Targeted serum metabolomic profiling and machine learning approach in Alzheimer's disease using the Alzheimer's disease diagnostics clinical study (ADDIA) cohort

Abstract

Background

Metabolic biomarkers can potentially be used for early diagnosis, prognostic risk stratification and/or early treatment and prevention of individuals at risk to develop Alzheimer's disease (AD).

Objective

Our goal was to evaluate changes in metabolite concentration levels associated with AD to identify biomarkers that could support early and accurate diagnosis and therapeutic interventions by using targeted mass spectrometry and machine learning approaches.

Methods

Serum samples collected from a total of 107 individuals, including 55 individuals diagnosed with AD and 52 healthy controls (HC) enrolled previously to ADDIA cohort were analyzed using the biocrates AbsoluteIDQ^® p400 HR kit metabolite and lipid panel. Several machine learning models including Least Absolute Shrinkage and Selection Operator (LASSO), Partial Least Squares (PLS), Random Forest, and XGBoost were trained to classify AD and HC. Repeated cross-validation was used to ensure performance evaluation.

Results

The LASSO and PLS models showed the strongest classification performance on the test set, achieving area under the ROC curve (AUC) values of 0.84 and 0.90, respectively. A refined model based on only the top 5 metabolites maintained strong performance, and the inclusion of Apolipoprotein E (APOE) genotype information notably improved classification accuracy, particularly by reducing false negatives in AD cases.

Conclusions

These results highlight important metabolic signatures that could help to reduce misdiagnosis and support the development of metabolomic panels to detect AD. The combination of multiple serum metabolic biomarkers and APOE genotyping can significantly improve classification accuracy and potentially assist in making non-invasive, cost-effective diagnostic approach.

Introduction

Alzheimer's disease (AD) is a progressive neurodegenerative disease and represents the most common cause of neurocognitive disorder worldwide. ¹ AD is clinically manifested by a gradual decline in memory and cognitive function that becomes more pronounced as the disease progresses. ² At the down of anti-amyloid immunotherapies and more than ever before, the early detection of AD is crucial for timely intervention and effective treatment of the disease. However, due to the heterogeneity of disease progression, early detection of AD remains difficult. ³ To date, most diagnostic methods are based on clinical assessment, neuroimaging and cerebrospinal fluid (CSF) analysis, which are either invasive or costly, limiting their routine use.^4,5 Blood-based biomarkers have been recently developed with encouraging performances. ⁶

First introduced in 1998, ⁷ metabolomics is a rapidly evolving discipline providing additional insights into the study of biochemical processes and disease mechanisms. By profiling metabolites in body fluids such as blood, urine, or CSF, metabolic changes associated with diseases such as AD can be identified. The appeal of blood metabolomics lies in its minimal invasiveness and its potential for early detection of disease. The changes in metabolic profile can occur up to 25 years before the clinical symptoms of AD. ⁸ In the context of AD, metabolomics holds promise not only for diagnosis but also for improving clinical trials. The current low success rate in AD drug development is often due to late-stage interventions. ⁹ Identifying reliable metabolomic biomarkers and understanding their changes could facilitate early diagnosis and allow preventive and therapeutic interventions to be delivered at a stage where they are more likely to be effective.

Given the need for a non-invasive, cost-effective and scalable approach, serum derived from peripheral blood was used for this study as it combines accessibility with a greater likelihood of finding pathophysiological biomarkers in AD.^10,11 Several studies have shown that the use of metabolomics has already enabled the detection of subtle, disease-specific patterns in serum that reflect underlying AD-related metabolic changes.^12–14 Although the concentrations of blood biomarkers are lower due to the blood-brain barrier, recent advances in metabolomic profiling enable the detection of significant disease-related metabolic changes. ¹⁴ Furthermore, machine learning models, which are useful for dealing with complex datasets, have shown great potential for improving the accuracy of blood biomarker-based diagnostics.^15–17

The aim of this study was to identify serum biomarkers, particularly metabolites, detect and differentiate AD from healthy control (HC) using an approach that integrates metabolomic analysis and machine learning techniques.

