A liquid biopsy-based predictive model identifies potential candidates for bladder preservation after neoadjuvant chemotherapy in muscle-invasive bladder cancer

Detection of circulating tumor DNA (ctDNA) and urinary tumor DNA (utDNA) after neoadjuvant chemotherapy (NAC) has been shown to predict treatment response and outcome in muscle-invasive bladder cancer. Pathological complete response (pCR) is associated with improved survival outcomes and serves as a proxy for treatment response. We hypothesize that combining ctDNA, utDNA, and clinical variables in a predictive model could improve prediction of treatment response and outcome and potentially be used to identify candidates for bladder preservation. We analyzed a unique dataset of urine and plasma samples together with detailed clinical measures and outcomes from 167 patients with localized muscle-invasive bladder cancer treated with at least three cycles of cisplatin-based NAC followed by radical cystectomy. ctDNA and utDNA measurements were performed using a tumor informed whole genome sequencing approach (TrueMRD, Veracyte) before (at baseline) and after NAC. pCR was defined as ypT0N0M0 at the time of radical cystectomy. Gradient boosting models were trained on 70% of the data using clinical variables alone and in combination with utDNA status and/or ctDNA dynamics and baseline ctDNA levels as well as imaging results from CT scans during NAC. Testing was done on the remaining 30% of data. Model performance was evaluated using the area under the precision-recall curve (AUCPR), and associations with ctDNA-free survival were investigated using Kaplan-Meier analyses. For prediction of pCR, a model including baseline clinical variables as well as utDNA status, ctDNA dynamics, baseline ctDNA levels and imaging results gave an AUCPR of 0.93 and a negative predictive value (NPV) of 0.80. Using this model, patients with predicted pCR had a significantly better 1-year ctDNA-free survival (HR = 0.2, p = 0.03). A model including clinical variables, utDNA status, ctDNA dynamics and baseline ctDNA levels had a similar performance (AUCPR = 0.90; NPV = 0.82). A model including only the baseline clinical variables performed poorly (AUCPR = 0.54; NPV = 0.50). Adding either imaging results (AUCPR = 0.52; NPV = 0.48), utDNA status (AUCPR = 0.81; NPV = 0.65), or ctDNA dynamics and baseline ctDNA levels (AUCPR = 0.73; NPV = 0.62) did only make minor improvements. Our findings demonstrate that integration of ctDNA dynamics and utDNA status, as well as imaging results into predictive models significantly enhances prediction of response to NAC and enables a clinically meaningful stratification of patients. Future independent testing in clinical trials is required, but our results support the development of liquid biopsy-guided bladder-sparing strategies in muscle-invasive bladder cancer.