Abstract
Alzheimer’s disease (AD) is associated with heterogeneous atrophy
patterns. We employed a semi-supervised representation learning
technique known as Surreal-GAN, through which we identified two latent
dimensional representations of brain atrophy in symptomatic mild
cognitive impairment (MCI) and AD patients: the “diffuse-AD” (R1)
dimension shows widespread brain atrophy, and the “MTL-AD” (R2)
dimension displays focal medial temporal lobe (MTL) atrophy.
Critically, only R2 was associated with widely known sporadic AD
genetic risk factors (e.g., APOE ε4) in MCI and AD patients at
baseline. We then independently detected the presence of the two
dimensions in the early stages by deploying the trained model in the
general population and two cognitively unimpaired cohorts of
asymptomatic participants. In the general population, genome-wide
association studies found 77 genes unrelated to APOE differentially
associated with R1 and R2. Functional analyses revealed that these
genes were overrepresented in differentially expressed gene sets in
organs beyond the brain (R1 and R2), including the heart (R1) and the
pituitary gland, muscle, and kidney (R2). These genes were enriched in
biological pathways implicated in dendritic cells (R2), macrophage
functions (R1), and cancer (R1 and R2). Several of them were “druggable
genes” for cancer (R1), inflammation (R1), cardiovascular diseases
(R1), and diseases of the nervous system (R2). The longitudinal
progression showed that APOE ε4, amyloid, and tau were associated with
R2 at early asymptomatic stages, but this longitudinal association
occurs only at late symptomatic stages in R1. Our findings deepen our
understanding of the multifaceted pathogenesis of AD beyond the brain.
In early asymptomatic stages, the two dimensions are associated with
diverse pathological mechanisms, including cardiovascular diseases,
inflammation, and hormonal dysfunction—driven by genes different from
APOE—which may collectively contribute to the early pathogenesis of AD.
All results are publicly available at
[122]https://labs-laboratory.com/medicine/.
Subject terms: Personalized medicine, Predictive markers
Introduction
Alzheimer’s disease (AD) is the most common cause of dementia in older
adults and remains incurable despite many pharmacotherapeutic clinical
trials, including anti-amyloid drugs [[123]1, [124]2] and anti-tau
drugs [[125]3]. This is largely due to the complexity and multifaceted
nature of the underlying neuropathological processes leading to
dementia. The research community has embraced several mechanistic
hypotheses to elucidate AD pathogenesis [[126]4–[127]6]. Among these,
the amyloid hypothesis has been dominant over the past decades and has
proposed a dynamic biomarker chain: extracellular beta-amyloid (Aβ)
triggers a cascade that leads to subsequent intracellular
neurofibrillary tangles, including hyperphosphorylated tau protein (tau
and p-tau), neurodegeneration, including medial temporal lobe atrophy,
and cognitive decline [[128]7, [129]8]. However, the amyloid hypothesis
has been re-examined and revised due to substantial evidence that
questions its current form [[130]8–[131]10]. While amyloid remains
critical to AD development, the amyloid cascade model has been
continually refined as other biological factors are discovered to
influence the pathway from its accumulation to cell death.
Cardiovascular dysfunction has been widely associated with an increased
risk for AD[[132]11]. There is also growing evidence that inflammatory
[[133]10–[134]12] and neuroendocrine processes [[135]5, [136]13]
influence pathways of amyloid accumulation and neuronal death. The
inflammation hypothesis claims that microglia and astrocytes release
pro-inflammatory cytokines as drivers, by-products, or beneficial
responses associated with AD progression and severity
[[137]12–[138]14]. The neuroendocrine hypothesis, first introduced in
the context of aging [[139]15], has been extended to AD [[140]16],
where it proposes that neurohormones secreted by the pituitary and
other essential endocrine glands can affect the central nervous system
(CNS), which subsequently contribute to developing AD. For example,
Xiong and colleagues [[141]17] recently found that blocking the action
of follicle-stimulating hormone in mice abrogates the AD-like phenotype
(e.g., cognitive decline) by inhibiting the neuronal C/EBPβ–δ-secretase
pathway. These findings emphasize the need to further elucidate early
brain and body changes well before they lead to irreversible clinical
progression [[142]18].
Recent advances in artificial intelligence (AI), especially deep
learning (DL), applied to magnetic resonance imaging (MRI), showed
great promise in biomedical applications [[143]19, [144]20]. DL models
discover complex non-linear relationships between phenotypic and
genetic features and clinical outcomes, thereby providing informative
imaging-derived endophenotypes [[145]21]. In particular, AI has been
applied to MRI to disentangle the neuroanatomical heterogeneity of AD
with categorical disease subtypes [[146]22–[147]26]. The genetic
underpinnings [[148]17, [149]27, [150]28] of this neuroanatomical
heterogeneity in AD are also complex and heterogeneous. The most recent
large-scale genome-wide association study [[151]28] (GWAS: 111,326 AD
vs. 677,633 controls) has identified 75 genomic loci, including APOE
genes, associated with AD. However, such case-control group comparisons
conceal genetic factors that might contribute differentially to
different dimensions of AD-related brain change. More importantly, the
genetic variants that contribute to the initiation and early
progression of brain change in younger and asymptomatic individuals are
poorly understood.
In this study, we utilize a novel semi-supervised deep learning
approach, Surreal-GAN, to characterize the neuroanatomical
heterogeneity of the disease. Unlike our previous model, Smile-GAN
[[152]22], which categorized subtypes, Surreal-GAN generates multiple
continuous latent dimensional representations, simultaneously
accounting for spatial and temporal disease heterogeneity, similar to
what was accomplished in a previous unsupervised clustering model known
as Sustain [[153]24]. These multi-dimensional scores reflect the
co-expression level of respective brain atrophy dimensions in the same
patient; this is biologically plausible, as brain diseases like AD
often progress continuously over a long disease trajectory. Refer to
the method (Surreal-GAN for disease heterogeneity) and Supplementary
eMethod [154]1 for methodological details of Surreal-GAN, comparisons
to other subtyping methods, and strengths of semi-supervised
representation learning. We hypothesized that genetic variants,
potentially unrelated to APOE genes, contribute to early manifestations
of multiple dimensions of brain atrophy in early asymptomatic stages.