Methods

Sampling and laboratory analysis

The serum samples from ADDIA cohort (NCT03030586) were analyzed using the AbsoluteIDQ^® p400 HR kit (Biocrates Life Sciences AG). The extraction was conducted with a loading volume of 10 µL per sample and analyzed with Liquid Chromatography coupled to tandem Mass Spectrometry and Flow injection analysis tandem mass spectrometry. The data underwent quality control, statistical analysis, and machine learning (Least Absolute Shrinkage and Selection Operator (LASSO), Partial Least Squares (PLS), Random Forest, eXtreme Gradient Boosting (XGBoost), Naïve Bayes) to identify metabolomic signatures for AD classification (Figure 1).

Figure 1.

Targeted metabolomics workflow for identifying metabolic patterns in AD. AD: Alzheimer's disease; HC: Healthy controls; FIA-MS/MS: Flow injection analysis tandem mass spectrometry; LC-MS/MS: Liquid Chromatography coupled to tandem Mass Spectrometry; QC: Quality control.

Study participants and data retrieval

Biological samples were collected from 107 individuals and analyzed for 408 quantified metabolites. MetIDQ™ software, which is an integral part of the AbsoluteIDQ^® p400 HR kit, was used for peak selection, identification, and quantitative assessment. The concentrations of metabolites subjected to Flow injection analysis tandem mass spectrometry were automatically calculated by the software. The analyte peaks detected by Liquid Chromatography coupled to tandem Mass Spectrometry were integrated with Thermo Xcalibur software version 4.3. Concentrations [µmol/l] were determined from the calibration standards using the appropriate quantification method provided in the kit. This approach facilitated the retention of metabolites that exceeded the detection limit but fell below the quantification limit, as specified by the Biocrates software. All retrieved data were loaded, quality-assessed, data processing and fully analyzed using R and RStudio (programming language GNU R, version 4.4.1; RStudio version 2024.9.0.375).

Apolipoprotein E (APOE) genotyping

The APOE genotyping was performed for each sample using the CE-IVDR APO-Easy^® kits, providing accurate determination of the 6 APOE genotypes. ¹⁸

Data preparation and quality control

Missing values below the Limit of Detection were imputed using a logspline algorithm, and metabolites with >20% missing values were excluded. ¹⁹ Metabolite concentrations were normalized to Biocrates quality control level 2 samples, as recommended by the manufacturer. To account for potential confounding factors, linear regression was applied to correct for batch effects and biological variation. ²⁰

Statistical analysis and modeling

Patient metadata were summarized using descriptive statistics. Differential metabolite levels were assessed by Student's t-test, with log₂ fold changes, and p-values adjusted for false discovery rate (FDR) using the Benjamini–Hochberg method. ²¹ Metabolite discriminatory power was evaluated using the area under the curve (AUC).^22,23 Furthermore, Metabolites showing statistically significant differences between groups (p-value < 0.05; AUC > 0.60) were examined with Pearson correlation analysis ²⁴ and selected for enrichment pathway analysis.

Metabolite set enrichment analysis (MSEA)

To identify dysregulated metabolic pathways in AD, we performed overrepresentation analysis (ORA) of differentially expressed metabolites using MetaboAnalystR (v6.0). ²⁵

Machine learning (ML) modeling

The dataset was partitioned into a training set (70%) and a held-out test set (30%) using stratified sampling. Feature selection was conducted only on the training set using LASSO regression. Using the training set, we trained five classifiers: LASSO, Random Forest, Naïve Bayes, Partial Least Squares, and XGBoost.

Model hyperparameters were tuned via repeated stratified 5-fold cross-validation (20 repeats) on the training set, with fixed random seeds. The test set remained entirely unused during this phase.

The best performing model was refined by retaining only the top-ranked features based on variable importance scores. To assess the additive value of APOE, we retrained the model with and without APOE genotype, incorporating published odds ratios as prior knowledge reflecting the relative risk of developing AD compared to the reference genotype ε3/ε3.^26,27

The final model was evaluated exclusively on the held-out test set, which was completely unseen data and never involved in any training or selection steps.

