How leading labs turn cfDNA into a disease signal

A practical comparison of feature engineering, statistics, and machine learning across fragmentomics and methylation programs

The shared pattern

The leading programs differ in assay and commercial maturity, but their analytical logic is similar: preserve molecule-level information, compress millions of observations into biologically motivated features, control nuisance variation, train a model without information leakage, and test it in people who were not used to develop it.

Different molecular and statistical signals converge into an interpretable cfDNA fingerprint.

Five analytical styles

Program	Primary signal	Statistical/ML pattern	Transferable lesson
Johns Hopkins / DELFI	Genome-wide fragment-size and coverage profiles	Window-level feature engineering followed by supervised classifiers and held-out validation	Lock preprocessing and summarize the genome before modeling
CUHK plasma-DNA program	Fragment ends, sizes, nucleosome patterns, methylation, tissue of origin	Mechanism-led scores, probabilistic/deep-learning models, and independent cohorts	A feature should have a biological and technical model, not only predictive importance
UCLA methylation methods	Read-level CpG patterns and reference methylomes	Probabilistic mixture/tissue-localization models	Model plasma as a mixture and validate reference-panel transferability
Stanford single-molecule work	Methylation and physical length on the same molecule	Joint molecule-level features and multivariable classification	Preserve linkage between features when the platform allows it
Tagomics	Enriched methylation signal and atlas-based tissue deconvolution	Proprietary feature selection, deconvolution, and prediction	Enrichment may improve signal, but independent validation and transparent references still matter

These are families of approaches, not an assertion that all underlying code or training data are public. Commercial performance claims cannot be reproduced from a website alone.

What “fingerprinting” means statistically

A disease fingerprint is not a single heatmap. It is a pre-defined vector of measurements that may include:

• DMR methylation levels and pathway summaries;
• fragment-size quantiles and short/long ratios;
• mono-/di-nucleosomal peak characteristics;
• end-motif frequencies or diversity scores;
• TSS or open-chromatin coverage summaries;
• inferred tissue contributions; and
• sIgA, inflammatory, and clinical covariates.

The feature vector becomes credible only when the complete transformation—from raw reads to one row per participant—is reproducible and frozen before final evaluation.

The modeling sequence

flowchart LR
  R["Raw reads + assay data"] --> Q["Sample and batch QC"]
  Q --> F["Biology-led features"]
  F --> S["Scaling + feature selection<br/>inside resampling"]
  S --> M["Parsimonious model"]
  M --> V["Nested/held-out validation"]
  V --> C["Calibration, uncertainty,<br/>decision value"]

For this study

With 18 participants, the safest target is feature feasibility and effect estimation, not a production classifier. If a model is explored, use a very small locked feature panel, penalization, participant-level resampling, and bootstrap stability. All filtering, imputation, scaling, and feature selection must occur inside each training fold.

Useful baselines are often more informative than an elaborate neural network:

1. clinical covariates only;
2. methylation only;
3. fragmentomics only;
4. sIgA/inflammation only; and
5. a small integrated model.

Compare incremental performance, calibration, and stability—not just apparent AUC. A model that perfectly separates 18 discovery samples may simply encode batch, library depth, or an unstable feature.

Progression needs longitudinal data

Cross-sectional UC, CRC, and healthy groups can support discrimination hypotheses. They cannot show within-person progression. Progression analysis requires repeated samples and dated clinical outcomes; mixed-effects models, joint longitudinal–time-to-event models, or landmark prediction then become relevant. Until those data exist, call the output a candidate disease-state fingerprint, not a progression predictor.

ImportantThe strongest transferable lesson

The best laboratories do not begin with an algorithm. They begin with a measurement model: what created the feature biologically, what can distort it technically, and what independent data would falsify the claim.

See the technology comparison → · Read the package guide →