We first trained the Surreal-GAN model to define the AD dimensions to
test this hypothesis in the late symptomatic stages. We then examined
their expression back to early asymptomatic stages. In our previous
study [[155]29], we derived two neuroanatomical dimensions (R1 and R2)
by applying Surreal-GAN to the MCI/AD participants (target population)
and cognitively unimpaired (CU) participants (reference population)
from the Alzheimer’s Disease Neuroimaging Initiative study (ADNI
[[156]30]). Herein, we applied the trained model to three asymptomatic
populations and one symptomatic population: the general population
(N = 39,575; age: 64.12 ± 7.54 years) from the UK Biobank (UKBB
[[157]31]) excluding demented individuals; the cognitively unimpaired
population (N = 1658; age: 65.75 ± 10.90 years) from ADNI and the
Baltimore Longitudinal Study of Aging study (BLSA [[158]32]); the
cognitively unimpaired population with a family risk (N = 343; age:
63.63 ± 5.05 years) from the Pre-symptomatic Evaluation of Experimental
or Novel Treatments for Alzheimer’s Disease (PREVENT-AD [[159]33]); the
MCI/AD population (N = 1534; age: 73.45 ± 7.69 years) from ADNI and
BLSA. Refer to the method (Study design and populations) and Table
[160]1 for details of the definition of these populations.
Table 1.
Study characteristics.
Population Study Participant (N) Scan (N) Age (year) Sex/female CU AD
MCI proxy-AD
MCI/AD ADNI & BLSA 1534 7019 73.45 (54.27, 93.00) 888/58% 0 424 1110 NA
General UKBB 39,575 40,981 64.12 (44.56, 82.27) 18,625/47% 39,574^b 1
NA 10,189^c
CU ADNI & BLSA 1658 6143 65.75 (22.00, 80.00) 939/57% 1658 0 0 NA
CU with a family risk PREVENT-AD 343 1215 63.63 (55.13, 84.22) 243/71%
343^a NA NA 343^a
[161]Open in a new tab
We present the age with the mean, min, and max in each population. The
definition of cognitively unimpaired (CU)^a in PREVENT-AD, asymptomatic
participants^b in UKBB, and proxy-AD^c in UKBB are detailed as (a)
Participants (proxy-AD and CU with a family risk) from the PREVENT-AD
study were recruited with the following criteria: (i) being cognitively
normal, (ii) having a family history of AD, (iii) aging within 15 years
from the age of disease onset of their youngest relative, and (iv) no
history of neurological or psychiatric diseases; (b) The UKBB
participants (the general population) represent a general population
with healthy Aging and diseases (not AD, specifically). We excluded
those diagnosed with all sources of dementia (G30 in ICD-10 diagnoses,
see below). However, these asymptomatic participants might have
diagnoses of other illnesses or comorbidities based on ICD-10; (c)
Participants with proxy-AD in UKBB are defined by a family history of
AD with the following criteria: (i) illnesses_of_father_f20107 and (ii)
illnesses_of_mother_f20110.
Materials and methods
Study design and populations
The current study consists of four main populations (Table [162]1),
which were jointly consolidated by the iSTAGING [[163]34] and the AI4AD
consortia ([164]http://ai4ad.org/): the iSTAGING consortium
consolidated all imaging and clinical data; imputed genotyping data
were downloaded from UKBB; the AI4AD consortium consolidated the
whole-genome sequencing (WGS) data for the ADNI study. The iSTAGING
consortium is an NIH-funded effort that systematically and
statistically consolidates and harmonizes brain imaging data for the
study of aging and AD, including different ethnicity groups and
demographics, and covers the entire lifespan. The AI4AD consortium aims
to leverage the power of AI to study AD and aging, which also
consolidates WGS data across the USA. Supplementary eMethod [165]2
details each population’s definition and inclusion criteria. Our goal
is to consolidate and harmonize large-scale lifespan imaging data to
model the full spectrum of Alzheimer’s disease and assess how the
identified AD dimensions are expressed across various stages of the
disease, especially at the early stages.
Image preprocessing
All T1w-weighted MR images were first corrected for magnetic field
intensity inhomogeneity [[166]35]. A deep learning-based skull
stripping algorithm was applied for the removal of extra-cranial
material. In total, 145 anatomical regions of interest (ROIs) were
generated in gray matter (GM, 119 ROIs), white matter (WM, 20 ROIs),
and ventricles (6 ROIs) using a multi‐atlas label fusion method
[[167]36] (Supplementary eMethod [168]3). The 119 ROIs were
statistically harmonized by an extensively validated approach, i.e.,
ComBat-GAM [[169]37], using the entire imaging data of iSTAGING.
Supplementary eFig. [170]1 demonstrates the normality check for the
MUSE ROI (right accumbent area) before and after statistical
harmonization, illustrating that our statistical harmonization enhanced
the normality of ROIs across various studies. The harmonized MUSE ROIs
were then fit to Surreal-GAN to derive the dimensions.
Surreal-GAN for disease heterogeneity
Surreal-GAN [[171]29] (Supplementary eFig. [172]2) dissects underlying
disease-related heterogeneity via a deep representation learning
approach under the principle of semi-supervised learning [[173]22,
[174]23]. At a high level, its most fundamental novelty is that it
provides a continuous representation of the presence of multiple,
non-exclusive abnormal brain patterns in each individual, rather than
clustering individuals into one of many clusters, i.e., disease
subtypes. More specifically, several methodological advancements were
considered compared to its predecessor, Smile-GAN [[175]22]. First,
Surreal-GAN is to model neuroanatomical heterogeneity by considering
both spatial and temporal (i.e., disease severity) variation using only
cross-sectional MRI data. Secondly, Surreal-GAN disentangles the
neuroanatomical heterogeneity of AD by enabling patients to
simultaneously exhibit multiple distinct imaging patterns (i.e., high
scores for expressing all these patterns), resulting in
high-dimensional scores across multiple dimensions. Lastly, in contrast
to prior probability-based clustering methods like Smile-GAN,
Surreal-GAN operates without the constraint that all dimensional scores
must sum to 1. This allows for a more normal distribution of
dimensional scores suited for GWAS (Supplementary eFig. [176]3).
Further methodological details are elaborated upon in Supplementary
eMethod [177]1.
Alternative clustering techniques, such as Sustain [[178]24] and
Bayesian latent [[179]38] methods, are available for deciphering the
neuroanatomical heterogeneity in AD [[180]22–[181]26, [182]39].
Surreal-GAN distinguishes itself from these approaches based on
fundamental methodological distinctions, such as its utilization of
semi-supervised deep learning compared to the unsupervised approach of
others [[183]40]. Additionally, Surreal-GAN generates continuous
dimensions associated with distinct phenotypic outcomes, allowing the
simultaneous co-expression of multiple patterns instead of categorizing
patients into a single dominant subtype or stage, as seen in other
methods. Notably, the two dimensions, R1 and R2, displayed correlations
with the four subtypes generated by Smile-GAN, particularly R2
exhibited a correlation with P3 (reflecting medial temporal lobe
atrophy), and both R1 and R2 displayed correlations with P4
(representing global atrophy), as depicted in Supplementary eFig.
