A phenome-wide hunt for risk factors of Alzheimer’s disease: from metabolic clues to neuroimaging evidence

Abstract

Background

This study systematically investigated phenotypes causally associated with Alzheimer’s disease (AD) across the phenome and validated the findings at cognitive and neuroimaging levels using real-world clinical data.

Methods

We performed phenome-wide Mendelian Randomization (MR) analyses on genetic proxies for over 860 disease phenotypes to identify traits causally associated with AD. Lipid metabolism-related phenotypes identified through MR were further examined in the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset to assess associations with AD risk, brain structure, and cognition.

Results

MR analyses revealed a significant causal association between lipid metabolism, particularly low-density lipoprotein cholesterol (LDL-C), and the risk of AD (OR: 1.05, 95% CI: 1.03–1.07). In ADNI, higher LDL-C indicators correlated with increased AD risk, reduced hippocampal and entorhinal volumes, and poorer cognitive performance. Notably, elevated cholesterol-to-total lipid ratios in small LDL particles were negatively associated with the entorhinal-hippocampal complex. Among cognitively normal individuals, higher LDL-C indicators were associated with smaller hippocampus-amygdala transition area (HATA) and CA3 head volumes. In those with mild cognitive impairment, higher LDL-C was associated with reduced entorhinal surface area.

Conclusions

Our findings suggest that disrupted LDL-C metabolism may play a causal role in the development and progression of AD.

Supplementary information

The online version contains supplementary material available at 10.1186/s12967-026-07756-4.

Introduction

Alzheimer’s disease (AD), the leading cause of dementia in aging populations, manifests as progressive memory loss, language dysfunction, and cognitive decline [1]. Despite decades of research, effective therapies remain elusive, partly because many proposed therapeutic targets may reflect downstream consequences rather than upstream drivers of AD pathogenesis [2]. This underscores the critical need to disentangle causal relationships between modifiable risk factors and AD progression through robust epidemiological frameworks.

Among the various diseases comorbid with AD, disorders related to metabolism are frequently reported, including diabetes [3, 4], hypertension [5], obesity [6], and disturbances of lipid metabolism [7]. Reports indicate that the prevalence of metabolic syndrome (MetS) in the population aged 15 and above is approximately 24.5% to 34.7% [8, 9]. Therefore, achieving a comprehensive understanding of the causal impacts of MetS and related disease phenotypes on AD is crucial. The diagnosis of MetS is based on the presence of at least three of the following metabolic conditions [10]: 1) abdominal obesity; 2) hypertension; 3) high fasting plasma glucose; 4) elevated plasma triglycerides, and 5) reduced high-density lipoprotein cholesterol (HDL-C). Recent research suggests that MetS may be a risk factor for dementia [3, 11]. Utilizing the extensive population-based prospective cohort of the UK Biobank, Qureshi et al. identified an elevated correlation between MetS and the incidence of dementia. More pronounced associations were noted in individuals with four or more MetS components and those without the apolipoprotein E (APOE)-ε4 allele [11]. Machado-Fragua et al. [3] propose that the risk of dementia increases with each additional MetS component manifesting in midlife. Furthermore, they argue that this risk does not accumulate only after a threshold of three components. However, a meta-analysis based on longitudinal studies suggests that there is no statistically significant association between MetS and dementia or AD [12]. Hence, the causative links between MetS and AD remain inconclusive, and the specific contributions of different types of metabolic abnormalities in this process remain unclear. Currently, it remains unclear which specific MetS component primarily affects AD and how it influences brain phenotypes and cognitive function.

As a phenotype developing under the influence of multiple mechanisms, including genetics, proteins, and metabolism [2, 13], AD is an embodiment of complexity. This makes a comprehensive understanding of this disease a challenge that cannot be cracked through looking at one or a few risk factors or connected phenotypes. Inferring causal effects solely from observational studies of AD may not be feasible due to the presence of confounding variables and limitations in sample size. Mendelian randomization (MR) offers a powerful solution to these challenges. In recent years, MR methods have been widely applied in causal inference of neuropsychiatric diseases and MetS phenotypes [14–18]. The validity of MR analysis relies on three fundamental assumptions: (1) the relevance assumption, requiring strong association between genetic instruments and the exposure (typically with F-statistics > 10); (2) the independence assumption, stipulating that genetic variants are not associated with any confounder affecting both exposure and outcome; and (3) the exclusion restriction assumption, demanding that genetic instruments should affect the outcome solely through the exposure pathway without horizontal pleiotropy. By leveraging genetic variants as instrumental variables (IVs), MR minimizes confounding biases inherent in observational studies, enabling causal inference between exposures and outcomes [19], which provides an opportunity to uncover causal associations between various disease phenotypes and AD. However, existing MR studies have narrowly focused on isolated metabolic traits, neglecting the phenome-wide complexity of MetS-AD interactions. A systematic evaluation of causal pathways across the entire metabolic spectrum is urgently needed.

To address these gaps, we conducted a three-stage analytical framework integrating phenome-wide causal discovery, clinical validation, and neuroanatomical mapping (Fig. 1). First, in the discovery phase, we conducted hypothesis-agnostic phenome-wide MR screening of 868 phenotypes within the FinnGen cohort (n = 377,000), prioritizing exposures with robust genetic evidence for causal links to AD and further validating these associations through additional data. Following this, the associations between the target phenotype and the risks of AD were validated using data from real-world populations. Finally, employing the neuroimaging dataset, we explored the correlations between risk factors and subregional structures within the medial temporal lobe in populations with prodromal AD and normal cognitive function. This multi-level approach connects genetic data, clinical outcomes, and brain structure changes to identify how metabolic factors contribute to AD progression.

Fig. 1.

Overview of the study design. A: In the initial phase, we identified phenotypes causally linked to AD risk, focusing on metabolic disorders. Our emphasis was on screening key components related to metabolic syndrome, revealing a predominant role of lipid metabolism disruption in major MetS-associated phenotypes. Results from lipoprotein data underscored a significant association between AD risk and low-density lipoprotein cholesterol (LDL-C), highlighting lipid metabolism dysregulation in AD risk. B: In the second phase, we expanded our investigation to the ADNI database, exploring relationships between LDL-C metabolic indicators and AD risk, cognitive function, and brain phenotypes in the elderly population. C: In the third phase, we validated the causal link between LDL-C levels and the hippocampus using MR analysis. Leveraging real brain imaging data from ADNI, we further explored the association between LDL-C metabolic indicators and detailed brain phenotypes in the medial temporal lobe, both in mild cognitive impairment (MCI) and cognitively normal (CN) populations

Materials and methods

Study design and data sources

This study consists of three stages. Firstly, using the recently released large-scale FinnGen R9 Genome-wide association study (GWAS) datasets, we conducted a phenome-wide two-sample MR analysis. We identified several phenotypes causally associated with AD risk, with a primary focus on metabolic-related disorders. Furthermore, we investigated the key components of MetS and a variety of associated variables, revealing that lipid metabolism disruption plays a predominant role in AD. These findings support a hypothesis connecting circulating levels of low-density lipoprotein cholesterol (LDL-C), apolipoprotein B (ApoB), and apolipoprotein E (ApoE) with an elevated risk of AD. In the second stage of this study, we extended our investigation to the ADNI database. We aimed to dissect the relationships between lipid metabolic indicators associated with LDL-C and the risk of AD, cognitive function, and brain phenotypes in a real-world elderly cohort. Finally, in the third stage, we validated the causal relationship between LDL-C levels and the various structural features of the hippocampus through MR analysis. Simultaneously, utilizing real brain imaging data from the ADNI database, we further elucidated the association between LDL-C metabolic indicators and the brain phenotypes of the medial temporal lobe at the subfield levels, both in the patients with mild cognitive impairment (MCI) and in the cognitively normal (CN) subjects. Figure 1 illustrates the workflow of the current study.

