Frequency-Domain Transformation of cfDNA End-Motif Profiles Enhances Robust Cancer Detection

Background/Objectives: Cell-free DNA (cfDNA) end-motifs (EDMs) are promising fragmentomic features for noninvasive cancer detection; however, their diagnostic utility may be limited by background signals from abundant hematopoietic-derived cfDNA fragments. Existing EDM-based approaches, including the Motif Diversity Score (MDS) and classifiers based on raw motif frequencies, often show limited robustness across different datasets. Methods: To address this limitation, we developed a frequency-domain analytical framework based on the Discrete Fourier Transform (DFT), converting k-mer EDM frequency profiles into amplitude spectral features. We further constructed a stacking-based Ensemble Spectral Model (ESM) integrating multi-scale spectral features from 4-6-mer EDMs. Results: The framework was evaluated using 1782 plasma cfDNA samples from four independent studies comprising six datasets. Raw EDM profiles showed extremely high similarity between cancer and non-cancer samples (mean Spearman R = 0.999). Following DFT transformation, amplitude spectra showed improved separability between groups. Across datasets, the ESM achieved a mean AUC of 0.843, representing a 15.0% improvement over raw 4-mer EDM-based SVM models and a 56.4% improvement over the MDS. At 95% specificity, mean sensitivity reached 0.585, exceeding those of the raw EDM (0.418) and MDS (0.195). Frequency-guided motif attribution further linked spectral features to sequence-level motif patterns and potential regulatory programs. Conclusions: Frequency-domain transformation improves the representation of cfDNA EDM profiles and provides a robust analytical framework for cross-dataset cancer detection.