Discovery of novel metabolic biomarkers in blood serum for diagnosis of Alzheimer's disease

Abstract

Background

Blood metabolites have emerged as promising candidates in the search for biomarkers for Alzheimer's disease (AD), as evidence shows that various metabolic derangements contribute to neurodegeneration in AD.

Objective

We aim to identify metabolic biomarkers for AD diagnosis.

Methods

We conducted an in-depth analysis of the serum metabolome of AD patients and age, sex-matched cognitively unimpaired older adults using ultra-high-performance liquid chromatography-high resolution mass spectrometry. The biomarkers associated with AD were identified using machine learning algorithms.

Results

Using the discovery dataset and support vector machine (SVM) algorithm, we identified a panel of 14 metabolites predicting AD with a 1.00 area under the curve (AUC) of receiver operating characteristic (ROC). The SVM model was tested against the verification dataset using an independent cohort and retained high predictive accuracy with a 0.97 AUC. Using the random forest (RF) algorithm, we identified a panel of 13 metabolites that predicted AD with a 0.96 AUC when tested against the verification dataset.

Conclusions

These findings pave the way for an efficient, blood-based diagnostic test for AD, holding promise for clinical screenings and diagnostic procedures.

Keywords

Alzheimer's disease biomarker diagnosis neurodegenerative disease serum metabolome

Introduction

Alzheimer's disease (AD) is a leading cause of dementia, accounting for 60–80% of dementia cases.^1,2 By 2050, the global prevalence of all-cause dementia is projected to reach 113 million.² AD is neuropathologically characterized by the accumulation of hyperphosphorylated tau and aggregates of amyloid-β (Aβ) peptide in the brain. Clinically, patients with AD experience gradual and progressive cognitive decline and disability. Most treatments available for AD offer only temporary symptomatic relief without modifying the course of the disease.³ Recently, emerging early-stage anti-amyloid antibody therapies have shown very modest positive effects on disease progression. The consensus opinion in the scientific community emphasizes early detection and intervention. This emphasizes the need to identify reliable biomarkers, preferably blood-based, for AD.

The National Institute on Aging and Alzheimer's Association (NIA-AA) recognizes cerebrospinal fluid (CSF) proteins such as total Tau (T-Tau), phosphorylated Tau (p-Tau), and Aβ₄₂ as key AD biomarkers.^4,5 Blood tests, however, offer a more practical alternative to CSF tests, especially for frequent assessments in clinical trials.⁶ Yet, establishing consistent blood biomarkers for AD has been challenging. The compromised BBB in aging and AD suggests that blood metabolites might more accurately represent brain biochemical alterations.^7–11

Research increasingly associates the potential contribution of metabolic dysregulations to the pathogenesis of AD.^12,13 Metabolic abnormalities of glucose metabolism in the brains of MCI patients develop early in the process, predisposing neurons to energy and functional disruptions.^12,14,15 Conditions such as diabetes, obesity, and metabolic syndrome, along with lifestyle factors, including diet and sedentary behavior, contribute significantly to AD risk.^16–19 The microbiota-gut-brain axis has also gained attention in AD, with certain gut bacterial byproducts potentially being implicated in AD^20–23 Some of these factors have been detected using metabolic approaches.^20,21,24 Teruya and colleagues identified numerous metabolite panels in plasma associated with dementia, though the specificity of these biomarkers to AD remains unexplored.²⁵

The prevailing view in AD therapeutics suggests that early interventions may yield the most promising therapeutic outcomes. As such, accurate diagnosis of AD at its earliest stages through biomarkers becomes paramount. In pursuit of this objective, we endeavored to discover a set of metabolite biomarkers linked to AD diagnosis, utilizing serum samples from the Texas Alzheimer's Research and Care Consortium (TARCC) (https://www.txalzresearch.org).²⁶ To ascertain whether blood-derived metabolites could distinguish AD from cognitively unimpaired controls, we combined the precision of ultra-high-performance liquid chromatography-high resolution mass spectrometry (UPLC-HRMS) with advanced computational modeling techniques.

Methods

Ethical approval

This study was conducted under the approval of the University of Texas Medical Branch (UTMB) Institutional Review Board (IRB) protocol 21-0201 on August 15, 2021. In addition, this study complied with all applicable federal regulations governing the protection of human subjects.

Cohort

Serum samples were requested through the TARCC bio-banking facility. TARCC is a longitudinal collaborative research initiative between ten Texas medical research institutions.²⁷ TARCC aims to investigate factors involved in the development and progression of AD and improve early diagnosis, treatment, and prevention of AD. The participants of the TARCC study were subjects older than 55 years and recruited at dementia clinics of the TARCC member institutions. All centers followed a harmonized collection protocol. The diagnosis of AD was based on the clinical examination with a neuropsychological battery using NINCDS-ADRDA criteria,²⁸ while age and sex-matched controls showed no cognitive impairment and achieved a zero on the Clinical Dementia Rating (CDR) scale.²⁹ The exclusion criteria include any significant neurological disease other than AD, such as Parkinson's disease, vascular dementia, progressive nonfluent aphasia, primary progressive aphasia, or prion disease. Participants with psychiatric disorders were also excluded. Age, sex, medical history, history of cardiovascular disease (i.e., heart disease, hypertension, diabetes mellitus, and hyperlipidemia), and family history of dementia were recorded. The levels of T-Tau, p-Tau, and Aβ₄₂ in patients’ CSF and brain were unknown.

Study design

Serum samples from age and sex-matched AD patients and cognitively unimpaired older adults (Controls) were used. The study was conducted in two phases: the discovery phase and the verification phase. In the discovery phase, 20 serum samples were used (n = 20; controls, 10; AD, 10). In the verification phase, 57 serum samples were used (n = 57, controls, 30, AD, 27). The demographic data of the discovery and verification study are presented in Table 1. The workflow of biomarker discovery is shown in Figure 1. Briefly, the serum metabolome of AD patients and cognitively unimpaired older adults were analyzed with LC-MS to identify AD-associated alternation in the metabolome. The significant metabolite hits were then used to identify a panel of AD biomarker candidates using machine learning technology. Finally, the performance of the biomarker panel was validated in the verification study.

Figure 1.

The workflow of identifying metabolic biomarkers for AD diagnosis. The study was conducted in two phases: the discovery phase and the verification phase. In each phase, serum metabolome from AD patients and healthy controls were analyzed with UPLC-HRMS. Then, the biomarker candidates were identified using machine learning and further verified in the verification phase.

Table 1.

Demographic characteristics of the metabolomics studies.

Demographic characteristics	Discovery Study		Verification Study
Demographic characteristics	Control (n = 10)	AD (n = 10)	Control (n = 30)	AD (n = 27)
Age, years (SD)	68.60 (7.12)	73.00 (11.37)	68.40 (6.50)	69.19 (9.86)
Sex, female (%)	6 (60.0%)	4 (40.0%)	12 (40.0%)	10 (37.0%)
CDR Global Score (SD)	0.00 (0.00)	1.35 (0.74)	0.00 (0.00)	1.33 (0.75)
Total MMSE score (SD)	29.50 (0.70)	16.60 (5.79)	29.60 (0.56)	16.59 (7.66)
Dx Probable AD (%)	0 (0.0%)	10 (100.0%)	0 (0.0%)	27 (100.0%)
Dx Frontotemporal dementia (%)	0 (0.0%)	0 (0.0%)	0 (0.0%)	0 (0.0%)
Does subject meet criteria for dementia? (%)	0 (0.0%)	10 (100.0%)	0 (0.0%)	27 (100.0%)
Does subject have normal cognition? (%)	10 (100%)	0 (0.0%)	30 (100%)	0 (0.0%)
Systolic blood pressure sitting (SD)	124.30 (9.91)	132.33 (13.84)	125.83 (9.92)	131.31 (13.84)
Hypercholesterolemia (%)	5 (50.0%)	3 (30.0%)	12 (40.0%)	6 (22.2%)
Hypertension (%)	6 (60.0%)	5 (50.0%)	17 (56.7%)	10 (37.0%)
Dx Other major psychiatric illness (%)	0 (0.0%)	0 (0.0%)	0 (0.0%)	0 (0.0%)
Dx Parkinson's Disease (%)	0 (0.0%)	0 (0.0%)	0 (0.0%)	0 (0.0%)
Dx Progressive nonfluent aphasia (%)	0 (0.0%)	0 (0.0%)	0 (0.0%)	0 (0.0%)
Dx Possible vascular dementia (%)	0 (0.0%)	0 (0.0%)	0 (0.0%)	0 (0.0%)
Dx Other disorder (%)**	0 (0.0%)	0 (0.0%)	0 (0.0%)	0 (0.0%)
Dx Primary progressive aphasia (%)	0 (0.0%)	0 (0.0%)	0 (0.0%)	0 (0.0%)
Dx Prion disease (%)	0 (0.0%)	0 (0.0%)	0 (0.0%)	0 (0.0%)
A3_DADDEM (%)	2 (20.0%)	5 (50.0%)	7 (23.3%)	10 (37.0%)
A3_MOMDEM (%)	4 (40.0%)	3 (30.0%)	15 (50.0%)	12 (44.4%)
A3_SIBS (SD)	2.50 (1.84)	1.70 (1.63)	0.17 (1.68)	1.67 (1.35)
A3_SIBSDEM (%)	0 (0.0%)	1 (10.0%)	1 (3.3%)	1 (3.7%)

