🧬 biology

Static DNA-Methylation Burden Discriminates Hematologic-Malignancy Subtypes but Does Not License Per-Patient Trajectory Inference: A Public-Cohort Analysis and an Identifiability Framework

This study demonstrates that while static DNA-methylation burden and entropy can effectively classify hematologic malignancy subtypes and distinguish them from healthy donors using public cohort data, such cross-sectional measures are insufficient for inferring individual disease trajectories, necessitating prospective serial sampling to validate lead-time estimation.

Original authors: Frederic Scheer

Published 2026-07-23

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Frederic Scheer

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Body's Hidden Diary and the Mystery of the Missing Time Machine

Imagine your body is a bustling city, and inside every cell, there is a tiny, invisible diary written in a chemical code called DNA. Most of the time, this diary stays the same, recording the instructions for how to be a healthy blood cell. But sometimes, the city gets a little chaotic. Before a disease like blood cancer even shows up with symptoms, the cells start making strange, frantic notes in their diaries. They add extra marks or erase important ones. Scientists call this "DNA methylation." It's like a chemical highlighter pen that cells use to turn genes on or off. When the highlighter goes wild, it's often the very first sign that something is wrong, appearing months or even years before a doctor could ever find a lump or a fever.

The big question scientists are asking is: Can we read these messy diaries to catch the disease early? And even better, can we tell how fast the disease is coming? If we know a patient is sick, that's one thing. But if we could say, "You are moving toward cancer, and you have two years to fix it," that would be a medical superpower. This is the difference between taking a snapshot of a car crash and predicting the crash before the car even leaves the driveway. The paper you are about to read dives deep into this mystery, using a massive library of public data to see what we can actually prove right now, and what is just a wild guess.

The Snapshot vs. The Movie

This paper is like a detective who has a stack of photos from a crime scene but no video footage. The author, Frederic Scheer, wanted to solve a tricky puzzle: Can we look at a single snapshot of a person's blood (a "static" view) and tell them exactly how fast they are heading toward blood cancer?

The answer, according to this study, is a firm "No, not yet."

Here is the story of what the paper found, broken down into the good news and the hard truth.

The Good News: We Can Spot the Trouble

The researchers looked at thousands of blood samples from public databases. They focused on a specific type of blood cancer (like chronic myeloid leukemia) and compared the blood of sick people to healthy people. They measured something called "methylation burden."

Think of this like checking a room for mess. If a room is perfectly tidy, the "mess score" is zero. If someone has been throwing papers everywhere, the score goes up. The team found that they could build a very sharp "mess detector."

They created a test called qMethyl-48. It looks at 48 specific spots in the DNA diary.
When they tested this on 656 healthy people, the test was incredibly good at saying, "This person is fine." It was right 96.6% of the time.
When they tested it on people who already had certain blood cancers, it caught almost all of them, with a success rate between 94.9% and 100%.

So, the "snapshot" works great. If you take a picture of a patient's blood right now, this test can tell you if they are currently in a dangerous state. It can even tell you which kind of blood trouble they have. This is a powerful tool for sorting people into "high risk" and "low risk" groups right now.

The Hard Truth: We Can't Predict the Future (Yet)

Here is where the paper gets serious and stops us from getting too excited. The researchers argued that while we can take a great snapshot, we cannot use that single photo to predict the movie.

Imagine you see a car parked on a hill. You can measure how steep the hill is (that's the "static burden"). But if you only have one photo, you have no idea if the car is:

Parked with the brakes on.
Slowly rolling down.
About to roll down at full speed.
Or if it's actually rolling up the hill.

Some scientists have tried to guess the speed by saying, "Well, if the car is this far down the hill, and we assume it started at the top, it must be moving at X miles per hour." The author of this paper says: That is a circular trap. You are just guessing the speed based on the distance, and then using that guessed speed to prove the distance. It's like saying, "I know the car is moving fast because it's far away, and I know it's far away because it's moving fast."

The paper explains that to know the speed (velocity) or when the crash will happen (arrival time), you need a time axis. You need a video, not a photo. You need to see the same person's blood at least two or three times over months or years. Without that "time travel" data, any claim about how fast a disease is coming is just a guess, not a measurement.

The Blueprint for the Future

The paper doesn't just say "we can't do it"; it builds a blueprint for how we could do it in the future. The author suggests a new kind of study where doctors would take blood samples from the same people over and over again (serial sampling).

They set up a "checklist" or a series of gates that this future study must pass to prove it works:

Gate 1: Can we actually see a pattern in the changes over time?
Gate 2: Does the pattern separate people who get sick from those who don't?
Gate 3: Can we prove the changes are actually causing the disease and not just happening because of age or inflammation?
Gate 4: Is the data clean enough to trust?

If a future study passes all these gates, then we can finally say, "Yes, we can predict the future." Until then, the paper insists we must be honest: we have a great tool for checking the current risk, but we do not have a crystal ball for the future speed.

The Bottom Line

This paper is a masterclass in scientific honesty. It says, "We have a fantastic flashlight that can find the danger right now." But it also warns, "Don't pretend that flashlight can show you the path the danger will take tomorrow."

The "static" test (the snapshot) is ready to be used in labs today to help doctors sort out who needs extra care. But the "dynamic" prediction (the movie) is still a dream that requires a new kind of experiment to become real. The author argues that separating these two ideas is crucial: if we claim we can predict the future when we can't, people might stop trusting the very real, very useful tool we do have for the present.

