Urine cell-free DNA multi-omics to detect MRD and predict survival in bladder cancer patients

Abstract

Circulating tumor DNA (ctDNA) sensitivity remains subpar for molecular residual disease (MRD) detection in bladder cancer patients. To remedy this problem, we focused on the biofluid most proximal to the disease, urine, and analyzed urine tumor DNA in 74 localized bladder cancer patients. We integrated ultra-low-pass whole genome sequencing (ULP-WGS) with urine cancer personalized profiling by deep sequencing (uCAPP-Seq) to achieve sensitive MRD detection and predict overall survival. Variant allele frequency, inferred tumor mutational burden, and copy number-derived tumor fraction levels in urine cell-free DNA (cfDNA) significantly predicted pathologic complete response status, far better than plasma ctDNA was able to. A random forest model incorporating these urine cfDNA-derived factors with leave-one-out cross-validation was 87% sensitive for predicting residual disease in reference to gold-standard surgical pathology. Both progression-free survival (HR = 3.00, p = 0.01) and overall survival (HR = 4.81, p = 0.009) were dramatically worse by Kaplan-Meier analysis for patients predicted by the model to have MRD, which was corroborated by Cox regression analysis. Additional survival analyses performed on muscle-invasive, neoadjuvant chemotherapy, and held-out validation subgroups corroborated these findings. In summary, we profiled urine samples from 74 patients with localized bladder cancer and used urine cfDNA multi-omics to detect MRD sensitively and predict survival accurately.

Fig. 1. Pathologic complete response prediction using a random forest model based on urine tumor DNA.

a Urine was collected prospectively from 74 localized bladder cancer patients pre-operatively on the day of curative-intent radical cystectomy after physician’s-choice neoadjuvant treatment. Urine cell-free DNA was sequenced by uCAPP-Seq (for single nucleotide variants) and ULP-WGS (for genome-wide copy number alterations) and then correlated with residual tumor in the surgical resection specimen and with patient survival. This figure panel was created with BioRender.com. b SNV-derived maximum VAFs, c inferred tumor mutational burden, and d CNA-derived tumor fraction levels in urine cell-free DNA from patients with localized bladder cancer. Scatter plots display these three different urine cell-free DNA metrics, stratified by pathologic complete response status, with significance determined by the Mann–Whitney U-test. VAF and CNA-derived tumor fraction data are shown after square root transformation. e ROC analysis of random forest model integrating urine tumor DNA metrics and other pretreatment clinical variables (Supplementary Fig. 5). ROC curve demonstrating the model’s performance for predicting pCR after LOOCV (AUC = 0.80, p < 0.0001). f Stacked bar plot depicting NPV and PPV of the random forest model with LOOCV, with significance determined by the Fisher’s exact test. AUC area under the curve, cfDNA cell-free DNA, CNA copy number alteration, iTMB inferred tumor mutational burden, LOOCV leave-one-out cross-validation, max maximum, MRD molecular residual disease, NPV negative predictive value, pCR pathologic complete response, PPV positive predictive value, ROC receiver operating characteristic, SNV single nucleotide variant, Sqrt square root, TFx tumor fraction, uCAPP-Seq urine cancer personalized profiling by deep sequencing, ULP-WGS ultra-low-pass whole genome sequencing, VAF variant allele frequency.

Fig. 2. Survival analysis comparing urine MRD detection to pathologic analysis of the resection specimen.

Kaplan–Meier plots showing a progression-free survival and b overall survival stratified by MRD detection in urine, determined by the LOOCV random forest model (Supplementary Fig. 5). c Progression-free survival and d overall survival stratified by pCR determined by microscopic analysis of the radical cystectomy specimen. Survival times shown are relative to the time of radical cystectomy. p values were calculated by the log-rank test and HRs by the Mantel–Haenszel method. HR hazard ratio, LOOCV leave-one-out cross-validation, MRD molecular residual disease, pCR pathologic complete response.