. In addition, BDNF is both stored and released from platelets [98].

Recently, a meta-analytic review found human BDNF in either serum or plasma increases acutely after a single session of PA and that this effect intensified with training [99]. It is worth noting there are conflicting proposals about how PA intensity affects BDNF signaling. Several studies have found higher intensities acutely elicit more serum BDNF [100], improved memory [101], and increased functional connectivity in hippocampal-cortical and reward systems [102], possibly stimulated by increased sympathetic activation [103]. On the other hand, it has been argued that lower intensities, avoiding a stress response, are optimal for rapid increases in hippocampal BDNF mRNA in rats [104], and can acutely improve pattern separation coupled with increased hippocampal functional connectivity in humans [41]. Critically, only one of these studies included older adults [102], so it will be important to determine how these proposals generalize to aging.

With respect to training, independent meta-analyses have reported that >2 weeks of aerobic training increased resting circulating BDNF measured in either serum or plasma [99, 105]. Following a year of training, elevated serum BDNF was correlated with hippocampal volume [66] and default network functional connectivity [106]. In contrast, several cross-sectional studies have reported a negative relationship between fitness and serum BDNF [107-110]. The sources of these discrepancies remain unclear. However, there are technical considerations that may preclude human PA studies from yielding consistent results. For instance, the number of platelets may affect peripheral BDNF measurements [111]. In addition, the different ELISA kits used may affect reproducibility of results [112].

In tandem with BDNF, vascular endothelial growth factor (VEGF) and insulin-like growth factor (IGF-1) mediate PA effects. Specifically, VEGF, a potent mitogen of endothelial cells, is increased in the rodent hippocampus during PA [113]. VEGF plays a role in PA-induced angiogenesis [114, 115] and increased neurogenesis [116]. This increment may be associated with activation of lactate receptor HCAR1 in vessel walls, which mediate PA-induced increases in hippocampal VEGF A and angiogenesis [115]. Thus, VEGF may promote neurotransmission and neurovascular adaptations in response to increased metabolic demand associated with neuronal growth and network integration. IGF-1, on the other hand, is directly taken up from circulation into the cortex and hippocampus [117, 118]. In rodents, blocking peripheral IGF-1 blocks PA-induced increases in hippocampal BDNF [119] and adult neurogenesis [120]. In older humans, higher peripheral IGF-1 levels are associated with better cognition. However, with PA, levels of the growth factor have been shown to be increased, decreased or not changed [121]. The differences between studies may be associated with methodological challenges of these assays [122, 123], and illustrate the need for additional peripheral biomarkers of PA effects on the brain.

Identification of mechanisms underlying indirect activation of central growth factor signaling and neurogenesis via factors released from peripheral organs such as muscle (myokines), liver (hepatokines) and adipose tissue (adipokines) [124, 125], may lead to novel approaches to assess PA effects on cognitive function. An important regulator of muscle physiology is 5' adenosine monophosphate-activated protein kinase (AMPK). AMPK activation blocks energy-consuming processes and promotes ATP synthesis and glucose uptake [126]. AMPK also regulates transcription factors involved in muscle contractile processes such PPARδ[127] and PGC1α [128]. In rodents, PA activates AMPK and induces PGC1α expression in the muscle, elevating FNDC5 protein levels, which is secreted as irisin in circulation. Irisin may induce hippocampal BDNF mRNA expression [129] and is detected in human serum [130]. In addition, FNDC/irisin is neuroprotective in a mouse model of Alzheimer’s Disease [131]. Treatment of skeletal muscle cells in vitro with AMPK agonist AICAR also led to the identification of the myokine cathepsin B. In turn, administering recombinant cathepsin B to neural progenitor cells increased BDNF gene expression. Blocking cathepsin B during PA also occludes benefits on memory and neurogenesis in mice [132]. Moreover, in humans, 4 months of treadmill training increased plasma cathepsin B levels in association with improved visuospatial configural memory [132]. Recent research also established that β-hydroxybutyrate, a ketone synthesized in the liver and accumulating in the hippocampus during PA, activates BDNF gene promotors via inhibition of HDAC (histone deacetylases) [133]. Thus, the identification of central and peripheral factors may allow us to understand the mechanisms underlying the PA effects on cognitive function, which may also lead to methods for enhancing benefits (Table 1).