[184]4. These two dimensions capture the individual-level manifestation
of the two distinct imaging atrophy patterns, contrasting them with
healthy control groups within the ADNI study across the AD spectrum.
Brain and clinical variable associations
We performed brain-wide associations for the 119 GM ROIs. For baseline
brain-wide associations, linear regression models were fitted with R1
and R2 dimensions as independent variables, with each ROI as the
dependent variable, controlling for age, sex, intracranial volume
(ICV), and/or diagnosis as covariates.
We performed a two-step linear regression for longitudinal brain-wide
associations. First, we derived the individual-level age change rate
using a linear mixed-effects model. To this end, we included a
participant-specific random slope for age and random intercept; age and
sex were treated as fixed effects. Secondly, the same linear regression
model as in baseline brain associations was fitted with the age change
rate as the independent variable.
We also performed clinical variable association for all clinical
variables and neuropsychological testing available for each population,
using the same model in the baseline brain-wide associations.
Bonferroni correction of 119 GM ROIs was performed to adjust for the
multiple comparisons. We included various clinical variables across
different studies, including AI-derived imaging signatures, such as
SPARE-AD [[185]41], an imaging surrogate for AD atrophy patterns, and
SPARE-BA [[186]42] for brain aging-related atrophy. Other clinical
variables also included cognitive scores (e.g., the Rey Auditory Verbal
Learning Test (RAVLT)), modifiable risk factors (e.g., BMI), and CSF
biomarkers (Aβ42). The detailed 45 clinical variables are presented in
Supplementary Table [187]3.
Genetic analyses
Genetic analyses were performed for the WGS data from ADNI and the
imputed genotype data from UKBB [[188]43]. Our quality check protocol
is detailed in Supplementary eMethod [189]4. This resulted in 1487
participants and 24,194,338 SNPs in ADNI WGS data. For UKBB, we limited
our analysis to European ancestry participants, resulting in 33,541
participants and 8,469,833 SNPs [[190]44–[191]47].
Using UKBB data, we first estimated the SNP-based heritability using
GCTA-GREML [[192]48], controlling for confounders of age (at imaging),
age-squared, sex, age-sex interaction, age-squared-sex interaction,
ICV, and the first 40 genetic principal components, following a
previous pioneer study [[193]49]. In GWAS, we performed a linear
regression for each neuroanatomical dimension and included the same
covariates as in the heritability estimates. We adopted the genome-wide
P-value threshold (5 ×10^−8) in all GWAS. The annotation of genomic
loci (displayed by its top lead SNP) and gene mappings, prioritized
gene set enrichment, and tissue specificity analyses were performed
using FUMA ([194]https://fuma.ctglab.nl/, version: v1.3.8)
(Supplementary eMethod [195]5 and [196]6). A two-step procedure
(Supplementary eMethod [197]7) was performed to determine if an
annotated genomic locus or gene was associated with AD-related clinical
traits. We calculated the polygenic risk scores (PRS) [[198]50] using
both ADNI and UKBB genetic data (Supplementary eMethod [199]8).
Finally, we constructed a target-drug-disease network for these genes
associated with R1 and R2 to identify these “druggable genes”
(Supplementary eMethod [200]9).
Results
Two dominant dimensions of brain atrophy found in MCI and AD
In MCI/AD patients, the “diffuse-AD” dimension (R1) showed widespread
brain atrophy without an exclusive focus on the medial temporal lobe
(Fig. [201]1A and Supplementary eTable [202]1 for p values and effect
sizes). In contrast, the “MTL-AD” dimension (R2) displayed more focal
medial temporal lobe atrophy, prominent in the bilateral
parahippocampal gyrus, hippocampus, and entorhinal cortex (Fig.
[203]1A). All results, including p values and effect sizes (Pearson’s
correlation coefficient r), are presented in Supplementary eTable
[204]1. The atrophy patterns of the two dimensions defined in the
symptomatic MCI/AD population (Fig. [205]1A) were present in the
asymptomatic populations, albeit with a smaller magnitude of r
(Supplementary eTables [206]1, [207]4 and [208]8). We presented the age
distribution (Supplementary eFig. [209]5A), as well as the expression
of R1 and R2, along with the population-level difference for the four
populations in Supplementary eFig. [210]5B–D.
Fig. 1. The manifestation of the R1 and R2 dimensions of brain atrophy in the
MCI/AD population.
[211]Fig. 1
[212]Open in a new tab
A Brain association studies reveal two dominant brain atrophy
dimensions. A linear regression model was fit to the 119 GM ROIs at
baseline for the R1 and R2 dimensions. The −log[10](p value) of each
significant ROI (Bonferroni correction for the number of 119 ROIs:
−log[10](p value) >3.38) is shown. A negative value denotes brain
atrophy with a negative coefficient in the linear regression model. All
the statistics (r, Pearson’s correlation coefficient) are presented in
Supplementary Table [213]1. The brain maps denote the signed p value,
and the range of r for each dimension is also shown. Of note, the
sample size (N) for R1 and R2 is the same for each ROI. B Genome-wide
association studies demonstrate that the R2, but not R1, dimension is
associated with variants related to APOE genes (genome-wide p value
threshold with the red line: −log[10](p value) >7.30). We associated
each common variant with R1 and R2 using the whole-genome sequencing
data from ADNI. Gene annotations were performed via positional,
expression quantitative trait loci, and chromatin interaction mappings
using FUMA [[214]58]. We then manually queried whether they were
previously associated with AD-related traits in the GWAS Catalog
[[215]55]. Red-colored loci/genes indicate variants associated with
AD-related traits in previous literature. C Clinical association
studies show that the R2 dimension is associated to a larger extent
with AD-specific biomarkers, including SPARE-AD [[216]41], an imaging
surrogate to AD atrophy patterns, and APOE ε4, the well-established
risk allele in sporadic AD. The R1 dimension is associated to a larger
extent with aging (e.g., SPARE-BA [[217]42], an imaging surrogate for
brain aging) and vascular-related biomarkers (e.g., WML white matter
lesion). The same linear regression model was used to associate the R1
and R2 dimensions with the 45 clinical variables, including cognitive
scores, modifiable risk factors, CSF biomarkers, disease/condition
labels, demographic variables, and imaging-derived phenotypes. The
radar plot shows representative clinical variables; results for all 45
clinical variables are presented in Supplementary eTable [218]3. The
SPARE-AD and SPARE-BA scores are rescaled for visualization purposes.
