Fragmentomics-based cancer type and subtype classification in 60,000 cell-free DNA samples

Purpose: Tissue-of-origin prediction and tumor subtyping enable more precise treatment selection, especially for cancers of unknown primary (CUP) or without a defined subtype, though these remain a challenge in liquid biopsy applications. Current genomic approaches include methylation profiling or whole-genome sequencing, which require additional laboratory workflows on top of comprehensive genomic profiling, or may utilize multiple detected somatic alterations, which may not be present in low tumor fraction liquid biopsy samples. Cell-free DNA fragmentation patterns reflect tissue-specific chromatin architecture and gene expression programs. We evaluated whether fragmentomic features could enable cancer type and subtype classification on our comprehensive genomic profiling platform. Methods: We analyzed 60,000 FoundationOne ® Liquid CDx samples and extracted fragmentomic features for research use only across 10,510 genomic target regions. Disease classification was performed in lung, breast, prostate, and colorectal cancer using a feedforward neural network classifier with an 80/20 training-test split using cross-entropy loss. Subtype classifiers were developed in lung cancer for histological subtype and in breast cancer for hormone receptor status. Results: We developed a 4-disease classifier using fragmentomic features in samples with circulating tumor DNA (ctDNA) fraction of 1% or greater. A high AUC was achieved across diseases (lung: 0.95, breast: 0.97, prostate: 0.97, colorectal: 0.97), with performance maintained even at low shed level (ctDNA fraction of 1-2%, AUCs 0.94-0.98). The lung histological subtype classifier achieved >90% accuracy in distinguishing between adenocarcinoma, squamous cell carcinoma, and small cell carcinoma. A breast cancer hormone receptor status classifier achieved >95% accuracy using inferred status from genomic data. Conclusions: Fragmentomic features in liquid biopsy samples enable accurate cancer type and histological subtype classification at ≥1% tumor fraction without requiring somatic variant detection. This approach has promise to address unmet needs in precision oncology.