1104 - Urine Cell-Free DNA Analysis to Detect Tumor Tissue of Origin and Predict Pathologic Complete Response across Genitourinary Malignancies

P. S. Chauhan¹, I. Alahi¹, A. Panda², N. Colon³, J. Sheng³, R. Mueller⁴, N. Cayce³, A. L. Shiang⁵, P. K. Harris⁶, F. Qaium¹, E. Kim⁷, M. Reimers⁸, W. Smelser⁹, Z. Smith⁷, and A. A. Chaudhuri¹; ¹Department of Radiation Oncology, Mayo Clinic, Rochester, MN, ²University of Chicago, Chicago, IL, ³Washington University School of Medicine, Saint Louis, St. Louis, MO, ⁴Washington University School of Medicine in St. Louis, St. Louis, MO, ⁵Cedars-Sinai Medical Center,, Los Angeles, CA, ⁶Washington University School of Medicine, St. Louis, MO, ⁷Department of Surgery, Division of Urology, Washington University School of Medicine, St. Louis, MO, ⁸Washington University School of Medicine, Department of Medicine, Division of Medical Oncology, St. Louis, MO, ⁹Washington University in St. Louis, St Louis, MO

Purpose/Objective(s): Liquid biopsy for MRD detection from plasma may not be sensitive enough. Here, we analyzed urine, a noninvasive and genitourinary-relevant analyte, from patients with genitourinary (GU) cancers and utilized an approach that incorporates copy-number derived tumor fraction as well as fragmentomic profiling to sensitively detect MRD and predict pathologic cancer tissue-of-origin. Materials/

Methods: A total of 150 patients with GU malignancies and 34 healthy adults were enrolled into this prospective study. We acquired urine preoperatively from 90 bladder cancer (BC) patients (67% muscle-invasive) who underwent cystectomy and 18 patients with renal-cell cancer (RCC) who underwent nephrectomy. We also collected urine samples from 42 metastatic prostate cancer (mPC) patients. We performed whole genome sequencing of urine cfDNA from all 150 patients and 34 healthy adults. Tumor fraction (TFx) levels based on genome-wide copy number alterations was estimated using ichorCNA. In a subset of 51 BC patients, we inferred fragment size and genomic coverage of urine cfDNA using Picard Tools. The short-to-long (S/L) fragment ratio per bin was calculated by partitioning the genome into 100 kb bins and evaluating the ratio of GC-corrected short fragments (50-150 bp) to long fragments (151-250 bp) within each bin. A normalized genome-wide S/L score was calculated by averaging across all bins. End motif analysis was also performed and motif diversity score (MDS) based on Shannon entropy was calculated. Machine learning was performed using the random forest algorithm; model performance was validated using a held-out test set.
Results: Comparing TFx across the GU cohort, BC had the highest median TFx (4.4%) followed by RCC (3.7%) and mPC (2.7%). In our cohort of 90 BC patients, 43% of patients achieved pCR (n = 39) while 57% had residual disease detected in their surgical sample (no pCR; n = 51). Patients with no pCR had significantly higher copy number-derived TFx in urine compared to patients with pCR (median 11.8% vs 2.3%, p < 0.0001). TFx achieved an AUC of 0.89 for classifying no pCR versus pCR with sensitivity of 65% and specificity of 87% (by Youden’s J-statistic). In a subset of 51 low-risk individuals (healthy and BC pCR), urine cfDNA median fragment sizes were significantly longer than BC patients with no pCR (median 177 bp vs. 156 bp, p = 0.002). Healthy and BC pCR individuals also had significantly lower S/L ratio scores than no pCR BC patients (mean 0.98 vs. 1.5, p = 0.007). Based on urine cfDNA end motif profiles, a machine learning model utilizing the random forest algorithm achieved an AUC of 0.85 for identifying prostate cancer, 0.84 for bladder cancer, 0.71 for renal cell carcinoma, and 0.84 for healthy adults with an average AUC of 0.81 in a held-out validation cohort.
Conclusion: Genome-wide analysis of urine cell-free DNA with fragmentomics, copy number analysis and end motif profiling can predict pCR in bladder cancer patients, and predict tumor tissue of origin across genitourinary malignancies.