Can Plasma Proteins Predict Who Dies First? Answers from 848 People Followed for Eight and a Half Years in Taiwan

[This article is based on a publicly available published paper. The corresponding author, Dr. 陳亮恭, is also a consulting contact for a separate corporate case study (CET Taiwan subsidiary) in preparation on this site. This article draws only on published data and has no connection to any commercial or consulting relationship.]

At the end of the previous article, the evidence was arranged as a ladder: from animal experiments, to observational human studies, to intervention trials with clinical endpoints. Protein aging clocks that predict risk sit on the middle rungs — above in vitro cell work, below human intervention trials, in the category of observational predictive models.

This article continues along that path, but at a smaller scale: not a European or American cohort of tens of thousands, but 848 people in Taiwan followed for eight and a half years.

What the Study Did: 848 People, Eight and a Half Years, One Tube of Plasma

A team from Yang Ming Chiao Tung University and Taipei Veterans General Hospital (Huang et al., Aging Cell 2026) drew 848 participants from ILAS, the Integrative Longitudinal Aging Study, a Taiwanese longitudinal aging cohort. Mean age was 63.8 years (±9.2); 53% were women.

Each participant gave a single fasting plasma sample at baseline, then entered the follow-up period. Mean follow-up was 8.5 years (SD 1.2). By the end of follow-up, 92 participants had died — 10.8% of the cohort.

Those 848 samples were sent down two analytical pipelines:

• Proteins: Non-targeted analysis using nanoLC-nanoESI-MS/MS (LTQ Orbitrap Elite mass spectrometer). Proteins detected in more than 50% of samples were retained, leaving 516.
• Metabolites: LC-MS analysis (Waters Xevo G2-S QTof). Features with a missing rate above 20% were excluded; 532 were retained.

An external validation cohort of 231 participants was drawn from LAST (Longitudinal Aging Study in Taiwan, Wave 1, 2016–2019), with a median follow-up of 7.3 years and 17 deaths (7.4%).

Which Proteins Were Associated with Mortality?

The main analysis ran Cox proportional hazards models for each protein individually, adjusting for age, sex, years of education, smoking, alcohol use, history of hypertension, and history of diabetes.

Results:

• Nominal p < 0.05: 111 proteins
• FDR < 0.1 (lenient correction): 79 proteins
• FDR < 0.05: 47 proteins

Metabolite results were weaker: 55 nominal associations, none surviving FDR < 5%, and 27 surviving FDR < 30%.

Why Do Liver and Immune Signals Lead?

Of the 79 proteins with FDR < 0.1, 36 (46%) were “organ-preferential” — expressed at least 1.5-fold higher in a specific organ than in others, based on the GTEx database.

Among those 36 organ-preferential proteins, the distribution across organs was:

A clarification: the 28%/25%/17% figures use those 36 organ-preferential proteins as the denominator — not all 79, and not all 111. The paper’s Discussion section separately describes approximately 47% of the 111 as organ-preferential, which conflicts with the Results; Results §2.3 and Supplementary Figure S2 should be treated as the authoritative figures: 36 of 79, or 46%.

Why does the technical context matter here? The 1.5-fold organ-specificity threshold is more permissive than the 4-fold cutoff common in comparable studies. The liver is the primary source of plasma proteins, and mass spectrometry platforms naturally favor proteins present in high abundance in circulation — many of which are liver-derived. The prominence of liver signals has a plausible biological basis, but interpreting it requires keeping this technical backdrop in view.

Pathway analysis of the 79 FDR < 0.1 proteins using GO and IPA enrichment implicated inflammation, coagulation, superoxide scavenging and oxidative stress, protein folding, glucose metabolism, and signal transduction. For metabolites, pathway-level enrichment of nominal results (p < 0.05, uncorrected) pointed to caffeine metabolism, the pentose and glucuronate interconversion pathway, biopterin metabolism, and bile acid synthesis. These are exploratory findings — the metabolite input was nominally significant without multiple-testing correction — and should not be read as rigorously validated mechanistic pathways.

Building the Prediction Scores

The researchers used elastic-net regularization to guide feature selection, then built linear prediction scores from Cox models. The procedure: 50 random 1:1 train/test splits, with non-deceased participants undersampled in training sets to balance classes, followed by three-fold cross-validation. The top 20 features were selected based on mean absolute coefficients across iterations. A final linear score was then constructed using Cox coefficients from those selected features.

Top 5 features in ProteinScore:

KLKB1, PHOXF1 (gene symbols as reported; accuracy subject to confirmation), PTGFR, FETUB, APOD

Top 5 features in MetaboliteScore:

phenylacetylglutamine, DHEA-S, pipecolic acid, indoleacetic acid, 3-oxooctanoic acid

One distinction worth flagging: the paper’s Discussion refers separately to 3-hydroxyoctanoic acid and its HCAR3 signaling pathway. That is a different compound; the biological interpretation given there cannot be applied to 3-oxooctanoic acid.

How Well Do the Scores Predict?

Discovery Cohort (Internal)

All C-index figures below reflect internal performance within the discovery cohort. The same 848 participants used for feature selection were also used to evaluate discrimination. These are not external validation results.