Summary-level genetic data for AD, with cases totaling 71,880 and controls 383,378, were obtained from the GWAS study conducted by Jansen et al. [20]. We incorporated GWAS summary data for disease phenotypes from the FinnGen R9 database [19] as the exposure factors, specifically selecting those with a sample size exceeding 2000. We systematically excluded six categories of phenotypes (Supplementary Table S1): 1) medication-related phenotypes to avoid treatment confounding (n = 8); 2) reimbursement codes representing administrative rather than diagnostic categories (n = 2); 3) AD/dementia-related phenotypes to prevent analytical circularity (n = 10); 4) non-disease-related measurements such as anthropometric data (n = 4); 5) repetitive or overlapping phenotypes to eliminate redundancy (n = 9); and 6) death-related outcomes and other non-informative categories (n = 3). In addition to utilizing the available data from FinnGen, we also integrated summary-level GWAS genetic data of the waist circumference (cases = 455,545) provided by Jiang et al. [21]. To further elucidate the impact of several major blood lipid levels on AD, total cholesterol (TC), triglycerides (TG), HDL-C, LDL-C, and very-low-density lipoprotein cholesterol (VLDL-C) datapoints were incorporated using summary-level GWAS genetic data for these lipoproteins from studies conducted by Graham et al. [22] and Richardson et al. [23]. Refer to Supplementary Table S3 for detailed information on the GWAS datasets of metabolism-related phenotypes. ADNI (https://adni.loni.usc.edu/) is a longitudinal study initiated in 2003, focusing on adults aged 55 to 90 years. ADNI aims to comprehensively investigate the pathogenesis of AD, early diagnostic methods, and the natural progression of the disease by collecting multimodal neuroimaging, biomarker, clinical, and cognitive data. In the second and third stages, subjects’ diagnostic, lipid metabolism, cognitive, and neuroimaging data were all retrieved from the ADNI database.

Phenome-wide Mendelian randomization analysis

All disease phenotypes with a case sample size exceeding 2000 in the FinnGen Genomic Initiative database were included as exposure factors. We extracted single-nucleotide polymorphisms (SNPs) strongly associated with each of the disease phenotypes as IVs, with a significance threshold set at p < 5 × 10⁻⁸. When the number of IVs was less than 10, a P-value threshold was lowered to 1 × 10⁻⁵. The 1000 Genomes Project data from the European population was utilized as the reference panel for linkage disequilibrium (LD) clumping, with an LD threshold of r² = 0.001 within a 10,000 kb distance.

To address potential APOE-specific pleiotropy, which poses a unique challenge due to the APOE locus’s well-established pleiotropic effects on both lipid metabolism and AD pathogenesis, we verified that none of the selected SNPs were located within or in linkage disequilibrium with the APOE gene region. This exclusion ensures that the observed causal associations are not confounded by APOE-mediated pleiotropic pathways. For each exposure, a two-sample MR analysis was conducted to assess their causal relationships with AD outcomes, respectively. The primary analysis was executed using the inverse variance-weighing (IVW), supplemented by the weighted-median (WM) and MR-Egger techniques. We assessed the heterogeneity of SNPs using the Cochran’s Q test and I² statistics. Heterogeneity was considered to be present if the P-value < 0.05 and I² > 0.25. All the summary-level data used in this research were publicly accessible and obtained with the consent and ethical approval of the respective Review Boards.

Estimating the causal effects of MetS components on AD

The main findings from the phenome-wide MR analysis indicated a significant causal impact of various metabolic disorder phenotypes on AD. To validate their causal relationships with AD, we individually incorporated a total of 10 traits either as the primary components of MetS or as MetS-related traits and conducted two-sample MR analyses. Nine of those traits, including metabolic disorders, pure hypercholesterolemia, hyperlipidemia, disorders of lipoprotein metabolism and other lipidemias (DLMOL), obesity, hypertension, antihypertensive medication, type 2 diabetes, and diabetes with insulin treatment, were sourced from the FinnGen database [19]. Additionally, abdominal obesity, which is a key component of MetS, may be measured through waist circumference. Therefore, we also included the GWAS summary-level data for waist circumference provided by Jiang and colleagues [21] for further MR analysis.

Estimating the causal effects of circulating lipids and apolipoproteins on AD

Expanding upon the findings from the preceding 10 distinct MR analyses, our results strongly indicate a predominant causal role of lipid metabolism disorders in AD. Consequently, we proceeded to conduct a more detailed MR analysis using genomic data related to lipid metabolism. We obtained major lipid metabolism genetic data, including TC, TG, HDL-C, LDL-C, and VLDL-C from the GWAS Catalog database (https://www.ebi.ac.uk/gwas/). Following this, for each trait and AD, we performed two-sample MR analyses. Additionally, we also included and investigated some additional serum apolipoprotein-related variables, including apolipoprotein A1 (ApoA1), apolipoprotein B (ApoB), and apolipoprotein E (ApoE), to further assess the impact of these apolipoproteins on the risk of AD.

Validating the association between LDL-C indicators and AD risk in the ADNI database

We included baseline data from participants in ADNI 1, ADNI GO, and ADNI 2, encompassing demographic information, APOEε4 status, cognitive data, as well as summary-level imaging and metabolomics information. Data points included initial diagnosis by researchers at the baseline, relying on their medical history, Clinical Dementia Rating, Geriatric Depression Scale, Functional Activities Questionnaire scores, laboratory test results, and neuropsychological assessments. The diagnoses included normal cognition, MCI, AD, or other causes of dementia [24].

In the first stage of MR analysis, the levels of LDL-C and TC were identified as lipid metabolite components influencing AD. Matched 44 components of the lipidomic nuclear magnetic resonance (NMR) metabolite dataset were further extracted from the ADNI database [25]. Detailed information on the selected lipid metabolites can be found in Supplementary Table S4. Sample preparation, NMR data processing, and quality control information are all accessible in the ADNI database [26].

The relationship between LDL-C indicators, cognitive function, and brain phenotypes

The Wilcoxon rank-sum tests were employed for inter-group statistical analysis of 44 lipid metabolites to identify indicators with significant differences. Subsequently, univariate and multivariate logistic regression analyses were conducted to assess the impact of lipid metabolism indicators on the risk of AD. Additionally, summary-level cognitive function and brain phenotype data were extracted from the ADNI database and analyzed. Cognitive assessment data encompass a variety of scales, including the Clinical Dementia Rating Scale Sum of Boxes, Alzheimer’s Disease Assessment Scale Cognitive Subscale 13 (ADAS13), Alzheimer’s Disease Assessment Scale Cognitive Subscale Question 4, Mini-Mental State Examination, Rey Auditory Verbal Learning Test (RAVLT, comprising four sub-items: Immediate, Learning, Forgetting, and Perc Forgetting), Logical Memory Delayed Recall Total, and Digit Span Total Score. Regarding brain phenotypic data, indicators of structural changes in the medial temporal lobe, including hippocampal and entorhinal cortex volumes, which exhibit early and significant alterations in AD patients, were included. Additionally, the total intracranial volume (TIV) of the subjects was also incorporated for subsequent analyses.