Results

Demographic and clinical characteristics

The study included 55 AD patients and 52 HC from the ADDIA cohort. AD patients were significantly younger than HC, while sex distribution was balanced between groups (Table 1).

Table 1.

Demographical characteristics of participants.

Variable	AD (n = 55)	HC (n = 52)	Total (n = 107)	t/x2, p
Age
Mean ± SD	75 ± 7.9	78 ± 4.3	76 ± 6.6	−2.824, 0.005
Median	76	78	78
Q1–Q3	71–82	76–80	73–81
min–max	55–86	64–87	55–87
Sex (n)
Female	30 (55%)	29 (56%)	59 (55%)	0.016, 0.899
Male	25 (45%)	23 (44%)	48 (45%)
APOE
ε3/ε4	29 (53%)	4 (8%)	33 (31%)	48.94, < 0.001
ε4/ε4	12 (22%)	1 (2%)	13 (12%)
ε2/ε4	1 (2%)	0 (0%)	1 (1%)
ε3/ε3	12 (22%)	39 (75%)	51 (48%)
ε2/ε3	1 (2%)	8 (15%)	9 (8%)

AD: Alzheimer's disease; HC: healthy controls; SD: Standard deviation; APOE: Apolipoprotein E genotypes. p-value was calculated based on Welch's two-sample t-test or Pearson's chi-squared test.

APOE genotyping results

APOE genotype distribution differed significantly between AD and HC. High-risk genotypes were more prevalent in AD patients: ε3/ε4 and ε4/ε4. In contrast, the HC group showed higher frequency of the ε3/ε3 genotype and ε2/ε3 genotype. The ε2/ε4 genotype was rare (in AD only), and no ε2/ε2 genotypes were observed (Table 1).

Targeted metabolomics results

Metabolomic profiling revealed 22 metabolites exhibiting nominally significant differences (p < 0.05) between AD and HC. Of these, 18 metabolites demonstrated an AUC greater than 0.60 (Supplemental Table 1). Notably, PC(30:3), TG(55:8), and LPC(24:0) showed higher discriminatory potential. However, following correction for multiple testing, none of these metabolites retained statistical significance (q < 0.1; Supplemental Figure 2).

Correlations between significant metabolites and their classes

Metabolites within the same class tended to cluster together in the correlation network (Supplemental Figure 3A). Phosphatidylcholines, acylcarnitines, and sphingomyelins were the most abundant classes (Supplemental Figure 3B).

Enrichment analysis results

Pathway analysis identified six enriched metabolite sets associated with distinct biological pathways; however, none remained statistically significant after multiple testing correction (q < 0.1; Supplemental Figure 4, Supplemental Table 4).

ML model performance

Two feature sets were evaluated: (1) 57 metabolites derived from an initial panel of 400, and (2) a refined set of the top 5 metabolites, with and without APOE genotyping. Among tested algorithms, LASSO and PLS showed the strongest discriminative performance on the test set (Figure 2, Supplemental Tables 5 and 6).

Figure 2.

ROC curves of five machine learning models (ML) for distinguishing AD from HC in the test set. Models were trained on 57 serum metabolites. Curves represent Partial Least Squares (PLS, red), LASSO (olive-green), Random Forest (green), XGBoost (blue), and Naïve Bayes (purple).

Analysis of the feature importance revealed several high-ranking metabolites that contributed to the LASSO model prediction, with threonine (Thr) consistently among the top features (Figure 3). A refine model incorporating the top 5 metabolites further demonstrated that APOE inclusion enhanced classification accuracy, particularly in reducing AD misdiagnoses (Figure 4).

Figure 3.

Top 25 metabolites ranked by LASSO importance scores for distinguishing between AD and HC groups. The importance score represents the relative contribution of each metabolite in distinguishing between study groups based on LASSO model coefficients. Higher scores indicate greater relevance for classification. Metabolite abbreviations: Thr, threonine; Asp, aspartic acid; PC(x:y), phosphatidylcholine with x carbon atoms and y double bonds; PC-O(x:y), ether-linked (alkyl-acyl) phosphatidylcholine; SM(x:y), sphingomyelin; LPC(18:2), lysophosphatidylcholine; AC(x:y), acylcarnitine; AC(4:0-OH), hydroxybutyrylcarnitine; AC(16:1-OH), hydroxyhexadecenoylcarnitine; TG(x:y), triglyceride.

