https://www.nature.com/articles/s41591-025-04038-2
chatGPT:
Here is a concise review of the uploaded paper, “Shared and specific blood biomarkers for multimorbidity”.
Summary
The paper asks a useful geroscience question: are there blood biomarkers that track multimorbidity in general, and are there others that are specific to particular multimorbidity patterns? The authors studied 2,247 adults aged 60+ from the Swedish National Study on Aging and Care in Kungsholmen (SNAC-K), using 54 blood biomarkers spanning metabolic, inflammatory, vascular, neurodegenerative and organ-damage domains. They related these biomarkers to three different ways of defining multimorbidity: total disease count at baseline, latent multimorbidity patterns at baseline, and 15-year rate of disease accumulation. They then externally validated the longitudinal findings in 522 participants from the Baltimore Longitudinal Study of Aging.
The central result is that a small set of biomarkers was consistently associated with multimorbidity across definitions. The most reproducible “shared” biomarkers were GDF15, HbA1c, cystatin C, leptin and insulin, all positively associated with multimorbidity measures. For faster long-term disease accumulation, gamma-glutamyl transferase (GGT) was additionally positive and albumin was inversely associated. In effect, the paper argues that metabolic and immunometabolic dysregulation is a central biological signature of multimorbidity.
The authors also derived five multimorbidity patterns at baseline: Unspecific, Neuropsychiatric, Psychiatric and Respiratory, Sensory impairment and Anemia, and Cardiometabolic. Some biomarkers appeared broadly shared across patterns, while others showed more selective associations. For example, NfL was more strongly linked to the Neuropsychiatric pattern, and N-cadherin contributed more to the Cardiometabolic pattern. The paper’s PCA analysis suggests that although there is a shared multimorbidity signal, distinct biomarker subprofiles may underlie different disease clusters.
A strength of the study is that it does not treat multimorbidity as just “2+ diseases.” Instead, it examines disease burden, disease clustering, and disease accumulation over time. The external validation is also important: the longitudinal biomarker model generalized reasonably well to BLSA, with similar prediction error, which argues against a purely cohort-specific result.
What is novel here
The main novelty is not that markers like HbA1c or cystatin C correlate with poor health; that is already broadly known. The novelty is that the authors integrate many biomarkers across several biological domains and test them against multiple operationalizations of multimorbidity, including a longitudinal disease-accumulation phenotype rather than only a cross-sectional disease count. That is more ambitious and biologically meaningful than many prior multimorbidity biomarker studies.
A second novelty is the attempt to distinguish shared from pattern-specific biology. The paper argues that multimorbidity is partly driven by common mechanisms, especially metabolic/immunometabolic dysfunction, but also contains some disease-cluster-specific signals. That is conceptually useful because it bridges two views: multimorbidity as a general aging phenotype, and multimorbidity as a set of more specific syndromic constellations.
A third novelty is the external validation of the longitudinal model, which is more persuasive than stopping at internal selection. Many papers identify biomarker panels without any replication; this one at least tests whether the disease-accumulation signal transports to another aging cohort.
Critique
My overall view is that this is a good, thoughtful observational paper, but it is stronger as a biomarker mapping study than as a mechanistic paper.
The biggest limitation is causal ambiguity. The biomarkers were measured at baseline in people who already had varying disease burdens. So even when a biomarker predicts later accumulation, it may still be acting as a summary readout of already ongoing pathology rather than a driver of multimorbidity. The authors acknowledge this. That matters especially for biomarkers like GDF15, cystatin C, and albumin, which are highly integrative stress/illness markers rather than pathway-specific causal nodes.
A second limitation is that the strongest signal may partly reflect nonspecific illness severity rather than a distinct multimorbidity program. For example:
- GDF15 rises in many settings: mitochondrial stress, inflammation, renal dysfunction, cancer, frailty, cardiometabolic disease.
- Cystatin C is influenced by kidney function and broader systemic aging processes.
- HbA1c, insulin and leptin reflect metabolic state, adiposity, and insulin resistance, which are already upstream or downstream of many chronic diseases.
So the claim that metabolic dysregulation is central is plausible, but the data also fit a simpler interpretation: these markers identify people with higher systemic physiological dysregulation overall.
Third, there is some risk of circularity or confounding by disease definitions. If multimorbidity includes conditions like diabetes, CKD, obesity-related disease, heart disease and anemia, then biomarkers such as HbA1c, creatinine/cystatin C, leptin, insulin, hemoglobin are biologically close to the diagnostic space itself. That makes them clinically useful but can exaggerate the impression of deep mechanistic discovery. In other words, some associations may reflect the fact that the biomarker is already a near-measure of one or more constituent diseases.
Fourth, the cohort selection probably biases toward a healthier, higher-SES, mostly Swedish-origin sample, and the excluded participants with missing biomarkers were older, sicker and more impaired. That could affect both prevalence structure and biomarker associations. The validation cohort helps, but only partly, because BLSA is also not a broadly representative population sample.
Fifth, the biomarker platform is broad but still incomplete. The study lacks many things that would be very informative mechanistically: repeated biomarker measurements over time, richer omics, stronger mitochondrial markers, immune cell phenotyping, and markers tied to senescence burden, fibrosis, or extracellular matrix remodeling. The absence of longitudinal biomarker trajectories is especially important because multimorbidity is fundamentally dynamic.
Sixth, the statistical approach is sensible but still has interpretability limits. LASSO is good for correlated predictors, but selected variables can be unstable when biomarkers are collinear. The authors partly address this with resampling for pattern analyses and external validation for the longitudinal part, which is good, but the exact biomarker list should still be interpreted as a compressed predictive signature, not a definitive set of unique causal factors. The PCA is also descriptive: it summarizes covariance structure, but principal components are not biological pathways.
Seventh, some findings are biologically interesting but not yet convincing. The amyloid β42/40 inverse association across all multimorbidity patterns is intriguing, but the authors themselves note assay and interpretation issues. That result feels hypothesis-generating rather than solid.
Bottom line
This is a strong observational multimorbidity biomarker paper with three real contributions: it uses richer definitions of multimorbidity, identifies a reproducible shared biomarker core, and validates its longitudinal model in an external cohort. Its main message—that metabolic/immunometabolic dysregulation is a central shared substrate of multimorbidity—is credible and important.
But it does not yet show that these biomarkers are causal drivers, nor that they define a single unified mechanism of multimorbidity. The safest interpretation is: multimorbidity appears to have a strong shared systemic metabolic-stress signature, overlaid with smaller pattern-specific signals. That is clinically useful and geroscience-relevant, but still one step short of mechanistic proof.
If you want, I can also turn this into a claim-by-claim table with columns for claim, evidence, novelty, strengths, weaknesses, and my assessment.