Verification of association between LDL-C indicators and brain phenotypes

Initially, the MR method was employed to validate the causal impact of LDL-C levels on the volumes of the left and right hippocampus. The genetic data for hippocampal volumes were derived from a large-scale population-based GWAS conducted in 2021, utilizing the UK Biobank dataset [27]. At a more refined level, we utilized authentic 3D T1 brain structural data from the ADNI database to further investigate the impact of LDL-C levels on the entorhinal-hippocampal complex in individuals with prodromal AD (MCI) and those CN subjects from the ADNI database. Following the data quality assessment, we employed the FreeSurfer (version 7.1) [28] toolkit to execute the recon-all pipeline analysis. The FreeSurfer automated segmentation algorithm has demonstrated efficacy and dependability in quantifying the volume of hippocampal subregions [29]. For subjects who completed the recon-all pipeline, the bilateral entorhinal-hippocampal complex was segmented into 20 distinct subregions, including 19 hippocampal subregions and 1 entorhinal cortex, using the segmentHA_T1 function implemented in FreeSurfer [30]. Each participant’s segmentation was independently reviewed by two co-authors (DL and WS) by visual inspection, and any results deemed inaccurate were excluded. We took the mean of various variables from the bilateral entorhinal-hippocampal complex as the specific subfield phenotype for each participant. Subsequently, the brain subfield phenotypes and TIV values were obtained for further multiple linear regression analysis.

Statistical analysis

All analyses were conducted using R (version 4.2.2). In two-sample MR analyses, the results of the IVW method were considered the primary reference, while the MR-Egger and weighted median (WM) models were employed as supplementary techniques for sensitivity analysis. The MR-Egger regression intercept was employed to evaluate average directional pleiotropy. Heterogeneity was evaluated using both I² statistics and Cochran’s Q test, where I² > 0.25 and p < 0.05 indicated significant heterogeneity. When analyzing participant data from the ADNI cohort, Wilcoxon rank-sum tests were employed to compare group differences in age, education, TIV, and lipid metabolism indicators. Chi-square tests were employed for gender, race, and APOEε4 carrier status comparisons. The normality of variables was assessed using Shapiro-Wilk tests, and Winsorizing was employed as a tail-trimming method to enhance the robustness of the analysis. Logistic regressions were executed to evaluate the influence of various lipid indicators on AD risk. Lipid metabolism indicators that exhibited significant group differences in the Wilcoxon rank-sum test and showed significant associations in univariate logistic regression were subsequently included in multivariate logistic regression analyses. Multicollinearity among independent variables in regression analysis was assessed using the variance inflation factor (VIF), and variables with a VIF exceeding 5 were removed. In subsequent multivariate regression analyses related to cognitive function, gender, age, race, education, and APOEε4 carrier status were included as covariates. In brain phenotype multivariate regression analyses, in addition to the aforementioned variables, TIV was also included as a covariate. For entorhinal-hippocampal complex investigations, participants were stratified into high and low LDL-C groups. Group differences in brain phenotypes were assessed based on a general linear model, with gender, age, race, education, APOEε4 carrier status, and TIV included as covariates. Values of p < 0.05 were considered statistically significant. Multiple comparison correction of P-values was achieved using the false discovery rate (FDR) method. Throughout the “Results” section, uncorrected and FDR-corrected P-values are denoted as P-unadjusted and P-FDR, respectively.

Sensitivity analysis

Several sensitivity analyses were conducted to ensure robustness of our findings. We evaluated potential sample overlap between hippocampal dataset (from UK Biobank) and AD outcome dataset to minimize participant duplication bias. To address proxy-AD phenotype concerns, we repeated analyses using stringently defined clinical AD diagnosed cases only. Bidirectional Mendelian randomization was performed to assess reverse causation from AD to MetS-related phenotypes. Finally, we tested whether variables showing significant between-group differences in ADNI (e.g., education) affected the stability of our primary conclusions.

Ethics statement

This research utilized publicly accessible de-identified data from participant studies that had obtained approval from an ethics standards committee in the original research. No separate ethical approval was required in this study.

Results

Datasets and sample characteristics

In the initial phenome-wide MR analysis, a total of 868 phenotypes from the FinnGen database were included as exposure factors for the phenome-wide MR analysis, with the number of included cases ranging from 2003 to 377,277. Relevant data information can be found in Supplementary Table S2. Furthermore, a total of 814 participants were included from the ADNI dataset for the logistic regression analyses, including 314 individuals with AD and 500 control subjects with normal cognition. The demographic and clinical characteristics of those subjects are summarized in Table 1. Significant between-group differences were observed in educational level and APOEε4 carrier status between the AD and CN groups, while no significant differences were noted in terms of age, gender, race, and TIV between the two groups. After screening, 3D T1 brain structural data from a total of 713 participants (446 with MCI and 267 CN) successfully went through the FreeSurfer segmentation pipeline and were subsequently utilized for the analysis of the structural phenotypes within the entorhinal-hippocampal complex. The demographic characteristics of those subjects are presented in the Supplementary Table S5.

Table 1.

Demographic characteristics of the participants from ADNI dataset

	AD (N = 314)	CN (N = 500)	P value
Age, mean (SE), years	74.79 (0.44)	74.23 (0.26)	0.052
Gender, F/M, n	144/170	256/244	0.16
Race, White/Black/Asian/≥2	290/13/7/4	455/32/8/5	0.50
Education, mean (SE), years	15.14 (0.17)	16.41 (0.12)	3.11e-09***
APOEε4, 0/1/2, n	106/146/62	358/130/12	2.2e-16***
Total intracranial volume, cm3	1535.71 (10.32)	1512.22 (7.00)	0.20

Note: Values are expressed as the mean (SE). The p values for the comparison between AD and CN were determined using the Wilcoxon rank-sum test. The Chi-square test was utilized to examine associations among gender, race, and the presence of the APOEε4 risk allele. p values are presented without correction for multiple comparisons. Significance levels are denoted by asterisks (* p < 0.05, ** p < 0.01, *** p < 0.001). Abbreviations: AD, Alzheimer’s Disease; APOE, apolipoprotein E; CN, cognitively normal

Phenome-wide causal effects on AD

In the MR-IVW analyses, a total of 54 disease phenotypes were causally associated with AD (Table 2). To ensure instrument validity, we assessed the strength of all instrumental variables employed in our primary analyses. For the two-sample MR analysis examining the causal effect of lipid metabolic phenotypes on AD, the F-statistics demonstrated robust instrument strength (mean = 3443.3, median = 1330.0, minimum = 168.0), with all values substantially exceeding the conventional threshold of 10. Similarly, in the reverse MR analysis evaluating the effect of AD on MetS metabolic phenotypes, the F-statistics were consistently strong (mean = 126.8, median = 43.4, minimum = 30.6), well surpassing the minimum requirement. These robust F-statistics confirm strong instrument relevance and effectively mitigate concerns regarding weak instrument bias. Notably, 10 of these disease phenotypes were associated with complications of diabetes, including diabetic eye disease and diabetic nephropathy, with odds ratio (OR) values predominantly concentrated in the range of 0.99–1. However, results from large-sample two-sample MR indicated that neither type 1 (p = 0.347, case = 8,967) nor type 2 (p = 0.432, case = 57,698) diabetes had a causal impact on AD, indicating that the diabetes itself does not exert a causal contribution to AD. Additionally, six metabolic disease-related phenotypes showed a significant impact on AD risk, including metabolic disorders (OR: 1.08, 95% CI: 1.02–1.15, P-FDR = 0.007), pure hypercholesterolemia (OR: 1.05, 95% CI: 1.02–1.09, P-FDR = 0.003), hyperlipidemia (OR: 1.07, 95% CI: 1.01–1.13, P-FDR = 0.013), DLMOL (OR: 1.06, 95% CI: 1.01–1.12, P-FDR = 0.017), statin medication (OR: 1.02, 95% CI: 1.00–1.04, P-FDR = 0.03), and disorders of lipoprotein metabolism (OR: 1.02, 95% CI:1.00–1.05, P-FDR = 0.04). Other phenotypes that showed a significant causal impact on AD include disorders of the choroid and retina, mental disorders, and delirium (Fig. 2A). The detailed information for 868 disease phenotypes, along with the summary of MR results from all three methods (IVW, MR-Egger, and weighted median), is compiled in Supplementary Table S6. These results suggest that, at the phenome-wide level, the primary causal impact on AD risk is associated with disorders related to metabolic dysregulation.