PA regimens for brain health

One of the biggest questions about PA is what is the best regimen to improve memory? As in recent general PA guidelines, individual factors such as age, gender, weight, health history, suggest there is not one PA prescription for everyone as to how much or how long we should work out to improve our health [2], and we would agree the same applies to memory. Nevertheless, most of the human studies which investigated the benefits of PA showed that aerobic exercise at or above 60% of individualized maximum heart-rate (three times per week for about an hour each session, for more than three months), increased either brain or cognitive measures that otherwise decline with aging, especially when fitness was improved (see SI I-III). This is consistent with general recommendations for adults to get at least 150 minutes of moderate intensity PA per week [2]. Based on the reviewed evidence across species here, we further emphasize the importance of individually-determined intensity and monitoring fitness improvements for maximizing benefits to the brain and memory (see SI III). Moreover, short-term periods of PA at a broader range of intensities immediately enhances neural plasticity and memory [41, 134, 135]. This is consistent with the general guidelines’ suggestion to ideally spread PA throughout the week [2]. In the future, we may even understand how to maintain these benefits over the course of a day, which could perhaps motivate more people to override inactivity one day at a time. It is also clear from reviewed studies that consistent PA for weeks to months is needed for long-lasting improvements in hippocampal structural and functional plasticity, and memory function. Thus, overall, consistency is key, achieved through a lifestyle enriched in PA that gets the heart rate up, regardless of the specific regimen of choice.

Concluding remarks and future perspectives

The complexity of biological networks modified by PA requires model organisms that can isolate PA effects with tight experimental control. Animal models unravel plausible mechanisms that enable or constrain PA effects on memory, informing broader theoretical and biomarker development. Overall, data from animal models and humans indicate PA and fitness benefit functions that depend on intact hippocampal circuitry, such as relational memory and pattern separation. From our view, tasks tapping into these processes represent targeted “stress tests” for hippocampal circuitry, ideal for the next generation of intervention outcomes. The reviewed evidence supports the proposal that these tasks will be most sensitive to early system-level deterioration before advanced memory impairments.

Further, while exercise training designs can examine the outcomes of regular exercise, particularly in humans, they are limited in revealing the process of change. An acute paradigm examines rapid effects on cells and circuits during and acutely after PA, perhaps more directly indicating mechanisms of action. More broadly, the within-subject experimental control of acute and cross-over paradigms in humans brings massive advantages for allowing each individual to be their own control. It is also more feasible, as done in animal studies, to manipulate the system by not only enhancing but also blocking a proposed pathway while observing brain-behavior outcomes. While modifying an effect acutely does not guarantee it is necessary for training gains, converging evidence from acute and short timescales and long-term training could inspire strategies for mimicking or boosting exercise effects on the brain and memory.

With respect to long-term follow-up to clinical outcomes, experimental control in RCTs and clinical significance are difficult to optimize in a single study. Single-site RCTs with intensive biomarker measures are riddled with small and selective samples and many outcomes. Leveraging data-sharing and multi-site experimental studies will accelerate progress and enrich follow-up analyses for clinical outcomes. At the same time, an integrative theoretical model tying together critical biological variables in testable forms is necessary, coupled with the transparency of hypothesis testing with pre-registration. Exciting projects like the Molecular Transducers of Physical Activity Consortium (MoTrPAC,https://www.motrpac.org/), promise to provide rich datasets for this effort. Such theoretical and mechanistic advances in understanding aerobic PA effects provide a foundation for further determining how other PA modalities such as resistance exercise could boost memory benefits (Box 3). However, at this time, the data are strongest for aerobic PA to maximize memory protection in late-life, and support the recommendation that future RCTs maintain aerobic PA as a reference arm for comparison of multi-modal or alternative PA types. Moreover, through the complementary use of acute and training designs with long-term follow-up, such a process-based approach, would open the door to more rapid and personalized feedback about how to optimize one’s activity patterns for improved brain health.

Box 3: Can lifting weights also improve hippocampal memory?

