3.1. Characteristics of study participants
The biochemical parameters, clinical test results for cognitive performances, corresponding clinical parameters, and SIMOA measurements for both study groups have been summarised in Table 1. Briefly, the majority of the biochemical measurements were within the standard reference intervals. A few, but significant differences were also observed between the AD patients and cognitively healthy individuals, including a slightly higher age (p = 0.00001), higher LDH levels (p = 0.03), and lower glucose levels in the AD patient group (p = 0.01). Patients who required additional cognitive testing and paraclinical measurements were identified, where AD patients presented with low MMSE (20.0 ± 4.5) and ACE (58.0 ± 16.5) scores and a high FAQ (11.8 ± 6.2) score, whereas paraclinical tests demonstrated elevated levels of CSF tau, p-tau (81.7 ± 25.0 ng/L) and t-tau (520.4 ± 102.4 ng/L), and decreased levels of CSF Aβ (682.8 ± 216.3 ng/L) for some of the patients, indicating extracellular tau accumulation and intracellular Aβ build-up. Plasma measurements of markers for neuronal injury and AD hallmark proteins were included as additional clinical information. Generally, AD patients had significantly higher plasma levels of Aβ40 (p = 0.002), GFAP (p = 0.001), Nf-L (p = 0.00001), and p-tau181 (p = 0.00005) than healthy individuals; however, Aβ42 did not differ between the two groups (p = 0.5).
Table 1
Characteristics of study participants.
| Units | Con (n = 25) | AD (n = 25) | p-value | Reference interval |
Mean ± SD | Mean ± SD |
Demographics |
Age | Years | 66.6 ± 1.3 | 75.7 ± 8.2 | 0.00001 | - |
Male/female | n | 16 / 9 | 15 / 10 | - | - |
Biochemical measurements |
ALAT | U/L | 26.3 ± 8.6 | 22.3 ± 11.6 | 0.17 | 10.0–50.0 |
Albumin | g/L | 41.0 ± 1.9 | 41.5 ± 1.9 | 0.37 | 34–45 |
Carbamide | mmol/L | 5.8 ± 1.3 | 5.7 ± 1.5 | 0.77 | 3.1–8.1 |
Cholesterol | mmol/L | 5.4 ± 0.9 | 5.5 ± 1.1 | 0.88 | 4.2–8.5 |
Creatinine | µmol/L | 79.0 ± 10.2 | 83.4 ± 14.5 | 0.22 | 45–105 |
CRP | mg/L | 1.9 ± 1.4 | 2.2 ± 2.9 | 0.57 | < 8 |
Glucose | mmol/L | 6.4 ± 1.7 | 5.4 ± 0.9 | 0.01 | 4.2–7.8 |
Haemoglobin | mmol/L | 8.8 ± 0.7 | 8.5 ± 1.0 (n = 15) | 0.45 | 7.3–10.5 |
HDL | mmol/L | 1.5 ± 0.3 | 1.6 ± 0.4 | 0.35 | 0.7–1.9 |
LDL | mmol/L | 3.2 ± 0.8 | 3.3 ± 0.9 | 0.71 | 2.2–5.7 |
LDH | U/L | 170.2 ± 31.2 | 192.1 ± 38.7 | 0.03 | 105–255 |
Triglycerides | mmol/L | 1.5 ± 0.8 | 1.3 ± 0.8 | 0.34 | 0.6–3.9 |
Clinical parameters |
MMSE | - | - | 20.0 ± 4.5 | - | - |
ACE | - | - | 58.0 ± 16.5 (n = 21) | - | - |
FAQ | - | - | 11.8 ± 6.2 (n = 21) | - | - |
CSF Aβ | ng/L | - | 682.8 ± 216.3 (n = 9) | - | > 500 |
CSF p-tau | ng/L | - | 81.7 ± 25.0 (n = 9) | - | < 61 |
CSF t-tau | ng/L | - | 520.4 ± 102.4 (n = 9) | - | < 450 |
SIMOA |
Aβ40 | pg/mL | 95.1 ± 10.2 | 108.7 ± 17.4 | 0.002 | - |
Aβ42 | pg/mL | 5.3 ± 1.0 | 5.6 ± 1.3 | 0.5 | - |
GFAP | pg/mL | 88.6 ± 32.8 | 247.1 ± 277.9 | 0.001 | - |
Nf-L | pg/mL | 12.5 ± 4.4 | 36.9 ± 24.5 | 0.00001 | - |
p-tau181 | pg/mL | 1.8 ± 0.8 | 3.1 ± 1.3 | 0.00005 | - |
Demographics data of study participants together with biochemical measurements, cognitive test results, paraclinical measurements, and SIMOA measurements. Abbreviations; Aβ – Amyloid-β, ACE – Addenbrooke’s Cognitive Examination, AD – Alzheimer’s Disease, ALAT – Alanine transaminase, p-tau – Phosphorylated tau, CRP – C-reactive protein, CSF – Cerebrospinal fluid, FAQ – Functional Activities Questionnaire, GFAP – Glial fibrillary acidic protein, HDL – High-density lipoprotein, LDH – Lactate dehydrogenase, LDL – Low-density protein, MMSE – Mini-Mental State Examination, Nf-L – Neurofilament light, SD – Standard deviation, SIMOA – Single molecule array, t-tau – Total tau.