A C-index of 0.81 is respectable by the standards of comparable studies, but it was measured on the same data used to build the model. Optimistic bias is nearly unavoidable in that setting, which is exactly why external validation matters.

External Validation (LAST; 231 Participants, 17 Death Events)

The external validation used age and sex as the only adjusting variables — not the full set of seven clinical covariates used internally:

△ Borderline result (the authors’ own term). This does not meet the conventional p < 0.05 threshold, and no 95% confidence interval was reported.

Kaplan-Meier survival analyses split at the median: log-rank p = 0.11 for ProteinScore and p = 0.18 for MetaboliteScore — neither statistically significant.

The external validation set does not report a C-index. The paper’s Discussion explicitly acknowledges that performance declined relative to the discovery cohort, attributing part of this to differences between the two cohorts in demographic composition, educational attainment, environmental exposures, and lifestyle.

With only 17 death events, statistical power to confirm a model’s performance is inherently limited. The MetaboliteScore result is worth following up, but the scale of this validation is not sufficient to say the model has been validated.

Longevity Analysis: 183 People, Nothing Survived FDR

The study also asked a related question: do the molecular signals associated with mortality also relate to living an unusually long life?

Of the 848 participants, 119 (14.0%) met the study’s longevity threshold during follow-up — women reaching at least 85 and men at least 80. The longevity association analysis was not run on the full 848, however. It used a subset of 183 participants with known outcomes: those who had either reached the longevity threshold or had died, excluding anyone still alive but not yet old enough to qualify.

Within those 183 participants, logistic regression was run for each protein and metabolite. Twenty-four proteins and 40 metabolites showed nominal associations (p < 0.05), but none survived FDR correction.

This outcome is unsurprising. With 183 participants and hundreds of parallel tests, detecting a signal that clears FDR requires either a large effect size or a substantially larger sample — and neither condition holds here. The longevity analysis was an honest exploratory attempt; the sample just doesn’t have enough to say.

What Taiwanese Data Can and Cannot Contribute

Reading this study, it is hard not to ask: compared with large European or American cohorts enrolling tens of thousands, what is distinctive about this 848-person Taiwanese dataset?

What it can contribute: population representation.

The molecular composition of older adults is shaped by diet, lifestyle, genetic background, and environmental exposures. A model trained on Taiwanese dietary patterns (rice, seafood, soy products), urbanization patterns, and chronic disease epidemiology should, in principle, predict outcomes in older Taiwanese adults more reliably than one trained on Western cohorts and applied wholesale. This is where local data makes a genuine difference.

What it cannot contribute: causal inference.

This is an observational study. Blood was drawn at a single time point; death was tracked forward. It can identify which molecular levels were higher or lower in the years before death — but it cannot establish whether those molecules caused the deaths, whether they reflect the body compensating for ongoing decline, or whether they are simply markers of an adaptive process.

That question cannot be answered by study design, regardless of whether the cohort is Taiwanese or British, 848 people or 45,000. Causal claims require Mendelian randomization or intervention trials. Observational cohorts can only establish association.

The paper’s Discussion uses phrases such as “systems-level driver” and “potentially modifiable therapeutic target” — language that carries causal implications. This is the kind of directional speculation common in observational research, and the study design does not support those claims. To their credit, the authors acknowledge in the limitations section that they cannot distinguish causation, compensation, and adaptation. That is the interpretive boundary any reader of omics research should keep in mind.

Key Limitations

The authors list four limitations in the paper; additional considerations are incorporated below:

1. All-cause mortality: The 92 death events are not broken down by cause. It is not possible to determine which types of death the liver or immune signals predict most strongly.
2. Small external validation sample: Seventeen death events provide very limited statistical power. Demographic differences between the validation and discovery cohorts may also partly explain the performance decline.
3. Single time point: Only baseline plasma was collected. There is no way to trace how molecular signals change over time.
4. Design defines the scope of inference: Observational design establishes association only; it cannot distinguish cause, compensation, and adaptation.
5. Technical bias: Mass spectrometry platforms favor abundant liver-secreted proteins. The prominence of liver signals has a plausible technical explanation alongside any biological one.
6. Score construction complexity: ProteinScore was not defined directly from elastic-net coefficients. It emerged from a multi-step pipeline — 50 splits, feature selection, then a separately fitted Cox model. The overfitting risk in such procedures requires larger-scale external validation to rule out.

Where This Sits Relative to the First Article

Placing this back on the evidence ladder sketched in the first article: this study occupies the middle tier of observational human research, above animal models and below intervention trials.

Its contributions are:

• Older Taiwanese adults have a distinctive molecular mortality signature, with liver and immune signals most prominent in this dataset
• Protein and metabolite combinations discriminate better than clinical variables alone, at least in the discovery cohort
• The metabolite score held up better than the protein score in external validation and warrants further investigation

The questions it leaves open: can these molecular signals be modified through intervention? If they are modified, does mortality risk actually fall?

Those questions belong to the next rung on the ladder — intervention trials. Observational data, carried this far, has produced a hypothesis worth testing.

Paper source: Huang YL, et al. Liver‐ and Immune‐Enriched Molecular Signatures Associated With Mortality in Older Adults. Aging Cell. 2026;25(7):e70621. DOI 10.1111/acel.70621; PMID 42455656; PMC13372077. CC BY 4.0.