Figure 1. Longitudinal fragmentomics depends on prospectively defined time points, consistent preanalytics, and within-subject comparisons.

cfDNA fragmentomics can track genome-wide changes in fragment size, coverage, end motifs, and nucleosome-associated patterns without requiring every sample to contain a previously known mutation. Longitudinal research is challenging, however, because treatment, tissue injury, inflammation, blood-cell lysis, collection timing, library preparation, and sequencing depth can all change the fragmentome. A defensible study therefore needs a prespecified timeline, standardized plasma processing, technical and biological controls, locked feature definitions, and independent validation. This guide focuses on those design decisions for treatment-response and relapse research.

These services support research use only. CD Genomics does not provide clinical diagnosis or use fragmentomics results to direct individual treatment.

Key Takeaways

• Design the trajectory, not isolated time points. Baseline, early-treatment, end-of-treatment, and follow-up samples serve different research questions.
• Control preanalytics as study variables. Tube type, processing delay, centrifugation, storage, extraction, and library preparation can alter fragment features.
• Separate biological change from total cfDNA change. Fragment ratios and coverage patterns can shift when background cfDNA sources change.
• Lock features and thresholds before validation. Repeatedly tuning a model on the same longitudinal cohort overstates generalizability.
• Plan orthogonal anchors. Tumor fraction, variants, copy-number changes, imaging-derived research endpoints, or tissue data can help interpret fragmentomic trajectories.

Define the Longitudinal Question First

"Monitor response" can mean detecting an early molecular change, classifying an end-of-treatment state, describing rebound before a later study endpoint, or comparing trajectory patterns among cohorts. These are not interchangeable. The primary endpoint should specify the analysis window, comparator, fragmentomic feature set, and how repeated samples from the same participant are handled.

Researchers considering a sequencing workflow can review the cfDNA Fragmentomics WGS Service. The assay plan should still begin with a statistical estimand: for example, within-subject change from baseline to a fixed time point, slope over a defined interval, or time to a prespecified molecular rebound.

Build a Sampling Timeline Around Biology

Baseline should be collected close enough to treatment initiation to represent the pretreatment state, but not after a procedure that acutely changes cfDNA. When feasible, a second baseline can quantify short-term variability and identify unstable starting values. This is particularly helpful when the intended endpoint is a small early change.

Early on-treatment sampling

An early sample may capture rapid changes in cell death, tissue injury, immune activation, or tumor-derived cfDNA. Those processes can move in different directions. A dense pilot schedule can identify the informative window, after which a confirmatory study should use fixed collection days or protocol-defined windows.

End-of-treatment and recovery sampling

An end-of-treatment sample tests the cumulative state but may be influenced by the most recent dose or procedure. Adding a recovery sample helps distinguish a transient treatment-associated perturbation from a sustained trajectory. Record deviations in timing, dose, transfusion, surgery, infection, and other events that could change cfDNA composition.

Follow-up for relapse research

Relapse-oriented research requires an explicit definition of molecular rebound and an independent reference endpoint. Sampling intervals determine the maximum achievable lead-time resolution; a monthly schedule cannot establish a week-level change. Irregular follow-up and dropout may also be related to the participant's condition, so missingness cannot automatically be treated as random.

The cfDNA fragmentomics study-design guide covers general cohort architecture. In a longitudinal study, add a visit-window table that defines acceptable timing, repeat-collection rules, and which sample takes priority if several are available in one window.

Figure 2. Each collection window should have a biological rationale, an allowable timing range, and a prespecified analysis role.

Standardize Preanalytics Across Every Visit

Fragmentomics is sensitive to how blood becomes sequence data. Leukocyte lysis can add long genomic DNA, processing delay can change fragment distributions, and extraction or size selection can preferentially recover certain fragments. If preanalytic conditions shift with visit or outcome group, the study may learn workflow differences instead of biology.

Create one sample manual for all sites and visits. It should define blood volume, collection tube, mixing, storage temperature, maximum time to first centrifugation, centrifugation steps, plasma aliquot volume, freeze–thaw limit, extraction input, elution volume, and library method. Deviations should be captured as analyzable metadata rather than buried in notes.

Matched buffy-coat DNA is useful for identifying germline variants and clonal hematopoiesis when genomic features are interpreted alongside fragmentomics. Plasma blanks, extraction blanks, library controls, and reference materials help identify contamination and batch drift. The Liquid Biopsy Solutions page provides broader service context for coordinated plasma and genomic analyses.

Choose Fragmentomic Features Before Looking at Outcomes

The fragmentome can be represented in many ways: mono- and di-nucleosomal size distributions, short-to-long fragment ratios, genome-wide fragmentation profiles, coverage around transcription start sites, nucleosome footprints, fragment-end coordinates, end motifs, jagged ends, and regional entropy measures. Each feature family has different sensitivity to depth, library chemistry, and background cfDNA composition.

Low-pass Whole Genome Sequencing provides broad coverage for genome-wide size and coverage features. Targeted designs can achieve greater depth in selected regions but change the observable fragment population. EPIC-seq, for example, uses promoter-focused fragmentation features to infer gene-expression states [3]. A targeted feature should not be assumed to reproduce a WGS-derived metric.

Build a feature specification

Before model fitting, document:

• fragment inclusion rules, alignment quality, duplicate policy, and valid insert-size range;
• genomic bins or regions, blacklist handling, GC correction, and mappability filters;
• normalization denominator and whether sex chromosomes are included;
• feature aggregation across windows and minimum usable molecules;
• rules for batch correction and missing feature values;
• the exact software version and parameter set.

The cfDNA fragmentomics report-metrics guide helps connect these definitions to deliverables. Locked specifications prevent an apparent longitudinal signal from changing whenever software or filtering is updated.

