VARIANT ORIGIN PREDICTION FROM WITHIN-LOCUS CELL-FREE DNA FRAGMENTOMICS: A SELF NORMALIZED, SINGLE-SAMPLE CLASSIFIER DISCRIMINATING TUMOR-SOMATIC FROM HEMATOPOIETIC VARIANTS

Abstract

Background. Determining the biological origin of a mutation detected in plasma cell-free DNA (cfDNA) — whether it is tumor-somatic or hematopoietic (constitutional germline or clonal hematopoiesis, CHIP) — is a shared, unsolved obstacle across liquid-biopsy applications [1, 2]. Misassignment of origin degrades the specificity of targeted-therapy variant classification, inflates false-positive calls in multi-cancer early detection (MCED), and confounds molecular residual disease (MRD) monitoring [2, 3]. Matched white blood cell (WBC) sequencing resolves origin but increases assay cost and is frequently unavailable [1]. We developed a within-locus fragmentomic classifier that infers variant origin from a single plasma sample, without a matched normal. The classifier reads the fragment-length contrast between the mutant (ALT) and reference (REF) alleles at the same genomic position, which serves as a self-normalizing signal. Methods. In a retrospective single-center cohort of 226 histopathologically confirmed solid-tumor patients with matched buffy-coat germline and 35 healthy-donor plasma controls (73-gene hybrid-capture panel, ~20,000× pre deduplication depth, unique molecular identifier consensus, hg19), we computed per-variant ALT-versus-REF fragment-length cumulative distribution functions (CDFs). From these we derived eight self-normalized within locus features and trained an L2-regularized logistic regression classifier (somatic versus hematopoietic) with a pre-specified 0.50 decision threshold and a plasma variant allele frequency (VAF) ≥ 0.05 abstain floor. Performance was estimated by patient-grouped cross-validation and a held-out patient split, and calibration was assessed. Results. Using 1,943 quality-controlled variants (338 tumor-somatic, 1,605 hematopoietic) from 249 individuals, the classifier reached an out-of-fold area under the receiver-operating-characteristic curve (AUC) of 0.920 (sensitivity 83.1%, specificity 86.8% at 0.50) and, on the held-out patient set, an AUC of 0.962 (sensitivity 93.6%, specificity 85.8%); specificity on healthy-control plasma was 88.9%. Calibration was good (Brier score 0.103). Tumor-somatic alleles were systematically enriched in short (< 150 bp) fragments relative to the same-locus reference, whereas germline and healthy-control alleles overlaid the ~167 bp mononucleosomal reference profile. Conclusions. A self-normalized, within-locus fragmentomic classifier predicts cfDNA variant origin from a single plasma draw without matched WBC sequencing. It may help with targeted-therapy variant triage, suppressing hematopoietic false positives in early detection, and tumor-informed MRD tracking.