Table 2.

Summary of the significant IVW results in phenome-wide Mendelian randomization analysis

Exposure	N-IV	B (se)	OR (95%CI)	P-values
Disorders of choroid and retina	12	−0.032 (0.008)	0.97 (0.95–0.98)	8.53E-05
Proliferative diabetic retinopathy	11	−0.009 (0.002)	0.99 (0.99–1.00)	1.63E-04
Any mental disorder	19	0.423 (0.120)	1.53 (1.21–1.93)	4.24E-04
Diabetic retinopathy	13	−0.010 (0.003)	0.99 (0.98–1.00)	1.55E-03
Type 1 diabetes with coma	13	−0.008 (0.002)	0.99 (0.99–1.00)	1.95E-03
Pure hypercholesterolaemia	29	0.050 (0.017)	1.05 (1.02–1.09)	2.70E-03
Diabetic background retinopathy	13	−0.010 (0.004)	0.99 (0.98–1.00)	3.58E-03
Osteoporosis	50	−0.009 (0.003)	0.99 (0.98–1.00)	5.01E-03
Type1 diabetes, definitions combined, early onset	19	−0.006 (0.002)	0.99 (0.99–1.00)	5.35E-03
Degenerative macular diseases	31	−0.011 (0.004)	0.99 (0.98–1.00)	5.98E-03
Metabolic disorders	22	0.080 (0.029)	1.08 (1.02–1.15)	6.56E-03
Other mental disorders due to brain damage and dysfunction and to physical disease	18	0.102 (0.038)	1.11 (1.03–1.19)	7.47E-03
Corneal ulcer	20	−0.015 (0.006)	0.99 (0.97–1.00)	8.10E-03
Herpesviral infections	21	−0.009 (0.003)	0.99 (0.98–1.00)	8.71E-03
Delirium, not induced by alcohol and other psychoactive substances	18	0.103 (0.039)	1.11 (1.03–1.20)	8.78E-03
Generalized epilepsy	18	0.009 (0.004)	1.01 (1.00–1.02)	9.07E-03
Fracture of lower leg, including ankle	35	−0.015 (0.006)	0.99 (0.97–1.00)	9.30E-03
Diabetic hypoglycemia	11	−0.010 (0.004)	0.99 (0.98–1.00)	0.011
Single delivery by caesarean section	41	−0.014 (0.005)	0.99 (0.98–1.00)	0.012
Polymyalgia rheumatica	25	−0.012 (0.005)	0.99 (0.98–1.00)	0.012
Hyperlipidaemia, other/unspecified	12	0.067 (0.027)	1.07 (1.01–1.13)	0.013
Any mental disorder, or suicide, or psychic disorders complicating pregnancy, partum or puerperium or nerve system disorders	47	0.012 (0.005)	1.01 (1.00–1.02)	0.014
Nonorganic sleeping disorders	19	0.012 (0.005)	1.01 (1.00–1.02)	0.014
Endometriosis	26	0.011 (0.005)	1.01 (1.00–1.02)	0.015
Disorders of lipoprotein metabolism and other lipidaemias	37	0.060 (0.025)	1.06 (1.01–1.12)	0.017
Vitreous bleeding	28	−0.010 (0.004)	0.99 (0.98–1.00)	0.017
Glomerular diseases	23	−0.015 (0.006)	0.98 (0.97–1.00)	0.017
Zoster (herpes zoster)	20	0.011 (0.005)	1.01 (1.00–1.02)	0.018
Malignant neoplasm of rectum	31	−0.006 (0.003)	0.99 (0.99–1.00)	0.019
Memory loss	25	0.075 (0.032)	1.08 (1.01–1.15)	0.019
Diabetic nephropathy	11	−0.009 (0.004)	0.99 (0.98–1.00)	0.021
Primary coxarthrosis, bilateral	11	−0.016 (0.007)	0.98 (0.97–1.00)	0.021
Diabetic neuropathy	35	−0.007 (0.003)	0.99 (0.99–1.00)	0.023
Hydrocele	35	−0.008 (0.003)	0.99 (0.99–1.00)	0.025
Agranulocytosis	18	−0.009 (0.004)	0.99 (0.98–1.00)	0.025
Nontoxic multinodular goitre	39	0.005 (0.002)	1.00 (1.00–1.01)	0.027
Other neurological diseases	28	0.021 (0.010)	1.02 (1.00–1.04)	0.027
Statin medication	153	0.022 (0.010)	1.02 (1.00–1.04)	0.03
Seropositive rheumatoid arthritis, strict definition	12	−0.007 (0.003)	0.99 (0.99–1.00)	0.032
Focal epilepsy, strict definition	23	0.010 (0.005)	1.01 (1.00–1.02)	0.033
Autoimmune diseases excluding thyroid diseases, strict definition	40	−0.014 (0.007)	0.99 (0.97–1.00)	0.034
Carcinoma in situ of breast	28	−0.006 (0.003)	0.99 (0.99–1.00)	0.036
Hypothyroidism, strict autoimmune	124	−0.007 (0.003)	0.99 (0.99–1.00)	0.037
Inguinal hernia, bilateral	36	0.006 (0.003)	1.01 (1.00–1.01)	0.038
Vasomotor rhinitis	22	−0.007 (0.003)	0.99 (0.99–1.00)	0.038
Migraine	63	−0.012 (0.006)	0.99 (0.98–1.00)	0.04
Disorder of lipoprotein metabolism, unspecified	32	0.025 (0.012)	1.02 (1.00–1.05)	0.04
Heart failure, strict	82	0.010 (0.005)	1.01 (1.00–1.02)	0.041
Coxarthrosis, primary	27	−0.013 (0.006)	0.99 (0.98–1.00)	0.042
Lichen planus, including avohilmo	53	−0.006 (0.003)	0.99 (0.99–1.00)	0.042
Type 1 diabetes with ketoacidosis	13	−0.005 (0.002)	1.00 (0.99–1.00)	0.045
Alcohol withdrawal state with delirium	25	−0.006 (0.003)	0.99 (0.99–1.00)	0.045
Other disorders of skin and subcutaneous tissue, not elsewhere classified	21	−0.010 (0.005)	0.99 (0.98–1.00)	0.048
Unspecified acute lower respiratory infection	32	−0.009 (0.004)	0.99 (0.98–1.00)	0.049

Note: The above results are all significant findings from the two-sample Mendelian Randomization (MR) analysis using the inverse variance weighted (IVW) method, with P-values reported as the original values. AD, Alzheimer’s disease; OR: odds ratio; CI, confidence interval; N-IV, number of instrumental variables

Fig. 2.

