06 Oct 2022

Can the many metabolites coursing through your blood foretell the diabetes, heart disease, or dementia diagnosis that awaits you sometime down the road? Yes, according to researchers led by Ulf Landmesser and Roland Eils, Berlin Institute of Health at Charité-Universitätsmedizin Berlin, and John Deanfield, University College London. In the September 22 Nature Medicine, they reported that patterns of serum metabolites predicted the onset of 23 diseases over 10 years. When combined with age and sex, these signatures performed as well as, or better than, clinical prediction models, such as cardiovascular risk scores.

Disease-specific signatures were spotted among blood metabolites.
These signatures predicted the onset of 23 diseases.
Combining metabolites, age, and sex rivaled clinical prognosticators.

“The hope is that from one blood sample we can simultaneously forecast the risk of multiple common diseases and empower people to adjust their lifestyle to delay, or even prevent, illness,” said Eils.

“This is a very nice example of the strength of metabolomics [and] shows the possibility of increasing its power by combining it with readily available clinical measures,” wrote Miroslava Cuperlovic-Culf, National Research Council, Ottawa, Canada. Nicholette Allred, Wake Forest University School of Medicine, Winston-Salem, North Carolina, agreed. “This work demonstrates proof of concept for the clinical utility of metabolomics in risk stratification of complex diseases,” she noted (comments below).

To assess a person’s risk of a disease, clinicians routinely use prediction models that include demographics, lifestyle factors, and blood analysis. For example, the Atherosclerotic Cardiovascular Disease (ASCVD) score uses age, sex, blood pressure, smoking status, and plasma cholesterol levels to predict a person’s risk of heart disease over the next decade. However, these types of scores predict one disease. Could a test be devised to foresee multiple looming morbidities, and do so more accurately than current models?

Co-first authors Thore Buergel and Jakob Steinfeldt at the Berlin Institute of Health turned to the UK Biobank. This epidemiological project has collected data on half a million people aged 40-69 at baseline, and is tracking their age-related health conditions including neurodegenerative ones. A subset of 118,000 volunteers, median age 58, had their baseline blood tested using a proprietary proton NMR protocol that measures 168 common metabolites, mostly lipids and lipoproteins. Everyone's medical records spanned an average of 12 years of follow-up, documenting the incidence of 24 conditions, including six kinds of cancer, Type 2 diabetes, heart disease, Parkinson’s, and all-cause dementia. All told, this amounts to 1.4 million person-years of follow-up.

To find correlations lurking in this mountain of data, Buergel and colleagues developed a neural network, a type of machine-learning algorithm that spots complex patterns with minimal human interference. The scientists fed baseline metabolite values and follow-up diagnoses for each participant into the algorithm, then tasked it to identify metabolite patterns present only in people who developed one of the diseases. For all except breast cancer, unique metabolite signatures emerged.

Which metabolites most strongly foretold disease? For dementia, it was high cholesterol, glucose, glutamine, creatinine, and acetylated glycoproteins. The last two are markers of kidney function and inflammation, respectively. In contrast, high levels of the serum protein albumin, the lipid linoleic acid, leucine, and other branched-chain amino acids protected against dementia. Henrik Zetterberg, University of Gothenburg, Sweden, wondered what the metabolic profiles were for specific dementias and how they relate to the known pathology markers. “[It will be important] to determine if the metabolic changes are upstream or downstream of proteinopathy and other neurodegenerative brain changes,” he wrote (comment below).

**Predictive Power.** Compared to three clinical prediction models, adding metabolomics data improved the accuracy of 10-year disease risk predictions as measured by AUCs (y-axis). *[Courtesy of Buergel et al., Nature Medicine, 2022.]*

The closer a person’s metabolite profile resembled a given disease signature, the higher their odds of getting it within the next decade. For example, the top 10 percent of participants whose metabolite profile most closely matched that of the dementia signature were six times likelier to develop it than those in the bottom decile. People in the top decile for the cerebral stroke and diabetes signatures were 10-fold and 62 times likelier, respectively, to develop those conditions.

For most people, however, their metabolite signature underperformed compared to the three currently used clinical prediction models: age and sex; ASCVD scores; and a clinical panel made up of body-mass index, lifestyle factors, and a slew of diagnostic blood work such as cholesterol, glucose, and white blood cell tests. For a few diseases, such as Type 2 diabetes, the metabolic signatures alone were almost as accurate as the clinical panel (see image above), but for most other conditions, including dementia and Parkinson’s, they were weaker.

What about combining metabolites with prediction models? For dementia, this improved prediction only a tad, nudging the area under the curve for the clinical panel from 0.68 to 0.685 (see image). “We will have to add brain-derived biomarkers and other -omics to the models to achieve clinically meaningful risk predictions for dementia,” suggested Cristina Legido-Quigley, King’s College London.

For a dozen diseases, however, adding the metabolite data did sharpen predictions. This suggests that from a single blood sample, NMR metabolite analysis could simultaneously predict a person’s likelihood of getting any of these diseases, which include cardiovascular disease. This could save time and money in the clinic, the authors argue, and allow clinicians to prescribe preventative therapies or lifestyle changes.—Chelsea Weidman Burke

Comments

- Henrik Zetterberg
  University of Gothenburg
- Posted: 06 Oct 2022
This is such a heroic analysis effort underscoring the importance of metabolic changes in neurodegenerative dementias, as well as a large number of other major human diseases, in multiple cohorts.

Branched-chain amino-acid changes are now well-replicated in dementias, plus a number of additional, more novel observations were made.

The next steps in this research should include examination of metabolic profiles in specific neurodegenerative dementias, and also assessing their relationship with established biomarkers for underlying brain pathologies. An important future research topic will be to determine if the metabolic changes are upstream or downstream of the onset of proteinopathy accumulation and other neurodegenerative brain changes.