The gray-colored circle lines indicate the p value threshold in both
directions (Bonferroni correction for the 45 variables: −log[10](p
value) >2.95). A positive/negative −log[10](p value) value indicates a
positive/negative correlation (beta). The transparent dots represent
the associations that do not pass the Bonferroni correction; the
blue-colored dots and red-colored dots indicate significant
associations for the R1 and R2 dimensions, respectively.
At baseline, the R1-dominant group had 25.72% AD patients (N = 222 out
of 863); the R2-dominant group consisted of 30.10% AD patients (202 out
of 671). Within a 7-year follow-up period, MCI participants from both
the R1-dominant and R2-dominant groups progressed to AD, with the
R2-dominant group exhibiting a higher proportion of AD patients (40%
vs. 25%) (Supplementary eFig. [219]6A, B); the two dominant dimensions
developed independently throughout the 7-year follow-up period
(Supplementary eFig. [220]6C, D).
APOE genes are associated with R2 but not with R1 in the MCI/AD population
In GWAS, the R2 dimension, but not R1, was associated with
well-established AD genomic loci (rs429358, chromosome: 19, 45411941;
minor allele: C, p value: 1.05 × 10^−11) and genes (APOE, PVRL2,
TOMM40, and APOC1) (Fig. [221]1B). The details of the identified
genomic locus and annotated genes are presented in Supplementary eTable
[222]2. The PRS of AD showed a slightly stronger positive association
with the R2 dimension [r = 0.11, −log[10](p value) = 3.14] than with
the R1 dimension [r = 0.09, −log[10](p value) = 2.31, Supplementary
eFig. [223]7]. The QQ plots of the baseline GWAS are presented in
Supplementary eFig. [224]8.
Clinical profiles of the R1 and R2 dimensions in the MCI/AD population
Clinical association studies correlated the two dimensions with 45
clinical variables and biomarkers. Compared to the R1 dimension, R2
showed associations, to a larger extent than R1, with SPARE-AD and
RAVLT. SPARE-AD quantifies the presence of a typical imaging signature
of AD-related brain atrophy, which has been previously shown to predict
clinical progression in both CU and MCI individuals [[225]41]. RAVLT
measures episodic memory, a reliable neuropsychological phenotype in
AD, which is also correlated with medial temporal lobe atrophy
[[226]51, [227]52]. The R1 dimension was associated to a greater extent
with 1) SPARE-BA, which captures the individualized expression of
advanced brain age from MRI [[228]42]; 2) white matter lesions (WML),
which are commonly associated with vascular risk factors and cognitive
decline [[229]53], and 3) whole-brain uptake of 18F-fluorodeoxyglucose
(FDG) PET, which is a biomarker of brain metabolic function and
atrophy. Both dimensions were positively associated with cerebrospinal
fluid (CSF) levels of tau and p-tau and negatively associated with the
CSF level of Aβ42 [[230]54] (Fig. [231]1C), as well as the whole-brain
standardized uptake value ratio of 18F-AV-45 PET (Supplementary eTable
[232]3). Results for all 45 clinical variables, including cognitive
scores, modifiable risk factors, CSF biomarkers, disease/condition
labels, demographic variables, and imaging-derived phenotypes, are
presented in Supplementary eTable [233]3 for p values and effect sizes
(i.e., beta coefficients).
Clinical profiles of the R1 and R2 dimensions in the general population
Brain association studies confirmed the presence of the two atrophy
patterns in the general population (Fig. [234]2A and Supplementary
eTable [235]4 for p values and effect sizes). In clinical association
studies, the R1 dimension was significantly associated, to a larger
extent than R2, with cardiovascular (e.g., triglycerides) and diabetes
factors (e.g., Hba1c and glucose), executive function (TMT-B),
intelligence, physical measures (e.g., diastolic blood pressure),
SPARE-BA [−log[10](p value) = 236.89 for R1 and −46.35 for R2] and WML
[−log[10](p value) = 120.24 for R1 and 2.06 for R2]. In contrast, the
R2 dimension was more significantly associated with SPARE-AD
[−log[10](p value) = 136.01 for R1 and 250.41 for R2] and prospective
memory (Fig. [236]2B). Results for all 61 clinical variables, including
cardiovascular factors, diabetic blood markers, social demographics,
lifestyle, physical measures, cognitive scores, and imaging-derived
phenotypes, are presented in Supplementary eTable [237]5 for p values
and effect sizes.
Fig. 2. The expression of the R1 and R2 dimensions in the general population.
[238]Fig. 2
[239]Open in a new tab
A Brain association studies confirm the presence of the two dimensions
in the general population: the R1 dimension shows widespread brain
atrophy, whereas the R2 dimension displays focal medial temporal lobe
atrophy. p value and effect sizes (r, Pearson’s correlation
coefficient) are presented in Supplementary eTable [240]4. The brain
maps denote the signed p value, and the range of r for each dimension
is also shown. Of note, the sample size (N) for R1 and R2 is the same
for each ROI. B Clinical association studies further show that the R2
dimension is associated with prospective memory, and the R1 dimension
is associated with several cognitive dysfunctions, cardiovascular risk
factors (e.g., triglycerides), and diabetes (e.g., HbA1c). The same
linear regression models were used to associate the R1 and R2
dimensions with the 61 clinical variables, including cardiovascular
factors, diabetic blood markers, social demographics, lifestyle,
physical measures, cognitive scores, and imaging-derived phenotypes.
The radar plot shows representative clinical variables; all other
results are presented in Supplementary eTable [241]5. The gray circle
lines indicate the p value threshold in both directions (Bonferroni
correction for the 61 variables: −log[10](p value) >3.08). A
positive/negative −log[10](p value) value indicates a positive/negative
correlation (beta). Transparent dots represent the associations that do
not pass the Bonferroni correction; the blue-colored dots and
red-colored dots indicate significant associations for the R1 and R2
dimensions, respectively. C Genome-wide association studies demonstrate
that the R2 dimension is associated to a larger extent with genomic
loci and genes previously associated with AD-related traits in the
literature (genome-wide p value threshold with the red line: −log[10](p
value) >7.30). Each genomic locus is represented by its top lead SNP.
