Establishment of a 12-gene expression signature to predict colon cancer prognosis

Abstract

A robust and accurate gene expression signature is essential to assist oncologists to determine which subset of patients at similar Tumor-Lymph Node-Metastasis (TNM) stage has high recurrence risk and could benefit from adjuvant therapies. Here we applied a two-step supervised machine-learning method and established a 12-gene expression signature to precisely predict colon adenocarcinoma (COAD) prognosis by using COAD RNA-seq transcriptome data from The Cancer Genome Atlas (TCGA). The predictive performance of the 12-gene signature was validated with two independent gene expression microarray datasets:GSE39582includes 566 COAD cases for the development of six molecular subtypes with distinct clinical, molecular and survival characteristics;GSE17538is a dataset containing 232 colon cancer patients for the generation of a metastasis gene expression profile to predict recurrence and death in COAD patients. The signature could effectively separate the poor prognosis patients from good prognosis group (disease specific survival (DSS): Kaplan Meier (KM) Log Rankp= 0.0034; overall survival (OS): KM Log Rankp= 0.0336) inGSE17538. For patients with proficient mismatch repair system (pMMR) inGSE39582, the signature could also effectively distinguish high risk group from low risk group (OS: KM Log Rankp= 0.005; Relapse free survival (RFS): KM Log Rankp= 0.022). Interestingly, advanced stage patients were significantly enriched in high 12-gene score group (Fisher’s exact testp= 0.0003). After stage stratification, the signature could still distinguish poor prognosis patients inGSE17538from good prognosis within stage II (Log Rankp = 0.01) and stage II & III (Log Rankp= 0.017) in the outcome of DFS. Within stage III or II/III pMMR patients treated with Adjuvant Chemotherapies (ACT) and patients with higher 12-gene score showed poorer prognosis (III, OS: KM Log Rankp= 0.046; III & II, OS: KM Log Rankp= 0.041). Among stage II/III pMMR patients with lower 12-gene scores inGSE39582, the subgroup receiving ACT showed significantly longer OS time compared with those who received no ACT (Log Rankp= 0.021), while there is no obvious difference between counterparts among patients with higher 12-gene scores (Log Rankp= 0.12). Besides COAD, our 12-gene signature is multifunctional in several other cancer types including kidney cancer, lung cancer, uveal and skin melanoma, brain cancer, and pancreatic cancer. Functional classification showed that seven of the twelve genes are involved in immune system function and regulation, so our 12-gene signature could potentially be used to guide decisions about adjuvant therapy for patients with stage II/III and pMMR COAD.