View all comments by Henrik Zetterberg

- Nicholette Allred
  Wake Forest School of Medicine
- Posted: 06 Oct 2022
Exploring the clinical utility of metabolomics for stratifying risk in cardiometabolic disease is a logical next step in our efforts to combat the rising burden of diseases in the population. Metabolites, which are small molecule intermediates resulting from biological processes, are dynamic biomarkers that have been broadly used in association and prediction studies. They may be useful to inform upon dysregulated processes and provide information on disease pathophysiology and targeted interventions. Buergel et al. coupled metabolomics with machine-learning approaches to simultaneously inform upon multi-disease risk in the UK Biobank population, with replication in four additional cohorts. Results from this work illustrated that metabolite profiles were associated with incident rates of common disease, and were equal to, or outperformed, established clinical predictors of common disease. This demonstrates proof of concept for the clinical utility of metabolomics in disease risk stratification of complex diseases and creates an opportunity for refinement toward broader implementation.

As the authors note, metabolite coverage on the NMR platform is relatively narrow and lipid-focused. This platform pales in comparison to the thousands of metabolites detected, quantified, and reported in the literature using complementary approaches, e.g., mass spectrometry, and sample types. This is further compounded by the dynamic nature of metabolites, which are subject to rapid changes in response to biologic perturbation. Thus, standardized assessment, e.g., fasting state not often achieved in clinic, would be clinically warranted to optimize informativeness and allow for comparison. Beyond metabolic state, confounders such as diet, medications, environment, etc., need to be considered as these could impact the underlying biology and influence risk models.

Toward broader implementation, it should be noted that the volunteer study population, while an excellent clinical and research resource, were generally healthier than the U.K. population, e.g., older, decreased CVD risk factors, higher socioeconomic status, etc., and lack diversity in age, race/ethnicity, etc. Moreover, cardiometabolic diseases, such as those included herein, are not mutually exclusive and could bias risk models.

Moving forward, these observations motivate additional research questions. With a focus on precision medicine, broad generalizability needs to be assessed both in terms of populations and conditions assessed. And, once we identify risk, how can these profiles be used to prioritize, monitor, and target disease interventions with the broad goal of improving metabolic health?

View all comments by Nicholette Allred

- Miroslava Cuperlovic-Culf
  National Research Council, Canada
- Posted: 07 Oct 2022
This paper is a very nice example of the power of NMR metabolomics, primarily achieved thanks to the availability of extremely large datasets. This publication highlights the advantages of NMR metabolomics for diagnostics, and shows the potential to increase the power of metabolomics by combining it with readily available clinical measures.

The authors showed that the addition of quantitative, whole-blood, proton NMR metabolomics to other clinical measures improved risk factor assessment for 23 out of 24 common conditions including metabolic, vascular, respiratory, musculoskeletal, and neurological diseases, as well as cancer. Combining clinical markers with NMR metabolomics data is a very interesting way to boost the power of prediction models, and NMR metabolomics provides sufficient sample size for these analyses. Utilization of a deep, residual, multitask neural network to simultaneously learn predictors for different conditions at the same time, and from the same dataset, is a very interesting approach applied here in a novel way.

I was very surprised with the low performance of disease-specific metabolic state analysis in risk trajectory assessment for breast cancer, particularly when compared with prostate cancer. Also, high predictor power for Type 2 diabetes, dementia, and heart failure event rate achieved using only age and sex information, with only minor improvement obtained by adding other parameters, including metabolomics, is very interesting and somewhat unexpected.

Diagnostic metabolomics has been taking its rightful place over the last decade and this work is an excellent addition to many other examples of the power of metabolic profiling for different applications. In this work NMR measurements provided concentrations for 168 metabolic markers that are all used in the development of machine-learning models for risk assessment. Future applications of this particular model for diagnostics would require measurement, assignment, and quantification standardization that can be achieved if appropriate NMR equipment is in place. Perhaps this work will encourage addition of NMR to clinical laboratories of the future.

View all comments by Miroslava Cuperlovic-Culf

- Cristina Legido Quigley
  King’s College London/Steno Diabetes Center Copenhagen
- Posted: 07 Oct 2022
Having more detailed molecular data, which in this case is 150 data points made up of a few small molecules and many lipoprotein particles of different sizes, helps with assessing the risk of developing some diseases such as diabetes. Using this molecular platform for the prediction is similar to using clinical lipids, glucose, fatty acids, ketones, and omegas, so it could be particularly useful for diseases where metabolic syndrome or lipids are known risk predictors. The addition of machine learning for the prediction might exploit additional hidden biochemistry and therefore help the prediction further. But my takeaway was that, for dementia, the results of this particular panel of molecules and lipoproteins were not conclusive. The results of the metabolic state didn’t exceed the risk provided by sex and age, and hazard ratios could not be validated in an additional cohort.

It surprised me that many of the molecules and lipoproteins, when plotted in box plots, seem to vary in the same manner. This could be biological but also due to the NMR technology not being specific enough, as the authors mentioned in discussing the limitations of the study. And, if measurements are indeed correlated, this would impact the machine-learning modeling and the metabolic state calculation. It would be great to compare these NMR measurements with other technologies to better understand these observations.

I think of metabolomics as being a snapshot of metabolism and this is very useful to assess health, however, it is early days and this work involved a great number of participants but very few molecules and lipoprotein data. I see the value of measuring metabolic risk, or state, for patients, but for dementia we will have to add brain-derived biomarkers and other omics data to the models to achieve clinically meaningful risk predictions.

View all comments by Cristina Legido Quigley