Development and validation of postoperative circulating tumor DNA combined with clinicopathological risk factors for recurrence prediction in patients with stage I-III colorectal cancer

BACKGROUND: Circulating tumor DNA (ctDNA) detection following curative-intent surgery could directly reflect the presence of minimal residual disease, the ultimate cause of clinical recurrence. However, ctDNA is not postoperatively detected in ≥ 50% of patients with stage I-III colorectal cancer (CRC) who ultimately recur. Herein we sought to improve recurrence risk prediction by combining ctDNA with clinicopathological risk factors in stage I-III CRC. METHODS: Two independent cohorts, both consisting of early-stage CRC patients who underwent curative surgery, were included: (i) the discovery cohort (N = 124) with tumor tissues and postoperative plasmas for ctDNA determination; and (ii) the external validation cohort (N = 125) with available ctDNA results. In the discovery cohort, somatic variations in tumor tissues and plasmas were determined via a 733-gene and 127-gene next-generation sequencing panel, respectively. RESULTS: In the discovery cohort, 17 of 108 (15.7%) patients had detectable ctDNA. ctDNA-positive patients had a significantly high recurrence rate (76.5% vs. 16.5%, P < 0.001) and short recurrence-free survival (RFS; P < 0.001) versus ctDNA-negative patients. In addition to ctDNA status, the univariate Cox model identified pathologic stage, lymphovascular invasion, nerve invasion, and preoperative carcinoembryonic antigen level associated with RFS. We combined the ctDNA and clinicopathological risk factors (CTCP) to construct a model for recurrence prediction. A significantly higher recurrence rate (64.7% vs. 8.1%, P < 0.001) and worse RFS (P < 0.001) were seen in the high-risk patients classified by the CTCP model versus those in the low-risk patients. Receiver operating characteristic analysis demonstrated that the CTCP model outperformed ctDNA alone at recurrence prediction, which increased the sensitivity of 2 year RFS from 49.6% by ctDNA alone to 87.5%. Harrell's concordance index, calibration curve, and decision curve analysis also suggested that the CTCP model had good discrimination, consistency, and clinical utility. These results were reproduced in the validation cohort. CONCLUSION: Combining postoperative ctDNA and clinical risk may better predict recurrence than ctDNA alone for developing a personalized postoperative management strategy for CRC. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1186/s12967-023-03884-3.