Peripheral serum metabolomic profiles inform central cognitive impairment

The incidence of Alzheimer's disease (AD) increases with age and is becoming a significant cause of worldwide morbidity and mortality. However, the metabolic perturbation behind the onset of AD remains unclear. In this study, we performed metabolite profiling in both brain (n = 109) and matching serum samples (n = 566) to identify differentially expressed metabolites and metabolic pathways associated with neuropathology and cognitive performance and to identify individuals at high risk of developing cognitive impairment. The abundances of 6 metabolites, glycolithocholate (GLCA), petroselinic acid, linoleic acid, myristic acid, palmitic acid, palmitoleic acid and the deoxycholate/cholate (DCA/CA) ratio, along with the dysregulation scores of 3 metabolic pathways, primary bile acid biosynthesis, fatty acid biosynthesis, and biosynthesis of unsaturated fatty acids showed significant differences across both brain and serum diagnostic groups (P-value < 0.05). Significant associations were observed between the levels of differential metabolites/pathways and cognitive performance, neurofibrillary tangles, and neuritic plaque burden. Metabolites abundances and personalized metabolic pathways scores were used to derive machine learning models, respectively, that could be used to differentiate cognitively impaired persons from those without cognitive impairment (median area under the receiver operating characteristic curve (AUC) = 0.772 for the metabolite level model; median AUC = 0.731 for the pathway level model). Utilizing these two models on the entire baseline control group, we identified those who experienced cognitive decline in the later years (AUC = 0.804, sensitivity = 0.722, specificity = 0.749 for the metabolite level model; AUC = 0.778, sensitivity = 0.633, specificity = 0.825 for the pathway level model) and demonstrated their pre-AD onset prediction potentials. Our study provides a proof-of-concept that it is possible to discriminate antecedent cognitive impairment in older adults before the onset of overt clinical symptoms using metabolomics. Our findings, if validated in future studies, could enable the earlier detection and intervention of cognitive impairment that may halt its progression.


Checklist of Supporting Information
. Brain metabolic pathways and serum metabolic pathways composition and alterations Figure S2. PDS of metabolic pathways across Braak scores in brain Figure S3. PDS of metabolic pathways across CERAD scores in brain      Table S1. List of cognitive performance tests Table S2. Associations between identified metabolites/ratio and cognitive performance tests adjusting for age, gender, years of education, APOE ε4, BMI Table S3. Logistic regression of metabolite marker panel-based RF score to discriminate NCI (converters) vs. NCI (non-converters) adjusting for gender, years of education, APOE ε4, and BMI   Table S4. PDS of metabolic pathways differentially expressed in participants with cognitive decline Table S5. Associations between identified metabolic pathways and cognitive performance tests adjusting for age, gender, years of education, APOE ε4, and BMI Table S6. Logistic regression of metabolic pathway panel-based RF score to discriminate NCI (converters) vs. NCI (non-converters) adjusting for gender, years of education, APOE ε4, and BMI Table S7. Levels of detected metabolites across clinical groups in brain samples Table S8. Levels of detected metabolites across clinical groups in serum samples Table S9. PDS of mapped metabolic pathways across clinical groups in brain samples Table S10. PDS of detected metabolic pathways across clinical groups in serum samples Table S11. Levels of identified metabolites in samples with both brain and blood metabolomics data Table S12. Mixed effects model of metabolite marker panel-based RF score adjusting for age, gender, years of education, APOE ε4, and BMI Table S13. Associations between identified metabolites/ratio and cognitive performance domains adjusting for age, gender, years of education, APOE ε4, and BMI. Table S14. P values and Q values of identified metabolites across clinical groups in brain samples using ordinal regression. Table S15. P values and Q values of identified metabolites across clinical groups in serum samples using logistic regression. Table S16. P values and Q values of identified pathway across clinical groups in brain samples using ordinal regression. Table S17. P values and Q values of identified pathway across clinical groups in serum samples using logistic regression. Table S18. P values identified pathway across clinical groups in serum samples adjusting for other potential confounders using logistic regression.

Serum sample preparation
We mixed an aliquot of 50 μL serum with 150 µL methanol, and vortexed the mixture for 2 min , let it stand for 10 min and centrifuged at 4 °C for 10 min. We then transferred 160 µL of the supernatant to a clean tube and vacuum dried the remaining aliquot re-dissolved it with a matched amount of acetonitrile (0.1% formic acid) and added water (0.1% formic acid) to a volume of 40 μL. The supernatant after the centrifugation was used for UPLC-TQMS and GC-TOFMS analysis. A mixture of 20 μL from the final supernatant of each sample was prepared for use as pooled quality control (QC) samples.

Brain sample preparation
We weighted 30 mg brain tissue, homogenized it with 75 µL of 50% precooled methanol using a Bullet Blender Tissue Homogenizer (Next Advance, Inc., Averill Park, NY) for 3 min. After centrifugation at 4 °C for 15 min, we transferred the supernatant to a clean tube. We performed the second step extraction by adding precooled methanol and chloroform mixture (3:1) to the residue followed by centrifugation. We combined the supernatant with the previous one and vortexed the mixture for 5 min and performed centrifugation for 15 min. The supernatant was used for UPLC-TQMS and GC-TOFMS analysis. A mixture of 20 μL from the final supernatant of each sample was prepared for use as pooled QC samples.

Quality control Procedure
Previously prepared QC samples were run between every ten sample injections. For each metabolite in QC samples, we calculated the relative standard deviations (RSDs), which was less than 15% for batches of samples.       Individual bile acid was adjusted by the total bile acids concentration (i.e., % of total bile acids).
All abundance values were log10 transformed.