CDR: Clinical Dementia Rating; MMSE: Mini-Mental State Exam; AD: Alzheimer's disease, MCI, mild cognitive impairment; **Dx Other major psychiatric illness: logopenic, anomic, transcortical, word deafness, syntactic comprehension, or motor speech disorder; A3_DADDEM: father with dementia; A3_MOMDEM: mother with dementia; A3_SIBS: number of siblings; A3_SIBSDEM: number of siblings with dementia.

Serum metabolome analysis with ultra-high-performance liquid chromatography-high resolution mass spectrometry (UPLC-HRMS)

An aliquot of 20 µL of serum from each subject was deproteinized by acetonitrile precipitation (1:4 v:v). The deproteinized serum samples were dried using a nitrogen evaporator and resuspended in 0.1% formic acid for LC-MS/MS analysis. Samples were analyzed using an Easy-nLC 1000 UHPLC system coupled with a high-resolution, high-mass accuracy Q Exactive orbitrap mass spectrometer (Thermo Scientific). The metabolites were separated on a reverse-phase nano-HPLC column with a 20-min linear gradient from 2–95% mobile phase B (0.1% formic acid-90% acetonitrile) in mobile phase A (0.1% formic acid), followed by 30-min 90% mobile phase B. The Orbitrap was operated in data-dependent MS/MS mode with 70,000 resolution (FWHM) at the full scan and 17,500 (FWHM) at MS/MS. Each sample was analyzed two or three times.

Metabolomic data analysis

The MS data were analyzed with Compound Discoverer™ 3.3 software (Thermo Scientific) for metabolite identification and quantification. For metabolite identification, elemental compositions were predicted using accurate mass data, with compounds identified using the mzCloud mass spectral library and MS/MS information. mzCloud is a foremost mass spectral database that aids analysts in characterizing compounds even in cases where they may not be found in the library (https://www.mzcloud.org/).³⁰ Where there was no match from mzCloud, ChemSpider was used. ChemSpider is chemical structure database that is free of cost and allows connection to rapid text and structure probing of more than 100 million structures found in hundreds of sources of data (http://www.chemspider.com).³¹ For results with a ChemSpider match, mzLogic was used to rank the identifications by the likelihood of a match. mzLogic is an algorithm for data analytics that merges the fragmentation data found inside the online mzCloud advanced mass spectral library with other data in various structural databases that are online. This permits you to choose thousands of structural possibilities and sort them numerically based on sub-structural data and spectral similarity (https://www.thermofisher.com/us/en/home/industrial/mass-spectrometry/liquid-chromatography-mass-spectrometry-lc-ms/lc-ms-software/mass-spectral-libraries/mzlogic-data-analysis.html).³² The metabolites identified with less than two structure annotation sources or mass error greater than five ppm were excluded from further analysis. The proposed identification results were then checked manually to eliminate false positives. Finally, the peak areas under the curve were extracted and integrated using the default setting of Compound Discoverer.

Statistical analysis

Statistical analyses were performed using Perseus Software 1.6.15.0.^33,34 The discovery and verification datasets were analyzed independently. First, the intensity of each metabolite was log2 transformed. Then the mean of log2 transformed intensities of the technical replicates were calculated, and the group differences in metabolite MS intensity were analyzed using Student's t-test.

Hierarchical clustering and principal component analysis (PCA) were performed using Perseus Software 1.6.15.0.^33,34 The unsupervised hierarchical clustering and heat map were based on the log2-transformed MS intensity of metabolites. The rows of the heat map indicate the metabolites and the columns indicate the samples. The log2 intensity of each metabolite was z-score normalized for each row. Hierarchical clustering of the z-normalized log2 intensity was performed using Euclidean distances between means. The number of clusters was set at 300.

Classification analysis and feature selection

Classification analysis and feature selection were performed using the machine learning module of Perseus Software 1.6.15.0.^33,34 Support vector machine (SVM) modeling with radial basis function (RBF) kernel was performed by 4-fold cross-validation. Duplicated features were removed based on feature optimization ranking prior to model building. Features that are medications and their metabolites were also excluded from model building. Measurements from the discovery cohort were used as the “training” data set, and those from the verification cohort as the “test” data set. Model performance was evaluated by analysis of the AUC of the ROC curve, where sensitivity (true positive) versus 1-specificity (false positive) was plotted and assessment of classification accuracy was carried out.

In order to validate the analysis, intensities were standardized before the SVM model was run on the training dataset (Discovery cohort) with prefiltered metabolites, and predictions on the model using the test dataset (Verification cohort) were compared. To also look at individual metabolites, logistic regression models were fit for each one of the metabolites. Significant interactions between pairs of metabolites were searched using the rFSA algorithm.³⁵ Logistic regression analyses and SVM models were run on JMP version 16.1, SAS Institute Inc., Cary, NC. All other analyses were run using Rstudio version 2022.12.0 and R version 4.2.1. The significance for p-value was set at 0.05. Similarly, t-test with Permutation q-value <0.05 was used to control for false positive discovery rate in a multiple comparisons problem.

Results

Study subjects

The discovery cohort participants comprised ten cognitively unimpaired older adults and ten AD patients (Table 1). The cognitively unimpaired older adults had a mean age of 68.60 ± 7.12, a mean CDR global score of 0.00 ± 0.00, and a mean Mini-Mental State Examination (MMSE) score of 29.50 ± 0.70. Six of the cognitively unimpaired older adults were women. The ten AD patients had a mean age of 73.00 ± 11.37, a mean CDR global score of 1.35 ± 0.74, and a mean MMSE score of 16.60 ± 5.79. Four of the AD patients were women. The verification cohort participants comprised 30 cognitively unimpaired older adults and 27 AD patients (Table 1). The cognitively unimpaired older adults had a mean age of 68.40 ± 6.50, a mean CDR global score of 0.00 ± 0.00, and a mean MMSE score of 29.60 ± 0.56. Twelve of the cognitively unimpaired older adults were women. The 27 AD patients had a mean age of 69.19 ± 9.86, a mean CDR global score of 1.33 ± 0.75, and a mean MMSE score of 16.59 ± 7.66. Ten of the AD patients were women. None of the discovery and verification cohort participants had any other types of dementia or other neurological diseases, such as Parkinson's disease.

Identification of blood-based metabolites that differentiate AD and cognitively unimpaired older adults

We analyzed blood samples using UPLC-HRMS to identify metabolites that can distinguish between Alzheimer's patients and cognitively normal older adults. Our results were as follows:

Discovery Phase (20 participants: 10 controls, 10 AD patients): 7075 metabolites were detected. Of these, 1414 showed significant differences between AD patients and controls (p < 0.05) as indicated in Figure 2A and 2B. Some metabolites were detected multiple times in each sample. Using Support Vector algorithms, we ranked them by their distinguishing ability and retained only the highest-ranking instance of each metabolite. This left 750 unique significant metabolites differentiating AD from cognitively unimpaired age matched older adults.

