Composition-Adjusted Pathway Enrichment from Cell-Free DNA Fragment Lengths

Blood plasma carries short pieces of DNA released by dying cells, called cell-free DNA. The length of each piece records whether its region of the genome was tightly packed or open in the cell it came from. Active promoters, the switch regions just upstream of genes, are open, so they are cut more often and leave shorter, sparser fragments. In principle this means a routine low-coverage blood sequencing run already contains a readout of which genes, and so which biological pathways, were active in the cells shedding DNA. The obvious way to use that readout is to score fragmentation at every promoter, rank genes, and run a standard pathway test. That approach has a flaw: plasma DNA is a mixture dominated by blood cells, so a pathway can look "activated" only because the mix of contributing cell types shifted. Inflammation adds neutrophil DNA, a tumor dilutes the blood-cell share, and immune pathways then light up with no change in regulation inside any cell. This is the old problem of cell mixtures hiding what individual cell types are doing, here showing up as false pathway calls.

We propose a pathway test for cell-free DNA that first predicts how fragmentation at every promoter should look given the sample's estimated cell-type composition. It then asks whether a gene set's fragments deviate from that prediction. The pathway is the natural unit for a second reason as well. At 0.1–2X coverage a single promoter holds only a few fragments, but pooling fragments over the hundreds of promoters in a gene set gives usable counts. Grouping by pathway is therefore how the signal becomes measurable, not only a summary added at the end. The deliverable is a calibrated test that separates "this pathway's promoters are more open than the cell mixture explains" from "the cell mixture changed," together with a benchmark that measures this separation directly.

Core Claim

Pathway enrichment computed naively from cell-free DNA fragmentation will mostly rediscover composition shifts. Anchoring the expected promoter fragmentation to reference chromatin-accessibility maps, weighted by per-sample cell-type proportions, removes these false calls while keeping real regulatory signals such as lineage transcription-factor programs. It also works on a single sample, without a cohort to regress against.

Where This Sits

The fragmentation biology comes from established work. The roughly 167 bp nucleosomal peak and the shortening of tumor-derived fragments are described in Inferring Chromatin Architecture from Cell-Free DNA Fragmentation. Jensen et al. report that shorter fragments are more likely to be unmethylated, and FinaleMe turns fragment length and position into methylation calls. FinaleMe also warns of degraded performance in CpG-poor regions. That warning constrains our design: our null distribution must match genes on CpG and GC content.

Two lines of work already aggregate fragmentation over defined sets of loci, which shows that the pooling idea works.

  • DELFI profiles fragment-size ratios in large genomic windows. It separates small-cell from non-small-cell lung cancer through ASCL1 binding sites, reporting an AUC of 0.98.
  • The NMF fragment-length signature work breaks length histograms into tumor and normal components and 250 kb bin factors.

Neither method maps its signal onto curated pathways. Both treat composition only as a single tumor-versus-normal axis.

The method we reuse for composition comes from LIONHEART. It correlates bias-corrected cell-free DNA coverage with open-chromatin maps from 487 cell and tissue types and shows that blood-cell dilution drives much of its pan-cancer signal. LIONHEART stops at a diagnostic score. We use its composition estimate as the baseline that pathway signals must exceed. cfDecon reaches finer deconvolution from methylation reads, but it needs bisulfite conversion, which our WGS-only setting avoids.

On the enrichment side, classic ORA and GSEA take gene lists or ranked statistics. funkea bundles tests such as S-LDSC that estimate enrichment conditional on baseline annotations. That conditioning logic is what we carry over to fragmentomics, with cell-type composition as the baseline. gtCLIP collapses redundant significant pathways; we use it only as an optional reporting step. POPPY and Count Bridges handle composition-aware interpretation and deconvolution for RNA. They motivate the problem but cannot read cell-free DNA.

The research file found no work that tests pathway enrichment from cell-free DNA fragmentation, with or without composition adjustment. The contribution here is not "the first fragmentomic GSEA." It is the claim, and a demonstration, that such an analysis is uninterpretable unless composition is conditioned out.

Research Program

For each sample $s$ and promoter $g$ (±1 kb around the transcription start site), we compute two GC-corrected statistics: the short-fragment fraction (fragments under 150 bp) and the relative coverage depletion. We then estimate cell-type proportions $\hat\pi_{s,c}$ by fitting LIONHEART-style correlations to the 487 reference maps. The fit is constrained to non-negative weights over a reduced set of lineages, so the result can be read as a set of proportions.

From healthy-control plasma we learn a calibration function $f$ that maps reference accessibility $A_{g,c}$ (cell type $c$ at promoter $g$) to the expected fragmentation statistic. The composition-predicted expectation and the residual are

$$ \hat y_{s,g} = f\Big(\sum_c \hat\pi_{s,c}\,A_{g,c}\Big), \qquad r_{s,g} = y_{s,g} - \hat y_{s,g}, $$

where $y_{s,g}$ is the observed statistic.

A gene set $S$ is scored by pooling the fragments in its promoters rather than averaging noisy per-gene residuals. The resulting statistic $z_{s,S}$ compares the pooled residual against residuals of size-matched random gene sets drawn from the same strata of CpG density, GC content and baseline accessibility. This gives a single-sample enrichment p-value. For case-control questions, residual scores are compared across groups.