The causal impact of MetS and lipoprotein-related components on AD. A: the volcano plot illustrates disease phenotypes that exhibit a causal impact on AD at the phenotypic level. The horizontal and vertical axes represent the beta values and -log10(P) values from MR analysis, respectively. In the plot, cyan dots represent phenotypes with beta values less than 0, while red dots represent phenotypes with beta values greater than 0. Grey dots indicate phenotypes with non-significant p values. B: the first section of the forest plot presents assessments of the causal relationship between key components of metabolic syndrome (MetS). These components include metabolic disorders, obesity, waist circumference, hypertension, and type 2 diabetes, among others. The second section illustrates evaluations of the causal relationship between several major lipid metabolism indicators (TC, TG, HDL-C, LDL-C, VLDL-C) and AD. The P-values were adjusted using the false discovery rate (FDR) correction. Abbreviations: N-IV, number of instrumental variables; OR, odds ratio; CI, confidence interval; DLMOL, disorders of lipoprotein metabolism and other lipidaemias; MR, Mendelian randomization; TC, total cholesterol; TG, triglyceride; VLDL/LDL/HDL, very low/low/high-density lipoprotein

Causal effects of MetS components on AD

MetS has been considered a potential contributor to the risk of dementia or AD. To further explore whether the features of metabolic dysregulation identified in phenome-wide MR analysis, in fact, reflect MetS, we conducted additional MR analysis incorporating 10 representative disease traits, each with a number of cases exceeding 10,000. The IVW results revealed that exposure to metabolic disorders was causally associated with AD. However, no significant impacts on AD were observed for waist circumference, obesity, hypertension, antihypertensive medication, type 2 diabetes, and insulin treatment. Figure 2B depicts the effects of MetS components on AD. Despite the observed presence of heterogeneity in Cochran’s Q and I² statistics, some of the I² statistics indicated moderate heterogeneity (I² < 0.5), and the random-effects IVW method can balance this heterogeneity [31]. Furthermore, the MR-Egger intercept test revealed no indications of horizontal pleiotropy, indicating that heterogeneity did not introduce bias into the results. The complete set of outputs from all MR analysis methods (IVW, MR-Egger, and weighted median), including odds ratios, confidence intervals, P-values, and corresponding heterogeneity tests, is available in Supplementary Table S7. The above results indicate that, among the various metabolic dysregulations associated with MetS, a leading role in influencing AD is played by lipid metabolism.

Causal impacts of major lipids and lipoproteins on AD

To further elucidate which specific type of lipoproteins drives the risk of AD, we conducted additional analyses using GWAS summary-level data based on circulating lipoproteins. The findings indicate that the levels of LDL-C (OR: 1.05, 95% CI: 1.03–1.07, P-FDR = 2.77E-05) and TC (OR: 1.03, 95% CI: 1.01–1.05, P-FDR = 7.01E-03) significantly contribute to AD onset, whereas HDL-C, VLDL-C, and TG show no significant impact on AD. Furthermore, in addition to the levels of serum ApoE (OR: 1.05, CI: 1.02–1.07, P-FDR = 1.68E-05), those for ApoB (OR: 1.09, CI: 1.02–1.17, P-FDR = 8.61E-03) also significantly increase the risks of AD (Fig. 2). The complete results from all MR analysis methods (IVW, MR-Egger, and weighted median), including corresponding heterogeneity tests, are available in Supplementary Table S8.

Significant association between LDL-C indicators and AD risk in ADNI database

To further investigate the impact of LDL-C and TC indicators on AD, we conducted analyses using real-world population data collected in the ADNI database. Among 44 characteristics of the lipid metabolism, three LDL-C related indicators exhibited significant P-values in both Wilcoxon rank-sum tests and univariate logistic regression analyses, while also showing low collinearity (VIF < 5). Hence, subsequent analysis was centered on three LDL-C related indicators, including free cholesterol to total lipids ratio in large LDL particles (L-LDL-FC-PCT, OR: 1.19, 95% CI: 1.04–1.38, P-FDR = 0.03) (Fig. 3A), phospholipids to total lipids ratio in large LDL particles (L-LDL-PL-PCT, OR: 1.37, 95% CI: 1.05–1.79, P-FDR = 0.035) (Fig. 3B), and cholesterol to total lipids ratio in small LDL particles (S-LDL-C-PCT, OR: 1.12, 95% CI: 1.02–1.23, P-FDR = 0.03) (Fig. 3C). In multivariate logistic regression, the impact of L-LDL-FC-PCT on AD remained significant (OR: 1.20, 95% CI: 1.00–1.43, P-unadjusted = 0.047). Table 3 summarizes the analyses of different variables entered in univariate and multivariate logistic regression. Similar to the findings of MR analysis, results from real population data also indicate a significant elevation in the risk of AD associated with LDL-C indicators.

Fig. 3.

Association of low-density lipoprotein metabolic markers with cognitive and brain phenotypes. A-C: data analysis based on the ADNI dataset indicates that among the metabolic indicators associated with LDL-C, three indicators exhibit a significant correlation with the risk of AD in logistic regression models. Additionally, these three indicators show significant between-group differences, identified as L-LDL-FC-PCT (A), L-LDL-PL-PCT (B), and S-LDL-C-PCT (C), respectively; D-E: the regression analysis results indicate a correlation between L-LDL-FC-PCT and cognitive scores (ADAS 13 and RAVLT immediate), and this association remains significant even after FDR correction; the regression analysis results indicate a correlation between L-LDL-FC-PCT and ADAS 13 (D) and RAVLT immediate (E), and these associations remains significant after FDR correction. F-G: In the medial temporal lobe structure, the S-LDL-C-PCT indicator exhibits a significant correlation with the volumes of the hippocampus (F) and entorhinal cortex (G). H-I: additional validation results from Mendelian randomization reveal a significant influence of LDL-C on the volumes of both the left (H) and right (I) hippocampus. J-K: based on the median of S-LDL-C-PCT in CN populations used as a cutoff, participants were classified into low and high groups. Subsequently, the inter-group differences in hippocampal subfield and entorhinal cortex measures were statistically analyzed. The results indicate significant differences in bilateral CA3 head (J) and HATA (K) among the CN group

Table 3.