Figure 2.

Identification of blood-based metabolites that differentiate AD and cognitively unimpaired older adults. (A) Volcano plot of metabolites quantified in the discovery study. Red, metabolites were significantly elevated in the serum of AD patients; Green, metabolites were significantly depleted in the serum of AD patients; Grey, non-significant metabolites. (B) Heatmap of significant metabolites identified in the discovery study. (C) Volcano plot of metabolites quantified in the discovery study. Red, metabolites were significantly elevated in the serum of AD patients; Green, metabolites were significantly depleted in the serum of AD patients; Grey, non-significant metabolites. (D) Heatmap of significant metabolites identified in the discovery study. (E) Venn diagram of the significant metabolites identified in the discovery and verification studies. (F) Comparison of the profiles of the significant metabolites identified in both discovery and verification studies.

Verification Phase (57 participants: 30 controls, 27 AD patients): We detected 10,595 metabolites. 5563 metabolites showed significant differences (Student's t-test, p < 0.05). This was notably higher than in the discovery phase. To ensure robustness in our findings, we adjusted for multiple comparisons using a Permutation-based method, aiming to maintain a false discovery rate (FDR) under 5%. This resulted in 5136 significant metabolites, as shown in Figure 2C and 2D. After accounting for repeated metabolites, similar to the earlier process, 908 unique significant metabolites remained.

Comparison of discovery and verification phases

Three hundred fifty-one metabolites were found to be significant in both phases (see Figure 2E). Out of these, 162 metabolites (the upper-right and lower-left quarters of Figure 2 F) exhibited consistent differences in fold change between the AD and control groups across both studies. As shown in Figure 2 F, most of these 162 metabolites had consistent fold changes between the Discovery and Verification phases, but a few showed discrepancies. For instance, in the Discovery study, Hexylbenzene was found to be 504-fold more abundant in AD compared to healthy controls, whereas in the Verification study, it was only 13.86-fold higher. This may be due to the high variability of biomarker levels across the human population and the considerable molecular heterogeneity of individual AD patients.

After excluding five metabolites related to medications, the final list consisted of 157 metabolites. These metabolites were then used to create computational diagnostic models for AD.

Classification feature optimization

Classification with support vector machine algorithm

We employed the SVM algorithm with an RBF kernel, combined with 4-fold cross-validation, to rank the 157 metabolite features initially identified in the discovery study. The prediction error rate corresponding to the number of features utilized for AD prediction is depicted in Figure 3A. Notably, employing the foremost 14 features resulted in the minimal error rate of 5%. Utilizing additional features did not enhance this rate. When we performed principal component analysis (PCA) on the abundances of these top 14 metabolites, the AD cohort was discernibly segregated from the cognitively unimpaired older group, as shown in Figure 3B. Hence, these 14 metabolites were chosen as potential biomarkers for AD diagnosis. Their abundance profiles in the serum of both AD patients and healthy older adults are presented in Figure 3C and Table 2.

Figure 3.

Classification feature optimization. (A) Classification error rate versus the number of features used in the discovery study. (B) PCA analysis of subjects in the discovery studies using the top 14 features identified by the SVM algorithm and 4-fold cross-validation. (C) Profiles of the top 14 metabolite features in the discovery study. (D) Profiles of the top 14 metabolite features in the verification study. (E) Classification error rate versus the number of features used in the verification study. (F) PCA analysis of subjects of the verification studies using the top 14 features identified in the discovery study.

Table 2.

The profiles of biomarker candidates of SVM model in the discovery and verification studies.

Name	Formula		Discovery(n = 20, AD, 10; Control, 10)		Verification(n = 57, AD, 27; Control, 30)
Name	Formula	HMDB ID	Fold change (AD/Control)	p	Fold change (AD/Control)	p
Traumatic acid	C12H20O4	HMDB0000933	8.82	0.00064	3.07	0.00072
Hexylbenzene	C12H18	HMDB0061815	504.21	0.00003	13.86	0.00002
Androstenedione	C19H26O2	HMDB0000053	5.38	0.00043	1.49	0.00741
Ionene	C13H18	HMDB0059826	3.64	0.00030	2.41	0.02019
Benzene	C6H6	HMDB0001505	−1.85	0.00615	−2.18	0.00341
Dk3970000	C14H20O2	HMDB0013817	2.60	0.00080	2.56	0.00048
Ethylparaben	C9H10O3	HMDB0032573	−0.38	0.00052	−0.49	0.00037
M-Cresol	C7H8O	HMDB0002048	−1.71	0.02414	−26.95	0.00000
3360	C7H10O	HMDB0031541	−1.81	0.00007	−8.56	0.00000
1-methylimidazole	C4H6N2	HMDB0243940	−2.40	0.03449	−6.33	0.00017
2-Furoic acid	C5H4O3	HMDB0000617	−2.18	0.02694	−1.76	0.00594
γ-Oxo-3-pyridinebutanal	C9H9NO2	HMDB0004186	2.07	0.00033	1.71	0.00138
Isoquinoline	C9H7N	HMDB0034244	2.18	0.02462	2.11	0.03542
Linalyl butyrate	C14H24O2	HMDB0030427	3.22	0.00138	2.65	0.03340

In the verification phase, we assessed if these 14 metabolites could differentiate the AD cohort from the cognitively unimpaired older adults. Their profiles in this phase are illustrated in Figure 3D and Table 2, revealing a congruence with the discovery phase profiles. The prediction error in the verification study, utilizing these 14 metabolite profiles and SVM with an RBF kernel combined with 4-fold cross-validation, was 0.0%, as visualized in Figure 3E. A PCA based on these 14 metabolites’ abundances distinctly segregated the AD cohort from the cognitively healthy older adults, as displayed in Figure 3F.

Further, using the discovery dataset, we trained a model and subsequently tested it on the verification dataset. The efficacy of the resulting models in disease differentiation was evaluated using ROC curves, which graphically represent sensitivity against (1-specificity) for variable discrimination thresholds.³⁶ The AUC signifies a model's discriminative capacity between outcome groups. The ROCs for individual metabolites from both datasets are shown in Figure 4. In the discovery dataset, 10 out of the 14 metabolites attained an AUC exceeding 0.8 for AD prediction, whereas in the verification dataset, only four metabolites achieved this threshold. Incorporating all 14 metabolites considerably augmented model efficacy. The AUC values were 1.00 for the training dataset and 0.97 for the testing dataset, showcasing a robust performance in AD prediction. The confusion matrices for the models across both datasets can be found in Table 3, highlighting an accuracy rate of 100% for the training dataset and 86% for the testing dataset. Additionally, we investigated potential interaction terms for synergistic effects. Significant interactions were discerned between Hexylbenzene and Ethylparaben (p = 0.031), Traumatic Acid and 2,6-Di-tert-butylbenzoquinone (Dk3970000) (p = 0.01), and benzene and 3-Methyl-2-cyclohexen-1-one (3360) (p = 0.026).

Figure 4.

Performance of SVM classification model. Classification model with 14 metabolic features using the discovery dataset and tested with the verification dataset. The ROC and AUC of each individual metabolite and the combination of 14 features were shown.

Table 3.

Confusion matrices for the model in both datasets.

Training Dataset				Testing Dataset
		Predicted				Predicted
	Total n = 20	Control n = 10	AD n = 10		Total n = 57	Control n = 30	AD n = 27
Actual	Control 10	10	0	Actual	Control 30	26	4
Actual	AD 10	0	10	Actual	AD 27	4	23

Associations of metabolites with gender

We stratified the AD group and cognitively unimpaired older adult group (control) by sex. As shown in Table 4, in the female group, nine out of the 14 metabolites significantly differed between AD and cognitively unimpaired older adults, and in the male group, 11 out of the 14 metabolites significantly differed between AD and cognitively unimpaired older adults. Some of the metabolite biomarkers were not found to be significant in each gender group, probably due to the reduced sample size. In the group of cognitively unimpaired older adults, the abundance of the 14 metabolites shows no significant difference between females and males. However, in the AD group, the level of androstenedione in male AD patients was significantly higher than in female AD patients. Taken together, androstenedione is the only biomarker within the biomarker panel that correlates with gender. We also noted that this analysis did not take into consideration whether any subjects were taking androstenedione or undergoing hormonal replacement therapy. Androstenedione is more commonly used by young athletes to improve performance, and our age groups would generally not fall into this category. Some women may have received hormonal replacement therapy; however, we do not know whether any of these patients received hormonal therapy, as TARCC did not collect this information. We will investigate this gender effect using a well-controlled cohort in the future.