The R1 dimension identified 8 (blue-colored in bold) out of the 49
mapped genes associated with AD-related traits. The R2 dimension
identified 13 (red-colored in bold) out of 40 mapped genes associated
with AD-related traits. Gene annotations were performed via positional,
expression quantitative trait loci, and chromatin interaction mappings
using FUMA (Supplementary eTable [242]6 for all mapped genes)
[[243]58]. The genomic loci and mapped genes were manually queried in
the GWAS Catalog [[244]55] to determine whether they were previously
associated with AD (newly identified or not). * denotes that the
genomic locus is newly identified. D Besides AD-related traits, the
genes and genomic loci in the two dimensions were also associated with
other clinical traits, including inflammation, neurohormones, and
imaging-derived phenotypes, as shown in the literature from the GWAS
Catalog [[245]55]. The flowchart first maps the genomic loci and genes
(left) identified in the two dimensions onto the human genome (middle).
It then links these variants to any clinical traits identified in
previous literature from the GWAS Catalog (right). In the middle of the
human genome, we show chromosomes 1 to 22 (above to below); the blue
and red-colored genes are AD-related for the R1 and R2 dimensions,
respectively. The black-colored genes (C) are not annotated. INF
inflammation, PD psychiatric disorder, PM physical measure; “New”
(corresponding to the newly identified loci/genes in C) indicates that
the locus or gene was not associated with any traits in the literature.
DSST Digit Symbol Substitution Test, TMT Trail Making Test, CRP
C-reactive protein, AD Alzheimer’s disease, PD Parkinson’s disease, INF
inflammation, IDP imaging-derived phenotype.
Twenty-four genomic loci and 77 genes unrelated to APOE are associated with
the R1 and R2 dimensions in the general population
GWAS identified 24 genomic loci, 14 of which are newly identified (not
previously associated with any traits in GWAS Catalog), and 77
positionally and functionally mapped genes unrelated to APOE associated
with R1 or R2. In particular, the R1 dimension was significantly
associated with 11 genomic loci and 49 genes. Eight genes (blue-colored
genes in Fig. [246]2D) were previously associated with AD-related
traits; 12 newly identified loci/genes have not been previously
associated with any clinical traits. The R2 dimension was significantly
associated with 13 genomic loci and 40 annotated genes. 13 genes
(red-colored genes in Fig. [247]2D) were associated with AD-related
traits; 8 loci/genes were newly identified (Fig. [248]2C and
Supplementary eTable [249]6). These genomic loci and genes were also
associated with many clinical traits in the literature from the GWAS
Catalog [[250]55]. These included hormones (e.g., sex hormone-binding
globulin measurement vs. CCKN2C), inflammatory factors (e.g.,
macrophage inflammatory protein 1b measurement vs. CDC25A),
imaging-derived phenotypes (e.g., cerebellar volume measurement from
MRIs vs. DMRTA2), and psychiatric disorders (e.g., unipolar depression
vs. ASTN2) (Fig. [251]2D). Details of the GWAS Catalog results are
presented in Supplementary eFile [252]1. The Manhatton and QQ plots of
the baseline GWAS are presented in Supplementary eFig. [253]9. The LDSC
[[254]56] intercept of the two GWASs was close to 1, indicating no
substantial genomic inflation (R1 = 1.0032 ± 0.0084;
R2 = 1.023 ± 0.0084). Furthermore, our main GWASs using European
ancestry were robust in three sensitivity check analyses: split-sample,
sex-stratified, and mixed-effect [[255]57] linear model analyses.
Detailed results are presented in Supplementary eText [256]1 and
Supplementary eFile [257]2–[258]4.
The two dimensions were significantly heritable in the general
population based on the SNP-based heritability estimates (R1:
h^2 = 0.49
[MATH: ± :MATH]
0.02; R2: h^2 = 0.55
[MATH: ± :MATH]
0.02). The PRS of AD showed a marginally positive association with the
R2 dimension [−log[10](p value) = 1.42], but not with the R1 dimension
[−log[10](p value) = 0.47 < 1.31] in this population.
Genes associated with the R1 and R2 dimensions are overrepresented in organs
beyond the brain in the general population
Tissue specificity analyses test whether the mapped genes are
overrepresented in differentially expressed gene sets (DEG) in one
organ/tissue compared to all other organs/tissues using different gene
expression data [[259]58]. The genes associated with the R1 dimension
were overrepresented in the caudate, hippocampus, putamen, amygdala,
substantia nigra, liver, heart, and pancreas; the genes associated with
the R2 dimension were overrepresented in the caudate, hippocampus,
putamen, amygdala, anterior cingulate, pituitary, liver, muscle,
kidney, and pancreas (Fig. [260]3A and Supplementary eFig. [261]10).
Genes in DEG over-expressed in the heart were only associated with R1,
while those in DEG over-expressed in the pituitary gland, muscle, and
kidney were unique in R2. The expression values of every single gene
for all tissues are presented in Supplementary eFig. [262]11.
Fig. 3. Tissue specificity and biological pathway enrichment analysis of the
R1 and R2 dimensions in the general population.
[263]Fig. 3
[264]Open in a new tab
A Tissue specificity analyses show that genes associated with the two
dimensions of neurodegeneration are overrepresented in organs/tissues
beyond the human brain (R1 and R2). The unique overrepresentation of
genes in differentially expressed gene sets (DEG) in the heart (R1) and
the pituitary gland, muscle, and kidney (R2) may imply the involvement
of inflammation [[265]12, [266]75, [267]76] and neurohormone
dysfunction [[268]15–[269]17], respectively. The GENE2FUNC [[270]58]
pipeline from FUMA was performed to examine the overrepresentation of
prioritized genes (Fig. [271]2C) in pre-defined DEGs (up-regulated,
down-regulated, and both-side DEGs) from different gene expression
data. The input genes (Fig. [272]2C) were tested against each DEG using
the hypergeometric test. We present only the organs/tissues that passed
the Bonferroni correction for multiple comparisons. B Gene set
enrichment analysis shows that genes associated with the two dimensions
are enriched in different biological pathways. For example, genes
associated with the R1 dimension are implicated in down-regulated
macrophage functions, which have been shown to be associated with
inflammation [[273]13]. In contrast, the R2 dimension is enriched in AD
hallmarks (e.g., hippocampus), AD-related gene sets, and the pathway
involved in dendritic cells, which may regulate amyloid-β-specific
T-cell entry into the brain [[274]60]. Both dimensions are enriched in
gene sets involved in cancer, which may indicate overlapped genetic
underpinnings between AD and cancer [[275]59]. The GENE2FUNC [[276]58]
pipeline from FUMA was performed to examine the enrichment of
prioritized genes (Fig. [277]2C) in pre-defined gene sets.
Hypergeometric tests were performed to test whether the input genes
were overrepresented in any pre-defined gene sets. Gene sets were
obtained from different sources, including MsigDB [[278]95] and GWAS
Catalog [[279]55]. We show the significant results from gene sets
defined in the GWAS Catalog, curated gene sets, and immunologic
signature gene sets. All results are shown in Supplementary eTable
[280]7). C The target-drug-disease network for R1 and R2-associated
genes provides great potential for drug discovery and repurposing.
