Time-weighted ctDNA dynamics for precision monitoring of relapse risk in colon cancer.

Abstract

3608 Background: Effective monitoring for relapse is a critical component of post-surgical care in colorectal cancer (CRC), particularly for patients who remain at risk of recurrence despite curative-intent treatment. Circulating tumor DNA (ctDNA) is a powerful biomarker for detecting minimal residual disease (MRD) and predicting relapse with high sensitivity and specificity. However, its predictive value varies over time, with negative results closer to surgery being less reliable than results obtained later. Here we introduce a novel time-weighted approach to ctDNA monitoring using a tumor-informed assay (Signatera), assigning greater predictive power to negative results collected further from surgery. Methods: A Bayesian logistic regression model using 1,246 Signatera serial measurements was developed to predict recurrence risk across 167 patients with early-stage colon cancer, with time-weighted ctDNA dynamics as the primary predictive feature. Negative ctDNA values were assigned greater predictive power based on their temporal distance from surgery. Secondary covariates included clinical stage and adjuvant treatment status. Time-weighted ctDNA was calculated as the product of the ctDNA level at each timepoint and an inverse time factor (1/( t +1)), where t represents the weeks since surgery. The weighted values were aggregated for each patient to compute cumulative and average time-weighted ctDNA levels, which served as inputs to the model. Survival analysis was performed to evaluate recurrence-free survival (RFS), stratified by MRD status. Results: Tumor-informed ctDNA levels were measured longitudinally, with a median of 7 timepoints per patient (range, 2-16) collected over a median follow-up of 2.5 years. Stage distribution was 50% stage III (n = 83), 44% stage II (n = 74), and 6% stage I (n = 10). Mismatch repair deficiency was observed in 22 patients (13.2%). Adjuvant chemotherapy was administered in 102 patients (61.1%), and 16 patients (9.6%) experienced recurrence. Survival analysis revealed a significantly worse recurrence-free survival for MRD-positive patients compared to MRD-negative patients (HR = 4.2, 95%CI:2.8–6.4, Log-rank p < 0.0001). The Bayesian logistic regression model demonstrated robust predictive performance, with a posterior probability of recurrence < 5% for patients with three consecutive negative ctDNA results obtained > 6 months after surgery. Conversely, the model assigned a > 90% probability of recurrence for patients with persistent ctDNA positivity beyond the initial 3-month post-operative window. Conclusions: Time-weighted ctDNA dynamics demonstrated promising predictive capability for CRC recurrence. Our findings suggest that incorporating the temporal context of ctDNA measurements and leveraging the increasing reliability of negative results over time could refine risk stratification and improve personalized care strategies for CRC patients.