Results of univariable and multivariable logistic regression model in ADNI cohort

Characteristics	Univariable Logistic Regression		Multivariable Logistic Regression
	OR (95% CI) P		OR (95% CI) P
Age	1.01 (0.99–1.03)	0.24	1.03 (1.01–1.06)	0.02*
Gender (male vs. female)	1.24 (0.93–1.64)	0.14	1.46 (1.03–2.09)	0.04*
Race (vs. Asian)
White	1.20 (0.72–2.03)	0.50	0.37 (0.11–1.30)	0.11
More than one	1.28 (0.31–4.86)	0.72	0.35 (0.05–2.39)	0.28
Black or African American	0.63 (0.32–1.20)	0.17	0.16 (0.04–0.70)	0.01*
Education	0.86 (0.81–0.90)	2.80E-09***	0.84 (0.79–0.89)	1.35E-08***
APOEε4	3.98 (3.11–5.14)	3.36E-27***	4.13 (3.18–5.41)	1.06E-25***
Lipoprotein indicator
L-LDL-FC-PCT	1.19 (1.04–1.38)	0.014*	1.20 (1.00–1.43)	0.047*
L-LDL-PL-PCT	1.37 (1.05–1.79)	0.022*	1.23 (0.89–1.70)	0.22
S-LDL-C-PCT	1.12 (1.02–1.23)	0.015*	1.07 (0.96–1.20)	0.20

Note: p values are presented without correction for multiple comparisons. Significance levels are denoted by asterisks (* p < 0.05, ** p < 0.01, *** p < 0.001). Abbreviations: AD, Alzheimer’s Disease; APOE, apolipoprotein E; CN, cognitively normal; OR, odds ratio; CI, confidence interval; LDL, low-density lipoprotein; L-LDL-FC-PCT, free cholesterol to total lipids ratio in large LDL; L-LDL-PL-PCT, phospholipids to total lipids ratio in large LDL; S-LDL-C-PCT, cholesterol to total lipids ratio in small LDL

LDL-C indicators, cognitive function, and brain structure

To delve deeper into the associations of the three LDL-C indicators and the cognitive function, as well as other AD-related brain phenotypes, we conducted additional multivariate regression analyses. The adjusted results indicate significant correlations between L-LDL-FC-PCT and cognitive scores, including those collected using ADAS13 (β = 1.11, R² = 0.176, P-FDR = 0.044) (Fig. 3D) and RAVLT immediate (β = −1.34, R² = 0.214, P-FDR = 0.044) tests (Fig. 3E). Simultaneously, the levels of S-LDL-C-PCT were negatively correlated with the volume of the hippocampus (β = −61.57, R² = 0.30, P-unadjusted = 0.013) (Fig. 3F) and entorhinal cortex (β = −38.54, R² = 0.23, P-unadjusted = 0.026) (Fig. 3G) across the entire cohort. These results suggest that the levels of LDL-C may have a significant negative impact on cognitive function and the structures of the medial temporal lobe.

Verification of association between LDL-C and AD-characteristic brain regions

To further validate the relationship between LDL-C levels and the volumes of the hippocampal areas, two-sample MR analysis was executed. The MR results indicate a significant negative impact of LDL-C levels on the total volume of the left hippocampus (OR: 0.95, 95% CI: 0.91–1.00, P-FDR = 0.035) (Fig. 3H) and the right hippocampus (OR: 0.95, 95% CI: 0.91–1.00, P-FDR = 0.034) (Fig. 3I). After removing the covariate effects, further analysis based on real brain segmentation results reveals that, in the CN population, individuals in the high S-LDL-C-PCT group exhibit smaller bilateral mean volumes of the CA3 head (t = −2.38, P-unadjusted = 0.018) (Fig. 3J) and the hippocampus-amygdala transition area (HATA, t = −2.30, P-unadjusted = 0.022) (Fig. 3K). In sensitivity analyses based on unilateral brain regions, these results remained significant, except for the right HATA, which showed a P-unadjusted approaching significance at 0.078. Similarly, a comparable trend in the volumes of HATA (t = − 1.28, P-unadjusted = 0.20) and CA3 head (t = −1.45, P-unadjusted = 0.15) was observed in the MCI group, although the respective P-values had not reached significance threshold. Interestingly, in the MCI population, participants with high S-LDL-C-PCT levels had consistently smaller surface areas of the entorhinal cortex (t = −2.35, P-unadjusted = 0.019). Subfield-level analysis of the entorhinal-hippocampal complex indicated that the effects of LDL-C indicators on HATA and CA3 seem to precede those on other areas and are more pronounced. These results further validate the negative impact of LDL-C on the medial temporal lobe of the brain in general and on its entorhinal-hippocampal complex in particular.

Sensitivity analysis

We adopted a conservative estimation approach, assuming maximum possible sample overlap of approximately 8.72% of the outcome samples. Sensitivity analysis based on online sample power assessment tools (https://sb452.shinyapps.io/overlap/) showed that under maximum overlap proportion, the estimated bias approaches zero, and the Type I error rate remains at the nominal significance level (0.05) [32]. The sensitivity analysis using stringent clinical AD diagnoses demonstrated that the results for ApoE and ApoB remained consistent with our primary findings. For LDL-C (OR: 1.26, 95% CI: 0.96–1.64, p = 0.091), the P-value approached statistical significance, and the direction of its main effects in the IVW analysis remained concordant with results obtained when proxy-AD outcomes were included (Supplementary Table S9). Further bidirectional MR analysis results indicated that among MetS-related phenotypes, the genetic predisposition to AD itself significantly increased the risk of being diagnosed with lipid metabolism disorders. In contrast to its effects on lipid metabolism, AD showed no association with hypertension risk, while demonstrating inverse associations with obesity (OR: 0.75, 95% CI: 0.57–0.98, p = 0.038) and type 2 diabetes (T2D) (OR: 0.73, 95% CI: 0.64–0.84, p = 7.46E-06)(Supplementary Table S10). To evaluate the robustness of our findings regarding potential confounding by education level, we performed a sensitivity analysis excluding education from the multivariate model. When education was omitted, gender effects on lipid metabolism became non-significant (OR: 1.11, 95% CI: 0.79–1.54, p = 0.55), while the effects of age and APOEε4 were enhanced. Among the key lipid metabolites, L-LDL-FC-PCT approached statistical significance (OR: 1.11, 95% CI: 0.94–1.32, p = 0.08), and S-LDL-C-PCT became significant (OR: 1.07, 95% CI: 0.96–1.20, p = 0.04)(Supplementary Table S11). These results indicate that while education level affected the performance of the model overall, the core associations between LDL-related metabolites and AD risk remained when education levels were excluded, thus, supporting the robustness of our primary findings.

Discussion

Here we present the results of a hypothesis-free phenome-wide scan, which has explored associations between genetic components of a set of potential risk factors and AD. Our findings indicate that the phenotypes causally impacting AD gravitate towards the metabolic domain. Further exploration of the relationships between various metabolic phenotypes associated with MetS and the risks of AD suggested that the susceptibility to AD is driven by lipid metabolism abnormalities, particularly those related to LDL-C levels. These findings were validated using real population data from the ADNI. Notably, AD and lipid metabolism disorders exhibited a complex bidirectional relationship, forming a vicious cycle where each condition potentially exacerbates the other. Additionally, we also observed a significant negative correlation between LDL-C related indicators, the participants’ cognitive function and the phenotypes of medial temporal lobe of the brain. Overall, our study, which was launched as an unbiased phenome-wide approach, has narrowed down its scope into a detailed dissection of a significant association between LDL-C indicators and AD-related disease phenotypes. This comprehensive chain of evidence suggests a plausible relationship between specific, LDL-C related aspects of the dysregulation of lipid metabolism and the susceptibility to AD which typically manifests itself later in life.