R1-annotated “druggable genes” were developed for cardiovascular
diseases, various cancers, and inflammation, whereas R2-annotated
“druggable genes” were developed for diseases of the nervous system
(e.g., Parkinson’s disease). For the target-drug-disease network, the
5^th level of the Anatomical Therapeutic Chemical (ATC) code is
displayed for the DrugBank database [[281]96], and the disease name
defined by the International Classification of Diseases (ICD-11) code
is showed for the Therapeutic Target Database [[282]97]. The human
anatomy was created with [283]https://www.biorender.com/.
Genes associated with the R1 and R2 dimensions are enriched in key biological
pathways in the general population
Genes associated with the two dimensions were enriched in different
biological pathways. Genes associated with the two dimensions were
implicated in several types of cancer, including up-regulation of
fibroblast, breast cancer, and neuroblastoma tumors (Fig. [284]3B),
which indicate a certain extent of genetic overlaps and shared pathways
that may explain the intriguing inverse relationship between AD and
cancer [[285]59]. Genes associated with the R1 dimension were
implicated in pathways involved in the down-regulation of macrophages
(Fig. [286]3B), which are involved in the initiation and progression of
various inflammatory processes, including neuroinflammation and AD
[[287]13]. Inflammation is also known to be associated with vascular
compromise and dysfunction. This further concurs with the stronger
cardiovascular profile of R1, especially with increased WML and
predominant SPARE-BA increases. Genes associated with the R2 dimensions
were enriched in pathways involved in AD onset, hippocampus-related
brain volumes, and dendritic cells (Fig. [288]3B). In particular,
dendritic cells may regulate amyloid-β-specific T-cell entry into the
brain [[289]60], as well as the inflammatory status of the brain
[[290]61]. The gene set enrichment analysis results are presented in
Supplementary eTable [291]7.
Genes associated with the R1 and R2 dimensions show potential for drug
discovery and repurposing
We queried whether these 77 genes associated with R1 and R2 are
“druggable genes” from the constructed target-drug-disease network—the
target genes express proteins to bind drug-like molecules, and the drug
is at any stage of the clinical trial. For the 49 R1-annotated genes, 9
genes were targets for 15 drugs and drug-like molecules, treating
various cancer, inflammation, and cardiovascular dysfunctions. For the
40 R2-annotated genes, 6 genes were targets for 7 drugs developed for
diseases of the nervous system, such as Parkinson’s (Fig. [292]3C). The
pharmacological mechanisms targeted by these identified drugs are
largely related to the pathogenesis of AD in previous literature. For
example, FDA-approved Niacin [R1; target gene: NNMT; Anatomical
Therapeutic Chemical (ATC) code: C10AD02] is a B vitamin used to treat
various deficiencies and diseases in the cardiovascular system,
including myocardial infarctions [[293]62], hyperlipidemia [[294]63],
and coronary artery disease [[295]64]. Interestingly, a recent study
[[296]65] showed that Niacin detained AD progression in a 5xFAD mice
model. The niacin receptor HCAR2 modulates microglial response to
amyloid deposition, ultimately alleviating neuronal loss and cognitive
decline. Other drugs for potential drug repurposing of AD are the
FDA-approved Docetaxel (R1; target gene: MAP4; ATC: L01CD02) and
Paclitaxel (R1; target gene: MAP4; ATC: L01CD01), which both target
various cancers, including breast cancer and metastatic prostate
cancer. The intriguing inverse relationship between AD and cancer has
long been established, but the underlying shared etiology remains
unclear [[297]43]. One hypothesis was that microtubule-associated
protein tau—a pathological biomarker of AD—was associated with
resistance to Docetaxel in certain cancer treatments [[298]66]. In
addition, Docetaxel impacted the blood-brain barrier function of breast
cancer brain metastases [[299]67]. Another drug called KM-819 (R2;
target gene: FAF1) is currently in Phase 1 for a clinical trial of
Parkinson’s disease [[300]68], which aims to suppress
α-synuclein-induced mitochondrial dysfunction [[301]69], consistent
with the mitochondrial hypothesis [[302]70] of AD. To sum up, R1 and R2
show distinct landscapes of the “druggable genome” [[303]71] on drug
discovery and repurposing [[304]72] for future clinical translation.
The longitudinal rate of change in the R2 dimension, but not R1, is
marginally associated with the APOE ε4 allele, tau in cognitively unimpaired
individuals
Using cognitively unimpaired participants from ADNI and BLSA,
longitudinal brain association studies showed that the rate of change
in the R1 dimension was associated with the change of brain volume in
widespread brain regions. In contrast, the rate of change in the R2
dimension was associated with the change of brain volume in the focal
medial temporal lobe (Fig. [305]4A and Supplementary eTable [306]8 for
p values and effect sizes). This further indicates that the two
dominant patterns discovered cross-sectionally also progress in
consistent directions longitudinally. The two dimensions were not
associated with CSF biomarkers (Aβ42, tau, and p-tau) and the APOE ε4
allele (rs429358) at baseline [−log[10](p value) < 1.31)]. The rate of
change of the R2 dimension, but not R1, was marginally [nominal
threshold: −log[10](p value) >1.31] associated with the APOE ε4 allele,
the CSF level of tau, and p-tau (Fig. [307]4B and Supplementary eTable
[308]9 for p values and effect sizes), but they did not survive the
Bonferroni correction [−log[10](p value) = 2.95]. The longitudinal rate
of change of both dimensions was negatively associated [−log[10](p
value) >2.95] with the total CSF level of Aβ42.
Fig. 4. The longitudinal rate of change in R1 and R2 in the cognitively
unimpaired population.
[309]Fig. 4
[310]Open in a new tab
A Longitudinal brain association studies show that the R1 dimension
exhibits longitudinal brain volume decrease in widespread brain
regions, whereas the R2 dimension displays longitudinal brain volume
decrease in the focal medial temporal lobe. We first derived the rate
of change of the 119 GM ROIs and the R1 and R2 dimensions using a
linear mixed effect model; a linear regression model was then fit to
the rate of change of the ROIs, R1, and R2 to derive the beta
coefficient value of each ROI. A negative value denotes longitudinal
brain changes with a negative coefficient of the rate of change in the
linear regression model. p value and effect sizes (r, Pearson’s
correlation coefficient) are presented in Supplementary eTable [311]8.
