Biomarker Clocks in Medicine: Turning Blood Tests into Time Predictions
How 'disease clocks' convert biomarker levels into timelines, why Alzheimer's p-tau217 is a strong fit, and where this paradigm breaks down.
On this page
The shift: from “positive/negative” to “where are you on the timeline?”
Most biomarker conversations collapse into a binary: positive (pathology present) or negative (not detected).
But clinicians and trialists often need a different answer:
“How far along is this person — and what is the likely time window to a clinical transition?”
That’s where disease clocks come in.
A recent Nature Medicine study built plasma %p-tau217 clocks to estimate the age of biomarker positivity and model the likely onset of symptomatic Alzheimer’s disease. Primary source: https://www.nature.com/articles/s41591-026-04206-y
Disease Clocks: Biomarkers as Time
A disease clock is a way of translating a biomarker level into an estimate of where someone is on a disease timeline—for example, “years since biomarker positivity,” rather than just “positive vs negative.”
What it does
- Anchors people to a biological event (like crossing a biomarker threshold).
- Aligns trajectories so individuals can be compared in “disease time.”
- Supports forecasting of clinical transitions (like symptom onset) at the group level.
Why it’s useful
- Reduces noise created by relying on chronological age alone.
- Helps identify who is likely to convert within a trial window (e.g., 3–5 years).
- Makes staging intuitive: “years since positivity” is easier to reason about than raw units.
Alzheimer’s example
In Alzheimer’s research, plasma %p-tau217 can be modeled longitudinally to estimate the age a person likely became biomarker-positive. That estimated “positivity age” can then be used to model the expected age of symptom onset (with meaningful error margins).
Key limitation
A clock is not a personal countdown timer. Prediction error, co-pathologies, and population differences can be large—especially when applied to individuals rather than cohorts.
What a “disease clock” really is
A disease clock is a mapping:
biomarker level → inferred disease time
Instead of staging people by chronological age, clocks stage them by time since a biological event (for example, crossing a biomarker positivity threshold).
This gives you a “disease-time coordinate system” that often looks like:
- years before positivity (negative time)
- year 0 (positivity)
- years after positivity (positive time)
When the biology supports it, this alignment can reduce heterogeneity and improve forecasting.
Why p-tau217 is a strong candidate for clock modeling
Alzheimer’s disease is unusually “clockable” because:
- pathology unfolds over many years
- biomarker trajectories show relative consistency once initiated
- there is a meaningful sequence of biological changes
In the study, %p-tau217 was used because it tracks Alzheimer’s biology across the preclinical and early symptomatic phases and correlates with amyloid and tau PET.
The key finding most people miss: age changes the conversion window
A headline like “predicts symptom onset within ~3–4 years” is attention-grabbing, but the deeper point is this:
The time from biomarker positivity to symptom onset appears shorter in older individuals.
That means a biomarker value does not translate to a single universal “time-to-symptoms.” The system context (age, co-pathology burden, resilience) changes what the same biology means clinically.
This is also why simplistic “plaque = dementia” interpretations fail. Pathology matters, but clinical onset is a systems outcome.
Where this fits in AI in Health
Even without deep learning, this is a core AI-style maneuver:
- Learn progression dynamics from longitudinal data
- Convert a raw measurement into an inferred latent coordinate (“disease time”)
- Predict an outcome using the coordinate system rather than raw features alone
This is a pattern you’ll see across predictive medicine:
- sepsis trajectories
- kidney function decline modeling
- oncology progression curves
- neurodegeneration staging
Limits: when clocks break
Clocks don’t generalize automatically.
They work best when:
- biomarker change is consistent across individuals within a usable range
- training data spans enough time to learn dynamics
- the clinical transition is well-defined
They break down when:
- biology is highly heterogeneous
- preclinical windows are short
- biomarkers reflect current damage more than a long, staged accumulation
Could this work for ALS?
In principle, yes — but with a different goal.
ALS often has:
- shorter presymptomatic windows
- greater heterogeneity in onset and progression
- biomarkers that behave more like “injury meters”
So the useful model in ALS tends to be:
- trajectory prediction (rate of decline)
- near-term milestone prediction rather than a long-range “when will symptoms start” clock.
Why this matters (now): trial design
The near-term, high-confidence impact of clocks is not personal forecasting.
It’s trial enrichment:
- enrolling people likely to convert within the study window
- increasing power
- shortening timelines
That’s how you move from “biomarkers are interesting” to “biomarkers change the economics of prevention trials.”
FAQ
Is a biomarker clock the same as an AI model?
Not necessarily. A biomarker clock can be built using statistical or mathematical modeling without machine learning. The AI-relevant idea is the transformation of raw measurements into an inferred disease-time coordinate.
Why are biomarker clocks useful in clinical trials?
Trials often need endpoints within a fixed window. Clocks can enrich enrollment with people closer to conversion, improving efficiency.
Could this work for fast diseases like ALS?
Sometimes, but ALS clocks are usually about decline rate and near-term outcomes, not decades-ahead onset timing.