Table 4.

Association of 14 metabolite biomarkers with gender. The t-test p-values <0.05 was bolded.

Metabolites	FemalesAD (n = 10) versus Control (n = 12)		MalesAD (n = 17) versus Control (n = 18)		ControlsFemale (n = 12) versus Male (n = 18)		ADFemale (n = 10) versus Male (n = 17)
Metabolites	Fold change	p	Fold change	p	Fold change	p	Fold change	p
Traumatic acid	3.87	0.007	2.64	0.032	−0.45	0.524	0.10	0.872
Hexylbenzene	13.63	0.025	13.97	<0.001	−0.09	0.778	−0.13	0.944
Androstenedione	1.16	0.463	1.72	0.006	−0.18	0.52	−0.74	0.023
Ionene	1.30	0.642	3.51	0.014	0.41	0.514	−1.02	0.280
Benzene	−3.28	0.002	−1.71	0.136	0.07	0.88	−0.87	0.151
Dk3970000	2.80	0.008	2.40	0.019	−0.41	0.29	−0.19	0.784
Ethylparaben	1.76	0.122	2.21	0.001	0.03	0.902	−0.30	0.579
M-Cresol	−38.56	<0.001	−21.69	<0.001	0.19	0.708	−0.65	0.409
3360	−7.53	0.002	−9.38	<0.001	−0.49	0.384	−0.17	0.838
1-methylimidazole	−14.44	0.011	−3.94	0.002	−0.29	0.457	−2.17	0.116
2-Furoic acid	0.69	0.238	−1.97	0.014	−0.14	0.747	0.30	0.482
γ-Oxo-3-pyridinebutanal	1.41	0.070	1.91	0.008	−0.18	0.39	−0.62	0.173
Isoquinoline	2.06	0.226	2.16	0.090	0.34	0.679	0.28	0.657
Linalyl butyrate	2.35	0.219	2.79	0.095	−0.65	0.494	−0.90	0.351

Classification with random forest

We used the discovery phase dataset to apply Breiman and Cutler's RF algorithm to analyze 157 metabolite features identified in the discovery and verification studies and identified 20 features based on importance. Table 5 shows the profiles of these 20 metabolites in the serum of AD patients and cognitively unimpaired older adults and their importance in the RF model. We examined the prediction error using the profiles of 20 metabolites. As shown in Figure 5A, in the discovery study, the model yielded 0.0% prediction error rate with top five features, and increasing the number of features did not change the error rate. The prediction error rate for the verification study was higher than the discovery rate. The error rate with the top five features was 17.5% and the lowest error rate was 12.3% with top 16 features. Next, we trained the model using the discovery dataset and tested the trained model with the verification dataset. In this analysis, the performance of the resulting models in disease classification was examined using ROC curves. Figure 5B shows the AUC of ROCs. As the number of features increased, the AUC improved and the model with 13 features yielded the highest AUC of 0.96. Among the 13 features, six are overlapped with the features of SVM model, including Hexylbenzene, 3-Methyl-2-cyclohexen-1-one (3360), γ-Oxo-3-pyridinebutanal, DK3970000, and Androstenedione.

Figure 5.

Performance of RF classification model. (A) Prediction error rate of RF model. (B) AUC of ROCs of RF model.

Table 5.

The profiles of biomarker candidates of RF model in the discovery and verification studies.

Name	Formula	HMDB ID	Discovery(n = 20, AD, 10; Control, 10)		Verification(n = 57, AD, 27; Control, 30)		Importance
Name	Formula	HMDB ID	Fold change (AD/Control)	p	Fold change (AD/Control)	p	Importance
Hexylbenzene^a	C12H18	HMDB0061815	504.21	0.00003	13.86	0.00002	0.798396
N-Nonanoylglycine	C11H21NO3	HMDB0013279	2.69	0.00303	1.69	0.02339	0.37687
Ionene^a	C13H18	HMDB0059826	3.64	0.00030	2.41	0.02019	0.29151
2401/delta-Dodecalactone	C12H22O2	HMDB0037116	3.58	0.00168	2.45	0.00641	0.289255
3360/3-Methyl-2-cyclohexen-1-one^a	C7H10O	HMDB0031541	−1.81	0.00007	−8.56	0.00000	0.289138
(5Z,8Z)-5,8-Tetradecadienoic acid	C14H24O2	HMDB0340799	3.44	0.02998	1.85	0.00173	0.257661
FG7175000	C9H7NO		2.69	0.00175	1.55	0.00498	0.212159
γ-Oxo-3-pyridinebutanal^a	C9H9NO2	HMDB0004186	1.10	0.00033	1.71	0.00138	0.211026
DK3970000/2,6-Di-tert-butylbenzoquinone^a	C14H20O2	HMDB0013817	2.60	0.00080	2.56	0.00048	0.201963
3,4-Dimethyl-5-pentyl-2-furanpropanoic acid	C14H22O3		3.99	0.00723	3.98	0.00004	0.164183
Androstenedione^a	C19H26O2	HMDB0000053	5.38	0.00043	1.49	0.00741	0.15991
DL-carvone	C10H14O	HMDB0035089	1.93	0.00543	2.73	0.00000	0.157633
Decanoic acid	C10H20O2	HMDB0000511	3.05	0.00370	2.65	0.00001	0.140389
10-Undecenoic acid	C11H20O2	HMDB0033724	2.91	0.01305	2.72	0.00508	0.138158
MFCD00009868	C16H28O2	HMDB0340889	3.32	0.00355	2.96	0.00004	0.134775
4-Vinylcyclohexene	C8H12	HMDB0246592	−1.95	0.00185	−10.11	0.00019	0.131989
1-Phenyl-2-hexanone	C12H16O	HMDB0094661	2.00	0.00132	3.19	0.00013	0.130808
2781/Gamma-nonalactone	C9H16O2	HMDB0031531	2.24	0.00642	2.16	0.00853	0.129112
2,2-Methylenebisfuran	C9H8O2	HMDB0032947	2.82	0.00237	1.86	0.00002	0.121297
Allylcyclohexane	C9H16	HMDB0061920	2.14	0.00217	2.75	0.00309	0.12

Features that overlapped with the features in SVM model.

Based on the chemical structures, the metabolites features in SVM and RF models can be grouped into benzenoids (benzene; m-cresol; ethylparaben, hexylbenzene, ionene), organoheterocyclic compounds (2-furoic acid; isoquinoline), prenol lipids (2,6-di-tert-butylbenzoquinone, 2,6-Di-tert-butylbenzoquinone; (R)-Carvone), sterol lipids (androstenedione), fatty acyls (traumatic acid, gamma-nonalactone; Decanoic acid; Undecylenic acid), lipids and lipid-like molecules (Linalyl butyrate), cyclohexenones (3-Methyl-2-cyclohexen-1-one), imidazoles (1-methylimidazole), and aryl alkyl ketones (gamma-Oxo-3-pyridinebutanal).