Model Within-Subject Change Without Losing Cohort Context

Repeated samples from one participant are correlated. Treating them as independent observations produces overly narrow uncertainty and can leak participant-specific patterns into validation. Use subject-aware splits and statistical models that represent repeated measures.

Absolute values, deltas, slopes, and nadirs

Absolute feature values may be useful when baseline samples are unavailable, but they are vulnerable to stable between-person differences. A baseline-normalized delta controls some of that variation. Slopes use several visits but depend on interval spacing. Nadir-to-rebound metrics can fit relapse research but require a prespecified nadir definition and enough post-treatment samples.

Cancer-treatment monitoring research using cfDNA fragmentomes has demonstrated the feasibility of analyzing serial low-pass WGS samples [1], while longitudinal low-coverage WGS during high-dose radiotherapy illustrates how treatment itself can influence cfDNA features [2]. These studies reinforce the need for standardized timing and cautious attribution.

Prevent data leakage

All samples from one participant must remain in the same training, tuning, or validation partition. Feature selection, normalization parameters, batch correction, and thresholds should be learned from the training data and then frozen. If visits from one person appear in both training and testing sets, the model may recognize the person rather than the biological trajectory.

External validation should reproduce the intended collection schedule and preanalytic workflow as closely as possible. When that is impossible, transportability analysis should test whether feature distributions shift by site, tube, batch, or cohort.

Interpret Fragmentomics With Orthogonal Anchors

Fragmentomic change is not automatically tumor-specific. Treatment-associated tissue injury, infection, inflammation, exercise, hematopoietic shifts, and sample handling can alter cfDNA. Orthogonal measurements can help evaluate competing explanations.

Useful anchors include mutation-based ctDNA, copy-number-derived tumor fraction, methylation, protein biomarkers, imaging-derived research measurements, pathology from scheduled procedures, and tissue or blood-cell transcriptomics. Multi-modal cfDNA studies have shown that genomic and fragmentomic signals can provide complementary information [5]. The multi-signal liquid biopsy guide discusses how to combine evidence layers without assuming that every signal has the same source.

An orthogonal result does not need to match fragmentomics at every visit. Discordance can be biologically informative, but it should be investigated against timing, assay sensitivity, tumor fraction, and preanalytics. Define in advance whether the fragmentomic endpoint is confirmatory, supportive, or exploratory relative to other measures.

Figure 3. Orthogonal anchors help distinguish a fragmentomic trajectory from changes in total cfDNA, sample handling, or background tissue contributions.

Design Response and Relapse Analyses Differently

Response-oriented studies usually compare a defined early or end-of-treatment time point with baseline. Relapse-oriented studies evaluate change after a post-treatment reference state. Combining these questions in one endpoint can obscure the relevant window and inflate the number of analytical choices.

Response research

Specify the earliest time at which a biological change is plausible, the collection window, and the external research endpoint used for comparison. If several early time points are explored, reserve a later cohort to confirm the selected window. Account for treatment class because different interventions can produce distinct cell-death and inflammatory kinetics. A rectal-cancer study using cfDNA fragmentomics to predict pathological response provides one example of a response-oriented endpoint tied to a defined treatment setting [6].

Relapse research

Define the post-treatment reference, minimum follow-up, molecular rebound rule, confirmation requirement, and censoring strategy. A single high value may be a technical outlier; requiring a repeat rise can improve specificity but delays the event. Both consequences belong in the study plan.

Serial cfDNA research in multiple myeloma has shown how longitudinal molecular measurements can identify emerging treatment failure [4], while fragmentomic studies have linked multi-modal patterns with recurrence-related outcomes [5]. These findings motivate longitudinal designs but do not remove the need for independent validation in the intended cohort.

The fragmentomics versus targeted cancer panels guide compares genome-wide and mutation-focused evidence. A combined design can use mutation tracking as an anchor while preserving fragmentomics as a separately validated signal.

Plan QC, Failure Rules, and Missing Data

QC thresholds should be established before outcomes are analyzed. Typical categories include plasma input, extracted cfDNA yield, library complexity, usable paired reads, mapping rate, duplication, contamination, insert-size profile, genome-bin coverage, GC bias, and feature completeness. A sample can pass general sequencing QC yet fail a specific fragmentomic feature because that feature needs more usable molecules.

Predefine whether a failed visit is repeated, excluded, or retained with a missing value. Longitudinal models should not silently replace missing visits with the last observed value. Document why a sample is missing, because collection failure, inadequate plasma, low yield, and participant dropout have different implications.

Batch controls should be trended across plates and sequencing runs. When a reference material changes distribution, investigate the batch before applying a computational correction. Computational harmonization cannot prove that a biological fragment distribution survived inconsistent sample handling.

Pilot acceptance ranges should be established on the same sample type and workflow intended for the main study. A threshold imported from another laboratory, tube type, or library method may reject valid material or accept unstable data. Trend continuous QC values instead of relying only on pass/fail labels, because gradual drift can precede a visible batch failure.

Specify the Final Data Package

A complete longitudinal package connects each sequence file to its participant, visit window, preanalytic record, QC status, and feature vector. It should include a locked analysis manifest, sample-level QC, normalized and unnormalized features, subject-level trajectories, group summaries, model coefficients or versioned code, and uncertainty estimates.

The report should clearly label exploratory findings, prespecified endpoints, and independently validated results. For every trajectory figure, show actual collection times rather than assuming equally spaced visits. Include the number at risk or number of evaluable samples over time so attrition remains visible.

Figure 4. A decision-ready report preserves visit timing, preanalytic metadata, QC, trajectories, model version, and validation status.

FAQ

For research use only. Not for use in diagnostic procedures or individual treatment decisions.