Model performance comparison.

Colorectal cancer (CRC) has the second highest incidence rate among all cancers in Korea, with approximately 30% of patients with regional CRC experiencing recurrence. Understanding the genetic drivers of recurrence is essential for early detection and targeted treatment. Therefore, many studies have focused on genetic analysis using tumor-normal matched samples, as this approach provides more comprehensive insights. However, tumor-only samples are far more common in clinical practice because of the difficulty in obtaining normal tissues, making developing robust methods for analyzing tumor-only data a pressing need. This study aimed to investigate the genetic variations associated with CRC recurrence using tumor-only whole-exome sequencing data from 200 Korean patients with stage III CRC. By applying stringent filtering using public databases including Genome Aggregation Database (gnomAD), Exome Aggregation Consortium (ExAC), Single Nucleotide Polymorphism Database (dbSNP), 1000 Genomes Project (1000G), Korean Variant Archive 2 (KOVA2), and Korean Reference Genome Database (KRGDB), we identified 221 statistically significant mutations across 195 genes with distinct distributions between the recurrence and non-recurrence groups. Furthermore, statistical analysis of the clinical data revealed that the T-category, N-category, and preoperative carcinoembryonic antigen levels were correlated with CRC recurrence. Moreover, we identified nine networks through protein-protein interaction analysis and identified networks with high feature importance. We also developed a CRC recurrence prediction model using PyCaret, which achieved an area under the curve (AUC) of 0.77. Our findings highlight the importance of robust variant filtering in tumor-only sample analyses and provide insights into the genetic landscape of CRC recurrence in the Korean population.