Classification with endogenous metabolites

Some of the biomarkers in Tables 2 and 5 are either exclusive to plants or are not naturally occurring (exposome). The presence of these metabolites in human blood is due to diet or environmental exposure. Previous research has demonstrated that exogenous metabolites and environmental exposures play a substantial role in AD.^37–42 This underscores the importance of identifying, quantifying, and characterizing the entire spectrum of endogenous and exogenous metabolites in AD biological specimen samples and examining the biochemical processes linked to AD onset and evolution. However, we are interested in exploring the feasibility of creating an effective computational model for diagnosing AD using only endogenous metabolites. We excluded metabolites that are exclusively from plants or are not naturally occurring based on the Human Metabolome Database (HMBD) ontology section (https://hmdb.ca).⁴³ The remaining 56 endogenous metabolites were utilized for computational modeling, with the discovery dataset serving as the training dataset and the verification dataset as the testing dataset. Initially, we employed the random forest algorithm to assess the importance of each metabolite. Subsequently, the top 14 metabolites (Table 6) were analyzed using the SVM algorithm. As shown in Figure 6, increasing the number of features improves the area under the curve (AUC) of receiver operating characteristics (ROC) and reduces the prediction error rate. The model with four features - N-Nonanoylglycine, delta-Dodecalactone, 2,6-Di-tert-butylbenzoquinone, and 3-Methyl-2-cyclohexen-1-one - exhibits the best AUC. Incorporating more features did not notably enhance the model's performance.

Figure 6.

Performance of SVM classification model using endogenous metabolites. The AUC of ROCs and the prediction error rates are shown.

Table 6.

The profiles of biomarker candidates of SVM model using endogenous metabolites.

Name	Formula	HMDB ID	Discovery(n = 20, AD, 10; Control, 10)		Verification(n = 57, AD, 27; Control, 30)		Rank
Name	Formula	HMDB ID	Fold change (AD/Control)	p	Fold change (AD/Control)	p	Rank
N-Nonanoylglycine	C11H21NO3	HMDB0013279	2.69	0.003	1.69	0.023	1
delta-Dodecalactone	C12H22O2	HMDB0037116	3.58	0.002	2.45	0.006	2
2,6-Di-tert-butylbenzoquinone	C14H20O2	HMDB0013817	2.60	<0.001	2.56	<0.001	3
3-Methyl-2-cyclohexen-1-one	C7H10O	HMDB0031541	−1.81	<0.001	−8.56	<0.001	4
LINALYL BUTYRATE	C14H24O2	HMDB0030427	3.22	0.001	2.65	0.033	5
γ-Oxo-3-pyridinebutanal	C9H9NO2	HMDB0004186	2.07	<0.001	1.71	0.001	6
LILAC ACETALDEHYDE	C9H10O	HMDB0029639	1.81	0.002	3.18	<0.001	7
Traumatic Acid	C12H20O4	HMDB0000933	8.82	<0.001	3.07	<0.001	8
Homovanillin	C9H10O3	HMDB0005175	2.69	0.01	1.72	0.004	9
2-Tridecenal	C13H24O	HMDB0037275	2.70	0.008	2.55	0.022	10
Androstenedione	C19H26O2	HMDB0000053	5.38	<0.001	1.49	0.007	11
2-Acetyl-3,5-dimethylfuran	C8H10O2	HMDB0032159	1.78	0.01	2.80	<0.001	12
3-oxopalmitic acid	C16H30O3	HMDB0010733	2.71	0.004	1.54	0.02	13
8-Hydroxy-5,6-octadienoic acid	C8H12O3	HMDB0031101	2.67	0.02	1.55	0.01	14

Discussion

Peripheral blood tests for AD have emerged as a more convenient alternative, especially for repeated sampling in clinical trials.⁶ However, current blood-based tests, even those that assess Tau and Aβ levels, are limited by the scope of metabolites they can evaluate. This suggests the need for expanded paradigms of blood diagnostics for AD. The aim of this study is to discover novel blood biomarkers that could differentiate AD from cognitively unimpaired older adults with high specificity and sensitivity, laying the foundation for developing a blood diagnostic for AD. Accordingly, we integrated high-resolution mass spectrometry (HRMS)-based metabolomics technology and machine learning algorithms for developing biomarker identification and classification model. HRMS is one of the most commonly used technologies in metabolomics studies and offers excellent sensitivity and accuracy in metabolite identification.^44,45 In recent years, classification with machine learning algorithms has been widely used to achieve diagnostic and prognostic goals.^46–48 It is a data-driven approach where the algorithm is trained with labeled samples, each represented by a set of features, and the resulting classification model can predict the label of a sample from its features. In our study, the labels correspond to AD patients and cognitively unimpaired older adults, and the features are the abundance of metabolites in serum. The biomarker candidates that differentiated AD from cognitively unimpaired older adults were first identified using a discovery cohort, then these candidates were verified with a much larger verification cohort. In both discovery and verification studies, the dataset acquired in the discovery study was fed to the supervised machine-learning algorithms, and the resulting classification models were tested with the verification dataset. Because the number of significant metabolite features identified in the discovery and verification studies was larger than that which warrants further investigation, we performed feature reduction as commonly used in classification model development to narrow the list of potential biomarkers down to a more reasonable number by removing noisy data, redundant variables, or low-reproducible data.⁴⁹ The resulting 157 significant metabolite features were used for supervised machine learning. Both SVM and RF were performed, and each classifier selected a panel of biomarker candidates that accurately differentiated AD from cognitively unimpaired older adults. In comparison to newly discovered blood biomarkers for AD classification, our 14-metabolite SVM model or 13-metabolites RF model are accurately on par with the plasma p-tau181 (AUC = 0.85–0.98),^50,51 plasma p-tau217 (AUC = 0.89–0.96),⁵² a 19-protein panel (AUC = 0.96–0.98),⁵³ and outperformed a 5-metabolite panel (AUC = 0.84)⁵⁴ and a 2-metablites + age panel (AUC = 0.72).⁵⁵ Importantly, we rigorously validated the biomarker panels identified in this study using an independent set of samples and confirmed their excellent specificity and sensitivity in diagnosing AD.

Some of metabolic features identified by SVM and RF algorithms are overlapped, including Hexylbenzene, 3-Methyl-2-cyclohexen-1-one (3360), γ-Oxo-3-pyridinebutanal, DK3970000, and Androstenedione. Several studies previously identified potential plasma or CSF AD metabolic biomarkers classified as benzene and substituted derivatives.^56–58 However, none of this study's five benzenoid biomarkers have been reported previously. Among the five benzenoids, Hexylbenzene has the best predictive power, given that the AUC of ROCs in the Discovery and Verification studies are 1.0 and 0.8, respectively. Hexylbenzene has been identified from healthy human body fluids.⁵⁹ Our study shows that hexylbenzene was significantly elevated in the serum of AD patients, and the mechanism linking increased hexylbenzene levels with AD requires further studies.

Androstenedione is a steroidal hormone that is naturally occurring and made in the gonads and adrenal glands of both sexes. It is the common precursor of male and female sex hormones and can be transformed into testosterone and estrogen depending on the tissue type. Our study found that androstenedione was elevated in the serum of AD patients. Interestingly, we found a significant sex- difference in the correlation between androstenedione level with cognitive capacity. In the female group, the abundance of androstenedione shows no significant difference between AD and cognitively unimpaired aged adults. In the male group, androstenedione level was significantly elevated compared to cognitively unimpaired aged adults. In addition, in the AD group, the level of androstenedione in male AD patients was significantly higher than in female AD patients. Vaňková et al. also found that the level of androstenedione in AD patients was elevated compared to healthy controls.⁶⁰ However, they found that androstenedione was higher in women with AD than controls but not in men.⁶⁰ This inconsistency may be because the level of androstenedione is sex and age-dependent, and there is a significant age difference between AD patients and healthy controls (73.7 versus 66.4) in Vaňková ‘s study. To further validate the sex difference in the correlation of androstenedione level with cognitive capacity, we will need to use a larger cohort with age-, sex-matched participants who are free of concomitant androstenedione disorders, such as hyperplasia, congenital adrenal hyperplasia, and polycystic ovary syndrome. Although the role of androstenedione in AD pathogenesis is not fully understood, an in vivo study shows that elevated serum androstenedione levels correlate with memory impairment.⁶¹ Androstenedione administration altered performance within various domains measuring cognitive function. Specifically, it impaired spatial reference memory and the ability to maintain performance with increasing memory demand.⁶¹