3.2. Validation of metabolic signatures for Alzheimer’s Disease diagnostics
To validate the metabolic signature identified in our discovery study, NMR spectroscopy was applied to measure the concentration of serum metabolites in our validation study cohort. Three prediction models were tested for their performance based on AUC, accuracy, PPV, and NPV. These models included sPLS-DA, random forest, and XGBoost (Table 2). Based on these criteria, sPLS-DA showed the highest performance with five selected metabolites building the model (pyruvic acid, valine, histidine, isoleucine, and glutamine), while random forest performed the second best with four selected metabolites (histidine, valine, pyruvic acid, and glutamine), and lastly the XGBoost with three metabolites (histidine, pyruvic acid, and valine). Thus, sPLS-DA was selected as our data's most optimal validation model.
Table 2
Performance of prediction models.
Model | AUC | 95% CI | Accuracy | PPV | NPV | Features |
sPLS-DA | 0.89 | 0.79–0.98 | 0.86 | 0.88 | 0.85 | 5 |
Random Forest | 0.87 | 0.77–0.97 | 0.76 | 0.84 | 0.71 | 4 |
XGBoost | 0.84 | 0.73–0.95 | 0.74 | 0.72 | 0.76 | 3 |
Three validation models and their diagnostic performance; sparse-partial least squared discriminant analysis, random forest, and extreme gradient boosting. Abbreviations; AUC – Area under the curve, CI – Confidence interval, NPV – Negative predictive value, PPV – Positive predictive value, sPLS-DA – Sparse-partial least squared discriminant analysis, XGBoost – Extreme gradient boosting.
The validation model showed a small overlap between the patient and control groups, as seen in the scores plot of the measured serum samples (Fig. 1A). Based on the validated model, five metabolites significantly contribute to sample grouping, accounting for 44% of the group variation (Fig. 1B). Consequently, the model had an AUC performance of 0.89 (95% CI = 0.79–0.98) for discriminating AD patients from cognitively healthy individuals (Fig. 1C). Furthermore, the model had an accuracy = 0.86, PPV = 0.88, and NPV = 0.85, indicating its diagnostic value. Interestingly, when adding the significantly altered proteins (Aβ40, GFAP, Nf-L, and p-tau181) to the validation model, improved its diagnostic performance, resulting in an AUC of 0.97 (95% CI = 0.93–1.00) with an accuracy of 0.94, PPV of 0.96, and NPV of 0.92 (Fig. 1D).
Furthermore, the selected panel of five metabolites was correlated against the clinical data and markers of neuronal damage to determine their possible association with neurodegenerative diseases (Fig. 2). Most metabolites exhibited a negative correlation to Nf-L and p-tau181, with valine showing a moderate Pearson’s correlation of -0.51 and − 0.48, respectively. Pyruvic acid and isoleucine also showed moderate correlations with cognitive scoring tests FAQ (ρ = -0.5) and MMSE (ρ = 0.43).
Five of the metabolites measured in both the discovery and validation data sets were significantly altered when comparing healthy and diseased individuals (Table 3). These five metabolites exhibited identical changes in both the discovery and validation studies.
Table 3
Common significantly altered metabolites.
Discovery study |
Metabolite [mmol/L] | Con | AD | FC | p-value | FDR |
Mean | SD | Mean | SD |
Pyruvic acid | 0.032 | 0.007 | 0.026 | 0.004 | -0.2 | 0.03 | 0.14 |
Valine | 0.118 | 0.019 | 0.092 | 0.011 | -0.2 | 0.002 | 0.01 |
Leucine | 0.149 | 0.039 | 0.119 | 0.016 | -0.2 | 0.049 | 0.15 |
Histidine | 0.037 | 0.002 | 0.032 | 0.002 | -0.1 | 0.0002 | 0.003 |
Isoleucine | 0.045 | 0.009 | 0.037 | 0.006 | -0.2 | 0.03 | 0.14 |
Validation study |
Pyruvic acid | 0.118 | 0.031 | 0.079 | 0.027 | -0.3 | 0.00002 | 0.0001 |
Valine | 0.275 | 0.052 | 0.214 | 0.040 | -0.2 | 0.00003 | 0.0001 |
Leucine | 0.111 | 0.035 | 0.078 | 0.017 | -0.3 | 0.0001 | 0.0005 |
Histidine | 0.120 | 0.024 | 0.096 | 0.023 | -0.2 | 0.0006 | 0.002 |
Isoleucine | 0.065 | 0.026 | 0.049 | 0.016 | -0.2 | 0.02 | 0.03 |
Five metabolites were dysregulated in serum samples between cognitively affected and healthy individuals, sorted according to p-valuein from the validation cohort. Abbreviations; AD – Alzheimer’s Disease, Con – Healthy controls, FC – Fold change, SD – Standard deviation.