The brain maps denote the signed p value, and the range of r for each
dimension is also shown. Of note, the sample size (N) for R1 and R2 is
the same for each ROI. B The rate of change, not the baseline
measurement, in the two dimensions is negatively associated with the
CSF level of Aβ42 (Bonferroni correction for the 45 variables:
−log[10](p value) >2.95). The rate of change in the R2 dimension, not
the R1 dimension, was marginally (−log[10](p value) >1.31) associated
with the CSF level of tau and p-tau, and APOE ε4. All other clinical
associations are presented in Supplementary eTable [312]9. The
gray-colored circle lines indicate different p value thresholds in both
directions (Bonferroni correction for the 45 variables: −log[10](p
value) >2.95 and the nominal p value threshold: −log[10](p value)
>1.31). A positive/negative −log[10](p value) value indicates a
positive/negative correlation (beta). Transparent dots represent the
associations that do not pass the nominal p value threshold [log[10](p
value) = 1.31]; the blue-colored dots and red-colored dots indicate
significant associations [log[10](p value) >1.31] for the R1 and R2
dimensions, respectively.
We tested these associations using cognitively unimpaired individuals
with a high risk of AD based on their family history from the
PREVENT-AD cohort. Similarly, at baseline, the two dimensions were not
associated with CSF biomarkers or the APOE ε4 allele (rs429358). The
longitudinal rate of change in the R2 dimension, but not R1, was
marginally [nominal threshold: −log[10](p value) >1.31] associated with
the APOE ε4 allele [−log[10](p value) = 1.92], the CSF level of tau
[−log[10](p value) = 1.65], and p-tau [−log[10](p value) = 1.66].
Longitudinal brain association studies also confirmed the longitudinal
progression of the two dimensions in the MCI/AD population
(Supplementary eFig. [313]12A). The rates of change in the two
dimensions were both associated with APOE ε4 [−log[10](p value) = 12.54
for R1 and 9.05 for R2] in GWAS (Supplementary eFig. [314]12B), and
related to CSF levels of tau [−log[10](p value) = 16.47 for R1 and 9.73
for R2], p-tau [−log[10](p value) = 19.13 for R1 and 10.81 for R2], and
Aβ42 [−log[10](p value) = 13.64 for R1 and 13.55 for R2] (Supplementary
eFig. [315]12C).
Discussion
The current study leveraged a deep semi-supervised representation
learning method to establish two predominant dimensions in the
symptomatic MCI/AD population, which were independently found to be
expressed, to a lesser degree, in three asymptomatic populations. In
particular, the R1 dimension represented a “diffuse-AD” atrophy
pattern: varying degrees of brain atrophy throughout the entire brain.
In contrast, the R2 dimension showed an “MTL-AD” atrophy pattern: brain
atrophy predominantly concentrated in the medial temporal lobe (Fig.
[316]1A). Importantly, only R2 was found to be significantly associated
with genetic variants of the APOE genes in MCI/AD patients.
Furthermore, our study examined early manifestations of the R1 and R2
dimensions in asymptomatic populations with varying levels of AD risks
and their associations with genetics, amyloid plaques and tau tangles,
biological pathways, and body organs. We identified that 24 genomic
loci, 14 of which are GWAS identified 24 genomic loci, 14 of which are
newly identified, and 77 annotated genes contribute to early
manifestations of the two dimensions. Functional analyses showed that
genes unrelated to APOE were overrepresented in DEG sets in organs
beyond the brain (R1 and R2), including the heart (R1) and the
pituitary gland (R2), and enriched in several biological pathways
involved in dendritic cells (R2), macrophage functions (R1), and cancer
(R1 and R2). Several of these genes were “druggable genes” for cancer
(R1), inflammation (R1), cardiovascular diseases (R1), and diseases of
the nervous system (R2). Longitudinal findings in the cognitively
unimpaired populations showed that the rate of change of the R2
dimension, but not R1, was marginally associated with the APOE ε4
allele, the CSF level of tau, and Aβ42 (R1 and R2). Our findings
suggested that diverse pathologic processes, including cardiovascular
risk factors, neurohormone dysfunction, and inflammation, might occur
in the early asymptomatic stages, supporting and expanding the current
amyloid cascade (Fig. [317]5) [[318]7, [319]8].
Fig. 5. Genes unrelated to APOE influence early manifestations of R1 and R2.
[320]Fig. 5
[321]Open in a new tab
Genes unrelated to APOE and overrepresented in organs beyond the human
brain are associated with early manifestations of the R1 (diffuse-AD)
and R2 (MTL-AD) dimensions, which capture the heterogeneity of
AD-related brain atrophy. For visualization purposes, we display the
two genes with the highest expression values in the tissue specificity
analyses for each organ/tissue. The black arrow line emulates the
longitudinal progression trajectory along these two dimensions. The
positions of beta-amyloid, tau, and APOE (increasing APOE-mediated
progression) indicate the time point when they are associated with the
two dimensions. The blue/red gradient-color background indicates a
higher influence of APOE-related genes (left to right; early to late
stages). The brain atrophy patterns are presented in the 3D view. In
early asymptomatic stages, the R1-related genes are implicated in
cardiovascular diseases and inflammation; the R2-related genes are
involved in hormone-related dysfunction. Critically, longitudinal
progression of the dimension demonstrates an impact of the APOE genes
in early asymptomatic stages in R2, but this longitudinal effect occurs
only in late symptomatic stages in R1. These results suggest that
comorbidities (e.g., cardiovascular conditions) or normal aging in R1
may alter or delay the trajectory of neurodegeneration in early
asymptomatic stages; APOE-related genes may play a pronounced role in
the acceleration and progression in late symptomatic stages for both
dimensions. Of note, the underlying pathological processes that
initiate and drive the progression of the two dimensions are not
mutually exclusive. Hence, both R1 and R2 can be co-expressed in the
same individual. In addition, the two dimensions can also be affected
by other AD hypotheses, such as the mitochondrial hypothesis [[322]70]
and the metabolic hypothesis [[323]86]. MTL medial temporal lobe. The
human anatomy was created with [324]https://www.biorender.com/.
AD has been regarded as a CNS disorder. However, increasing evidence
has indicated that the origins or facilitators of the pathogenesis of
AD might involve processes outside the brain [[325]6]. For example,
recent findings revealed that gut microbiota disturbances might
influence the brain through the immune and endocrine system and the
bacteria-derived metabolites [[326]73, [327]74]. Our findings support
the view that multiple pathological processes might contribute to early
AD pathogenesis and identify non-APOE genes in the two dimensions
overrepresented in tissues beyond the brain (e.g., the heart, pituitary
gland, muscle, and kidney). Pathological processes may be involved in
different cells, molecular functions, and biological pathways,
exaggerating amyloid plaque and tau tangle accumulation and leading to
the downstream manifestation of neurodegeneration and cognitive
decline.