The impact of MetS on AD has been a subject of ongoing debate [3, 11, 12, 33–35], with recent findings leaning towards acknowledging a more significant association [3, 11, 36, 37]. In the study of the UK Biobank’s large population-based prospective cohort, Qureshi et al. found associations between MetS and the risk of dementia, with a stronger correlation observed in patients with 4 or 5 MetS components, and in the cohort of APOEε4 non-carriers [11]. According to Machado-Fragua et al., the risk of dementia increases with each additional MetS component presented in midlife, rather than after accumulating any three components; thus, this association is seen well before an individual’s health profile reaches conformance to a formal MetS definition [3]. However, a meta-analysis based on longitudinal studies suggests that MetS does not correlate with dementia or AD [12]. A primary reason for the observed divergence in these conclusions is the limited ability of observational studies to effectively assess the impact of comorbidities and possible reverse causation between metabolic characteristics and the diseases. Additionally, even if all the participants of a certain cohort meet the diagnostic criteria for MetS, their specific metabolic components remain heterogeneous, thus making comparisons between different studies difficult or impossible. Our MR results did not find that T2D could significantly increase AD risk, which is similar to previous MR research findings [38], but contradicts observational studies. These discrepancies may be caused by the role that hypoglycemic drugs play in this process [35]. Our study used RCT-like MR as a primary analytic means, and, therefore, provided a more reliable answer to the riddle of MetS, AD, and the concurrent disorders of the lipid metabolism in general, and LDL-C abnormalities in particular. Our findings also suggest that, when assessing the future cognitive function of individuals with MetS, it is essential to incorporate LDL-C indicators into the comprehensive evaluation of dementia risk alongside the existing diagnostic indicators.

While several metabolic phenotypes demonstrated statistically significant associations with AD (e.g., statin medication OR = 1.02, lipoprotein metabolism disorders OR = 1.02), the modest effect sizes warrant careful interpretation. At the individual level, these effects appear small; however, they may hold population-level significance given the high prevalence of these exposures. Additionally, MR estimates reflect lifelong causal effects, and cumulative exposure over decades may contribute meaningfully to AD risk. Importantly, these findings provide etiological insights into AD mechanisms rather than immediate clinical guidance, highlighting biological pathways that merit further investigation.

The ε4 allele of the APOE gene, recognized as the most potent genetic risk factor for AD, has long been a focal point of research. The primary functions of the ApoE protein include its direct roles in lipid metabolism, cholesterol transport, and modulation of immune-related functions [39]. Nevertheless, the precise mechanisms by which mutations in its encoding gene APOE contribute to the onset and progression of AD remain unclear. Recent studies have highlighted a close association between the ε4 variant of APOE and disturbances in the lipid metabolism within the brain tissue [40, 41]. In turn, APOEε4-related changes in the lipid contents lead to the aberrant activation of microglial cells and the local immune dysfunction, thereby exacerbating the progression of AD. Additionally, reduction of the levels of cholesterol ester within brain cells is capable of preventing AD-related brain damage and the resultant behavioral changes [40]. Our findings indicate that a significant increase in LDL-C levels posing a risk of AD. Furthermore, notable inverse correlations between the LDL-C levels, the assessments made using classical cognitive scales of AD and the medial temporal lobe brain phenotypes corroborate our assertion.

Even when the number of APOEε4 risk alleles is considered as a covariate, these results remain significant and suggest that lipid metabolism disturbances may not only be a manifestation of the pre-existing mutation in the apolipoprotein E encoding gene but may also be an important driving factor in the occurrence and progression of AD by itself. A recent clinical study conducted by Royall and colleagues found that statin drugs may have a protective effect on the severity of dementia in APOEε4 carriers [42]. These results are consistent with our study’s conclusions, indicating that statin medications may potentially slow the progression of dementia symptoms by reducing cholesterol levels and mitigating AD risk factors. These findings provide new insights calling for early cognitive assessment and lipid-formula targeting interventions in individuals carrying the APOEε4 risk gene.

The ApoE cascade hypothesis posits that the biochemical and biophysical characteristics of ApoE and its variants contribute to a series of cellular and tissue-wide events, leading to an increased risk of AD [43]. Nevertheless, in addition to ApoE, recent research has been gradually revealing the significant association between ApoB and AD [44–46]. In patients with early-onset AD, Wingo et al. conducted extensive targeted sequencing of APOB, a gene recognized for influencing the levels of LDL-C, and uncovered the links between rare coding variants in APOB and elevated LDL-C levels [44], and these associations were independent of APOEε4. Moreover, ApoB has also been reported as a novel marker for early tau pathology in AD, when measured in cerebrospinal fluid. Additionally, these levels show a significant correlation with cognitive function and the phenotypes of the entorhinal cortex [46]. Consistent with the aforementioned research, our phenome-wide screening results suggest that LDL-C may serve as a pivotal metabolic intermediate role in the etiology of AD.

Notably, both LDL-C and TC were highlighted by MR analysis as two significant causal drivers of AD. Therefore, we integrated all lipidomic data related to TC and LDL-C for subsequent validation analysis of the ADNI dataset. Indicators related to TC did not show significance in either intergroup comparisons or regression analysis, and these indicators had a higher VIF in the multivariate logistic regression model as well. Considering that TC encompasses LDL-C ^[44], our results suggest that LDL-C is a primary driver of the disorders of lipid metabolism, and that the links between TC and AD revealed in the first round of MR analysis actually reflect the embedded effects of LDL-C.

For LDL-C, this study identified several relevant and significant NMR components spanning from the free cholesterol to total lipids ratio in large LDL, to the phospholipids to total lipids ratio in large LDL, and to the cholesterol to total lipids ratio in small LDL. These findings align with previous report of Zarzar and colleagues that the free cholesterol to total lipids ratio in LDL is associated with AD progression [47]. Our findings regarding the causal role of LDL cholesterol in AD are consistent with recent work by Hu et al., who similarly demonstrated a significant causal association between LDL cholesterol and AD risk using Mendelian randomization [48]. This convergence of independent findings strengthens the evidence for LDL cholesterol as a modifiable risk factor in AD. However, our study provides unique contributions through a comprehensive phenome-wide screening approach, neuroimaging validation demonstrating effects on AD-characteristic brain regions, and integration of real-world clinical data. While Hu et al. examined the direct relationship between lipid metabolites and AD diagnosis, our research expands this understanding by exploring the associations across multiple dimensions, including cognitive domains, brain structural changes, and clinical phenotypes, contributing to a more comprehensive picture of how lipid metabolism may influence various aspects of the AD spectrum.

It is important to note that the changes in the ratio of cholesterol to phospholipids in LDL reflect alterations in cell membranes. Further studies are needed to elucidate the specific biological mechanisms connecting these lipid metabolism indicators to the progression of AD, possibly through the changes in the membrane stability and fluidity.

The impacts of LDL-C on AD were further substantiated by the findings made in cognitive-behavioral studies and brain phenotype datasets. Higher ADAS13 scores indicate poorer overall cognitive abilities, while RAVLT immediate scores reflect the short-term memory capacity. Accordingly, the disturbances of the lipid metabolism positively contributed to ADAS13 scores and negatively to RAVLT immediate scores, thus, feeding into the cognitive decline associated with AD. Additionally, the thickness of the medial temporal lobe, a likely structural mediator through which various risk factors contribute to AD-associated cognitive impairment [49, 50] was causally affected by the genetic components defining the lipid formula. Under physiological conditions, lipoproteins regulate neurobehavioral functions in the hippocampus and other brain regions through processes that are mediated by lipoprotein receptors [51, 52]. A significant negative correlation between the increased cholesterol to total lipids ratio in small LDL particles and the thickness of the structures of the entorhinal cortex and hippocampus was observed. Moreover, in the frame of the subregional assessments performed in normal controls, the volumes of both bilateral HATA and CA3 heads were susceptible to the influence of cholesterol ratios in small LDL particles. In the MCI population, this differences leveled off, possibly because the structural atrophy in these regions has already occurred pre-MCI, for example, due to the effects of certain early pathological factors such as the APOEε4 risk gene [53] and deposits of Tau pathology [54]. These inherent or early-existing risk factors may contribute to the loss of significance of the difference in subregional volumes when patients with high and low LDL-C are compared. In the MCI cohort, however, the inter-group differences between high and low LDL-C subgroups remain significant for the surface area of the entorhinal cortex. This observation potentially signifies that the pathogenic mechanisms of the cortical thinning in the spectrum of AD differ from those contributing to a slower age-related decline which is observed in control populations. Combining LDL-C indicators with morphological changes in the various areas of entorhinal cortex and hippocampus may contribute to further identifying individuals at risk for the prodromal stages of AD in populations with normal cognition and MCI and facilitate early intervention at the level of lipid metabolism.

