Colorectal cancer affects the colon or rectum and is a common global health issue, with 1.1 million new cases occurring yearly. The study aimed to identify gene signatures for the early detection of CRC using machine learning (ML) algorithms utilizing gene expression data. The TCGA-CRC and GSE50760 datasets were pre-processed and subjected to feature selection using the LASSO method in combination with five ML algorithms: Adaboost, Random Forest (RF), Logistic Regression (LR), Gaussian Naive Bayes (GNB), and Support Vector Machine (SVM). The important features were further analyzed for gene expression, correlation, and survival analyses. Validation of the external dataset GSE142279 was also performed. The RF model had the best classification accuracy for both datasets. A feature selection process resulted in the identification of 12 candidate genes, which were subsequently reduced to 3 (CA2, CA7, and ITM2C) through gene expression and correlation analyses. These three genes achieved 100% accuracy in an external dataset. The AUC values for these genes were 99.24%, 100%, and 99.5%, respectively. The survival analysis showed a significant logrank p-value of 0.044 for the final gene signatures. The analysis of tumor immunocyte infiltration showed a weak correlation with the expression of the gene signatures. CA2, CA7, and ITM2C can serve as gene signatures for the early detection of CRC and may provide valuable information for prognostic and therapeutic decision making. Further research is needed to fully understand the potential of these genes in the context of CRC.
Keywords: colorectal cancer; correlation; feature selection; gene expression; gene signatures; machine learning.