Urine cfDNA 4-mer end-motif signatures outperform other fragmentomic features for tissue-of-origin classification in genitourinary cancers

Introduction: Urine cell-free DNA (cfDNA) is a promising ultra-noninvasive analyte for genitourinary cancers (GU) cancers. In this study, we investigated urine cfDNA fragmentomic features and evaluated their ability to predict the tumor tissue of origin. Methods: A total of 204 GU cancer patients and 34 healthy adults were enrolled in this prospective study. Preoperative urine was collected from 89 bladder cancer (BC) patients (64% muscle-invasive) undergoing cystectomy, 65 renal cell carcinoma (RCC) patients undergoing nephrectomy and 50 metastatic prostate cancer (mPC) patients. Urine cfDNA was isolated and sequenced at 5x genome-wide coverage on a NovaSeq S4 flow cell. To identify malignant tissue of origin, we analyzed somatic copy number alterations (CNAs), genome-wide fragment length ratios, and 4-mer end motifs. CNAs and tumor fraction (TFx) were quantified using ichorCNA in 1-Mb windows, while chromosome arm-level fragment count z-scores and fragment length ratios were calculated in 5-Mb bins across the genome. Relative frequencies of all 256 possible 4-mer end motifs were calculated for each sample. Individual logistic regression models were developed based on each fragmentomic feature using leave-one-out cross validation (LOOCV). Outputs from these models were combined into multi-feature XGBoost models. Cohort-specific end-motif enrichments were identified using Kruskal-Wallis tests with FDR correction. Motifs were then hierarchically clustered based on their cohort-wise median frequencies to define groups of motifs with shared fragmentation patterns. Results: Across urine cfDNA in the GU cohort, BC showed the highest mean TFx at 10.0%, followed by mPC at 3.6% and RCC at 2.7%. A machine learning model based on 4-mer end motifs achieved a mean ROC AUC of 0.933 with classification accuracies of 63% for healthy, 87% for BC, 72% for mPC, and 83% for RCC. This model outperformed models based on ichorCNA (AUC = 0.767), chromosome arm fragment-count z-scores (AUC = 0.827), and fragment length ratios (AUC = 0.674). Combining end motifs with other fragmentomic features did not improve prediction accuracy. Multi-feature models incorporating end motifs with ichorCNA, chromosome arm-level z-scores, or fragment-length ratios yielded mean AUCs between 0.908-0.924, lower than the end-motif only model. Kruskal-Wallis tests identified 179 of 256 end motifs (79%) as significantly different across healthy, BC, mPC, and RCC. Hierarchical clustering of cohort-level median motifs revealed distinct end motif patterns: Motifs beginning with CT were enriched in RCC; AAAA was strongly enriched in BC; and motifs beginning with CA, AC, and AG were enriched in mPC. Conclusion: Urine cfDNA 4-mer end motifs capture GU cancer-specific biological patterns that enable accurate tissue-of-origin prediction, underscoring their promise as a diagnostic tool.