Figure 4.

Performance comparison of LASSO models using the top 5 metabolites for distinguishing AD from HC, with and without inclusion of APOE genotyping as a feature. (A) Confusion matrix for the model excluding APOE. (B) Confusion matrix for the model including APOE. The results are shown for the test set.

Discussion

Our metabolomics-based machine learning approach demonstrates competitive performance compared to established diagnostic methods. The combination of targeted metabolomics with machine learning algorithms and APOE genotyping achieves diagnostic accuracies comparable to current gold standard biomarkers, while offering practical advantages for clinical implementation.

The classification performance of our study aligns closely with other blood-based biomarker studies. The LASSO model achieved an AUC 0.84 of while PLS reached 0.90, performance metrics that are competitive with recent blood biomarker research. These results compare favorably to plasma p-tau217 biomarkers, which have demonstrated AUCs between 0.91 to 0.96 for detecting AD pathology,^28–30 though our approach provides comparable diagnostic capability using a different molecular class of biomarkers.

Recent large-scale studies of blood biomarkers show similar performance ranges. For instance, a comprehensive study by Stamate et al. using plasma metabolites achieved AUCs of 0.85–0.88 using machine learning approaches including XGBoost and deep learning. ³¹ Our metabolomics panel performance of AUC 0.90 demonstrates superior diagnostic capability within this established range. Notably, studies combining multiple metabolites typically achieve better performance than single biomarkers, supporting our multi-analyte approach.^32,33

Our classification accuracy compares reasonably with CSF biomarker performance, though CSF remains the current diagnostic standard. CSF biomarkers typically achieve sensitivities of 80–97% and specificities of 70–83% for AD diagnosis.^34–36 The sensitivity from our study range of 70.59–82.35% and specificity up to 87.5% (PLS model) positions our approach within the lower-to-middle range of CSF biomarker performance, which is notable given the non-invasive nature of serum sampling.

The CSF Aβ₄₂/tau ratios, considered among the most robust CSF measures, demonstrate AUCs of 0.93–0.94 for identifying AD. ³⁴ While our metabolomics approach achieved a lower AUC of 0.90, this represents competitive performance considering the accessibility advantages of serum-based testing over lumbar puncture procedures.

Amyloid positron emission tomography (PET) imaging, the current gold standard for detecting amyloid pathology, demonstrates high diagnostic accuracy with sensitivities of 88–98% and specificities of 80–95%.^35,36 Our metabolomics approach, while achieving lower sensitivity, offers practical advantages including lower cost, greater accessibility, and the absence of radiation exposure. The potential for the serum-based approach to serve as an initial screening tool before confirmatory PET imaging represents a valuable clinical utility.

Studies directly comparing blood biomarkers to amyloid PET show that top-performing blood tests can achieve comparable diagnostic accuracy.^28,30 The performance of our metabolomics panel suggests it could serve a similar screening role, particularly when enhanced with APOE genotyping.

Our results must be interpreted within the context of current clinical diagnostic limitations. Primary care physicians achieve only 61% diagnostic accuracy for AD using standard clinical evaluation, while dementia specialists reach 73% accuracy. ³⁷ Our metabolomics approach, with accuracies of 72–82%, substantially outperforms primary care diagnosis and approaches specialist-level performance.

The clinical misdiagnosis rates for AD range from 17–30% in specialist settings, with sensitivity ranging from 70.9–87.3% and specificity from 44.3–70.8%. ³⁸ Our approach demonstrates competitive performance within these ranges while providing objective, reproducible measurements that could reduce diagnostic variability.

