Sensitivity of portal vein ctDNA versus peripheral blood in predicting postoperative liver metastasis adjuvant therapy in colorectal cancer.

3139 Background: The detection of minimal residual disease (MRD) in colorectal cancer (CRC) is crucial for predicting postoperative recurrence, particularly liver metastasis. While peripheral blood (PB) liquid biopsy is widely used, portal vein blood (PVB) may offer higher sensitivity due to direct drainage of the hepatic circulation. Methods: We prospectively enrolled 297 CRC patients undergoing curative surgery. Intraoperative portal vein blood (PVB) and matched preoperative peripheral blood (PB) samples were collected simultaneously. Cell-free DNA was sequenced using a targeted 689-gene panel (1.5 Mb). Variants were filtered to retain only those with increased allele frequency (AF) in PVB relative to PB or exclusive detection in PVB (AF ≥1%), followed by stringent germline and noise removal. Binary mutation matrices were constructed for both blood sources. Machine learning (random forest, SVM, logistic regression, gradient boosting) was used to identify predictive gene signatures separately for PVB and PB. Model performance was compared using five-fold cross-validation. Results: PVB-derived ctDNA detected a significantly higher number of tumor-specific mutations compared to PB (mean 3.2 vs. 1.4 mutations per patient, p<0.001). PVB contained significantly higher ctDNA concentration (median 8.4 ng/mL vs 2.1 ng/mL in PB, p<0.001). A 20-gene panel selected from PVB data demonstrated superior predictive accuracy for liver metastasis compared to PB-derived markers. When validated in the same cohort, the PVB panel achieved an AUC of 0.849 (95% CI: 0.802–0.891) versus 0.714 (95% CI: 0.653–0.771) for PB-based prediction (p=0.003). Sensitivity for predicting liver metastasis was 70.0% for PVB versus 48.5% for PB. The PVB panel identified 26.6% of mutation carriers as high-risk, with 100% specificity (no recurrence in mutation-negative patients). In contrast, the PB model failed to achieve comparable risk stratification, with lower positive predictive value and higher false-negative rates. Conclusions: Portal vein blood ctDNA analysis is significantly more sensitive and accurate than peripheral blood in predicting postoperative liver metastasis in CRC. PVB-based liquid biopsy provides superior risk stratification, which could better inform adjuvant therapy decisions—identifying high-risk patients for intensified surveillance or treatment, while reducing overtreatment in low-risk patients. These findings support the clinical integration of PVB sampling for MRD detection in CRC surgical practice.