Cell-free DNA fragmentomics as an enabler of tumor-of-origin classification using a commercial targeted cfDNA panel.

3059 Background: Analysis of cell-free DNA (cfDNA) fragmentation patterns (“fragmentomics”) has emerged as a powerful approach for cancer detection and classification, but most prior studies rely on whole-genome or research-grade sequencing. The feasibility and performance of fragmentomics using commercial targeted cfDNA next-generation sequencing (NGS) panels used in routine clinical care remain unclear. We evaluated whether fragmentomics metrics derived from the Tempus xF targeted cancer exon panel could accurately classify tumor type. Methods: We analyzed cfDNA sequencing data from 1,950 patients across 10 cancer types sequenced on the Tempus xF clinical NGS panel (105 genes), split a priori into training (n=1,455) and validation (n=495) cohorts. Exon-level fragmentomics features were extracted, including fragment size distributions, Shannon entropy, normalized read depth, proportions of short fragments, fragment size bins, motif diversity, and chromatin-context–specific entropy metrics. Elastic-net regression models were trained to classify tumor type using 10-fold cross-validation, with area under the receiver-operating characteristic curve (AUROC) as the primary endpoint. Model performance was independently evaluated in the validation cohort, including stratification by circulating tumor DNA (ctDNA) fraction. Results: In cross-validation of the training cohort, a combined fragmentomics model achieved a median AUROC of 0.790 across cancer types. In independent validation, the combined model maintained robust performance with a median AUROC of 0.794. Prostate, breast, and ovarian cancers demonstrated the highest discrimination (validation AUROCs up to 0.952 and 0.936 for prostate and breast cancer, respectively). Performance improved substantially in samples with higher tumor burden: among validation samples with ctDNA fraction ≥10% (85.2% of cases), the median AUROC increased to 0.815 compared with 0.719 in low ctDNA samples. Conclusions: Fragmentomics analysis applied to an existing commercial targeted cfDNA panel enables accurate tumor-of-origin classification without additional sequencing or assays. These findings demonstrate a scalable strategy to extend the diagnostic capabilities of routine clinical cfDNA testing and support the integration of fragmentomics into real-world liquid biopsy workflows, particularly in patients with sufficient ctDNA fraction.