The recommendation to “be more active” has broad appeal practically. However, different types of PA may be more effective than others or work through complementary mechanisms informing therapeutic strategies. To begin answering these questions, RCTs have examined effects of resistance training and also multi-modal training that combine elements of aerobic, resistance, and flexibility or balance exercise. Instead of individualized heart rate training zones used for aerobic training, resistance training (RT) interventions determine resistance intensity referenced to the most you can lift a single time as a “1 repetition maximum” (e.g. for each of 6 exercises, 2 sets of 8 repetitions at 50% to 80% repetition maximum). A recent meta-analysis reported that, similar to aerobic training, RT does not improve episodic memory assessed with neuropsychological tests [155]. Some data hint resistance training could improve relational memory task performance in a female mild cognitive impairment population, when aerobic exercise would not [156]. However, the same study found maintained hippocampal volume after aerobic but not RT [12]. Overall, the evidence suggests aerobic and RT affect cognition via different pathways, as RT interventions also do not increase resting serum BDNF [105]. Supportive of complementary benefits, the meta-analysis found multi-modal training may be more effective for episodic memory (4 studies, g=0.48) than aerobic (7 studies, g=0.05) or RT (4 studies, g=0.07). Rodent models provide key insight to comparing pathways detectable with biomarkers. Using the ladder climbing model as RT, and treadmill training as aerobic exercise in rodents, 8 weeks of both types of training improved spatial learning and memory [157]. Whereas both types of PA increased the levels of proteins associated with synaptic plasticity (synapsin 1 and synaptophysin) [157], improved memory correlated with activation of distinct signaling pathways, suggesting aerobic and RT elicit distinct molecular mechanisms to improve memory processes. Treadmill training showed increased IGF-1, BDNF, TrkB, and calcium/calmodulin-dependent kinase II (β-CaMKII) in the hippocampus, whereas resistance training increased IGF-1 levels peripherally and in the hippocampus, as well as activation of its receptor signaling pathway (Akt protein) in the hippocampus. The peripheral IGF-1 increment after RT in this study is in agreement with a study in humans, where 24 weeks of RT increased peripheral IGF-1 levels [158]. Across animal and human studies, while direct comparisons are limited, evidence suggests RT could boost- but not replace- effects of aerobic exercise on memory.

Highlights.

• A sedentary lifestyle increases the risk of memory deterioration and Alzheimer’s disease. Unfortunately, there is no treatment for memory decline. However, animal studies indicate physical activity benefits hippocampus-dependent memory through biological mechanisms likely conserved in humans.

• Translating the promise for physical activity to human memory improvement or maintenance has been challenging and the best modes of activity and cognitive outcomes remain unclear.

• We propose a cross-species approach to bring insight to sensitive hippocampal memory tasks and a mechanistic foundation for physical activity-induced memory improvement in the long run.

Glossary

Adult neurogenesis

The adult brain contains two regions that can give rise to adult-born neurons in rodents: the subventricular zone of the lateral ventricles gives rise to new olfactory bulb neurons and the subgranular zone of the hippocampal dentate gyrus produces new granule cells. In humans, olfactory neurogenesis stops at birth [159], while the time-course of hippocampal neurogenesis remains under investigation [43, 44].

Blood oxygen level-dependent (BOLD) signal

A signal detected with fMRI indicating a reduction in deoxyhemoglobin that is known to couple with presynaptic activity of neuronal populations, and peaks approximately 4-6 seconds after neuronal activity onset.

Brain-derived neurotrophic factor (BDNF)

A protein supporting neuronal growth and repair, synaptic function, synaptic plasticity, and cellular homeostasis. BDNF is transcribed and expressed in the brain with high concentrations in the hippocampus, cortex, and hypothalamus. In peripheral tissues it is expressed in muscle, vascular endothelial cells, bone marrow megakaryocytes, and stromal and immune cells.

Cardiorespiratory fitness

Capacity to convert oxygen to physical work. The gold-standard measure is a graded maximal exercise test, which increases workload until exhaustion while measuring expired gases, resulting in a measure of the amount of oxygen one can use per minute per kilogram of body weight (mL/kg/min), to generate physical work. In our review, “fitness” refers specifically to cardiorespiratory fitness.

Default network

A hippocampal-cortical network that fragments with aging, particularly for medial temporal lobe and prefrontal connections (Figure 1C). The “Default” term refers to observations that the network is most active when individuals are not directed to think about anything and presumably revert to a default state, including mind wandering, mental simulation, and recollection of episodic memories.

Episodic memory

Memory for how events, places, and people come together in the episodes of day-to-day life. Proposed to rely on relational binding, the process of rapidly and obligatorily binding elements of experience across dimensions of space and time and their interactions.

Pattern completion

The computational process by which incomplete memory formations are reconstructed by partial sensory inputs that are part of the previously stored representations. This process helps to balance perceptual stability which allows accurate generalization in the presence of noise or complexity of external environments [30, 160].

Pattern separation

The computational process by which similar memories are stored and accessible as distinct representations in the brain. Empirical and computational evidence suggest the dentate gyrus enables pattern separation via sparse coding patterns. The behavioral expression of pattern separation processes is more generally referred to as mnemonic discrimination (e.g., Figure 1B and Figure 2B,C).

Physical activity (PA)

Bodily movement produced by the skeletal muscles that increase energy expenditure beyond resting levels. PA varies by type, frequency, duration, and intensity for both lifestyle-related activity and physical exercise that is planned, structured, and to improve physical performance.

Theta rhythm:

A slow rhythmic oscillatory activity (ranging the 4-8 Hz frequency) of the local field potential primarily found in the hippocampal structure. Its role has been known as critical to voluntary movement, spatial navigation, and memory processes (by coordinating neuronal ensembles within the medial temporal lobe) in animals and humans [for review, see 161].