These results show that in analyzing CSF, mapping the concentrations of small molecules in the LC-ECA platform produces initial prediction levels similar to those produced by well validated CSF neuropathology-related protein biomarkers for discriminating between clinically diagnosed AD and cognitively normal participants. Further, our data suggests that metabolomics markers alone may offer relatively high discriminatory ability to distinguish the clinical AD phenotype from controls, and the lack of correlation with MMSE suggests the markers may have potential applicability also in mild stages where there is relatively high need for accurate markers.
The current sample size is limited in its ability to truly build and evaluate potential predictive models without further evaluation and validation in larger studies; however, it is interesting that a model from a predictive resampling scheme with only 2 metabolites could be built that has high predictive potential. These results were found with a conservative screening approach that removed metabolites that are potential confounders due to differential drug response and also ApoE status.
While the potential predictive performance of the 15-65.533 and 8-93.65 metabolites is promising, their unknown identity makes biological interpretation difficult. This limits current interpretability, but it does demonstrate the potential of new, high-throughput technologies to identify novel metabolite markers that may be of biological and clinical relevance. Studies are in progress to further characterize and identify them, from both a bioinformatics perspective and a biochemical perspective. This will be helpful in enhancing our understanding of pathophysiologic mechanisms and the interrelationships between small molecules and AD-associated proteins.
Initial analyses to explore correlations with 15-65.533 and 8-93.65 show that they are significantly associated with several known metabolites that may help reveal their biological relevance (Additional file 6: Table S1). The correlation structure and pathway analyses of the electrochemistry metabolites collected in the current data are thoroughly discussed in a previous manuscript . Correlations of the two unknowns with other metabolites in the study were tested in a post-hoc analysis to try to interpret the potential biological relevance of the unknown metabolite (with all metabolites, including those that were associated with drug response). For example, 15-65.533 is highly correlated with methionine (P = 1E-6), which is also correlated with disease status. Additionally, 15-65.533 is also correlated with several other known metabolites, including Indole-3-propionic acid (I-3-PA), Kynurenine (KYN), Indole-3-Acetic Acid and Guanosine. While the mechanisms and potential pathway relationships need to be evaluated in future studies, there is previous evidence of the involvement of these known metabolites in disease etiology. For example, I-3-PA has been shown to be involved in neuronal damage and oxidative stress in the brain [48, 49], KYN is a major route of tryptophan metabolism and has been implicated in pathogenesis of several neuropsychiatric diseases. Several metabolites within this pathway were involved in the pathogenesis of AD in previous studies [50, 51]. The other unknown metabolite, 8-93.65, is strongly correlated with MET (P = 3E-12) and glutathione (GSH, P = 3E-5), both of which are elevated in AD . MET is the precursor for homocysteine and cysteine, which plays a critical role in GSH synthesis. The potential link of this pathway with AD along with details of the data analysis has been discussed in our previous manuscript .
In addition to the disease-metabolite associations, we also identified drug-associated metabolites. While not the primary goal of the current study, these associations may reveal interesting biology of drug metabolism and/or response. Again, as with the disease-associated metabolites, the majority of the drug-associated metabolites are compounds of unknown structure. Interestingly, one of the known metabolites, I-3-PA, which was associated with drug use, was associated with 15-65.533, potentially indicating/reinforcing a shared mechanism between the drug targets and disease etiology. This supports the potential of the metabolites to identify potential new drug targets by revealing insights into disease susceptibility.
We also identified two metabolites that were associated with ApoE genotype;
Phosphoethanolamine is an ethanolamine derivative that is used for synthesis of sphingomyelins that we previously implicated in AD , and monopalmitin (glycerol 1-palmitate) a lipid implicated in membrane integrity and stability and an energy storage source.
Should our findings be replicated and validated in a larger study of pathologically confirmed AD, it may lead to a clinically useful test. Similar studies in patients with MCI are underway to determine whether these combined models are capable of identifying distinct subgroups of MCI patients. Longitudinal follow-up of MCI patients can then determine whether the “AD-like” biomarker profile predicts the progression of cognitive decline and thus identifies a subgroup at high risk for developing dementia. Such participant groups are likely candidates for clinical trials of agents for slowing the progression of cognitive decline. In addition, these studies will also need to be extended to patients with other types of dementias, such as frontal temporal lobar dementia, Parkinson’s disease dementia and Lewy body dementias, to assess the specificity of these metabolites to pathologic AD vs. other dementias.
In addition to the above-stated limitations of sample size and current interpretability of results with metabolites of unknown structure, our AD participants were clinically diagnosed and we therefore did not have autopsy confirmation as a gold standard. Since our objective was to compare the performance of metabolomics markers against standard CSF Ab42 and tau measures, we did not use these measures to define pathologic AD. Therefore, it is likely that about 20% of our AD participants did not meet criteria for pathologic AD and about 20% of controls might harbor preclinical disease, which perhaps explains the observed accuracy of standard Ab42 and tau markers.
Additionally, the lack of correlation between the metabolites selected in the prediction models and MMSE highlights the potential of these metabolites as predictors of overall clinical AD diagnosis, and not necessarily a specific stage. Given our sample included both mild and moderate severity patients, studies of just mildly impaired patients would be more informative of the utility of the markers for early AD. Future studies should investigate these complex relationships, and interrogate potential associations amongst metabolites and other clinical and biological mediators and contributors to mild stages of AD.