The potential of metabolomic profiling predict individual multidisease outcomes
In a recent study published in Nature Medicine, researchers explored the potential of a nuclear magnetic resonance (NMR) spectroscopy-based metabolomic platform to estimate the risks for the onset of several medical conditions.
Background
Risk stratification is central to disease prevention. Over
the past decade, increasingly complex information on an individual’s phenotype
has become available beyond conventional demographic and laboratory
information. While blood metabolites such as cholesterols are established
clinical predictors for cardiovascular disease risk, many more have been linked
to common disease phenotypes. In recent years, studies have moved beyond
associations of individual markers by linking metabolomic profiles to aging,
disease onset and mortality, appreciating the human blood metabolome as a
direct reflection of the physiological state.
Prompt identification and prevention of risk factors associated with the development of medical conditions are critical. In recent times, metabolomic analyses have been performed to identify high-risk individuals; however, the metabolic data has been considered inadequate for incident disease risk estimation. nuclear magnetic resonance (NMR) spectroscopy enables rapid and relatively cost-effective molecular assessments compared to other metabolomic techniques such as mass spectroscopy.
About the study
In the present study, researchers investigated whether NMR
spectroscopy-based serological metabolomic profiles could reflect the true
physiological state of individuals and add to clinical biomarkers for
estimating the risks of developing 24 conditions, including vascular,
metabolic, respiratory, neurological, and musculoskeletal disorders and cancers
across 22 centers.
A neural network (NN) was trained to simultaneously learn
metabolomic states (METs) specific for medical disorders from 168 metabolic
markers quantitatively assessed among 117,981 individuals with a follow-up of
1,400,000 million individual years from the United Kingdom (UK) Biobank (BB).
To validate the model’s findings, four different cohorts were analyzed, i.e.,
the Rotterdam Study cohort, the Whitehall II cohort, the PROspective Study of
Pravastatin in the Elderly at Risk (PROSPER) cohort, and the Leiden Longevity
PAROFF study cohort were analyzed using the same 1H-NMR metabolomics assay.
In addition, three BBMRI-NL consortium cohorts were
analyzed. The markers included fatty acid and amino acid metabolites associated
with fluid balance and carbohydrate metabolism. Their association with
conventionally used clinical markers such as creatinine, albumin, and glucose
was assessed. Cox proportional hazard (CPH) modeling was used for the analysis,
and the hazard ratios (HRs) and odds ratios (ORs) were calculated.
To maximize the generalizability of the study findings, data
was partitioned spatially by recruitment centers. After obtaining final
estimations, test set predictions were Whitehall II aggregated for further
analyses. The data were analyzed by sex and age only (Age+Sex), by American
Heart Association (ASCVD)-based cardiovascular estimators, and PANEL estimators
(including >30 estimators with data on physical examination, laboratory
measurements, and lifestyle habits.
The estimators were further validated by CAIDE and FINDRISC
scores for dementia and type II diabetes (T2D), respectively. Furthermore, the
NMR estimates were correlated with rates of incident medical conditions in the
observation period and compared to clinical information based on C-index delta
values. SHAP (shapely additive explanation) values were determined for all 24
diseases investigated and UMAP (uniform manifold approximation and projection)
analysis was performed to identify which metabolites affected disease risks the
most.
Results
The median age of the sample population was 58 years, of
which 54% were women, and the participants were followed up for a median of 12
years, with 1,435,340 overall individual years. The METs were associated with
incident event rates of conditions investigated, except breast cancer, and for
a 10-year estimation, the MET and Age+Sex predictor combination equaled or
outperformed other estimators.
Further, MET data were additive to clinical estimators for
eight medical conditions, including T2D, heart failure, and dementia. High OR
values for T2D, abdominal aortic aneurysm (AAA), and heart failure were 62, 14,
and 11, respectively. On the contrary, OR values were low for cerebral stroke,
major adverse cardiac event (MACE), atrial fibrillation, all-cause dementia,
and COPD (chronic obstructive pulmonary disease), and further lower for asthma
and glaucoma.
The metabolomic state contained significantly lesser
estimative data compared to clinical estimators for glaucoma, cataract, and
cancers of the colon, skin, prostate, and rectal tissues and had a greater
estimative value than the ASCVD and Age+Sex estimators for kidney disorders,
T2D, and hepatic disorders. All tested models calibrated well in the UKBB
cohort, and the four external cohort findings validated significant
discriminative improvements by adding MET to the Age+Sex estimator data for
COPD, T2D, coronary heart disease (CHD), atrial fibrillation, and heart
failure.
Additionally, the C-index values indicated that MET data
significantly added to the comprehensive PANEL estimators for eight diseases,
including COPD, T2D, MACE, CHD, renal disorders, heart failure, and dementia.
The discriminatory gains generally translated into utility gains. After MET
data adjustments for comprehensive clinical predictors, the adjusted HR
estimates for T2D (HRPANEL 2.5, HRge+Sex 3.8), heart failure (HRPANEL 1.5,
HRAge+Sex 1.8), all-cause dementia (HRPANEL 1.5, HRAge+Sex 1.6), MACE (HRPANEL
1.4, HRAge+Sex 1.6), or COPD (HRPANEL 1.4, HRAge+Sex 1.6) showed a clear
distinction between the trajectories of incident risks.
MET HRs were validated externally with Age+Sex adjustments
for CHD, COPD, heart failure, all-cause dementia, and atrial fibrillation.
High-impact metabolites identified included glycine, tyrosine, and glutamine,
carbohydrate metabolism metabolites, albumin, creatinine, glycoprotein
acetylation (GlycA), acetoacetate, and acetone. However, besides albumin,
creatinine, glucose, creatinine/cystatin C, and lipids, strong correlations
were not observed between the NMR data and PANEL estimators.
Remarkable associations were observed between creatinine
with AAA, glucose, and T2D and GlycA with COPD and lung cancer. Albumin,
creatinine, leucine, tyrosine, and glutamine were identified as the most
predominant contributors to the estimated risk for all-cause dementia. Of
interest, high-risk individuals’ attribution profiles were consistently
dominated by low linoleic acid (LA), albumin, docosahexaenoic acid (DHA),
glycine, and histidine levels.
Conclusion
Overall, the study findings highlighted the potential of
NMR-based metabolomic profiling for informing on the incident risks of several
medical disorders simultaneously.
Journal reference:
Buergel, T. et al. (2022) "Metabolomic profiles predict
individual multidisease outcomes", Nature Medicine. doi:
10.1038/s41591-022-01980-3. https://www.nature.com/articles/s41591-022-01980-3
No comments