Accelerated ageing of body organs from blood proteins

A review and commentary on published paper: Oh, H. et al. Organ aging signatures in the plasma proteome track health and disease. Nature 624, 164–172 (2023).  https://www.nature.com/articles/s41586-023-06802-1

Ageing is a fact of life, as is radically unequal ageing. Consider David Attenborough still making glittering docuseries in his 90s, or Picasso generating the Las Meninas masterwork in his 70s in homage to childhood idol Velazquez. Most of us will not age like that; indeed, many will see their daily function and independence increasingly challenged by a host of age-related illnesses.

It turns out that the body’s organs and tissues also age unequally. The lucky few will benefit from their major organs staying in fine shape; at the other extreme, several organs may undergo accelerated ageing. This is the notion that the typical rate or pattern of change that accompanies chronological age in these organs can manifest earlier than otherwise welcome.

All sorts of combinations are theoretically possible. A 60-year old woman may have the cardiac profile of a typical 70-year old (a 10-year unwelcome age gap) but the “brain age” of a 50-year old (10-years thank-you-very-much).

The idea of organ-specific age gaps being computable at the individual level across the population has until recently been a pipe dream. It is simply not possible to sample directly from some of the key organs in-principle (e.g., brain tissue or cardiac muscle), and not feasible at scale for others (e.g., lung or liver).

This is where the exciting work of Se-Hwee Oh and colleagues promises to overturn the status quo. They identify protein signatures in the blood argued to be unique to different organs, specifically to the ageing of individual organs.
Let’s see what they did and then review what it may mean.

How did they compute organ age-gaps?

First, the researchers started off with blood plasma samples from about 5000 individuals, collated from different large cohort studies. The SomaScan proteomic assay was then applied to quantify ~5000 different proteins from the blood – a critical technical step since all the data acrobatics from this point hinges on the reliability of this information. To their credit, the researchers started off by eliminating about 200 proteins whose levels were not reliably estimated across repetitions.

From the ~4800 remaining proteins we get to our first major edit. Using the Human Protein Atlas, they define a protein as “organ-enriched” if its expression level is 4-times higher in one organ than any other. Why 4-times? No idea… there is no justification, so let’s just go with the flow. This returns ~850 (18%) organ-enriched proteins (different subsets for 11 different organs: brain, heart, lungs, etc) that enters the next step of the data pipeline. There is also an “organismal” subset of ~4000 proteins defined that have no particular organ-specificity.

Prior work have successfully used general plasma protein profiles to estimate chronological age, but the major novelty in this work is to use these organ-enriched protein profiles to predict age. People from one of the cohort studies (N~1200) were used to apply a particular type of machine learning (LASSO model), producing a weighted combination of enriched proteins for each organ that best predict age.

This model was then tested in the other 4 unseen cohort studies. Each organ-specific model, and the organismal model, was found to statistically predict years of age. Hence, for each person, and for each organ, it is possible to calculate a residual or deviation from the model’s line of best fit – in other words, how far away from the average proteomic profile a person’s organ plasma pattern is for their particular age.

Each organ then accrues an age-gap score (in years), a notional value as to whether for their age the person’s organs are above or below the expected proteomic profile.

What did they find about organ-specific age-gaps?

What do you think the distribution of organ age-gaps may have looked like? I expected these to be highly inter-correlated but they were not (r=0.29). This is the first interesting insight. At face value, typical changes in organ-specific blood proteins over the lifespan are relatively unique and there is no strong underlying ‘g’ factor at play. Biologically, it appears our organs age rather independently (at least according to the blood-visible proteins in this dataset).

‘Extreme’ organ agers were defined using a conventional 2-standard deviation cut-off and a similar pattern emerged. About 18% of individuals had a single dominant organ undergoing accelerated ageing, and less than 2% had an accelerated profile in multiple organs. If all of this plays out (i.e., replicable and generalizable; more on that later) then I can see this simple notion as having value in ICU settings where transition to multiple organ failure is of major clinical significance.

What do organ age-gaps correlate to?

It is perhaps no surprise that extreme proteomic agers for a particular organ were associated with a higher frequency of disease and disorder characteristic of that organ. For example, the “kidney ageotype” was associated with metabolic diseases (diabetes, obesity, hypercholesteromia etc) and the heart ageotype with atrial fibrillation and heart attack.  

There is an element of circularity here since organ enrichment was defined using Human Protein Atlas annotations that are heavily biased in the first place by presence of disease in the index organ (diabetes researchers tend to study the kidney and pancreas and not the lungs). More fundamentally, there is a root issue here that is left unaddressed: it is almost impossible to parse “pure biological ageing” from “age-related disease”. Is an accelerated ageing profile actually accelerated biology or just the accumulation of subclinical or undiagnosed disease? Are the two things even separable? 

Putting aside that thorny issue, there were also some surprises. If one selects for only those people with no known disease or clinical abnormal test at baseline (i.e., super healthy individuals) then every single year equivalent of the accelerated heart ageotype was linked to a ~25% increased risk of heart failure at 15 years follow up. This suggests the possibility of identification of at-risk people a long time before the advent of clinically visible changes – the perfect subset of people for the targeting of lifestyle based interventions.

