Multivariate MR biomarkers better predict cognitive dysfunction in mouse models of Alzheimer's disease

Abstract

To understand multifactorial conditions such as Alzheimer's disease (AD) we need brain signatures that predict the impact of multiple pathologies and their interactions. To help uncover the relationships between pathology affected brain circuits and cognitive markers we have used mouse models that represent, at least in part, the complex interactions altered in AD, while being raised in uniform environments and with known genotype alterations. In particular, we aimed to understand the relationship between vulnerable brain circuits and memory deficits measured in the Morris water maze, and we tested several predictive modeling approaches. We used in vivo manganese enhanced MRI traditional voxel based analyses to reveal regional differences in volume (morphometry), signal intensity (activity), and magnetic susceptibility (iron deposition, demyelination). These regions included hippocampus, olfactory areas, entorhinal cortex and cerebellum, as well as the frontal association area. The properties of these regions, extracted from each of the imaging markers, were used to predict spatial memory. We next used eigenanatomy, which reduces dimensionality to produce sets of regions that explain the variance in the data. For each imaging marker, eigenanatomy revealed networks underpinning a range of cognitive functions including memory, motor function, and associative learning, allowing the detection of associations between context, location, and responses. Finally, the integration of multivariate markers in a supervised sparse canonical correlation approach outperformed single predictor models and had significant correlates to spatial memory. Among a priori selected regions, expected to play a role in memory dysfunction, the fornix also provided good predictors, raising the possibility of investigating how disease propagation within brain networks leads to cognitive deterioration. Our cross-sectional results support that modeling approaches integrating multivariate imaging markers provide sensitive predictors of AD-like behaviors. Such strategies for mapping brain circuits responsible for behaviors may help in the future predict disease progression, or response to interventions.

Introduction

A key question in Alzheimer's disease (AD) research is how pathology differentially and sequentially affects vulnerable brain circuits, thereby giving rise to behavioral changes. Although critically important, detailed descriptions of interactions between genes and structural and functional phenotypes are poorly described. However, these interactions dictate the vulnerability for cognitive dysfunction in the context of aging and AD related pathologies. Investigating the circuits and mechanisms underlying cognitive dysfunction is important for understanding what triggers the switch from normal aging to AD, what predicts rates of disease progression, and how patient-specific therapeutic strategies may be developed [1]. Therapies for neurodegeneration have been directed to individual targets, such as altered synaptic transmission, amyloid deposition, or abnormal tau phosphorylation - all well-demonstrated pathologies in AD [2] [3]. Additional factors, such as cardio vascular status [4] [5], insulin resistance [6], and diabetes [7] may also influence cognition [4]. Importantly, neuro-immunological mechanisms, interacting with systemic inflammatory mediators and obesity, are thought to also modulate AD pathology [5,8,9]. Since attacking AD pathologies separately has not yet provided effective strategies for prevention or reduction of cognitive damage, we need models that provide an integrated view of how multiple variables and risk factors contribute to system wide dysfunction. We currently lack the quantitative integrative models required to understand multifactorial conditions.

To help understand the causative links between the biological and cognitive substrates typical of AD, it is helpful to conceptualize the brain as a set of interacting regions forming a spatially distributed network [10]. Structural networks integrate effects from changes occurring at different scales (synapse, cells, circuits), which in turn modulate the properties of functional networks. Several large-scale networks have been mapped in the brain and characterized by distinct functional profiles, such as sensory perception, movement, attention and cognition [11] [12]. However, it is not well understood how brain sub-networks map to the cognitive domain. Understanding these relationships in the normal brain, and their alterations in disease may inform on the mechanisms underlying mild cognitive impairment (MCI) or dementias such as AD [13].

One strategy to help understand how circuits influence behavior is to link imaging to the clinical AD cognitive phenotypes. The progressive loss of cognitive memory is commonly diagnosed using tests such as the Mini–Mental State Examination (MMSE) [14], the Montreal Cognitive Assessment [15], and others [16]. Clinical populations of individuals diagnosed with AD have shown overlaps in the patterns of gray matter atrophy [17], Aß distribution [18], and axonal density changes [19], but there are also marked differences in brain atrophy [20] or tau pathology distribution [21]. Such differences may relate to population heterogeneity in terms of genetics, disease stage, or comorbidities. An alternative hypothesis to explore AD etiology is based on selective vulnerability of cells and axonal pathways favoring disease propagation. Imaging can provide in vivo biomarkers [22] [23] that are related to pathology as observed ex vivo, or to functional changes. To successfully link imaging to clinical phenotypes, we need to develop integrative models that explain the initiation, potentiation, and propagation of selective vulnerability in cells and networks that underlie AD processes, in relation to risk factors.