Studies combining metabolomics with genetic markers consistently show improved performance over metabolomics approach alone, supporting our integrated strategy. ³⁹ The inclusion of APOE genotyping improved the AUC of our refined model from 0.62 to 0.74, representing a clinically significant enhancement. Although this remains a modest improvement, it aligns with established literature showing APOE ε4 as a strong genetic risk factor with positive predictive values of 77–100% for AD diagnosis.^40,41 However, the trade-off between sensitivity and specificity observed in our study reflects known challenges with APOE testing, where ε4 carriership can lead to false positives, particularly in certain populations. ³⁵

The identified metabolite patterns align well with established AD pathophysiology. The dysregulation of phosphatidylcholines, sphingomyelins, and triglycerides observed in our study is consistent with extensive lipidomic research in AD.^42–44 Specifically, our finding of altered PC(30:3) and LPC(24:0) levels corresponds with previous studies showing phosphatidylcholine species changes that achieve AUCs of 0.786–0.828 for AD detection. ⁴²

The decreased threonine levels in AD cohort align with amino acid studies showing altered levels in cognitive impairment.^45,46 While threonine itself has not been extensively studied as an AD biomarker, its connection to neurotransmitter synthesis and the finding of phospho-tau at threonine 231 as a CSF biomarker suggests biological relevance to AD pathophysiology. ⁴⁷

Our choice of LASSO and PLS algorithms demonstrates competitive performance compared to other machine learning approaches in metabolomics. Recent studies comparing multiple algorithms for metabolomics data show similar performance ranges, with XGBoost, random forest, and deep learning achieving AUCs of 0.85–0.88.^31,48 The feature selection capabilities of LASSO provide interpretability advantages over black box approaches while maintaining competitive diagnostic performance.

However, the constraints of serum metabolomics for clinical use must be acknowledged. The lower sensitivity compared to CSF and PET biomarkers may result in missed cases, particularly in early-stage disease. In addition, the complexity of metabolomic analysis requires specialized laboratory infrastructure that may limit immediate widespread adoption.

Limitations and future directions

Despite its strengths, our study has certain limitations. Firstly, the sample size remains moderate, which underscores the need for further validation in larger independent cohorts to ensure the robustness and generalizability of the findings. Secondly, while targeted metabolomics provides precise quantification of specific metabolites, untargeted metabolomics could help to uncover additional AD-related metabolic pathways, thus broadening the scope of our analysis. Lastly, longitudinal studies tracking metabolite concentration changes over time would be essential to improve our understanding of both disease classification and progression, enhancing the predictive power of biomarkers.

Future research integrating larger datasets from independent cohorts and advanced analytical approaches will be essential to translate and validate these findings into clinically actionable tools for complementing the accurate AD detection and monitoring. Examining metabolite ratios, thresholds, or overall patterns, rather than isolated compounds, will be considered to provide a deeper understanding of AD-related metabolomic changes. An increased focus will be placed on feature selection strategies to mitigate misclassification while maintaining the benefits of APOE integration to pave the way for more precise and reliable diagnostic tools in clinical settings. In closing, though traditional statistical methods have their merits, our study demonstrated the importance of machine learning in dealing with the complexities of AD diagnostics, preparing the way for a new era in diagnostic strategies.

Conclusion

These preliminary results emphasize the effectiveness of metabolomic serum biomarkers in combination with machine learning models for the discovery of non-invasive biomarkers to better distinguish AD from HC. This study demonstrates distinct metabolomic profiles across AD and HC, particularly emphasizing lipid metabolism, mitochondrial function, and neurotransmitter regulation as essential components of neurodegenerative pathology, providing insights into potential metabolic changes underlying AD. Our findings also highlight key metabolic disturbances in AD, particularly in lipid, sphingomyelin, and amino acid metabolism. These alterations provide valuable insights into AD pathophysiology, offering potential biomarkers for early diagnosis that could ultimately improve clinical outcomes and treatment strategies. Furthermore, integrating genetic risk factors (APOE) enhances classification accuracy, suggesting that multi-biomarker approaches could greatly improve AD diagnostics and risk stratification. This synergy between genetic and metabolic markers underscores the multifactorial nature of AD, showing the advantages of combining these approaches in enhancing diagnostic accuracy and providing a more thorough understanding of the disease.