3.3. Metabolic alterations in the validation cohort
To extrapolate novel metabolic information, serum samples from the validation study were examined for significantly altered metabolites between the groups. This brought the total number of significantly different metabolites between the groups to sixteen, with fourteen of these being significant after FDR correction. Eleven metabolites were previously identified in the discovery cohort and the remaning five were novel metabolites not previously identified; 3-hydroxybutyric acid, citric acid, lactic acid, lysine, and succinic acid (Table 4).
Table 4
Significantly altered metabolites in the validation cohort.
Metabolite [mmol/L] | Con | AD | FC | p-value | FDR |
Mean | SD | Mean | SD |
Tyrosine | 0.07 | 0.02 | 0.05 | 0.01 | -0.3 | 0.000006 | 0.0002 |
Pyruvic acid | 0.12 | 0.03 | 0.08 | 0.03 | -0.3 | 0.00002 | 0.0002 |
Valine | 0.27 | 0.05 | 0.21 | 0.04 | -0.2 | 0.00003 | 0.0003 |
Leucine | 0.11 | 0.03 | 0.08 | 0.02 | -0.3 | 0.0001 | 0.001 |
Lysine | 0.22 | 0.04 | 0.17 | 0.04 | -0.2 | 0.0002 | 0.001 |
Histidine | 0.12 | 0.02 | 0.10 | 0.02 | -0.2 | 0.0006 | 0.003 |
Glycerol | 0.19 | 0.05 | 0.28 | 0.11 | 0.5 | 0.001 | 0.007 |
3-Hydroxybutyric acid | 0.07 | 0.07 | 0.16 | 0.14 | 1.3 | 0.008 | 0.03 |
Acetoacetic acid | 0.01 | 0.01 | 0.02 | 0.02 | 1.5 | 0.01 | 0.03 |
Alanine | 0.50 | 0.07 | 0.44 | 0.09 | -0.1 | 0.01 | 0.03 |
Glucose | 6.68 | 1.88 | 5.55 | 1.05 | -0.2 | 0.01 | 0.03 |
Citric acid | 0.12 | 0.03 | 0.14 | 0.04 | 0.2 | 0.02 | 0.04 |
Isoleucine | 0.07 | 0.03 | 0.05 | 0.02 | -0.2 | 0.02 | 0.04 |
Succinic acid | 0.003 | 0.002 | 0.006 | 0.006 | 1.1 | 0.02 | 0.046 |
Phenylalanine | 0.06 | 0.01 | 0.05 | 0.01 | -0.1 | 0.04 | 0.07 |
Lactic acid | 2.26 | 0.54 | 1.96 | 0.52 | -0.1 | 0.049 | 0.1 |
Significantly altered metabolites measured in serum samples comparing cognitively affected with healthy individuals, sorted according to the p-value. Abbreviations; AD – Alzheimer’s Disease, Con – Healthy controls, FC – Fold change, FDR – False-discovery rate, SD – Standard deviation.
A network analysis was performed to investigate biological pathways for the significantly altered metabolites related to cognitive impairment (Fig. 3). Acetoacetate, a substrate for the TCA cycle (log2 FC = 1.4, p-value = 0.008), was the most elevated metabolite in AD. In contrast, pyruvate involved in glycolysis and gluconeogenesis (log2 FC = -0.6, p-value = 0.00002) was the most reduced metabolite in relation to AD. In addition, the metabolic pathways of Biopterin, Glycerophospholipid, Histidine, Lysine, and branch-chained amino acids (BCAAs); valine, leucine, and isoleucine, were also modified. This validation study identified and confirmed changes in histidine and BCAA metabolisms previously found in the discovery study; histidine (log2 FC = -0.3, p-value = 0.0006), isoleucine (log2 FC = -0.4, p-value = 0.02), leucine (log2 FC = -0.5, p-value = 0.00008), and valine (log2 FC = -0.4, p-value = 0.00002).