Several plausible biological mechanisms may underlie the observed association between LDL-C levels and medial temporal lobe atrophy. First, altered LDL particle composition can compromise blood-brain barrier (BBB) integrity [55], facilitating the entry of neurotoxic substances into brain parenchyma. Small, dense LDL particles, which are more prevalent in dyslipidemia, have enhanced capacity to penetrate endothelial barriers and promote oxidative stress within cerebral vasculature. This BBB disruption has been specifically linked to hippocampal dysfunction and accelerated cognitive decline in both animal models and human studies. Second, oxidized LDL particles serve as potent activators of inflammatory cells [56], triggering the release of pro-inflammatory cytokines that create a neurotoxic environment. The hippocampal region, with its high metabolic demands and dense vascular network, is particularly susceptible to this inflammatory damage and subsequent development of tau pathology [57]. Notably, the CA3 subfield and HATA have been reported to be highly sensitive to internal and external environmental changes [58, 59]. Collectively, this dyslipidemia-induced glial activation and chronic neuroinflammation may lead to synaptic dysfunction, dendritic spine loss, and ultimately neuronal death in hippocampal subfields, manifesting as the medial temporal lobe atrophy observed in our findings [40].

Our study has several limitations. Firstly, although the MR analysis of lipid metabolites and hippocampal volume was conducted solely for validation of observational findings, the potential sample overlap between exposure (UK Biobank hippocampal volume) and outcome datasets represents a significant limitation that may compromise causal inference. Sensitivity analysis results indicated that our findings were minimally affected by potential overlap effects. Secondly, while we strictly adhered to the three fundamental assumptions in MR analysis, we acknowledge the inherent limitations of the MR analytical approach itself, including probable pleiotropy across different lipid metabolism pathways and inadequate assessment of social and environmental factors. These findings need to be strengthened through integration with large-scale real-world case studies for further validating presented results. Thirdly, due to the significant overall hippocampal atrophy, which is typically observed in AD patients and to the lack of subregion templates specific to AD pathological states, conducting precise and accurate segmentation of hippocampal subregions is challenging [60, 61]. Because of that, this study could not proceed to further subregion segmentation using imaging data strictly specific to AD by exclusion of AD+ cohort. Future longitudinal structural data analyses, particularly those based on at-risk populations, may provide deeper and more dynamic insights into the impact of different LDL-C indicators on specific entorhinal-hippocampal subregions in individuals within the AD spectrum. Fourthly, although our bidirectional MR provides valuable evidence for causal inference, caution should be exercised when interpreting the “vicious cycle” between AD and lipid metabolism disorders. Shared genetic architecture may be an important contributor to the observed bidirectional associations. The possibility of reverse causation during early disease stages cannot be overlooked, and measurement bias may also affect the reliability of the results. A more balanced conclusion would be that current evidence suggests potential reciprocal influences between AD and lipid metabolism disorders, but further verification is needed through: (1) integration of longitudinal cohort data to track temporal sequences; (2) incorporation of functional genomics and mechanistic studies to elucidate underlying biological pathways. Only by synthesizing evidence from multiple sources can we more accurately understand the complex relationship between these conditions and provide reliable evidence for clinical interventions.

Conclusion

In conclusion, our study revealed causal effects of the various features of lipometabolic disturbance on AD. Our research provides evidence suggesting a potential role of LDL-C metabolism abnormalities in AD-related cognitive functions and brain phenotypes. These findings offer new insights for early diagnosis and intervention of AD. Further integration of omics technologies and animal research is essential to elucidate the biological pathways between lipid metabolism abnormalities and AD progression.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Abbreviations

AD

Alzheimer’s disease

ADAS13

Alzheimer’s Disease Assessment Scale Cognitive Subscale 13

ADNI

Alzheimer’s Disease Neuroimaging Initiative

APOB

Apolipoprotein B; APOE, apolipoprotein E

CN

Cognitively normal

DLMOL

Disorders of lipoprotein metabolism and other lipidemias

FDR

False discovery rate; GWAS, genome-wide association studies

HATA

Hippocampus-amygdala transition area

HDL-C

High-density lipoprotein cholesterol

IVs

Instrumental variables

IVW

Inverse variance-weighted

LDL-C

Low-density lipoprotein cholesterol

L-LDL-FC-PCT

Free cholesterol to total lipids ratio in large LDL

L-LDL-PL-PCT

Phospholipids to total lipids ratio in large LDL

MCI

Mild cognitive impairment

MetS

Metabolic syndrome

MR

Mendelian Randomization

NMR

Nuclear magnetic resonance

OR

Odds ratio

S-LDL-C-PCT

Cholesterol to total lipids ratio in small LDL

SNPs

Single nucleotide polymorphisms

TC

Total cholesterol

TG

Triglycerides

TIV

Total intracranial volume

VIF

Variance inflation factor

VLDL

Very-low-density lipoprotein

WM

Weighted-median

Author contributions

Conceptualization, Data curation and Formal analysis: Dongming Liu, Ancha Baranova, Wenxi Sun, Bing Zhang, and Fuquan Zhang. Funding acquisition: Dongming Liu. Methodology, Validation and Visualization: Dongming Liu, Ancha Baranova, and Fuquan Zhang. Writing - original draft: Dongming Liu, Ancha Baranova, and Wenxi Sun. Writing - review & editing: Wenxi Sun, Hongbao Cao, Bing Zhang, Fuquan Zhang. Supervision: Bing Zhang, and Fuquan Zhang. All authors have read and approved the final version of the manuscript. Dongming Liu, Ancha Baranova, and Wenxi Sun contributed equally to this work.

Funding

This work was funded by Basic Research Program of Jiangsu (BK20250234), the National Natural Science Foundation of China (82502317), the Jiangsu Funding Program for Excellent Postdoctoral Talent (2023ZB184), the China Postdoctoral Science Foundation (2023M741648), National Mentorship Training Program for Young Health Professionals in Suzhou (Qngg2022027), and Jiangsu Key Laboratory of Preventive and Translational Medicine for Major Chronic Non-communicable Disease (KJS2434).

Data available

All GWAS summary datasets used in this study are publicly available for download by qualified researchers. The clinical and lipid metabolism data analyzed in this study are governed by the following licenses/restrictions: Requests for data access should be made to the Alzheimer’s Disease Neuroimaging Initiative (ADNI). For inquiries regarding access to these datasets, please visit http://adni.loni.usc.edu/.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

The individual data utilized in this study were sourced from publicly available datasets, for which the original research had already obtained written informed consent. Stringent measures have been implemented to safeguard the privacy and confidentiality of all participants involved. As such, no additional consent for publication was required for this specific study.

Competing interests

None.