Epigenome-informed deep learning identifies tumor-derived cfDNA fragments and improves liquid biopsy diagnostics in the absence of genetic aberrations
Liquid biopsy holds great promise for cancer diagnosis and therapy guidance. However, the low mutational bur-den of many pediatric tumors poses challenges to the detection of circulating tumor DNA (ctDNA). Here we present DEFT, a high-resolution multimodal deep learning method for ctDNA detection that does not depend on genetic aberrations. DEFT predicts which cell-free DNA (cfDNA) fragments are tumor-derived, based on the integration of DNA methylation, cfDNA fragmentation, and genomic properties. DEFT’s fragment-level prediction unlocks ample training data for deep learning even among small cohorts, and enables head-to-head comparisons between tumor-derived and healthy-derived cfDNA fragments from the same patient for investigating cfDNA biology. We combined DEFT with a DNA methylation assay that conserves cfDNA fragmentation, reaching 96% and 90% sensitivity for detection of Ewing sarcoma at 95% specificity across 335 individuals in two multinational clinical trial cohorts. In summary, we present a new method for deep learning based cfDNA analysis and demonstrate its utility by developing liquid biopsy assays for diagnosing and monitoring pediatric tumors.
- Type: Cancer Genomics
- Archive: European Genome-phenome Archive (EGA)
Click on a Dataset ID in the table below to learn more, and to find out who to contact about access to these data
| Dataset ID | Description | Technology | Samples |
|---|---|---|---|
| EGAD50000003205 | Illumina NovaSeq 6000 | 318 |