The genetic and clinical underpinnings of the R1 dimension support
inflammation, as well as cardiovascular diseases, as a core pathology
contributing to AD [[328]12, [329]75, [330]76]. Genes associated with
the R1 dimension were previously associated with various
inflammation-related clinical traits (Fig. [331]2D), and enriched in
biological pathways involved in immunological response (e.g.,
up-regulation in macrophages [[332]77], Fig. [333]3B). In addition,
genes in this dimension were overrepresented in DEG sets in the heart
(Fig. [334]3A). Previous literature indicated that inflammation is
likely an early step that initiates the amyloidogenic pathway—the
expression of inflammatory cytokines leads to the production of
β-amyloid plaques [[335]13]. Several markers of inflammation are also
present in serum and CSF before any indications of Aβ or tau tangles
[[336]78]. For example, clusterin, a glycoprotein involved in many
processes and conditions (e.g., inflammation, proliferation, and AD)
induced by tumor necrosis factor (TNF), was present ten years earlier
than Aβ deposition [[337]79]. In addition, the R1 dimension was also
strongly associated with cardiovascular and diabetes biomarkers (Fig.
[338]2B). Inflammatory processes have been critical, well-established
risk factors for compromised cardiovascular function [[339]80], such as
coronary artery disease and the breakdown of the blood-brain barrier.
Our results corroborated the close relationships between AD,
cardiovascular diseases, and inflammation.
The genetic and clinical underpinnings of the R2 dimension support that
neuroendocrine dysfunction might be an early event contributing to the
pathogenesis of AD [[340]16, [341]17]. Genes in the R2 dimension were
previously associated with different hormone and pancreas-related
traits from GWAS Catalog (Fig. [342]2D); they were also overrepresented
in DEG in the pituitary and pancreas glands, muscle and kidney (Fig.
[343]3A), which are master glands or key organs in the endocrine system
[[344]81]. Previous literature suggested that neuroendocrine
dysfunction might contribute to AD development by secreting
neurohormonal analogs and affecting CNS function [[345]16]. For
example, luteinizing hormone-releasing hormone and follicle-stimulating
hormone in serum or neurons were associated with the accumulation of Aβ
plaques in the brain [[346]17, [347]82, [348]83]. However, early
experimental studies on antagonists of Luteinizing hormone-releasing
hormone and growth hormone-releasing hormone in animal models of AD
have shown promising but not entirely convincing evidence [[349]16].
Taken together, neurodegeneration in the R2 dimension represents an
AD-specific phenotype that might be driven by hormonal dysfunction,
leading to rapid accumulation of amyloid plaques, and was potentially
accelerated by the APOE ε4 allele—the rate of change in R2, but not R1,
was associated with the APOE ε4 allele in cognitively unimpaired
individuals (Fig. [350]4B).
The hypothesized implications above of the R1 and R2 dimensions on
inflammation, cardiovascular functions, and neuroendocrine dysfunctions
are not mutually exclusive and may collectively contribute to AD
pathogenesis. It has been shown that dysregulation of the
hypothalamic-pituitary-gonadal axis is associated with dyotic
signaling, modulating the expression of TNF and related cytokines in
systemic inflammation, and the induction of downstream
neurodegenerative cascades within the brain [[351]84, [352]85]. These
studies hypothesized that the neuroendocrine dysfunction and the
inflammation mechanism might be the upstream and downstream
neuropathological processes along the disease course of AD [[353]16].
That is, the loss of sex steroids and the elevation of gonadotropins
might lead to a higher level of inflammatory factors in the brain.
Finally, other competing hypotheses may also play a role in developing
AD in early asymptomatic stages, including the mitochondrial hypothesis
[[354]70], the metabolic hypothesis [[355]86], and the tau hypothesis
[[356]3].
The NIA-AA framework [[357]87] claims that AD is a continuum in which
AD pathogenesis is initiated in early asymptomatic cognitively
unimpaired stages and progresses to amyloid-positive and tau-positive
(A+T+) in late symptomatic stages [[358]87]. Our findings are
consistent with this framework and elucidate the cross-sectional and
longitudinal associations of the two dimensions with genetic and
clinical markers from early asymptomatic to late symptomatic stages. In
early asymptomatic stages, the rates of change in the two dimensions
are both associated with amyloid. However, only the R2 dimension, not
R1, is marginally associated with the APOE ε4 allele and the CSF level
of tau (Fig. [359]4B). In contrast, in late symptomatic stages, the
rates of change in the two dimensions are both associated with the APOE
ε4 allele, CSF levels of tau, p-tau, and amyloid. Our findings suggest
that comorbidities or normal aging in R1 may alter the rate or
trajectory of neurodegeneration at early asymptomatic stages, but
APOE-related genes might play a more pronounced role in the
acceleration and progression during late symptomatic stages for both
dimensions (Fig. [360]5).
Several recent studies [[361]88–[362]92], as detailed in an insightful
overview by Luo et al., collectively provide a comprehensive
transcriptomics and epigenomics atlas depicting AD progression at the
single-cell level [[363]93]. Similarly, researchers have also proposed
a new theory suggesting that Alzheimer’s disease may not only be a
brain disorder but could also be considered an autoimmune disease
[[364]94]. These studies highlight the involvement of microglia-related
inflammation, lipid metabolism, and mitochondrial dysfunction. This
substantiates the primary hypothesis in our study: the two dimensions
are linked to diverse pathological mechanisms, encompassing
cardiovascular diseases, inflammation, and hormonal dysfunction,
potentially driven by genes beyond APOE.
Limitations
This study has several limitations. Firstly, there is a need for
longitudinal data from the general population, as exemplified by the UK
Biobank, to provide further validation for the hypotheses proposed to
cover the entire AD spectrum in the same population. Secondly, it is
essential to extend the generalization of the current GWAS findings to
include underrepresented ethnic groups, going beyond the European
ancestry populations.
Outlook
In conclusion, the current study used a novel deep semi-supervised
representation learning method to establish two AD dimensions. Our
findings support that those diverse pathological mechanisms, including
cardiovascular diseases, inflammation, hormonal dysfunction, and
involving multiple organs, collectively affect AD pathogenesis in
asymptomatic stages. These novel biomarkers may serve as instrumental
variables to guide future treatments in the early asymptomatic stages
of AD, targeting multi-organ dysfunction beyond the brain.
Supplementary information
[365]Supplementary eFiles^ (334.6KB, xlsx)
[366]Supplementary Materials^ (7.5MB, docx)
Acknowledgements