Colorectal cancer is one of the top leading causes of cancer mortality worldwide, especially in China. However, most of the current treatments are invasive and can only be applied to very few cancers. The earlier a malignant tumor is diagnosed, the higher the patient's survival rate. In this study, we proposed a computational framework to identify highly-reliable and easierly-detectable biomarkers capable of secreting into blood, urine and saliva by integrating transcriptomics and proteomics data at the system biology level. First, a large number of transcriptome data were processed to identify candidate biomarkers for colorectal cancer. Second, three classified models are constructed to predict biomarkers for colorectal cancer capable of secreting into blood, urine and saliva, which are effective disease diagnosis media to facilitate clinical screening. Then biological functions and molecular mechanisms of the candidate biomarkers of colorectal cancer are inferred utilizing multi-source biological knowledge and literature mining. Furthermore, the classification power of different combinations of candidate biomarkers is verified by machine learning models. In addition, the targeted drugs of the predicted biomarkers are further analyzed to provide assistance for clinical treatment of colorectal cancer. In this paper, our proposed computational model not only provides the effective candidate biomarkers ESM1, CTHRC1, AZGP1 for colorectal cancer capable of secreting into blood, urine and saliva, but also helps to understand the molecular mechanism of colorectal cancer. This computational framework can span the huge gap between transcriptome and proteomics, which can easily be applied to the biomarker research for other types of tumor.