1-Methylimidazole was found to be lower in AD serum and a significant metabolite candidate in the SVM model of the discovery and verification study (Table 2; Table 4, Row 10). Recently, a link between 1-Methylimidazole and AD pathology was proposed. The connection points to how Aβ binds copper (II) with high affinity in its N-terminal region.⁶² This is in line with earlier studies that reported the presence of Aβ-Cu interactions in the formation of toxic oligomers.^63,64 In accordance with those previous reports, a recent study synthesized and characterized two new 1-methylimidazole-containing N-acylhydrazonic metallophores (X1TMP and X1Benz). X1TMP contained the 34,5-trimethoxybenzoyl moiety present in the structure of mescaline, a natural hallucinogenic protoalkaloid found in various species of cacti. X1TMP and X1Benz are novel 1-methylimidazole-containing N-acylhydrazones that may represent agents of the future in the treatment of AD.⁶⁵

We also identified isoquinoline to be higher in AD serum and to be a significant metabolite in the SVM model (Table 2; Table 4, Row 13). Protoberberines are the most studied of eight subgroups of isoquinoline alkaloids (IAs), where cholinesterase activity relevant to AD has been identified. The plant Berberis vulgaris is a known rich source of quaternary protoberberine alkaloids and acetylcholinesterase inhibitors berberine and palmatine. These have been extensively studied with regard to AD in the past decade and may hold novel therapeutic options for the future.^66,67 In another study, Dauricine (DAU), a benzyl tetrahydroisoquinoline alkaloid isolated from the root portion of Menispermum dauricum DC was tested in an AD model. The study investigated the effects of DAU on Neuro2a (N2a) cells that were transfected with amyloid precursor protein (Swedish mutant). DAU treatment reduced AD-like pathology by inhibiting AβPP processing, reducing Aβ accumulation, and mitigating the hyperphosphorylation of tau via PP2A, p35/25, and CDK5 pathways. These findings suggest that DAU and other isoquinolines may have therapeutic utility in future AD treatment.⁶⁸

5,8 tetradecadienoic acid (Table 5, Row 6), was found to be a biomarker candidate in the RF model. An earlier study identified a panel of five plasma metabolites that allowed to differentiate between amnestic mild cognitive impairment (aMCI) and control subjects. The panel included 5,8 tetradecadienoic acid together with thymine, arachidonic acid, 2-aminoadipic acid, and N,N-dimethylglycine. In order to determine the exact role of 5,8 tetradecadienoic acid in MCI, additional comparative studies between AD, MCI, and healthy controls in serum and plasma will be necessary.⁶⁹

It has been demonstrated that the metabolism of medium chain fatty acids (MCFAs) containing 6 to 12 carbon atoms occurs in astrocytes. Octanoic acid (C8) and decanoic acid (capric acid) (C10) are the MCFAs that often play an important role in this.⁷⁰ Our study also identified decanoic acid as a biomarker candidate following the RF model analysis (Table 5, Row 13). In addition, it has been found that the supplementation of the same MCFAs C8 and C10 has shown a positive effect on cognition in human MCI and AD studies.^71–73 Furthermore, a recent report showed an improvement in antioxidative status as well as a reduction in ROS levels in human (SH-SY5Y) and murine Neuro2a cells supplemented with phosphatidylcholine (PC) plus the addition of decanoic acid.⁷⁴ It has been proposed that decanoic acid, as well as other MCFAs may affect cell energetics by improving mitochondrial function. Also, the activation of peroxisome proliferator-activated receptor gamma (PPARγ) by decanoic acid has been shown to generate mitochondrial biogenesis and prevent neuronal death induced by glucose deprivation.⁷⁵ This evidence suggests that decanoic acid and other MCFAs may be effective in reducing oxidative stress, promoting neuronal health and improving cognition levels. Accordingly, we also found decanoic acid to be an associated biomarker candidate in our RF model for the discovery and verification studies. Overall, the significant associations with decanoic acid suggest that further studies should be performed in both models and in humans.

Our RF model also found 10-Undecenoic acid (undecylenic acid), a naturally occurring unsaturated fatty acid, to be a candidate biomarker (Table 5, Row 14). Recently, 10-Undecenoic acid was discovered to inhibit μ-calpain, a factor that bridges Aβ- and hyperphosphorylated tau-mediated pathways. It showed neuroprotective effects by inhibiting Aβ oligomerization and Aβ fibrillation. 10-Undecenoic acid reversed neuronal cell death induced by Aβ. Therefore, future studies that consider the mechanism of μ-calpain inhibition by 10-Undecenoic acid would be beneficial in order to determine if it is a feasible treatment option for the future.⁷⁶

4-Vinylcyclohexene (VCH) was another significant candidate biomarker found in the RF model (Table 5, Row 16). Studies have often used VCH and the 4-4-Vinylcyclohexene diepoxide metabolite (VCD) for induction of ovarian failure in mice. Importantly, both occupational chemicals have been identified as having carcinogenic potential based on studies conducted in mice and rats. Consequently, it was found that VCH and VCD caused ovarian follicle and uterine atrophy in mice exposed for 13 weeks.⁷⁷ Recently, VCD was utilized in a study that measured cognitive and neurological deficits in female APOE-TR mice who received a high-fat diet and had ovarian failure induced through VCD injections. They found that VCD-induced ovarian failure in conjunction with APOE4 homozygosity led to an increase in both cognitive and neurological deficits.⁷⁸ Another study conducted in APOE4 VCD-treated mice found improved recognition memory (NOR) (p < 0.05), but no change in spatial working memory (Y-maze) in mice that received DHA-enriched fish oil.⁷⁹ VCD (and VCH) application in AD menopause model studies in mice will require further and innovative designs in conjunction with APOE phenotype status. This will help to determine if genetic and occupational risk mechanisms are gender related in AD.

Traumatic acid, a fatty acid derivative, was first isolated and synthesized by English, Bonner, and Haagen-Smit in 1939.⁸⁰ It is a known wound hormone which causes parenchymatous tissues in plants (bean pods and potato tubers) to proliferate.⁸¹ Recently, Traumatic acid was tested in cytotoxicity assessments in which it was found that it may act as a compound that delays the stimulatory effect of pesticides on breast cancer cells.^81,82 In our study, the biological significance of Traumatic acid could be diet related. However, more specifically designed studies will be required in order to determine if our finding is associated with the consumption of Traumatic acid found in food sources.

This study lays the foundations for future multi-site analytical and clinical validation of these biomarkers and classification models. Once validated with larger sample sets covering a broad patient population, this panel and classification model will enable timely diagnosis of AD objectively and quantitatively so that clinicians can start therapeutic intervention early and monitor the efficacy of treatment. One limitation of this study is that we did not examine whether this biomarker panel can differentiate AD from other types of dementia, such as behavioral variant frontotemporal dementia or vascular dementia, or whether there are phenotypes- and genotype-difference in the performance of this classification model. We will study this topic in future studies. Another limitation is the slight gender difference between the groups (60% women in the control group versus 40% women in the AD group). It is hoped that future studies conducted in a larger and more carefully selected “n” could help to address this discrepancy given that these valuable samples were generously provided to us by a collaborative bio-banking facility.

Footnotes

Acknowledgments

We would like to thank the Texas Alzheimer's Research and Care Consortium (TARCC) for providing the serum samples.

ORCID iDs

Yingxin Zhao

Alejandro Villasante-Tezanos

Ernesto G. Miranda-Morales

Xiang Fang

Author contributions

Yingxin Zhao (Conceptualization; Data curation; Methodology; Software; Validation; Writing – original draft; Writing – review & editing); Alejandro Villasante-Tezanos (Data curation; Methodology; Software; Writing – original draft; Writing – review & editing); Ernesto G Miranda-Morales (Investigation; Methodology; Writing – original draft; Writing – review & editing); Miguel Pappolla (Conceptualization; Funding acquisition; Validation; Writing – original draft; Writing – review & editing); Xiang Fang (Conceptualization; Funding acquisition; Validation; Writing – original draft; Writing – review & editing).

Funding

This work was supported by National Institutes of Health Grants NIA (1R01AG077253) (to YZ), NIA (1R21AG066060) (to XF and MAP), NIA R61AG075725 (to XF), Texas Alzheimer's Research and Care Consortium (TARCC) research grant (to XF), and John Sealy Distinguished Chair in Alzheimer's diseases (XF).

