Eight sources compare plasma cfDNA fragmentomic workflows on retrospective case-control cohorts.
None is a prospective screening study and none measures clinical benefit. Cristiano et al. 2019, the DELFI developers, report 73% sensitivity (152 of 208 cancers) at a reported 98% specificity for DELFI in repeated cross-validation against 215 healthy individuals. Hou et al. 2024 re-implemented ten published fragmentation patterns and DELFI with one support vector machine each on the same cohort.
They are independent of every pattern except IFS, which their corresponding author developed. In 10 x 10-fold cross-validation, pan-cancer AUC within open chromatin regions ranged from 0.8741 (fragment length) to 0.9736 (5' end motif), and their DELFI re-implementation reached 0.9575. Rankings changed on independent cohorts.
On 46 lung cancers and 385 healthy individuals, with a model trained on 12 lung cancers, the end-motif model fell to AUC 0.5355 and sensitivity 0.0652 at 95% specificity, while fragment size distribution reached 0.9386 and 0.7174. Hou et al.'s supplementary pan-cancer and liver cancer sensitivities conflict with other tables of the same file and are not shown. Wang et al. 2026, the UNITE developers, stratified cancers by ichorCNA tumour fraction in pooled public and new 0.1x data that include samples of the DELFI and Jiang et al. cohorts used above.
For cancers with an estimated tumour fraction of 3% or less, cross-validated AUC was 0.878 for all five features combined, 0.873 for fragment length, 0.857 for the SD of length counts, 0.736 for copy number, 0.715 for the C/T end-motif ratio and 0.651 for the short/long ratio, against 0.603 for a classifier using the ichorCNA tumour fraction alone, which detected 0.058 of these cancers at 95% specificity. For cancers with an estimated fraction above 3% and up to 10%, the ichorCNA classifier reached 0.983, above the combined model (0.962) and copy number alone (0.926). The extracted UNITE tables give sensitivity at a fixed specificity only for the ichorCNA comparator; the fragmentomic models' values are in a sheet not yet extracted.
Jamshidi et al. 2022 (Cancer Cell), by the GRAIL developers, scored fragment length and fragment endpoint classifiers from 30x cfDNA whole-genome sequencing on the same CCGA substudy 1 participants as their methylation, copy-number and variant classifiers, at a 98% specificity threshold set post hoc within each set. On the validation set of 464 cancers and 362 non-cancers the fragment length classifier detected 29% (95% CI 25% to 34%) and the fragment endpoint classifier 18% (15% to 22%), against 34% for whole-genome methylation analysed double-blinded; both fragment classifiers were developed after the validation set had been unblinded, so these are not blinded validation results, and no tumour-fraction-stratified sensitivity is printed. Zeng et al. 2026 (Nature Cancer), by the cfMeDIP-seq developers, computed fragmentomic features from cfMeDIP-seq reads, which are enriched for methylated fragments rather than drawn from whole-genome sequencing, and report means over 10 fold models.
On an independent paired-end validation set, a fragment-ratio classifier reached AUC 0.848 and detected 0.533 of cancers at 99% specificity, and fragment ratios combined with methylation reached 0.954 and 0.766. The validation set has 17 healthy controls, so its 99% specificity allows no false positive among them, and Li-Fraumeni syndrome carriers without current cancer count as cancers in its sensitivities. Three of the five classifiers that include end motifs scored below AUC 0.5 there, which the authors attribute to healthy controls processed with a different laboratory protocol, so validation results for all five are not shown.
Helzer et al. 2025 (Nature Communications) compared twelve fragmentomic metrics adapted to targeted exon panels, including the authors' own fragment-size entropy metrics, their combination and Griffin, with one classifier and shared folds, on plasma from patients with metastatic cancer. The healthy-versus-cancer one-vs-rest AUROC ranged from 0.835 (open-chromatin entropy) to 0.989 (first-exon depth) in a UW cohort of 404 cancers and 27 healthy donors, with Griffin at 0.944, and from 0.951 to 1.000 in public GRAIL panel data from 151 cancers and 47 healthy donors. In a developer comparison on 0.1x whole-genome sequencing (Cell Reports Medicine 2026), a fragment-level language model (Fragmentia-AI) reached AUC 0.9561 on a held-out set of unstated size, against 0.9209 to 0.935 for the developers' re-implementation of DELFI features and 0.881 to 0.8863 for Griffin features under three classifiers.
The same company's whole-genome model (Molecular Biomedicine 2026), on an independent test cohort of 2,952 cancers and 3,044 non-cancer controls, reached AUC 0.910 for cancers with an ichorCNA tumour fraction below 3% and detected 0.46 of them at a cutoff set to 98% specificity in training (observed test specificity 0.97); among those low-fraction cancers it detected 0.391 at stage I/II (637 of 1,628) and 0.708 at stage III/IV (345 of 487). Its tumour-fraction strata are ichorCNA estimates, no comparator was run on its cohorts, and its training cohort has the same number of non-cancer controls (1,245) as the Cell Reports Medicine cohort, so the two studies may share samples. Three of these four studies report on their authors' own method, and in the fourth the authors' own entropy metrics are among those compared.
Cohorts, feature implementations, classifiers and validation designs differ, so no value is comparable across sources.