We will also build the benchmark that this question has lacked. It uses three kinds of semisynthetic plasma, all made from real reads:

  1. Composition-only mixtures. Healthy plasma is mixed at known ratios with high-tumor-fraction plasma, as in the serial-dilution design of Predicting tumour content of liquid biopsies, or with additional healthy samples whose blood-cell shares differ. Any pathway called here beyond the composition expectation is a false positive.
  2. Implanted-signal mixtures. Fragments at promoters of a randomly chosen gene set are resampled toward a shorter length distribution at controlled effect sizes, while composition is varied independently. These measure power.
  3. Real biological positive controls. The leading case is ASCL1 target genes in small-cell lung cancer from the DELFI data, which should survive adjustment because they reflect within-tumor regulation.

Baselines

The do-nothing reference is fragmentomic GSEA without adjustment. It ranks promoters by the same raw short-fragment and depletion statistics and runs standard GSEA and Fisher's exact test against the same pathway collection, the GO, KEGG and Reactome sets used by gtCLIP. This is the analysis most groups would run first. If it is already well calibrated on composition-only mixtures, our premise is wrong.

The strongest competitor is cohort-level covariate adjustment. Each promoter statistic is regressed on the estimated proportions $\hat\pi$ across samples, and GSEA is then run on the residuals. This is the standard fix for composition confounding in bulk RNA analysis, and it needs no reference maps. It will use the same proportion estimates, gene sets, matched-null construction and fragment pooling as our method, so the only difference is the source of the expectation: reference-anchored and per-sample, or fitted across the cohort.

The third baseline is LIONHEART itself, read as an interpretation rather than a classifier. It asks whether the 487 cell-type scores alone already explain the outcome-associated variation that pathway scores claim to capture. For diagnostic context, LIONHEART reports a pan-cancer ROC AUC of 0.80 ± 0.11 across 8 datasets, a leave-one-dataset-out average of 0.797 (weighted 0.823), and 0.859 on the external Zhu et al. cohort. Those figures come from LIONHEART and are the numbers any added pathway features must improve upon.

For the enrichment task itself, there is no reported result found on Lacuna for pathway-level enrichment from cell-free DNA fragmentation. The comparison is therefore internal to our benchmark.

Ablations isolate the key ingredient. One removes the composition term, leaving $\hat y$ as the healthy mean. One replaces reference-anchored $\hat y$ with the cohort regression. One drops fragment pooling in favor of per-gene ranking. One drops CpG/GC matching of the null.

Beating the baselines means four things. On composition-only mixtures, the false-positive rate at nominal 5% must stay near 5% while unadjusted GSEA is clearly inflated. Power under implanted signals must be at least equal to the cohort-regression baseline. The method must keep a well-calibrated test in single-sample use, where cohort regression cannot be fit. ASCL1-type positive controls must stay significant after adjustment.

Why not just…?

Why not just run GSEA on promoter fragmentation ranks? Because the dominant source of variation in plasma DNA is which cells contributed it. LIONHEART's own analysis shows blood-cell dilution carries much of the cancer signal, and Liao et al. describe large fragmentation shifts in sepsis. Lineage-marker pathways will therefore track composition. The composition-only mixtures measure this directly. If naive GSEA turns out to be calibrated there, we will report that the adjustment is unnecessary.

Why not just regress out estimated proportions across the cohort? This is the real threat. With large, balanced cohorts it may perform as well as our method, and we will say so if it does. It has two structural weaknesses that the experiments test. It needs many samples, so it cannot score one patient. When composition correlates with disease status, it can also absorb the disease signal itself. The implanted-signal mixtures vary that correlation deliberately to show where each approach fails. If cohort regression matches us everywhere, the contribution narrows to the benchmark and the pooled, matched-null test, and the paper will be framed that way.

Why not just interpret LIONHEART's cell-type scores and skip pathways? Cell-type proportions answer "who is shedding," not "what program is active." The question that matters for things like therapy-induced changes inside tumor cells is the second one. We test whether pathway residuals add predictive information beyond the 487 scores using leave-one-dataset-out validation. A null result would mean the pathway layer is redundant for diagnosis, although it could still help interpretation.

Why not just use methylation sequencing with standard enrichment? Bisulfite conversion destroys much of the scarce input DNA, and most clinical archives are plain WGS, as the FinaleMe authors argue. FinaleMe's 0.89 correlation with WGBS holds in CpG-rich regions. Promoters are mostly CpG-rich, which is exactly why a fragmentation-only promoter readout is plausible. Methylation-based enrichment would face the same composition confound in any case.

Decisive Experiments

Data come only from cohorts in the file. These are LIONHEART's 8 pan-cancer WGS datasets and the Zhu et al. validation set, the DELFI and LUCAS shallow-WGS samples, the 142 prostate-cancer sWGS samples, the serial-dilution series from the stage IV cohort, and the 712 serial colorectal MRD samples for longitudinal stability.

The first check is calibration. We plot quantiles of observed against expected p-values on composition-only mixtures for every method. Our method passes only if its p-values follow the uniform null while unadjusted GSEA departs from it.

The second check is power across coverage. Implanted-signal recovery is measured at 0.1X, 0.5X and 2X downsampling to show where pathway pooling makes single-sample calls possible.

The third check is biological. After adjustment, ASCL1 target sets must separate the two lung cancer types in the DELFI data. In the colorectal series, pathway residuals should move with mutation-defined tumor-positive status, using the labels of Combining mutation detection with fragmentomics features, and not with blood-cell share.

The fourth check is cross-cohort portability. Pathway calls must be reproducible under leave-one-dataset-out validation, given the pre-analytical sensitivity documented across these cohorts.

The main limitation is that the file lists no cohort with matched tissue RNA and plasma WGS. Ground truth for real (non-implanted) pathway activity therefore rests on the semisynthetic design and on lineage positive controls rather than on direct expression measurements.