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data availability

The data supporting the findings of this study are available within the article.

References

Knopman

Amieva

Petersen

, et al. Alzheimer disease. Nat Rev Dis Primers 2021; 7: 33.

Beiser

Breteler

MMB

, et al. The changing prevalence and incidence of dementia over time - current evidence. Nat Rev Neurol 2017; 13: 327–339.

Long

Holtzman

. Alzheimer disease: an update on pathobiology and treatment strategies. Cell 2019; 179: 312–339.

Jack

Jr. Bennett

Blennow

, et al. A/T/N: an unbiased descriptive classification scheme for Alzheimer disease biomarkers. Neurology 2016; 87: 539–547.

Jack

Jr. Bennett

Blennow

, et al. NIA-AA Research framework: toward a biological definition of Alzheimer's disease. Alzheimers Dement 2018; 14: 535–562.

Mattsson

Carrillo

Dean

, et al. Revolutionizing Alzheimer's disease and clinical trials through biomarkers. Alzheimers Dement (Amst) 2015; 1: 412–419.

Blennow

Zetterberg

. Biomarkers for Alzheimer's disease: current status and prospects for the future. J Intern Med 2018; 284: 643–663.

Lashley

Schott

Weston

, et al. Molecular biomarkers of Alzheimer's disease: progress and prospects. Dis Model Mech 2018; 11: dmm031781.

Simren

Leuzy

Karikari

, et al. The diagnostic and prognostic capabilities of plasma biomarkers in Alzheimer's disease. Alzheimers Dement 2021; 17: 1145–1156.

10.

Cullen

Leuzy

Janelidze

, et al. Plasma biomarkers of Alzheimer's disease improve prediction of cognitive decline in cognitively unimpaired elderly populations. Nat Commun 2021; 12: 3555.

11.

Voyle

Kim

Proitsi

, et al. Blood metabolite markers of neocortical amyloid-beta burden: discovery and enrichment using candidate proteins. Transl Psychiatry 2016; 6: e719.

12.

Zhang

Alshakhshir

Zhao

. Glycolytic metabolism, brain resilience, and Alzheimer's disease. Front Neurosci 2021; 15: 662242.

13.

Yassine

Finch

. APOE Alleles and diet in brain aging and Alzheimer's disease. Front Aging Neurosci 2020; 12: 150.

14.

de Leon

Convit

Wolf

, et al. Prediction of cognitive decline in normal elderly subjects with 2-[(18)F]fluoro-2-deoxy-D-glucose/poitron-emission tomography (FDG/PET). Proc Natl Acad Sci U S A 2001; 98: 10966–10971.

15.

Mosconi

De Santi

, et al. Hippocampal hypometabolism predicts cognitive decline from normal aging. Neurobiol Aging 2008; 29: 676–692.

16.

Wieckowska-Gacek

Mietelska-Porowska

Wydrych

, et al. Western diet as a trigger of Alzheimer's disease: from metabolic syndrome and systemic inflammation to neuroinflammation and neurodegeneration. Ageing Res Rev 2021; 70: 101397.

17.

Pillai

Bena

Bekris

, et al. Metabolic syndrome biomarkers relate to rate of cognitive decline in MCI and dementia stages of Alzheimer's disease. Alzheimers Res Ther 2023; 15: 54.

18.

Feng

Nyunt

, et al. Metabolic syndrome and the risk of mild cognitive impairment and progression to dementia: follow-up of the Singapore longitudinal ageing study cohort. JAMA Neurol 2016; 73: 456–463.

19.

Sabbagh

Perez

Holland

, et al. Primary prevention recommendations to reduce the risk of cognitive decline. Alzheimers Dement 2022; 18: 1569–1579.

20.

Horgusluoglu

Neff

Song

, et al. Integrative metabolomics-genomics approach reveals key metabolic pathways and regulators of Alzheimer's disease. Alzheimers Dement 2022; 18: 1260–1278.

21.

Schumacher-Schuh

Bieger

Borelli

, et al. Advances in proteomic and metabolomic profiling of neurodegenerative diseases. Front Neurol 2021; 12: 792227.

22.

Jia

Yang

Zhu

, et al. A metabolite panel that differentiates Alzheimer's disease from other dementia types. Alzheimers Dement 2022; 18: 1345–1356.

23.

Pappolla

Perry

Fang

, et al. Indoles as essential mediators in the gut-brain axis. Their role in Alzheimer's disease. Neurobiol Dis 2021; 156: 105403.

24.

Kalecky

German

Montillo

, et al. Targeted metabolomic analysis in Alzheimer's disease plasma and brain tissue in non-hispanic whites. J Alzheimers Dis 2022; 86: 1875–1895.

25.

Teruya

Chen

Kondoh

, et al. Whole-blood metabolomics of dementia patients reveal classes of disease-linked metabolites. Proc Natl Acad Sci U S A 2021; 118: e2022857118.

26.

Texas Alzheimer’s Research and Care Consortium. A Darrel K Royal Texas Alzheimer’s initiative, https://www.txalzresearch.org (2024, accessed 21 August 2024).

27.

Waring

O'Bryant

Reisch

, et al. The Texas Alzheimer's research consortium longitudinal research cohort: study design and baseline characteristics. Texas Public Health J 2008; 60: 9–13.

28.

McKhann

Drachman

Folstein

, et al. Clinical diagnosis of Alzheimer's disease: report of the NINCDS-ADRDA work group under the auspices of department of health and human services task force on Alzheimer's disease. Neurology 1984; 34: 939–944.

29.

Hughes

Berg

Danziger

, et al. A new clinical scale for the staging of dementia. Br J Psychiatry 1982; 140: 566–572.

30.

M/Z Cloud™. Advanced mass spectral database, https://www.mzcloud.org (2024, accessed 21 August 2024).

31.

ChemSpider. Search and share chemistry, https://www.chemspider.com (2024, accessed 21 August 2024).

32.

mzLogic. Data analysis Algorythm, https://www.thermofisher.com/us/en/home/industrial/mass-spectrometry/liquid-chromatography-mass-spectrometry-lc-ms/lc-ms-software/mass-spectral-libraries/mzlogic-data-analysis.html (2024, accessed 21 August 2024).

33.

Tyanova

Cox

. Perseus: a bioinformatics platform for integrative analysis of proteomics data in cancer research. Methods Mol Biol 2018; 1711: 133–148.

34.

Tyanova

Temu

Sinitcyn

, et al. The Perseus computational platform for comprehensive analysis of (prote)omics data. Nat Methods 2016; 13: 731–740.

35.

Lambert

Gong

Elliott

, et al. rFSA: an R package for finding best subsets and interactions. R J 2018; 10: 295–308.

36.

Zweig

Campbell

. Receiver-operating characteristic (ROC) plots: a fundamental evaluation tool in clinical medicine. Clin Chem 1993; 39: 561–577.

37.

Finch

Kulminski

. The Alzheimer's disease exposome. Alzheimers Dement 2019; 15: 1123–1132.

38.

Lefevre-Arbogast

Chaker

Mercier

, et al. Assessing the contribution of the chemical exposome to neurodegenerative disease. Nat Neurosci 2024; 27: 812–821.

39.

Sakowski

Koubek

Chen

, et al. Role of the exposome in neurodegenerative disease: recent insights and future directions. Ann Neurol 2024; 95: 635–652.

40.

Vardarajan

Kalia

Manly

, et al. Differences in plasma metabolites related to Alzheimer's disease, APOE epsilon4 status, and ethnicity. Alzheimers Dement (N Y) 2020; 6: e12025.

41.

Vignoli

Tenori

. NMR-based metabolomics in Alzheimer's disease research: a review. Front Mol Biosci 2023; 10: 1308500.

42.

Wang

Chen

, et al. Formaldehyde, epigenetics, and Alzheimer's disease. Chem Res Toxicol 2019; 32: 820–830.

43.

HMBD. The human metabolome database, https://hmdb.ca (2024, accessed 21 August 2024).