Neuroimaging approaches to map brain levels of behavioral descriptors have traditionally used voxel based statistical analyses (VBA) of deformation fields, structural and functional connectivity maps, vascular perfusion, or amyloid deposition and tau maps. But statistical approaches that pursue a dichotomous strategy, and aim to separate data according to image features do not necessarily explain the behavioral changes, nor do they disclose the biological processes underlying them. More recently predictive modeling approaches have been proposed to provide statistically relevant imaging correlates of memory changes spanning a continuum range, as observed in AD [24] [25] [26].

In this study, we have used a mouse model of AD to develop such predictive approaches. Mice provide tools for dissecting the contributions of genes on circuits and behavior. In particular, they provide homogenous populations, and can be tightly controlled for genetic and environmental factors, thus simplifying the problem of mapping brain circuits responsible for behavior. To establish and test a novel integrative predictive modeling approach one could choose among multiple mouse models. Most traditional models replicate one of the AD hallmarks (amyloid plaques, or tau tangles), but do not fully reflect the complex biology of AD. These transgenic mice express mutated APP [27] or PSEN1 [28], and the combination of mutations may accelerate and enhance the phenotype [29]. However, humans do not overexpress APP or PSEN1 to the same levels as in mice, and the role of beta amyloid deposition has not been fully clarified. More recently, models which also express hyperphosphorylated tau have been generated [30]. Still, most transgenic models fail to replicate other phenotypes seen in human AD, including neurodegeneration [31] [32]. Models of late onset, or sporadic AD are based on the genetic risk conferred by the presence of APOE4 alleles [33] [34] [35]. These models, while very promising in their ability to help us understand the etiology and progression of AD, require long times to express phenotypes, including behavioral deficits. None of these models addresses the differences between the mouse and the human innate immune systems, and the potentially important role of microglia in the development of AD [36]. For example the nitric oxide levels produced by immune activation of the NOS2 gene mouse are much higher than in humans [37]. In the APPSwDI^+/+/mNos2^−/− (CVN-AD) strain [38,39], mNos2 deletion makes the immune responses more similar. In conjunction with the Swedish, Dutch and Iowa mutations, this promotes an AD-like background required for studying the underlying mechanisms of pathological regional vulnerability. CVN-AD mice replicate multiple AD pathologies, including amyloid and tau deposition, neuronal loss, altered microglial activity with typical AD-like inflammatory patterns and deficits in memory and learning [38,40] [41] [37] [42] [43]. The appearance of cognitive deficits with aging in this strain mimics processes in humans with AD and can be assessed using the Morris water maze test. This behavioral test is commonly used to quantify the loss of spatial learning and memory in animal models of aging and AD [44,45].

To map behavioral changes to specific brain regions and networks we have used in vivo manganese enhanced MRI (MEMRI). Manganese ions (Mn²⁺) are paramagnetic and induce T1 shortening [46], enhancing tissue contrast [47]. Mn²⁺ has also been used to characterize trans-synaptic connectivity and axonal transport properties in rodents [48,49] [50]. Importantly, Mn²⁺ enters neural cells via voltage gated calcium channels and vesicular reuptake, presenting an alternative for task based fMRI in rodents [51] [52], while alleviating limitations due to the types of tasks that animals can perform in the magnet, or to anesthesia [53]. These strategies to characterize brain structure and function in CVN-AD mice in relation to age matched controls can identify vulnerable brain circuits responsible for behaviors typical of AD.

To identify vulnerable regions and networks that predict deficits in memory and learning, we used an integrative approach that was not linked to a single identified neuropathological mechanism, but was reflective of multiple concomitant factors. We evaluated how traditional mass-univariate analyses can predict behavior dysfunction, and followed with a multivariate approach involving dimensionality reduction. Eigenanatomy, a sparse dimensionality reduction method, was incorporated to extract brain regions responsible for changes in morphometry, signal intensity due to Mn²⁺ uptake (reflective of brain activity), and magnetic susceptibility (reflective of altered iron homeostasis and conducive to oxidative stress and inflammation) [54] [55]. However, these methods analyzed individual biomarkers separately. We have employed both a data-driven as well as a hypothesis-driven approach to associate imaging phenotypes with behavioral markers for cognitive status and to identify circuits vulnerable to AD like pathology. Eigenanatomy produced candidate regions and circuits, and we selected regions that appeared important based on one or more biomarkers, and confirmed by previous studies as relevant to AD. To predict cognitive dysfunction based on in vivo multivariate imaging markers we used sparse canonical correlation analysis (SCCA) [56,57]. SCCA selects regions so to maximize correlation among imaging and cognitive measures, in a supervised approach. The result is a network of regions that underlie changes in cognition, incorporated in a multivariate analysis. Our results provide insight into the relationships between structural networks and cognitive function in animal models of AD, supporting the value of multivariate approaches for humans with AD.