Technical Summary: Static DNA-Methylation Burden vs. Trajectory Inference

Problem Statement
The paper addresses the critical distinction between two interpretations of DNA methylation data in hematologic malignancies:

Static Classification: Determining if an individual's current methylation state is anomalous relative to a healthy baseline (a solvable classification problem).
Dynamic Trajectory Inference: Determining if that state is moving toward disease with a measurable velocity, arrival time, and lead time (a dynamical system problem).

While pre-diagnostic cohorts show genome-wide methylation deviations months to years before clinical diagnosis, the paper argues that the scientific community often erroneously conflates static deviation with dynamic motion. The core problem is that "trajectory language" (velocity, lead time) is easily asserted but structurally impossible to derive from cross-sectional (single-timepoint) or sparse two-point data, as these datasets lack the time axis required to define a derivative.

Methodology
The study utilizes a secondary analysis of publicly available, de-identified Illumina HumanMethylation450 and EPIC array data. No proprietary, electronic health record, or individual patient data were used.

Metrics Defined:
- Static Methylation Burden (MB): The mean absolute $\beta$ -deviation of a sample from a healthy whole-blood reference distribution.
- Regulatory Entropy (H): The normalized Shannon entropy of that deviation distribution.
Screening Tool: A locked, directional 48-CpG burden screen (qMethyl-48) was evaluated. The specific locus composition is withheld pending intellectual property prosecution. The screen scores a sample positive if $\ge$ 4 loci deviate.
Cohorts:
- Reference: 656 healthy whole-blood donors (GSE40279).
- Disease: Public cohorts for Chronic Myeloid Leukemia (CML), Chronic Myelomonocytic Leukemia (CMML), and Myeloproliferative Neoplasms (MPN).
Theoretical Framework: The paper applies State-Space Reconstruction theory (specifically Takens' embedding theorem and fractal delay-embedding) to analyze the identifiability of trajectory recovery. It posits that recovering dynamics requires a time series sampled with sufficient delays to reconstruct the system's invariant manifold, a condition not met by cross-sectional data.

Key Results

Static Discrimination: Static burden and entropy successfully discriminate WHO-2022 hematologic malignancy subtypes (specifically myeloid) from healthy donors.
Screen Performance: Against the 656 healthy donors, the qMethyl-48 screen achieved 96.6% specificity at its locked threshold. Sensitivity ranged from 94.9% to 100% across CML, CMML, and MPN cohorts.
Limitations: These results are strictly single-timepoint classification measures. They confirm the presence of a separable static signal but provide no evidence regarding the rate or direction of change.
Myeloid vs. Lymphoid: The results apply to myeloid subtypes. The paper explicitly notes that lymphoid diseases (e.g., CLL), which often exhibit global hypomethylation rather than the hypermethylation burden indexed by this screen, are expected to behave differently and are not covered by these claims.

Theoretical Contributions: The Identifiability Framework
The paper's primary theoretical contribution is a rigorous argument against trajectory inference from static data and a formal framework for what is required to earn such claims:

Circularity of Constructed Velocity: Defining velocity as $v = (\beta_{arrival} - \beta_{healthy}) / T_{lead}$ is algebraically a rescaled static burden. It measures no time derivative and creates a circular correlation between the assumed lead time and the calculated velocity.
Structural Obstacles: Trajectory quantities (velocity, arrival, conserved energy) are properties of dynamical systems observed over time. Cross-sectional data has no time axis; two-point data represents a degenerate minimum.
Recoverability Ceilings: Even with time series, state-space reconstruction has limits. Direct driving vs. indirect synchronization cannot be resolved without removing confounders (age, CHIP, cell composition, inflammation).
Noise Floor: Measurement noise (2–4%) sets a hard floor; velocity calculated as a difference of noisy $\beta$ -values over short intervals amplifies noise, bounding feasible sampling intervals.

Proposed Path Forward: A Serial-Sampling Program
To move from static risk stratification to dynamic trajectory inference, the paper proposes a prospective serial-sampling program with pre-specified "go/no-go" gates. This program is designed so that a negative result at any stage is a publishable finding that refutes the trajectory hypothesis. The five gates are:

Q1 (Embedding): Verify a stable embedding dimension is recoverable across subjects.
Q2 (Convergence): Demonstrate continuity-detection statistics separate progressors from non-progressors against a null.
Q3 (Attractor): Validate that individual trajectory limit sets coincide with a diagnosis centroid.
Q4 (Confounding): Prove the driving relationship survives the removal of age, CHIP, cell composition, and inflammation.
Q5 (Provenance): Confirm data provenance and that per-probe precision is below the noise floor.

Significance and Claims
The paper concludes with two distinct, non-conflated findings:

Immediate Utility: A static, single-timepoint methylation-burden classifier is currently supported by public data. It discriminates subtypes, validates externally, and is deployable as a laboratory-developed test for risk stratification.
Structural Limitation: The dynamic reading (velocity, lead time) is not licensed by cross-sectional data. The limitation is structural (absence of a time axis), not biological.

The significance of the work lies in separating "what is usable now" from "what must be prospectively earned." By framing trajectory recovery as a state-space-reconstruction problem and specifying a gated experimental design, the paper aims to prevent the field from making circular kinetic claims while providing a rigorous pathway to substantiate them if they are biologically real. The authors emphasize that a framework specifying the experiment capable of disproving the hypothesis is a strength, ensuring that future claims are based on pre-specified evidence rather than assertion.