44.

Heiles

. Advanced tandem mass spectrometry in metabolomics and lipidomics-methods and applications. Anal Bioanal Chem 2021; 413: 5927–5948.

45.

Letertre

MPM

Dervilly

Giraudeau

. Combined nuclear magnetic resonance spectroscopy and mass spectrometry approaches for metabolomics. Anal Chem 2021; 93: 500–518.

46.

Massetti

Russo

Franciotti

, et al. A machine learning-based holistic approach to predict the clinical course of patients within the Alzheimer's disease spectrum. J Alzheimers Dis 2022; 85: 1639–1655.

47.

Franciotti

Nardini

Russo

, et al. Comparison of machine learning-based approaches to predict the conversion to Alzheimer's disease from mild cognitive impairment. Neuroscience 2023; 514: 143–152.

48.

Wang

Zhang

, et al. Rapid biomarker screening of Alzheimer's disease by interpretable machine learning and graphene-assisted Raman spectroscopy. ACS Nano 2022; 16: 6426–6436.

49.

Spratt

. Statistical approaches to candidate biomarker panel selection. Adv Exp Med Biol 2016; 919: 463–492.

50.

Karikari

Pascoal

Ashton

, et al. Blood phosphorylated tau 181 as a biomarker for Alzheimer's disease: a diagnostic performance and prediction modelling study using data from four prospective cohorts. Lancet Neurol 2020; 19: 422–433.

51.

Karikari

Benedet

Ashton

, et al. Diagnostic performance and prediction of clinical progression of plasma phospho-tau181 in the Alzheimer's disease neuroimaging initiative. Mol Psychiatry 2021; 26: 429–442.

52.

Palmqvist

Janelidze

Quiroz

, et al. Discriminative accuracy of plasma phospho-tau217 for Alzheimer disease vs other neurodegenerative disorders. JAMA 2020; 324: 772–781.

53.

Jiang

Zhou

, et al. Large-scale plasma proteomic profiling identifies a high-performance biomarker panel for Alzheimer's disease screening and staging. Alzheimers Dement 2022; 18: 88–102.

54.

Shao

Ouyang

, et al. Alteration of metabolic profile and potential biomarkers in the plasma of Alzheimer's disease. Aging Dis 2020; 11: 1459–1470.

55.

Lin

Huang

, et al. A metabolomic approach to identifying biomarkers in blood of Alzheimer's disease. Ann Clin Transl Neurol 2019; 6: 537–545.

56.

Kaddurah-Daouk

Zhu

Sharma

, et al. Alterations in metabolic pathways and networks in Alzheimer's disease. Transl Psychiatry 2013; 3: e244.

57.

Trushina

Dutta

Persson

, et al. Identification of altered metabolic pathways in plasma and CSF in mild cognitive impairment and Alzheimer's disease using metabolomics. PLoS One 2013; 8: e63644.

58.

Gonzalez-Dominguez

Garcia-Barrera

Gomez-Ariza

. Application of a novel metabolomic approach based on atmospheric pressure photoionization mass spectrometry using flow injection analysis for the study of Alzheimer's disease. Talanta 2015; 131: 480–489.

59.

de Lacy Costello

Amann

Al-Kateb

, et al. A review of the volatiles from the healthy human body. J Breath Res 2014; 8: 014001.

60.

Vankova

Velikova

Vejrazkova

, et al. The role of steroidomics in the diagnosis of Alzheimer's disease and type 2 diabetes mellitus. Int J Mol Sci 2023; 24: 8575.

61.

Camp

Gerson

Tsang

, et al. High serum androstenedione levels correlate with impaired memory in the surgically menopausal rat: a replication and new findings. Eur J Neurosci 2012; 36: 3086–3095.

62.

Atrian-Blasco

Gonzalez

Santoro

, et al. Cu and Zn coordination to amyloid peptides: from fascinating chemistry to debated pathological relevance. Coord Chem Rev 2018; 375: 38–55.

63.

Savelieff

Lee

Liu

, et al. Untangling amyloid-beta, tau, and metals in Alzheimer's disease. ACS Chem Biol 2013; 8: 856–865. 20130318.

64.

Rana

Sharma

. Cu and Zn interactions with Abeta peptides: consequence of coordination on aggregation and formation of neurotoxic soluble Abeta oligomers. Metallomics 2019; 11: 64–84.

65.

Carvalho

Barbosa

Flores

, et al. New mescaline-related N-acylhydrazone and its unsubstituted benzoyl derivative: promising metallophores for copper-associated deleterious effects relief in Alzheimer's disease. J Inorg Biochem 2023; 238: 112033.

66.

Cahlikova

Vrabec

Pidany

, et al. Recent progress on biological activity of Amaryllidaceae and further isoquinoline alkaloids in connection with Alzheimer's disease. Molecules 2021; 26: 5240.

67.

Honorio

Sainimnuan

Hannongbua

, et al. Binding interaction of protoberberine alkaloids against acetylcholinesterase (AChE) using molecular dynamics simulations and QM/MM calculations. Chem Biol Interact 2021; 344: 109523.

68.

Liu

Chen

Zhou

, et al. The isoquinoline alkaloid dauricine targets multiple molecular pathways to ameliorate Alzheimer-like pathological changes in vitro. Oxid Med Cell Longev 2018; 2018: 2025914.

69.

Wang

Zhou

Huang

, et al. Plasma metabolite profiles of Alzheimer's disease and mild cognitive impairment. J Proteome Res 2014; 13: 2649–2658.

70.

Ameen

Freude

Aldana

. Fats, friends or foes: investigating the role of short- and medium-chain fatty acids in Alzheimer's disease. Biomedicines 2022; 10: 2778.

71.

Rebello

Keller

Liu

, et al. Pilot feasibility and safety study examining the effect of medium chain triglyceride supplementation in subjects with mild cognitive impairment: a randomized controlled trial. BBA Clin 2015; 3: 123–125.

72.

Avgerinos

Egan

Mattson

, et al. Medium chain triglycerides induce mild ketosis and may improve cognition in Alzheimer's disease. A systematic review and meta-analysis of human studies. Ageing Res Rev 2020; 58: 101001.

73.

Fortier

Castellano

Croteau

, et al. A ketogenic drink improves brain energy and some measures of cognition in mild cognitive impairment. Alzheimers Dement 2019; 15: 625–634.

74.

Mett

Muller

. The medium-chain fatty acid decanoic acid reduces oxidative stress levels in neuroblastoma cells. Sci Rep 2021; 11: 6135. 20210317.

75.

Mett

. The impact of medium chain and polyunsaturated omega-3-fatty acids on amyloid-beta deposition, oxidative stress and metabolic dysfunction associated with Alzheimer's disease. Antioxidants (Basel) 2021; 10: 1991.

76.

Lee

Eom

Kim

, et al. Neuroprotective effect of undecylenic acid extracted from Ricinus communis L. through inhibition of mu-calpain. Eur J Pharm Sci 2012; 46: 17–25.

77.

Kappeler

Hoyer

. 4-vinylcyclohexene Diepoxide: a model chemical for ovotoxicity. Syst Biol Reprod Med 2012; 58: 57–62.

78.

Pontifex

Martinsen

Saleh

RNM

, et al. APOE4 Genotype exacerbates the impact of menopause on cognition and synaptic plasticity in APOE-TR mice. FASEB J 2021; 35: e21583.

79.

Pontifex

Martinsen

Saleh

RNM

, et al. DHA-enriched fish oil ameliorates deficits in cognition associated with menopause and the APOE4 genotype in rodents. Nutrients 2022; 14: 1698.

80.

English

Jr. Bonner

Haagen-Smit

. Structure and synthesis of a plant wound hormone. Science 1939; 90: 329.

81.

Farmer

. Fatty acid signalling in plants and their associated microorganisms. Plant Mol Biol 1994; 26: 1423–1437.

82.

Jablonska-Trypuc

Wydro

Wolejko

, et al. Toxicological effects of traumatic acid and selected herbicides on human breast cancer cells: in vitro cytotoxicity assessment of analyzed compounds. Molecules 2019; 24: 1710.