In another surprising example, Alzheimer’s disease was associated with accelerated ageing in just about every organ except the brain. More on this interesting idea below.

Do brain age-gaps link to Alzheimer or dementia?

Brain age gaps were not reliably linked to AD diagnosis across the cohorts and so the team developed a new set of models built around the Clinical Dementia Rating (CDR) score, a standard measure of global cognitive function. More about the CDR in my earlier blog on lecanemab.

This time a CognitionBrain model was trained using one cohort to specifically select those proteomic features linked to CDR and chronologic age. This new model now generalised to one other cohort (but not the other three?), suggesting that AD was associated with two additional years of CogBrain ageing. Moreover, a 1SD increase in this ageotype was linked to a 2-point increase in the CDR within two years, a change of potential clinical significance.

How does this CogBrain proteomic profile compare to traditional AD predictors such as age, baseline CDR, polygenic risk score and the standard blood biomarker plasma Tau-181?  

Interestingly, in a multivariate model (all predictors simultaneously) run on just one cohort (why?) it was the CogBrain gap score that was most predictive after baseline CDR. I cannot for the life of me understand why education was not included in the model, a real faux pas given this is the single strongest sociodemographic predictor of AD dementia. This extends to the absence of any reporting of education levels for different cohorts in the Supplementary tables, a non-trivial issue that I will return to at the end.

Surprising combinations of blood protein may be hiding in the data linked to AD diagnosis. So what was in the mix? The profile included 49 proteins, of which 47 are linked to neurons and glia from single-cell brain RNAseq studies. The most strongly weighted proteins included synaptic complexins and neurexins, neurite-related strathmin and olfactomedin, and several others linked to astrocytes and oligodendrocytes. Overall, a mix of kind-of-expected proteins and novel ones.

What is surprising is the complete absence of bog-standard AD proteins such amyloid or tau. What to make of this? Are these below the limit of detection of the SomaScan platform, or simply not as strongly connected to AD diagnosis as commonly thought? The lack of discussion of this basic point was surprising.

Despite this, and provided there is a lot more analysis and follow up (and replication) to do, there is a solid lead here pointing to the presence of CNS-relevant proteins in human plasma that may be of diagnostic and predictive use.

And what of the intriguing notion of accelerated ageing in non-brain organs linked to AD? Replicating the above pipeline, agegaps in the CognitionArtery, CognitionPancreas, CognitionBrain and CognitionOrganismal models were linked to AD diagnosis in multiple cohorts, and seemed to generalise to general cognition in one healthy cohort.

Over the lifespan, it appeared the CognitionArtery and CognitionOrganismal changes occurred first, highlighted by changes in a tightly correlated cluster of proteins that include pleiotrophin, transgelin, WNT1 and several others expressed in the vascular system such as smooth muscle cells, pericytes and perivascular fibroblasts.

This is yet another piece of independent data suggesting a fundamental link between abnormal vascular biology and Alzheimer’s disease, sure to trigger further research down this path.

Is this clinic-ready? No.

This is exciting research with lots of potential areas for application. But there are plenty of landmines to avoid.

First, there is an element of arbitrariness in how organ-enriched was defined. For a general methodology that is trying to appear unbiased this is dissatisfying and could be overcome.

Second, the work relies on a particular proteomic platform and begs the question if the results are platform-dependent? Too much biomarker work has failed to translate to practice because of technical limitations in the analytical method. Even using the SomaScan platform, the threshold for retest reliability was set at a correlation of r= 0.4, quite low for an analytical method. Many of the protein measures would have differed significantly if analysed on a different day or different instrument, and this undermines confidence in the replicability of the outcomes.

Third, and most worrisome, is the narrow sociodemographic profile of the cohorts. Only the basic ageotype work was replicated across all cohorts and the cognition and AD work seems to have been validated in selective cohorts. Moreover, all of the cohorts were exceptionally Caucasian, white and highly educated: the mean education level of the LonGenity cohort was 17.5 years. The “average” person in this study was a university post graduate!  

It doesn’t take much imagination to see how the organ-specific ageing profile from this rarified sample will be quite different to the average person on the street. Hence, the models presented are predictive at best for only a small slice of the population.

How does this advance the field?

If one assumes that the vitality of different body organs can manifest and be quantified in circulating blood plasma that is already a big step. There is no good mechanistic explanation for this in the brain, classically thought to have a blood-brain-barrier specially designed to avoid this kind of spill-over. But that, like other examples of medical textbook folklore, may be proven wrong in the face of modern analytical methods.

Building on that, in the future it might be possible to get a general snapshot of the health of different body organs relative to your peers. In much the same way as the individual chemicals and proteins in blood are compared to normative values, in the future it may be possible to accurately contextualise complex blood protein signatures against the correct reference database to arrive at the “biological age” for your organs.

That could have profound impact on the practice of medicine – not just in the critical care space as alluded, but for the better targeting of preventative medicine strategies, for better drug prescribing practice, clinical trial stratification, the possibilities are endless.

Clever use of proteomics on blood is going to explode in the coming years with the burgeoning use of ML and AI. As for all these applications, human value will hinge on the quality of the raw data, representativeness of the training and validation sets, and a clear intent on end-use and clear eye on how to achieve it in the real world.

 By Professor Michael Valenzuela, CEO